1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature --include-generated-funcs --replace-value-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" "reduction_size[.].+[.]" "pl_cond[.].+[.|,]" --prefix-filecheck-ir-name _
2 // Test host codegen.
3 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK1
4 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s
5 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK2
6 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK3
7 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s
8 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK4
9 
10 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK5
11 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s
12 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK6
13 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK7
14 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s
15 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK8
16 
17 // Test target codegen - host bc file has to be created first.
18 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm-bc %s -o %t-ppc-host.bc
19 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK9
20 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o %t %s
21 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK10
22 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm-bc %s -o %t-x86-host.bc
23 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK11
24 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o %t %s
25 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK12
26 
27 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm-bc %s -o %t-ppc-host.bc
28 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK13
29 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o %t %s
30 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK14
31 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm-bc %s -o %t-x86-host.bc
32 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK15
33 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o %t %s
34 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK16
35 
36 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK17
37 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s
38 // RUN: %clang_cc1 -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK18
39 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK19
40 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s
41 // RUN: %clang_cc1 -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK20
42 
43 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK21
44 // RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s
45 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK22
46 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK23
47 // RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s
48 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK24
49 
50 // Test target codegen - host bc file has to be created first.
51 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm-bc %s -o %t-ppc-host.bc
52 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK25
53 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o %t %s
54 // RUN: %clang_cc1 -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK26
55 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm-bc %s -o %t-x86-host.bc
56 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK27
57 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o %t %s
58 // RUN: %clang_cc1 -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK28
59 
60 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm-bc %s -o %t-ppc-host.bc
61 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK29
62 // RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o %t %s
63 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK30
64 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm-bc %s -o %t-x86-host.bc
65 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK31
66 // RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o %t %s
67 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap  %s --check-prefix=CHECK32
68 
69 // expected-no-diagnostics
70 #ifndef HEADER
71 #define HEADER
72 
73 
74 
75 
76 // We have 8 target regions, but only 6 that actually will generate offloading
77 // code and have mapped arguments, and only 4 have all-constant map sizes.
78 
79 
80 
81 // Check target registration is registered as a Ctor.
82 
83 
84 template<typename tx, typename ty>
85 struct TT{
86   tx X;
87   ty Y;
88 };
89 
90 int global;
91 
92 int foo(int n) {
93   int a = 0;
94   short aa = 0;
95   float b[10];
96   float bn[n];
97   double c[5][10];
98   double cn[5][n];
99   TT<long long, char> d;
100 
101   #pragma omp target teams num_teams(a) thread_limit(a) firstprivate(aa) nowait
102   {
103   }
104 
105   #pragma omp target teams if(target: 0)
106   {
107     a += 1;
108   }
109 
110 
111   #pragma omp target teams if(target: 1)
112   {
113     aa += 1;
114   }
115 
116 
117 
118   #pragma omp target teams if(target: n>10)
119   {
120     a += 1;
121     aa += 1;
122   }
123 
124   // We capture 3 VLA sizes in this target region
125 
126 
127 
128 
129 
130   // The names below are not necessarily consistent with the names used for the
131   // addresses above as some are repeated.
132 
133 
134 
135 
136 
137 
138 
139 
140 
141 
142   #pragma omp target teams if(target: n>20)
143   {
144     a += 1;
145     b[2] += 1.0;
146     bn[3] += 1.0;
147     c[1][2] += 1.0;
148     cn[1][3] += 1.0;
149     d.X += 1;
150     d.Y += 1;
151   }
152 
153   const int nn = 0;
154   #pragma omp target teams shared(nn)
155   #pragma omp parallel firstprivate(nn)
156   (void)nn;
157   #pragma omp target teams firstprivate(nn)
158   #pragma omp parallel shared(nn)
159   (void)nn;
160   return a;
161 }
162 
163 // Check that the offloading functions are emitted and that the arguments are
164 // correct and loaded correctly for the target regions in foo().
165 
166 
167 
168 // Create stack storage and store argument in there.
169 
170 // Create stack storage and store argument in there.
171 
172 // Create stack storage and store argument in there.
173 
174 // Create local storage for each capture.
175 
176 
177 
178 // To reduce complexity, we're only going as far as validating the signature of the outlined parallel function.
179 
180 
181 void bazzzz(int n, int f[n]) {
182 #pragma omp target teams private(f)
183   ;
184 }
185 
186 template<typename tx>
187 tx ftemplate(int n) {
188   tx a = 0;
189   short aa = 0;
190   tx b[10];
191 
192   #pragma omp target teams if(target: n>40)
193   {
194     a += 1;
195     aa += 1;
196     b[2] += 1;
197   }
198 
199   return a;
200 }
201 
202 static
203 int fstatic(int n) {
204   int a = 0;
205   short aa = 0;
206   char aaa = 0;
207   int b[10];
208 
209   #pragma omp target teams if(target: n>50)
210   {
211     a += 1;
212     aa += 1;
213     aaa += 1;
214     b[2] += 1;
215   }
216 
217   return a;
218 }
219 
220 struct S1 {
221   double a;
222 
223   int r1(int n){
224     int b = n+1;
225     short int c[2][n];
226 
227     #pragma omp target teams if(target: n>60)
228     {
229       this->a = (double)b + 1.5;
230       c[1][1] = ++a;
231     }
232 
233     return c[1][1] + (int)b;
234   }
235 };
236 
237 int bar(int n){
238   int a = 0;
239 
240   a += foo(n);
241 
242   S1 S;
243   a += S.r1(n);
244 
245   a += fstatic(n);
246 
247   a += ftemplate<int>(n);
248 
249   return a;
250 }
251 
252 
253 
254 // We capture 2 VLA sizes in this target region
255 
256 
257 // The names below are not necessarily consistent with the names used for the
258 // addresses above as some are repeated.
259 
260 
261 
262 
263 
264 
265 
266 
267 
268 
269 
270 
271 
272 
273 
274 
275 
276 
277 
278 
279 // Check that the offloading functions are emitted and that the arguments are
280 // correct and loaded correctly for the target regions of the callees of bar().
281 
282 // Create local storage for each capture.
283 // Store captures in the context.
284 
285 
286 // To reduce complexity, we're only going as far as validating the signature of the outlined parallel function.
287 
288 
289 // Create local storage for each capture.
290 // Store captures in the context.
291 
292 
293 
294 
295 // To reduce complexity, we're only going as far as validating the signature of the outlined parallel function.
296 
297 // Create local storage for each capture.
298 // Store captures in the context.
299 
300 
301 
302 // To reduce complexity, we're only going as far as validating the signature of the outlined parallel function.
303 
304 #endif
305 // CHECK1-LABEL: define {{[^@]+}}@_Z3fooi
306 // CHECK1-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
307 // CHECK1-NEXT:  entry:
308 // CHECK1-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
309 // CHECK1-NEXT:    [[A:%.*]] = alloca i32, align 4
310 // CHECK1-NEXT:    [[AA:%.*]] = alloca i16, align 2
311 // CHECK1-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
312 // CHECK1-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
313 // CHECK1-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
314 // CHECK1-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
315 // CHECK1-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
316 // CHECK1-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
317 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
318 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
319 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
320 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
321 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED4:%.*]] = alloca i64, align 8
322 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
323 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
324 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
325 // CHECK1-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
326 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
327 // CHECK1-NEXT:    [[AA_CASTED7:%.*]] = alloca i64, align 8
328 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS9:%.*]] = alloca [1 x i8*], align 8
329 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS10:%.*]] = alloca [1 x i8*], align 8
330 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS11:%.*]] = alloca [1 x i8*], align 8
331 // CHECK1-NEXT:    [[A_CASTED12:%.*]] = alloca i64, align 8
332 // CHECK1-NEXT:    [[AA_CASTED14:%.*]] = alloca i64, align 8
333 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS16:%.*]] = alloca [2 x i8*], align 8
334 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
335 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
336 // CHECK1-NEXT:    [[A_CASTED21:%.*]] = alloca i64, align 8
337 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [9 x i8*], align 8
338 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [9 x i8*], align 8
339 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [9 x i8*], align 8
340 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
341 // CHECK1-NEXT:    [[NN:%.*]] = alloca i32, align 4
342 // CHECK1-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
343 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 8
344 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 8
345 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 8
346 // CHECK1-NEXT:    [[NN_CASTED38:%.*]] = alloca i64, align 8
347 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS40:%.*]] = alloca [1 x i8*], align 8
348 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS41:%.*]] = alloca [1 x i8*], align 8
349 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS42:%.*]] = alloca [1 x i8*], align 8
350 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
351 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
352 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
353 // CHECK1-NEXT:    store i16 0, i16* [[AA]], align 2
354 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
355 // CHECK1-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
356 // CHECK1-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
357 // CHECK1-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
358 // CHECK1-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP2]], align 4
359 // CHECK1-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
360 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
361 // CHECK1-NEXT:    [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
362 // CHECK1-NEXT:    [[TMP6:%.*]] = mul nuw i64 5, [[TMP5]]
363 // CHECK1-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP6]], align 8
364 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[__VLA_EXPR1]], align 8
365 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
366 // CHECK1-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_]], align 4
367 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
368 // CHECK1-NEXT:    store i32 [[TMP8]], i32* [[DOTCAPTURE_EXPR_2]], align 4
369 // CHECK1-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
370 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
371 // CHECK1-NEXT:    store i16 [[TMP9]], i16* [[CONV]], align 2
372 // CHECK1-NEXT:    [[TMP10:%.*]] = load i64, i64* [[AA_CASTED]], align 8
373 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
374 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
375 // CHECK1-NEXT:    store i32 [[TMP11]], i32* [[CONV3]], align 4
376 // CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
377 // CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
378 // CHECK1-NEXT:    [[CONV5:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED4]] to i32*
379 // CHECK1-NEXT:    store i32 [[TMP13]], i32* [[CONV5]], align 4
380 // CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED4]], align 8
381 // CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
382 // CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
383 // CHECK1-NEXT:    store i64 [[TMP10]], i64* [[TMP16]], align 8
384 // CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
385 // CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
386 // CHECK1-NEXT:    store i64 [[TMP10]], i64* [[TMP18]], align 8
387 // CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
388 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
389 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
390 // CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
391 // CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP21]], align 8
392 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
393 // CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
394 // CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP23]], align 8
395 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
396 // CHECK1-NEXT:    store i8* null, i8** [[TMP24]], align 8
397 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
398 // CHECK1-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i64*
399 // CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP26]], align 8
400 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
401 // CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
402 // CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP28]], align 8
403 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
404 // CHECK1-NEXT:    store i8* null, i8** [[TMP29]], align 8
405 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
406 // CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
407 // CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
408 // CHECK1-NEXT:    [[TMP33:%.*]] = load i16, i16* [[AA]], align 2
409 // CHECK1-NEXT:    store i16 [[TMP33]], i16* [[TMP32]], align 4
410 // CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
411 // CHECK1-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
412 // CHECK1-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
413 // CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
414 // CHECK1-NEXT:    [[TMP37:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
415 // CHECK1-NEXT:    store i32 [[TMP37]], i32* [[TMP36]], align 4
416 // CHECK1-NEXT:    [[TMP38:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
417 // CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8* [[TMP38]] to %struct.kmp_task_t_with_privates*
418 // CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 0
419 // CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP40]], i32 0, i32 0
420 // CHECK1-NEXT:    [[TMP42:%.*]] = load i8*, i8** [[TMP41]], align 8
421 // CHECK1-NEXT:    [[TMP43:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
422 // CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP42]], i8* align 4 [[TMP43]], i64 12, i1 false)
423 // CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 1
424 // CHECK1-NEXT:    [[TMP45:%.*]] = bitcast i8* [[TMP42]] to %struct.anon*
425 // CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 0
426 // CHECK1-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
427 // CHECK1-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP30]] to i8*
428 // CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP47]], i8* align 8 [[TMP48]], i64 24, i1 false)
429 // CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 1
430 // CHECK1-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
431 // CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP31]] to i8*
432 // CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP50]], i8* align 8 [[TMP51]], i64 24, i1 false)
433 // CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 2
434 // CHECK1-NEXT:    [[TMP53:%.*]] = bitcast [3 x i64]* [[TMP52]] to i8*
435 // CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP53]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
436 // CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 3
437 // CHECK1-NEXT:    [[TMP55:%.*]] = load i16, i16* [[AA]], align 2
438 // CHECK1-NEXT:    store i16 [[TMP55]], i16* [[TMP54]], align 8
439 // CHECK1-NEXT:    [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP38]])
440 // CHECK1-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
441 // CHECK1-NEXT:    [[CONV6:%.*]] = bitcast i64* [[A_CASTED]] to i32*
442 // CHECK1-NEXT:    store i32 [[TMP57]], i32* [[CONV6]], align 4
443 // CHECK1-NEXT:    [[TMP58:%.*]] = load i64, i64* [[A_CASTED]], align 8
444 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i64 [[TMP58]]) #[[ATTR3:[0-9]+]]
445 // CHECK1-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
446 // CHECK1-NEXT:    [[CONV8:%.*]] = bitcast i64* [[AA_CASTED7]] to i16*
447 // CHECK1-NEXT:    store i16 [[TMP59]], i16* [[CONV8]], align 2
448 // CHECK1-NEXT:    [[TMP60:%.*]] = load i64, i64* [[AA_CASTED7]], align 8
449 // CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
450 // CHECK1-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i64*
451 // CHECK1-NEXT:    store i64 [[TMP60]], i64* [[TMP62]], align 8
452 // CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
453 // CHECK1-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i64*
454 // CHECK1-NEXT:    store i64 [[TMP60]], i64* [[TMP64]], align 8
455 // CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
456 // CHECK1-NEXT:    store i8* null, i8** [[TMP65]], align 8
457 // CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
458 // CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
459 // CHECK1-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
460 // CHECK1-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
461 // CHECK1-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
462 // CHECK1:       omp_offload.failed:
463 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR3]]
464 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
465 // CHECK1:       omp_offload.cont:
466 // CHECK1-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
467 // CHECK1-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
468 // CHECK1-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
469 // CHECK1-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
470 // CHECK1-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
471 // CHECK1-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
472 // CHECK1-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
473 // CHECK1-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
474 // CHECK1-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
475 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
476 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
477 // CHECK1:       omp_if.then:
478 // CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
479 // CHECK1-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
480 // CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
481 // CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
482 // CHECK1-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
483 // CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
484 // CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
485 // CHECK1-NEXT:    store i8* null, i8** [[TMP79]], align 8
486 // CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
487 // CHECK1-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
488 // CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
489 // CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
490 // CHECK1-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
491 // CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
492 // CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
493 // CHECK1-NEXT:    store i8* null, i8** [[TMP84]], align 8
494 // CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
495 // CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
496 // CHECK1-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
497 // CHECK1-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
498 // CHECK1-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
499 // CHECK1:       omp_offload.failed19:
500 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
501 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
502 // CHECK1:       omp_offload.cont20:
503 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
504 // CHECK1:       omp_if.else:
505 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
506 // CHECK1-NEXT:    br label [[OMP_IF_END]]
507 // CHECK1:       omp_if.end:
508 // CHECK1-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
509 // CHECK1-NEXT:    [[CONV22:%.*]] = bitcast i64* [[A_CASTED21]] to i32*
510 // CHECK1-NEXT:    store i32 [[TMP89]], i32* [[CONV22]], align 4
511 // CHECK1-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED21]], align 8
512 // CHECK1-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
513 // CHECK1-NEXT:    [[CMP23:%.*]] = icmp sgt i32 [[TMP91]], 20
514 // CHECK1-NEXT:    br i1 [[CMP23]], label [[OMP_IF_THEN24:%.*]], label [[OMP_IF_ELSE30:%.*]]
515 // CHECK1:       omp_if.then24:
516 // CHECK1-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
517 // CHECK1-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
518 // CHECK1-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
519 // CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
520 // CHECK1-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
521 // CHECK1-NEXT:    store i64 [[TMP90]], i64* [[TMP96]], align 8
522 // CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
523 // CHECK1-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
524 // CHECK1-NEXT:    store i64 [[TMP90]], i64* [[TMP98]], align 8
525 // CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
526 // CHECK1-NEXT:    store i64 4, i64* [[TMP99]], align 8
527 // CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
528 // CHECK1-NEXT:    store i8* null, i8** [[TMP100]], align 8
529 // CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
530 // CHECK1-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
531 // CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
532 // CHECK1-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
533 // CHECK1-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
534 // CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
535 // CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
536 // CHECK1-NEXT:    store i64 40, i64* [[TMP105]], align 8
537 // CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
538 // CHECK1-NEXT:    store i8* null, i8** [[TMP106]], align 8
539 // CHECK1-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 2
540 // CHECK1-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
541 // CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP108]], align 8
542 // CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 2
543 // CHECK1-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
544 // CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP110]], align 8
545 // CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
546 // CHECK1-NEXT:    store i64 8, i64* [[TMP111]], align 8
547 // CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 2
548 // CHECK1-NEXT:    store i8* null, i8** [[TMP112]], align 8
549 // CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 3
550 // CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
551 // CHECK1-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
552 // CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 3
553 // CHECK1-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
554 // CHECK1-NEXT:    store float* [[VLA]], float** [[TMP116]], align 8
555 // CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
556 // CHECK1-NEXT:    store i64 [[TMP92]], i64* [[TMP117]], align 8
557 // CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 3
558 // CHECK1-NEXT:    store i8* null, i8** [[TMP118]], align 8
559 // CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 4
560 // CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
561 // CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
562 // CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 4
563 // CHECK1-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
564 // CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 8
565 // CHECK1-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
566 // CHECK1-NEXT:    store i64 400, i64* [[TMP123]], align 8
567 // CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 4
568 // CHECK1-NEXT:    store i8* null, i8** [[TMP124]], align 8
569 // CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 5
570 // CHECK1-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
571 // CHECK1-NEXT:    store i64 5, i64* [[TMP126]], align 8
572 // CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 5
573 // CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
574 // CHECK1-NEXT:    store i64 5, i64* [[TMP128]], align 8
575 // CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
576 // CHECK1-NEXT:    store i64 8, i64* [[TMP129]], align 8
577 // CHECK1-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 5
578 // CHECK1-NEXT:    store i8* null, i8** [[TMP130]], align 8
579 // CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 6
580 // CHECK1-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
581 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP132]], align 8
582 // CHECK1-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 6
583 // CHECK1-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
584 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP134]], align 8
585 // CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
586 // CHECK1-NEXT:    store i64 8, i64* [[TMP135]], align 8
587 // CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 6
588 // CHECK1-NEXT:    store i8* null, i8** [[TMP136]], align 8
589 // CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 7
590 // CHECK1-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
591 // CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 8
592 // CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 7
593 // CHECK1-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
594 // CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 8
595 // CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
596 // CHECK1-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 8
597 // CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 7
598 // CHECK1-NEXT:    store i8* null, i8** [[TMP142]], align 8
599 // CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 8
600 // CHECK1-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
601 // CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 8
602 // CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 8
603 // CHECK1-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
604 // CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 8
605 // CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
606 // CHECK1-NEXT:    store i64 16, i64* [[TMP147]], align 8
607 // CHECK1-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 8
608 // CHECK1-NEXT:    store i8* null, i8** [[TMP148]], align 8
609 // CHECK1-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
610 // CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
611 // CHECK1-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
612 // CHECK1-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
613 // CHECK1-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
614 // CHECK1-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
615 // CHECK1:       omp_offload.failed28:
616 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
617 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
618 // CHECK1:       omp_offload.cont29:
619 // CHECK1-NEXT:    br label [[OMP_IF_END31:%.*]]
620 // CHECK1:       omp_if.else30:
621 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
622 // CHECK1-NEXT:    br label [[OMP_IF_END31]]
623 // CHECK1:       omp_if.end31:
624 // CHECK1-NEXT:    store i32 0, i32* [[NN]], align 4
625 // CHECK1-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
626 // CHECK1-NEXT:    [[CONV32:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
627 // CHECK1-NEXT:    store i32 [[TMP154]], i32* [[CONV32]], align 4
628 // CHECK1-NEXT:    [[TMP155:%.*]] = load i64, i64* [[NN_CASTED]], align 8
629 // CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
630 // CHECK1-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i64*
631 // CHECK1-NEXT:    store i64 [[TMP155]], i64* [[TMP157]], align 8
632 // CHECK1-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
633 // CHECK1-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i64*
634 // CHECK1-NEXT:    store i64 [[TMP155]], i64* [[TMP159]], align 8
635 // CHECK1-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 0
636 // CHECK1-NEXT:    store i8* null, i8** [[TMP160]], align 8
637 // CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
638 // CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
639 // CHECK1-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
640 // CHECK1-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
641 // CHECK1-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
642 // CHECK1:       omp_offload.failed36:
643 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i64 [[TMP155]]) #[[ATTR3]]
644 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
645 // CHECK1:       omp_offload.cont37:
646 // CHECK1-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
647 // CHECK1-NEXT:    [[CONV39:%.*]] = bitcast i64* [[NN_CASTED38]] to i32*
648 // CHECK1-NEXT:    store i32 [[TMP165]], i32* [[CONV39]], align 4
649 // CHECK1-NEXT:    [[TMP166:%.*]] = load i64, i64* [[NN_CASTED38]], align 8
650 // CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
651 // CHECK1-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i64*
652 // CHECK1-NEXT:    store i64 [[TMP166]], i64* [[TMP168]], align 8
653 // CHECK1-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
654 // CHECK1-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
655 // CHECK1-NEXT:    store i64 [[TMP166]], i64* [[TMP170]], align 8
656 // CHECK1-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 0
657 // CHECK1-NEXT:    store i8* null, i8** [[TMP171]], align 8
658 // CHECK1-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
659 // CHECK1-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
660 // CHECK1-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
661 // CHECK1-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
662 // CHECK1-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
663 // CHECK1:       omp_offload.failed43:
664 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i64 [[TMP166]]) #[[ATTR3]]
665 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
666 // CHECK1:       omp_offload.cont44:
667 // CHECK1-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
668 // CHECK1-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
669 // CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
670 // CHECK1-NEXT:    ret i32 [[TMP176]]
671 //
672 //
673 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
674 // CHECK1-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
675 // CHECK1-NEXT:  entry:
676 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
677 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
678 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
679 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
680 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
681 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
682 // CHECK1-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
683 // CHECK1-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
684 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
685 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
686 // CHECK1-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
687 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
688 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
689 // CHECK1-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
690 // CHECK1-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
691 // CHECK1-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
692 // CHECK1-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
693 // CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
694 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
695 // CHECK1-NEXT:    ret void
696 //
697 //
698 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined.
699 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
700 // CHECK1-NEXT:  entry:
701 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
702 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
703 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
704 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
705 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
706 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
707 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
708 // CHECK1-NEXT:    ret void
709 //
710 //
711 // CHECK1-LABEL: define {{[^@]+}}@.omp_task_privates_map.
712 // CHECK1-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
713 // CHECK1-NEXT:  entry:
714 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
715 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 8
716 // CHECK1-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 8
717 // CHECK1-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 8
718 // CHECK1-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 8
719 // CHECK1-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
720 // CHECK1-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 8
721 // CHECK1-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 8
722 // CHECK1-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 8
723 // CHECK1-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 8
724 // CHECK1-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
725 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
726 // CHECK1-NEXT:    [[TMP7:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 8
727 // CHECK1-NEXT:    store [3 x i8*]* [[TMP6]], [3 x i8*]** [[TMP7]], align 8
728 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
729 // CHECK1-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 8
730 // CHECK1-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 8
731 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
732 // CHECK1-NEXT:    [[TMP11:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 8
733 // CHECK1-NEXT:    store [3 x i64]* [[TMP10]], [3 x i64]** [[TMP11]], align 8
734 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
735 // CHECK1-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 8
736 // CHECK1-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 8
737 // CHECK1-NEXT:    ret void
738 //
739 //
740 // CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry.
741 // CHECK1-SAME: (i32 signext [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
742 // CHECK1-NEXT:  entry:
743 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
744 // CHECK1-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
745 // CHECK1-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
746 // CHECK1-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
747 // CHECK1-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
748 // CHECK1-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
749 // CHECK1-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 8
750 // CHECK1-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 8
751 // CHECK1-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 8
752 // CHECK1-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 8
753 // CHECK1-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
754 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
755 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
756 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
757 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
758 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
759 // CHECK1-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
760 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
761 // CHECK1-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
762 // CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
763 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
764 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
765 // CHECK1-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
766 // CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
767 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
768 // CHECK1-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
769 // CHECK1-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
770 // CHECK1-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META14:![0-9]+]])
771 // CHECK1-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META17:![0-9]+]])
772 // CHECK1-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META19:![0-9]+]])
773 // CHECK1-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META21:![0-9]+]])
774 // CHECK1-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !23
775 // CHECK1-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !23
776 // CHECK1-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
777 // CHECK1-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
778 // CHECK1-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !23
779 // CHECK1-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
780 // CHECK1-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
781 // CHECK1-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
782 // CHECK1-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
783 // CHECK1-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
784 // CHECK1-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
785 // CHECK1-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !23
786 // CHECK1-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 8, !noalias !23
787 // CHECK1-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 8, !noalias !23
788 // CHECK1-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 8, !noalias !23
789 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i64 0, i64 0
790 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i64 0, i64 0
791 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i64 0, i64 0
792 // CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
793 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
794 // CHECK1-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
795 // CHECK1-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
796 // CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
797 // CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
798 // CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
799 // CHECK1:       omp_offload.failed.i:
800 // CHECK1-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
801 // CHECK1-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
802 // CHECK1-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !23
803 // CHECK1-NEXT:    [[TMP30:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !23
804 // CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
805 // CHECK1-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
806 // CHECK1-NEXT:    store i32 [[TMP31]], i32* [[CONV4_I]], align 4, !noalias !23
807 // CHECK1-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !23
808 // CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
809 // CHECK1-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
810 // CHECK1-NEXT:    store i32 [[TMP33]], i32* [[CONV6_I]], align 4, !noalias !23
811 // CHECK1-NEXT:    [[TMP34:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !23
812 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i64 [[TMP30]], i64 [[TMP32]], i64 [[TMP34]]) #[[ATTR3]]
813 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
814 // CHECK1:       .omp_outlined..1.exit:
815 // CHECK1-NEXT:    ret i32 0
816 //
817 //
818 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
819 // CHECK1-SAME: (i64 [[A:%.*]]) #[[ATTR2]] {
820 // CHECK1-NEXT:  entry:
821 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
822 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
823 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
824 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
825 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
826 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[A_CASTED]] to i32*
827 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
828 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
829 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]])
830 // CHECK1-NEXT:    ret void
831 //
832 //
833 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..2
834 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]]) #[[ATTR2]] {
835 // CHECK1-NEXT:  entry:
836 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
837 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
838 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
839 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
840 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
841 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
842 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
843 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
844 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
845 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
846 // CHECK1-NEXT:    ret void
847 //
848 //
849 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
850 // CHECK1-SAME: (i64 [[AA:%.*]]) #[[ATTR2]] {
851 // CHECK1-NEXT:  entry:
852 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
853 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
854 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
855 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
856 // CHECK1-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
857 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
858 // CHECK1-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
859 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
860 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP1]])
861 // CHECK1-NEXT:    ret void
862 //
863 //
864 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..3
865 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
866 // CHECK1-NEXT:  entry:
867 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
868 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
869 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
870 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
871 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
872 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
873 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
874 // CHECK1-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
875 // CHECK1-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
876 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
877 // CHECK1-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
878 // CHECK1-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
879 // CHECK1-NEXT:    ret void
880 //
881 //
882 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
883 // CHECK1-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
884 // CHECK1-NEXT:  entry:
885 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
886 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
887 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
888 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
889 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
890 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
891 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
892 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
893 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
894 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
895 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
896 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
897 // CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
898 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
899 // CHECK1-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
900 // CHECK1-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
901 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
902 // CHECK1-NEXT:    ret void
903 //
904 //
905 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..6
906 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
907 // CHECK1-NEXT:  entry:
908 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
909 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
910 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
911 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
912 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
913 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
914 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
915 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
916 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
917 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
918 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
919 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
920 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
921 // CHECK1-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
922 // CHECK1-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
923 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
924 // CHECK1-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
925 // CHECK1-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
926 // CHECK1-NEXT:    ret void
927 //
928 //
929 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
930 // CHECK1-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
931 // CHECK1-NEXT:  entry:
932 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
933 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
934 // CHECK1-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
935 // CHECK1-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
936 // CHECK1-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
937 // CHECK1-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
938 // CHECK1-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
939 // CHECK1-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
940 // CHECK1-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
941 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
942 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
943 // CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
944 // CHECK1-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
945 // CHECK1-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
946 // CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
947 // CHECK1-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
948 // CHECK1-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
949 // CHECK1-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
950 // CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
951 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
952 // CHECK1-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
953 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
954 // CHECK1-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
955 // CHECK1-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
956 // CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
957 // CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
958 // CHECK1-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
959 // CHECK1-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
960 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
961 // CHECK1-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
962 // CHECK1-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
963 // CHECK1-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
964 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
965 // CHECK1-NEXT:    ret void
966 //
967 //
968 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..9
969 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
970 // CHECK1-NEXT:  entry:
971 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
972 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
973 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
974 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
975 // CHECK1-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
976 // CHECK1-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
977 // CHECK1-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
978 // CHECK1-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
979 // CHECK1-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
980 // CHECK1-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
981 // CHECK1-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
982 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
983 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
984 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
985 // CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
986 // CHECK1-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
987 // CHECK1-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
988 // CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
989 // CHECK1-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
990 // CHECK1-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
991 // CHECK1-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
992 // CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
993 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
994 // CHECK1-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
995 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
996 // CHECK1-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
997 // CHECK1-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
998 // CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
999 // CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
1000 // CHECK1-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
1001 // CHECK1-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
1002 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
1003 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
1004 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
1005 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
1006 // CHECK1-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
1007 // CHECK1-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
1008 // CHECK1-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
1009 // CHECK1-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
1010 // CHECK1-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
1011 // CHECK1-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
1012 // CHECK1-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
1013 // CHECK1-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
1014 // CHECK1-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
1015 // CHECK1-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
1016 // CHECK1-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
1017 // CHECK1-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
1018 // CHECK1-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
1019 // CHECK1-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
1020 // CHECK1-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
1021 // CHECK1-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
1022 // CHECK1-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
1023 // CHECK1-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
1024 // CHECK1-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
1025 // CHECK1-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
1026 // CHECK1-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
1027 // CHECK1-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
1028 // CHECK1-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
1029 // CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
1030 // CHECK1-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
1031 // CHECK1-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
1032 // CHECK1-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
1033 // CHECK1-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
1034 // CHECK1-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
1035 // CHECK1-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
1036 // CHECK1-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
1037 // CHECK1-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
1038 // CHECK1-NEXT:    ret void
1039 //
1040 //
1041 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
1042 // CHECK1-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
1043 // CHECK1-NEXT:  entry:
1044 // CHECK1-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
1045 // CHECK1-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
1046 // CHECK1-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
1047 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
1048 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
1049 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
1050 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
1051 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
1052 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP1]])
1053 // CHECK1-NEXT:    ret void
1054 //
1055 //
1056 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..11
1057 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
1058 // CHECK1-NEXT:  entry:
1059 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1060 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1061 // CHECK1-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
1062 // CHECK1-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
1063 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1064 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1065 // CHECK1-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
1066 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
1067 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
1068 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
1069 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
1070 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
1071 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i64 [[TMP1]])
1072 // CHECK1-NEXT:    ret void
1073 //
1074 //
1075 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..12
1076 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
1077 // CHECK1-NEXT:  entry:
1078 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1079 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1080 // CHECK1-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
1081 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1082 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1083 // CHECK1-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
1084 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
1085 // CHECK1-NEXT:    ret void
1086 //
1087 //
1088 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
1089 // CHECK1-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
1090 // CHECK1-NEXT:  entry:
1091 // CHECK1-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
1092 // CHECK1-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
1093 // CHECK1-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
1094 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
1095 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
1096 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
1097 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
1098 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
1099 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP1]])
1100 // CHECK1-NEXT:    ret void
1101 //
1102 //
1103 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..15
1104 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
1105 // CHECK1-NEXT:  entry:
1106 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1107 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1108 // CHECK1-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
1109 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1110 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1111 // CHECK1-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
1112 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
1113 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[CONV]])
1114 // CHECK1-NEXT:    ret void
1115 //
1116 //
1117 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..16
1118 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
1119 // CHECK1-NEXT:  entry:
1120 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1121 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1122 // CHECK1-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
1123 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1124 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1125 // CHECK1-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
1126 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
1127 // CHECK1-NEXT:    ret void
1128 //
1129 //
1130 // CHECK1-LABEL: define {{[^@]+}}@_Z6bazzzziPi
1131 // CHECK1-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
1132 // CHECK1-NEXT:  entry:
1133 // CHECK1-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
1134 // CHECK1-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
1135 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
1136 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
1137 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
1138 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
1139 // CHECK1-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
1140 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
1141 // CHECK1-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
1142 // CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1143 // CHECK1-NEXT:    [[TMP3:%.*]] = bitcast i8** [[TMP2]] to i64*
1144 // CHECK1-NEXT:    store i64 [[TMP1]], i64* [[TMP3]], align 8
1145 // CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1146 // CHECK1-NEXT:    [[TMP5:%.*]] = bitcast i8** [[TMP4]] to i64*
1147 // CHECK1-NEXT:    store i64 [[TMP1]], i64* [[TMP5]], align 8
1148 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
1149 // CHECK1-NEXT:    store i8* null, i8** [[TMP6]], align 8
1150 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1151 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1152 // CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1153 // CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
1154 // CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1155 // CHECK1:       omp_offload.failed:
1156 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i64 [[TMP1]]) #[[ATTR3]]
1157 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1158 // CHECK1:       omp_offload.cont:
1159 // CHECK1-NEXT:    ret void
1160 //
1161 //
1162 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
1163 // CHECK1-SAME: (i64 [[VLA:%.*]]) #[[ATTR2]] {
1164 // CHECK1-NEXT:  entry:
1165 // CHECK1-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
1166 // CHECK1-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
1167 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
1168 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i64 [[TMP0]])
1169 // CHECK1-NEXT:    ret void
1170 //
1171 //
1172 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..19
1173 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR2]] {
1174 // CHECK1-NEXT:  entry:
1175 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1176 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1177 // CHECK1-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
1178 // CHECK1-NEXT:    [[F:%.*]] = alloca i32*, align 8
1179 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1180 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1181 // CHECK1-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
1182 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
1183 // CHECK1-NEXT:    ret void
1184 //
1185 //
1186 // CHECK1-LABEL: define {{[^@]+}}@_Z3bari
1187 // CHECK1-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
1188 // CHECK1-NEXT:  entry:
1189 // CHECK1-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
1190 // CHECK1-NEXT:    [[A:%.*]] = alloca i32, align 4
1191 // CHECK1-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
1192 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
1193 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
1194 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
1195 // CHECK1-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
1196 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
1197 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
1198 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
1199 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
1200 // CHECK1-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
1201 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
1202 // CHECK1-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
1203 // CHECK1-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
1204 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
1205 // CHECK1-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
1206 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
1207 // CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
1208 // CHECK1-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
1209 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
1210 // CHECK1-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
1211 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
1212 // CHECK1-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
1213 // CHECK1-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
1214 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
1215 // CHECK1-NEXT:    ret i32 [[TMP8]]
1216 //
1217 //
1218 // CHECK1-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
1219 // CHECK1-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
1220 // CHECK1-NEXT:  entry:
1221 // CHECK1-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
1222 // CHECK1-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
1223 // CHECK1-NEXT:    [[B:%.*]] = alloca i32, align 4
1224 // CHECK1-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
1225 // CHECK1-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
1226 // CHECK1-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
1227 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 8
1228 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 8
1229 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 8
1230 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 8
1231 // CHECK1-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
1232 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
1233 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
1234 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
1235 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
1236 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
1237 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
1238 // CHECK1-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
1239 // CHECK1-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
1240 // CHECK1-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
1241 // CHECK1-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
1242 // CHECK1-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
1243 // CHECK1-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
1244 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
1245 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_CASTED]] to i32*
1246 // CHECK1-NEXT:    store i32 [[TMP5]], i32* [[CONV]], align 4
1247 // CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
1248 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[N_ADDR]], align 4
1249 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP7]], 60
1250 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
1251 // CHECK1:       omp_if.then:
1252 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
1253 // CHECK1-NEXT:    [[TMP8:%.*]] = mul nuw i64 2, [[TMP2]]
1254 // CHECK1-NEXT:    [[TMP9:%.*]] = mul nuw i64 [[TMP8]], 2
1255 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1256 // CHECK1-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
1257 // CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 8
1258 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1259 // CHECK1-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
1260 // CHECK1-NEXT:    store double* [[A]], double** [[TMP13]], align 8
1261 // CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
1262 // CHECK1-NEXT:    store i64 8, i64* [[TMP14]], align 8
1263 // CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
1264 // CHECK1-NEXT:    store i8* null, i8** [[TMP15]], align 8
1265 // CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
1266 // CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
1267 // CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
1268 // CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
1269 // CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i64*
1270 // CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP19]], align 8
1271 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
1272 // CHECK1-NEXT:    store i64 4, i64* [[TMP20]], align 8
1273 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
1274 // CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
1275 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
1276 // CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
1277 // CHECK1-NEXT:    store i64 2, i64* [[TMP23]], align 8
1278 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
1279 // CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
1280 // CHECK1-NEXT:    store i64 2, i64* [[TMP25]], align 8
1281 // CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
1282 // CHECK1-NEXT:    store i64 8, i64* [[TMP26]], align 8
1283 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
1284 // CHECK1-NEXT:    store i8* null, i8** [[TMP27]], align 8
1285 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
1286 // CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
1287 // CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP29]], align 8
1288 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
1289 // CHECK1-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i64*
1290 // CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP31]], align 8
1291 // CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
1292 // CHECK1-NEXT:    store i64 8, i64* [[TMP32]], align 8
1293 // CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
1294 // CHECK1-NEXT:    store i8* null, i8** [[TMP33]], align 8
1295 // CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
1296 // CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
1297 // CHECK1-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 8
1298 // CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
1299 // CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
1300 // CHECK1-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 8
1301 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
1302 // CHECK1-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 8
1303 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 4
1304 // CHECK1-NEXT:    store i8* null, i8** [[TMP39]], align 8
1305 // CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1306 // CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1307 // CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
1308 // CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1309 // CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
1310 // CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1311 // CHECK1:       omp_offload.failed:
1312 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
1313 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1314 // CHECK1:       omp_offload.cont:
1315 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
1316 // CHECK1:       omp_if.else:
1317 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
1318 // CHECK1-NEXT:    br label [[OMP_IF_END]]
1319 // CHECK1:       omp_if.end:
1320 // CHECK1-NEXT:    [[TMP45:%.*]] = mul nsw i64 1, [[TMP2]]
1321 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP45]]
1322 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
1323 // CHECK1-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
1324 // CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP46]] to i32
1325 // CHECK1-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
1326 // CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP47]]
1327 // CHECK1-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
1328 // CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
1329 // CHECK1-NEXT:    ret i32 [[ADD4]]
1330 //
1331 //
1332 // CHECK1-LABEL: define {{[^@]+}}@_ZL7fstatici
1333 // CHECK1-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
1334 // CHECK1-NEXT:  entry:
1335 // CHECK1-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
1336 // CHECK1-NEXT:    [[A:%.*]] = alloca i32, align 4
1337 // CHECK1-NEXT:    [[AA:%.*]] = alloca i16, align 2
1338 // CHECK1-NEXT:    [[AAA:%.*]] = alloca i8, align 1
1339 // CHECK1-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
1340 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
1341 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
1342 // CHECK1-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
1343 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 8
1344 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 8
1345 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 8
1346 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
1347 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
1348 // CHECK1-NEXT:    store i16 0, i16* [[AA]], align 2
1349 // CHECK1-NEXT:    store i8 0, i8* [[AAA]], align 1
1350 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
1351 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
1352 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
1353 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
1354 // CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
1355 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
1356 // CHECK1-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
1357 // CHECK1-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
1358 // CHECK1-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
1359 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
1360 // CHECK1-NEXT:    store i8 [[TMP4]], i8* [[CONV2]], align 1
1361 // CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
1362 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
1363 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
1364 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
1365 // CHECK1:       omp_if.then:
1366 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1367 // CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
1368 // CHECK1-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
1369 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1370 // CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
1371 // CHECK1-NEXT:    store i64 [[TMP1]], i64* [[TMP10]], align 8
1372 // CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
1373 // CHECK1-NEXT:    store i8* null, i8** [[TMP11]], align 8
1374 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
1375 // CHECK1-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
1376 // CHECK1-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
1377 // CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
1378 // CHECK1-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
1379 // CHECK1-NEXT:    store i64 [[TMP3]], i64* [[TMP15]], align 8
1380 // CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
1381 // CHECK1-NEXT:    store i8* null, i8** [[TMP16]], align 8
1382 // CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
1383 // CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
1384 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP18]], align 8
1385 // CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
1386 // CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
1387 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP20]], align 8
1388 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
1389 // CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
1390 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
1391 // CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
1392 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 8
1393 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
1394 // CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
1395 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 8
1396 // CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
1397 // CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
1398 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1399 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1400 // CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1401 // CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
1402 // CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1403 // CHECK1:       omp_offload.failed:
1404 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
1405 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1406 // CHECK1:       omp_offload.cont:
1407 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
1408 // CHECK1:       omp_if.else:
1409 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
1410 // CHECK1-NEXT:    br label [[OMP_IF_END]]
1411 // CHECK1:       omp_if.end:
1412 // CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
1413 // CHECK1-NEXT:    ret i32 [[TMP31]]
1414 //
1415 //
1416 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
1417 // CHECK1-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
1418 // CHECK1-NEXT:  entry:
1419 // CHECK1-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
1420 // CHECK1-NEXT:    [[A:%.*]] = alloca i32, align 4
1421 // CHECK1-NEXT:    [[AA:%.*]] = alloca i16, align 2
1422 // CHECK1-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
1423 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
1424 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
1425 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
1426 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
1427 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
1428 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
1429 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
1430 // CHECK1-NEXT:    store i16 0, i16* [[AA]], align 2
1431 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
1432 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
1433 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
1434 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
1435 // CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
1436 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
1437 // CHECK1-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
1438 // CHECK1-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
1439 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
1440 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
1441 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
1442 // CHECK1:       omp_if.then:
1443 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1444 // CHECK1-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
1445 // CHECK1-NEXT:    store i64 [[TMP1]], i64* [[TMP6]], align 8
1446 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1447 // CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
1448 // CHECK1-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
1449 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
1450 // CHECK1-NEXT:    store i8* null, i8** [[TMP9]], align 8
1451 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
1452 // CHECK1-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i64*
1453 // CHECK1-NEXT:    store i64 [[TMP3]], i64* [[TMP11]], align 8
1454 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
1455 // CHECK1-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
1456 // CHECK1-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
1457 // CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
1458 // CHECK1-NEXT:    store i8* null, i8** [[TMP14]], align 8
1459 // CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
1460 // CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
1461 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 8
1462 // CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
1463 // CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
1464 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 8
1465 // CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
1466 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
1467 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1468 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1469 // CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1470 // CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
1471 // CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1472 // CHECK1:       omp_offload.failed:
1473 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
1474 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1475 // CHECK1:       omp_offload.cont:
1476 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
1477 // CHECK1:       omp_if.else:
1478 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
1479 // CHECK1-NEXT:    br label [[OMP_IF_END]]
1480 // CHECK1:       omp_if.end:
1481 // CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
1482 // CHECK1-NEXT:    ret i32 [[TMP24]]
1483 //
1484 //
1485 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
1486 // CHECK1-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
1487 // CHECK1-NEXT:  entry:
1488 // CHECK1-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
1489 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
1490 // CHECK1-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
1491 // CHECK1-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
1492 // CHECK1-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
1493 // CHECK1-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
1494 // CHECK1-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
1495 // CHECK1-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
1496 // CHECK1-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
1497 // CHECK1-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
1498 // CHECK1-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
1499 // CHECK1-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
1500 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
1501 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
1502 // CHECK1-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
1503 // CHECK1-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
1504 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
1505 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
1506 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
1507 // CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
1508 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
1509 // CHECK1-NEXT:    ret void
1510 //
1511 //
1512 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..22
1513 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
1514 // CHECK1-NEXT:  entry:
1515 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1516 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1517 // CHECK1-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
1518 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
1519 // CHECK1-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
1520 // CHECK1-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
1521 // CHECK1-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
1522 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1523 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1524 // CHECK1-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
1525 // CHECK1-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
1526 // CHECK1-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
1527 // CHECK1-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
1528 // CHECK1-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
1529 // CHECK1-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
1530 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
1531 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
1532 // CHECK1-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
1533 // CHECK1-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
1534 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
1535 // CHECK1-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
1536 // CHECK1-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
1537 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
1538 // CHECK1-NEXT:    store double [[ADD]], double* [[A]], align 8
1539 // CHECK1-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
1540 // CHECK1-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
1541 // CHECK1-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
1542 // CHECK1-NEXT:    store double [[INC]], double* [[A4]], align 8
1543 // CHECK1-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
1544 // CHECK1-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
1545 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
1546 // CHECK1-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
1547 // CHECK1-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
1548 // CHECK1-NEXT:    ret void
1549 //
1550 //
1551 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
1552 // CHECK1-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
1553 // CHECK1-NEXT:  entry:
1554 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
1555 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
1556 // CHECK1-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
1557 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
1558 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
1559 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
1560 // CHECK1-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
1561 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
1562 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
1563 // CHECK1-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
1564 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
1565 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
1566 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
1567 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
1568 // CHECK1-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
1569 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
1570 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
1571 // CHECK1-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
1572 // CHECK1-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
1573 // CHECK1-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
1574 // CHECK1-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
1575 // CHECK1-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
1576 // CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
1577 // CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
1578 // CHECK1-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
1579 // CHECK1-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
1580 // CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
1581 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
1582 // CHECK1-NEXT:    ret void
1583 //
1584 //
1585 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..24
1586 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
1587 // CHECK1-NEXT:  entry:
1588 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1589 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1590 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
1591 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
1592 // CHECK1-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
1593 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
1594 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1595 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1596 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
1597 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
1598 // CHECK1-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
1599 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
1600 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
1601 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
1602 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
1603 // CHECK1-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
1604 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
1605 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
1606 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
1607 // CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
1608 // CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
1609 // CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
1610 // CHECK1-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
1611 // CHECK1-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
1612 // CHECK1-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
1613 // CHECK1-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
1614 // CHECK1-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
1615 // CHECK1-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
1616 // CHECK1-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
1617 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
1618 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
1619 // CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
1620 // CHECK1-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
1621 // CHECK1-NEXT:    ret void
1622 //
1623 //
1624 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
1625 // CHECK1-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
1626 // CHECK1-NEXT:  entry:
1627 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
1628 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
1629 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
1630 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
1631 // CHECK1-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
1632 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
1633 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
1634 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
1635 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
1636 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
1637 // CHECK1-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
1638 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
1639 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
1640 // CHECK1-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
1641 // CHECK1-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
1642 // CHECK1-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
1643 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
1644 // CHECK1-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
1645 // CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
1646 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
1647 // CHECK1-NEXT:    ret void
1648 //
1649 //
1650 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..27
1651 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
1652 // CHECK1-NEXT:  entry:
1653 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1654 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1655 // CHECK1-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
1656 // CHECK1-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
1657 // CHECK1-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
1658 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1659 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1660 // CHECK1-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
1661 // CHECK1-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
1662 // CHECK1-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
1663 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
1664 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
1665 // CHECK1-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
1666 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
1667 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
1668 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
1669 // CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
1670 // CHECK1-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
1671 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
1672 // CHECK1-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
1673 // CHECK1-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
1674 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
1675 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
1676 // CHECK1-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
1677 // CHECK1-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
1678 // CHECK1-NEXT:    ret void
1679 //
1680 //
1681 // CHECK1-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
1682 // CHECK1-SAME: () #[[ATTR4]] {
1683 // CHECK1-NEXT:  entry:
1684 // CHECK1-NEXT:    call void @__tgt_register_requires(i64 1)
1685 // CHECK1-NEXT:    ret void
1686 //
1687 //
1688 // CHECK2-LABEL: define {{[^@]+}}@_Z3fooi
1689 // CHECK2-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
1690 // CHECK2-NEXT:  entry:
1691 // CHECK2-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
1692 // CHECK2-NEXT:    [[A:%.*]] = alloca i32, align 4
1693 // CHECK2-NEXT:    [[AA:%.*]] = alloca i16, align 2
1694 // CHECK2-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
1695 // CHECK2-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
1696 // CHECK2-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
1697 // CHECK2-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
1698 // CHECK2-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
1699 // CHECK2-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
1700 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
1701 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
1702 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
1703 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
1704 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR__CASTED4:%.*]] = alloca i64, align 8
1705 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
1706 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
1707 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
1708 // CHECK2-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
1709 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
1710 // CHECK2-NEXT:    [[AA_CASTED7:%.*]] = alloca i64, align 8
1711 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS9:%.*]] = alloca [1 x i8*], align 8
1712 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS10:%.*]] = alloca [1 x i8*], align 8
1713 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS11:%.*]] = alloca [1 x i8*], align 8
1714 // CHECK2-NEXT:    [[A_CASTED12:%.*]] = alloca i64, align 8
1715 // CHECK2-NEXT:    [[AA_CASTED14:%.*]] = alloca i64, align 8
1716 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS16:%.*]] = alloca [2 x i8*], align 8
1717 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
1718 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
1719 // CHECK2-NEXT:    [[A_CASTED21:%.*]] = alloca i64, align 8
1720 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [9 x i8*], align 8
1721 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [9 x i8*], align 8
1722 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [9 x i8*], align 8
1723 // CHECK2-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
1724 // CHECK2-NEXT:    [[NN:%.*]] = alloca i32, align 4
1725 // CHECK2-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
1726 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 8
1727 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 8
1728 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 8
1729 // CHECK2-NEXT:    [[NN_CASTED38:%.*]] = alloca i64, align 8
1730 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS40:%.*]] = alloca [1 x i8*], align 8
1731 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS41:%.*]] = alloca [1 x i8*], align 8
1732 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS42:%.*]] = alloca [1 x i8*], align 8
1733 // CHECK2-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
1734 // CHECK2-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
1735 // CHECK2-NEXT:    store i32 0, i32* [[A]], align 4
1736 // CHECK2-NEXT:    store i16 0, i16* [[AA]], align 2
1737 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
1738 // CHECK2-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
1739 // CHECK2-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
1740 // CHECK2-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
1741 // CHECK2-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP2]], align 4
1742 // CHECK2-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
1743 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
1744 // CHECK2-NEXT:    [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
1745 // CHECK2-NEXT:    [[TMP6:%.*]] = mul nuw i64 5, [[TMP5]]
1746 // CHECK2-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP6]], align 8
1747 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[__VLA_EXPR1]], align 8
1748 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
1749 // CHECK2-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_]], align 4
1750 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
1751 // CHECK2-NEXT:    store i32 [[TMP8]], i32* [[DOTCAPTURE_EXPR_2]], align 4
1752 // CHECK2-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
1753 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
1754 // CHECK2-NEXT:    store i16 [[TMP9]], i16* [[CONV]], align 2
1755 // CHECK2-NEXT:    [[TMP10:%.*]] = load i64, i64* [[AA_CASTED]], align 8
1756 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
1757 // CHECK2-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
1758 // CHECK2-NEXT:    store i32 [[TMP11]], i32* [[CONV3]], align 4
1759 // CHECK2-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
1760 // CHECK2-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
1761 // CHECK2-NEXT:    [[CONV5:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED4]] to i32*
1762 // CHECK2-NEXT:    store i32 [[TMP13]], i32* [[CONV5]], align 4
1763 // CHECK2-NEXT:    [[TMP14:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED4]], align 8
1764 // CHECK2-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1765 // CHECK2-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
1766 // CHECK2-NEXT:    store i64 [[TMP10]], i64* [[TMP16]], align 8
1767 // CHECK2-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1768 // CHECK2-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
1769 // CHECK2-NEXT:    store i64 [[TMP10]], i64* [[TMP18]], align 8
1770 // CHECK2-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
1771 // CHECK2-NEXT:    store i8* null, i8** [[TMP19]], align 8
1772 // CHECK2-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
1773 // CHECK2-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
1774 // CHECK2-NEXT:    store i64 [[TMP12]], i64* [[TMP21]], align 8
1775 // CHECK2-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
1776 // CHECK2-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
1777 // CHECK2-NEXT:    store i64 [[TMP12]], i64* [[TMP23]], align 8
1778 // CHECK2-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
1779 // CHECK2-NEXT:    store i8* null, i8** [[TMP24]], align 8
1780 // CHECK2-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
1781 // CHECK2-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i64*
1782 // CHECK2-NEXT:    store i64 [[TMP14]], i64* [[TMP26]], align 8
1783 // CHECK2-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
1784 // CHECK2-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
1785 // CHECK2-NEXT:    store i64 [[TMP14]], i64* [[TMP28]], align 8
1786 // CHECK2-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
1787 // CHECK2-NEXT:    store i8* null, i8** [[TMP29]], align 8
1788 // CHECK2-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1789 // CHECK2-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1790 // CHECK2-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
1791 // CHECK2-NEXT:    [[TMP33:%.*]] = load i16, i16* [[AA]], align 2
1792 // CHECK2-NEXT:    store i16 [[TMP33]], i16* [[TMP32]], align 4
1793 // CHECK2-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
1794 // CHECK2-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
1795 // CHECK2-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
1796 // CHECK2-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
1797 // CHECK2-NEXT:    [[TMP37:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
1798 // CHECK2-NEXT:    store i32 [[TMP37]], i32* [[TMP36]], align 4
1799 // CHECK2-NEXT:    [[TMP38:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
1800 // CHECK2-NEXT:    [[TMP39:%.*]] = bitcast i8* [[TMP38]] to %struct.kmp_task_t_with_privates*
1801 // CHECK2-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 0
1802 // CHECK2-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP40]], i32 0, i32 0
1803 // CHECK2-NEXT:    [[TMP42:%.*]] = load i8*, i8** [[TMP41]], align 8
1804 // CHECK2-NEXT:    [[TMP43:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
1805 // CHECK2-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP42]], i8* align 4 [[TMP43]], i64 12, i1 false)
1806 // CHECK2-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 1
1807 // CHECK2-NEXT:    [[TMP45:%.*]] = bitcast i8* [[TMP42]] to %struct.anon*
1808 // CHECK2-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 0
1809 // CHECK2-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
1810 // CHECK2-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP30]] to i8*
1811 // CHECK2-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP47]], i8* align 8 [[TMP48]], i64 24, i1 false)
1812 // CHECK2-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 1
1813 // CHECK2-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
1814 // CHECK2-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP31]] to i8*
1815 // CHECK2-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP50]], i8* align 8 [[TMP51]], i64 24, i1 false)
1816 // CHECK2-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 2
1817 // CHECK2-NEXT:    [[TMP53:%.*]] = bitcast [3 x i64]* [[TMP52]] to i8*
1818 // CHECK2-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP53]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
1819 // CHECK2-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 3
1820 // CHECK2-NEXT:    [[TMP55:%.*]] = load i16, i16* [[AA]], align 2
1821 // CHECK2-NEXT:    store i16 [[TMP55]], i16* [[TMP54]], align 8
1822 // CHECK2-NEXT:    [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP38]])
1823 // CHECK2-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
1824 // CHECK2-NEXT:    [[CONV6:%.*]] = bitcast i64* [[A_CASTED]] to i32*
1825 // CHECK2-NEXT:    store i32 [[TMP57]], i32* [[CONV6]], align 4
1826 // CHECK2-NEXT:    [[TMP58:%.*]] = load i64, i64* [[A_CASTED]], align 8
1827 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i64 [[TMP58]]) #[[ATTR3:[0-9]+]]
1828 // CHECK2-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
1829 // CHECK2-NEXT:    [[CONV8:%.*]] = bitcast i64* [[AA_CASTED7]] to i16*
1830 // CHECK2-NEXT:    store i16 [[TMP59]], i16* [[CONV8]], align 2
1831 // CHECK2-NEXT:    [[TMP60:%.*]] = load i64, i64* [[AA_CASTED7]], align 8
1832 // CHECK2-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
1833 // CHECK2-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i64*
1834 // CHECK2-NEXT:    store i64 [[TMP60]], i64* [[TMP62]], align 8
1835 // CHECK2-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
1836 // CHECK2-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i64*
1837 // CHECK2-NEXT:    store i64 [[TMP60]], i64* [[TMP64]], align 8
1838 // CHECK2-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
1839 // CHECK2-NEXT:    store i8* null, i8** [[TMP65]], align 8
1840 // CHECK2-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
1841 // CHECK2-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
1842 // CHECK2-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1843 // CHECK2-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
1844 // CHECK2-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1845 // CHECK2:       omp_offload.failed:
1846 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR3]]
1847 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1848 // CHECK2:       omp_offload.cont:
1849 // CHECK2-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
1850 // CHECK2-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
1851 // CHECK2-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
1852 // CHECK2-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
1853 // CHECK2-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
1854 // CHECK2-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
1855 // CHECK2-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
1856 // CHECK2-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
1857 // CHECK2-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
1858 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
1859 // CHECK2-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
1860 // CHECK2:       omp_if.then:
1861 // CHECK2-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
1862 // CHECK2-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
1863 // CHECK2-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
1864 // CHECK2-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
1865 // CHECK2-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
1866 // CHECK2-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
1867 // CHECK2-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
1868 // CHECK2-NEXT:    store i8* null, i8** [[TMP79]], align 8
1869 // CHECK2-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
1870 // CHECK2-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
1871 // CHECK2-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
1872 // CHECK2-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
1873 // CHECK2-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
1874 // CHECK2-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
1875 // CHECK2-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
1876 // CHECK2-NEXT:    store i8* null, i8** [[TMP84]], align 8
1877 // CHECK2-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
1878 // CHECK2-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
1879 // CHECK2-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1880 // CHECK2-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
1881 // CHECK2-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
1882 // CHECK2:       omp_offload.failed19:
1883 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
1884 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
1885 // CHECK2:       omp_offload.cont20:
1886 // CHECK2-NEXT:    br label [[OMP_IF_END:%.*]]
1887 // CHECK2:       omp_if.else:
1888 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
1889 // CHECK2-NEXT:    br label [[OMP_IF_END]]
1890 // CHECK2:       omp_if.end:
1891 // CHECK2-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
1892 // CHECK2-NEXT:    [[CONV22:%.*]] = bitcast i64* [[A_CASTED21]] to i32*
1893 // CHECK2-NEXT:    store i32 [[TMP89]], i32* [[CONV22]], align 4
1894 // CHECK2-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED21]], align 8
1895 // CHECK2-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
1896 // CHECK2-NEXT:    [[CMP23:%.*]] = icmp sgt i32 [[TMP91]], 20
1897 // CHECK2-NEXT:    br i1 [[CMP23]], label [[OMP_IF_THEN24:%.*]], label [[OMP_IF_ELSE30:%.*]]
1898 // CHECK2:       omp_if.then24:
1899 // CHECK2-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
1900 // CHECK2-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
1901 // CHECK2-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
1902 // CHECK2-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
1903 // CHECK2-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
1904 // CHECK2-NEXT:    store i64 [[TMP90]], i64* [[TMP96]], align 8
1905 // CHECK2-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
1906 // CHECK2-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
1907 // CHECK2-NEXT:    store i64 [[TMP90]], i64* [[TMP98]], align 8
1908 // CHECK2-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
1909 // CHECK2-NEXT:    store i64 4, i64* [[TMP99]], align 8
1910 // CHECK2-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
1911 // CHECK2-NEXT:    store i8* null, i8** [[TMP100]], align 8
1912 // CHECK2-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
1913 // CHECK2-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
1914 // CHECK2-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
1915 // CHECK2-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
1916 // CHECK2-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
1917 // CHECK2-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
1918 // CHECK2-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
1919 // CHECK2-NEXT:    store i64 40, i64* [[TMP105]], align 8
1920 // CHECK2-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
1921 // CHECK2-NEXT:    store i8* null, i8** [[TMP106]], align 8
1922 // CHECK2-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 2
1923 // CHECK2-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
1924 // CHECK2-NEXT:    store i64 [[TMP2]], i64* [[TMP108]], align 8
1925 // CHECK2-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 2
1926 // CHECK2-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
1927 // CHECK2-NEXT:    store i64 [[TMP2]], i64* [[TMP110]], align 8
1928 // CHECK2-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
1929 // CHECK2-NEXT:    store i64 8, i64* [[TMP111]], align 8
1930 // CHECK2-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 2
1931 // CHECK2-NEXT:    store i8* null, i8** [[TMP112]], align 8
1932 // CHECK2-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 3
1933 // CHECK2-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
1934 // CHECK2-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
1935 // CHECK2-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 3
1936 // CHECK2-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
1937 // CHECK2-NEXT:    store float* [[VLA]], float** [[TMP116]], align 8
1938 // CHECK2-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
1939 // CHECK2-NEXT:    store i64 [[TMP92]], i64* [[TMP117]], align 8
1940 // CHECK2-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 3
1941 // CHECK2-NEXT:    store i8* null, i8** [[TMP118]], align 8
1942 // CHECK2-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 4
1943 // CHECK2-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
1944 // CHECK2-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
1945 // CHECK2-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 4
1946 // CHECK2-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
1947 // CHECK2-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 8
1948 // CHECK2-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
1949 // CHECK2-NEXT:    store i64 400, i64* [[TMP123]], align 8
1950 // CHECK2-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 4
1951 // CHECK2-NEXT:    store i8* null, i8** [[TMP124]], align 8
1952 // CHECK2-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 5
1953 // CHECK2-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
1954 // CHECK2-NEXT:    store i64 5, i64* [[TMP126]], align 8
1955 // CHECK2-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 5
1956 // CHECK2-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
1957 // CHECK2-NEXT:    store i64 5, i64* [[TMP128]], align 8
1958 // CHECK2-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
1959 // CHECK2-NEXT:    store i64 8, i64* [[TMP129]], align 8
1960 // CHECK2-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 5
1961 // CHECK2-NEXT:    store i8* null, i8** [[TMP130]], align 8
1962 // CHECK2-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 6
1963 // CHECK2-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
1964 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP132]], align 8
1965 // CHECK2-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 6
1966 // CHECK2-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
1967 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP134]], align 8
1968 // CHECK2-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
1969 // CHECK2-NEXT:    store i64 8, i64* [[TMP135]], align 8
1970 // CHECK2-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 6
1971 // CHECK2-NEXT:    store i8* null, i8** [[TMP136]], align 8
1972 // CHECK2-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 7
1973 // CHECK2-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
1974 // CHECK2-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 8
1975 // CHECK2-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 7
1976 // CHECK2-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
1977 // CHECK2-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 8
1978 // CHECK2-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
1979 // CHECK2-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 8
1980 // CHECK2-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 7
1981 // CHECK2-NEXT:    store i8* null, i8** [[TMP142]], align 8
1982 // CHECK2-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 8
1983 // CHECK2-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
1984 // CHECK2-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 8
1985 // CHECK2-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 8
1986 // CHECK2-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
1987 // CHECK2-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 8
1988 // CHECK2-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
1989 // CHECK2-NEXT:    store i64 16, i64* [[TMP147]], align 8
1990 // CHECK2-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 8
1991 // CHECK2-NEXT:    store i8* null, i8** [[TMP148]], align 8
1992 // CHECK2-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
1993 // CHECK2-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
1994 // CHECK2-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
1995 // CHECK2-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1996 // CHECK2-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
1997 // CHECK2-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
1998 // CHECK2:       omp_offload.failed28:
1999 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
2000 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
2001 // CHECK2:       omp_offload.cont29:
2002 // CHECK2-NEXT:    br label [[OMP_IF_END31:%.*]]
2003 // CHECK2:       omp_if.else30:
2004 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
2005 // CHECK2-NEXT:    br label [[OMP_IF_END31]]
2006 // CHECK2:       omp_if.end31:
2007 // CHECK2-NEXT:    store i32 0, i32* [[NN]], align 4
2008 // CHECK2-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
2009 // CHECK2-NEXT:    [[CONV32:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
2010 // CHECK2-NEXT:    store i32 [[TMP154]], i32* [[CONV32]], align 4
2011 // CHECK2-NEXT:    [[TMP155:%.*]] = load i64, i64* [[NN_CASTED]], align 8
2012 // CHECK2-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
2013 // CHECK2-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i64*
2014 // CHECK2-NEXT:    store i64 [[TMP155]], i64* [[TMP157]], align 8
2015 // CHECK2-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
2016 // CHECK2-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i64*
2017 // CHECK2-NEXT:    store i64 [[TMP155]], i64* [[TMP159]], align 8
2018 // CHECK2-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 0
2019 // CHECK2-NEXT:    store i8* null, i8** [[TMP160]], align 8
2020 // CHECK2-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
2021 // CHECK2-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
2022 // CHECK2-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2023 // CHECK2-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
2024 // CHECK2-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
2025 // CHECK2:       omp_offload.failed36:
2026 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i64 [[TMP155]]) #[[ATTR3]]
2027 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
2028 // CHECK2:       omp_offload.cont37:
2029 // CHECK2-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
2030 // CHECK2-NEXT:    [[CONV39:%.*]] = bitcast i64* [[NN_CASTED38]] to i32*
2031 // CHECK2-NEXT:    store i32 [[TMP165]], i32* [[CONV39]], align 4
2032 // CHECK2-NEXT:    [[TMP166:%.*]] = load i64, i64* [[NN_CASTED38]], align 8
2033 // CHECK2-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
2034 // CHECK2-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i64*
2035 // CHECK2-NEXT:    store i64 [[TMP166]], i64* [[TMP168]], align 8
2036 // CHECK2-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
2037 // CHECK2-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
2038 // CHECK2-NEXT:    store i64 [[TMP166]], i64* [[TMP170]], align 8
2039 // CHECK2-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 0
2040 // CHECK2-NEXT:    store i8* null, i8** [[TMP171]], align 8
2041 // CHECK2-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
2042 // CHECK2-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
2043 // CHECK2-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2044 // CHECK2-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
2045 // CHECK2-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
2046 // CHECK2:       omp_offload.failed43:
2047 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i64 [[TMP166]]) #[[ATTR3]]
2048 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
2049 // CHECK2:       omp_offload.cont44:
2050 // CHECK2-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
2051 // CHECK2-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
2052 // CHECK2-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
2053 // CHECK2-NEXT:    ret i32 [[TMP176]]
2054 //
2055 //
2056 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
2057 // CHECK2-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
2058 // CHECK2-NEXT:  entry:
2059 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2060 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
2061 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
2062 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
2063 // CHECK2-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
2064 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2065 // CHECK2-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
2066 // CHECK2-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
2067 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2068 // CHECK2-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
2069 // CHECK2-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
2070 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
2071 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
2072 // CHECK2-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
2073 // CHECK2-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
2074 // CHECK2-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
2075 // CHECK2-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
2076 // CHECK2-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
2077 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
2078 // CHECK2-NEXT:    ret void
2079 //
2080 //
2081 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined.
2082 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
2083 // CHECK2-NEXT:  entry:
2084 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2085 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2086 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2087 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2088 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2089 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2090 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2091 // CHECK2-NEXT:    ret void
2092 //
2093 //
2094 // CHECK2-LABEL: define {{[^@]+}}@.omp_task_privates_map.
2095 // CHECK2-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
2096 // CHECK2-NEXT:  entry:
2097 // CHECK2-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
2098 // CHECK2-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 8
2099 // CHECK2-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 8
2100 // CHECK2-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 8
2101 // CHECK2-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 8
2102 // CHECK2-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
2103 // CHECK2-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 8
2104 // CHECK2-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 8
2105 // CHECK2-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 8
2106 // CHECK2-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 8
2107 // CHECK2-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
2108 // CHECK2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
2109 // CHECK2-NEXT:    [[TMP7:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 8
2110 // CHECK2-NEXT:    store [3 x i8*]* [[TMP6]], [3 x i8*]** [[TMP7]], align 8
2111 // CHECK2-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
2112 // CHECK2-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 8
2113 // CHECK2-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 8
2114 // CHECK2-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
2115 // CHECK2-NEXT:    [[TMP11:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 8
2116 // CHECK2-NEXT:    store [3 x i64]* [[TMP10]], [3 x i64]** [[TMP11]], align 8
2117 // CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
2118 // CHECK2-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 8
2119 // CHECK2-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 8
2120 // CHECK2-NEXT:    ret void
2121 //
2122 //
2123 // CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry.
2124 // CHECK2-SAME: (i32 signext [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
2125 // CHECK2-NEXT:  entry:
2126 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
2127 // CHECK2-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
2128 // CHECK2-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
2129 // CHECK2-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
2130 // CHECK2-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
2131 // CHECK2-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
2132 // CHECK2-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 8
2133 // CHECK2-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 8
2134 // CHECK2-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 8
2135 // CHECK2-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 8
2136 // CHECK2-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
2137 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
2138 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
2139 // CHECK2-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
2140 // CHECK2-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
2141 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
2142 // CHECK2-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
2143 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
2144 // CHECK2-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
2145 // CHECK2-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
2146 // CHECK2-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
2147 // CHECK2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
2148 // CHECK2-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
2149 // CHECK2-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
2150 // CHECK2-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
2151 // CHECK2-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
2152 // CHECK2-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
2153 // CHECK2-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META14:![0-9]+]])
2154 // CHECK2-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META17:![0-9]+]])
2155 // CHECK2-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META19:![0-9]+]])
2156 // CHECK2-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META21:![0-9]+]])
2157 // CHECK2-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !23
2158 // CHECK2-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !23
2159 // CHECK2-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
2160 // CHECK2-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
2161 // CHECK2-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !23
2162 // CHECK2-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
2163 // CHECK2-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
2164 // CHECK2-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
2165 // CHECK2-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
2166 // CHECK2-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
2167 // CHECK2-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
2168 // CHECK2-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !23
2169 // CHECK2-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 8, !noalias !23
2170 // CHECK2-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 8, !noalias !23
2171 // CHECK2-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 8, !noalias !23
2172 // CHECK2-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i64 0, i64 0
2173 // CHECK2-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i64 0, i64 0
2174 // CHECK2-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i64 0, i64 0
2175 // CHECK2-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
2176 // CHECK2-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
2177 // CHECK2-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
2178 // CHECK2-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
2179 // CHECK2-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
2180 // CHECK2-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
2181 // CHECK2-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
2182 // CHECK2:       omp_offload.failed.i:
2183 // CHECK2-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
2184 // CHECK2-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
2185 // CHECK2-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !23
2186 // CHECK2-NEXT:    [[TMP30:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !23
2187 // CHECK2-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
2188 // CHECK2-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
2189 // CHECK2-NEXT:    store i32 [[TMP31]], i32* [[CONV4_I]], align 4, !noalias !23
2190 // CHECK2-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !23
2191 // CHECK2-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
2192 // CHECK2-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
2193 // CHECK2-NEXT:    store i32 [[TMP33]], i32* [[CONV6_I]], align 4, !noalias !23
2194 // CHECK2-NEXT:    [[TMP34:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !23
2195 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i64 [[TMP30]], i64 [[TMP32]], i64 [[TMP34]]) #[[ATTR3]]
2196 // CHECK2-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
2197 // CHECK2:       .omp_outlined..1.exit:
2198 // CHECK2-NEXT:    ret i32 0
2199 //
2200 //
2201 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
2202 // CHECK2-SAME: (i64 [[A:%.*]]) #[[ATTR2]] {
2203 // CHECK2-NEXT:  entry:
2204 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2205 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
2206 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2207 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2208 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2209 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[A_CASTED]] to i32*
2210 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
2211 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
2212 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]])
2213 // CHECK2-NEXT:    ret void
2214 //
2215 //
2216 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..2
2217 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]]) #[[ATTR2]] {
2218 // CHECK2-NEXT:  entry:
2219 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2220 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2221 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2222 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2223 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2224 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2225 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2226 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2227 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
2228 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
2229 // CHECK2-NEXT:    ret void
2230 //
2231 //
2232 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
2233 // CHECK2-SAME: (i64 [[AA:%.*]]) #[[ATTR2]] {
2234 // CHECK2-NEXT:  entry:
2235 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2236 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
2237 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2238 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2239 // CHECK2-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
2240 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
2241 // CHECK2-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
2242 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
2243 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP1]])
2244 // CHECK2-NEXT:    ret void
2245 //
2246 //
2247 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..3
2248 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
2249 // CHECK2-NEXT:  entry:
2250 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2251 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2252 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2253 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2254 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2255 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2256 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2257 // CHECK2-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
2258 // CHECK2-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
2259 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
2260 // CHECK2-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
2261 // CHECK2-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
2262 // CHECK2-NEXT:    ret void
2263 //
2264 //
2265 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
2266 // CHECK2-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
2267 // CHECK2-NEXT:  entry:
2268 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2269 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2270 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
2271 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
2272 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2273 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2274 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2275 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2276 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2277 // CHECK2-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
2278 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
2279 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
2280 // CHECK2-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
2281 // CHECK2-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
2282 // CHECK2-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
2283 // CHECK2-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
2284 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
2285 // CHECK2-NEXT:    ret void
2286 //
2287 //
2288 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..6
2289 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
2290 // CHECK2-NEXT:  entry:
2291 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2292 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2293 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2294 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2295 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2296 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2297 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2298 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2299 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2300 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2301 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2302 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
2303 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
2304 // CHECK2-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
2305 // CHECK2-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
2306 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
2307 // CHECK2-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
2308 // CHECK2-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
2309 // CHECK2-NEXT:    ret void
2310 //
2311 //
2312 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
2313 // CHECK2-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
2314 // CHECK2-NEXT:  entry:
2315 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2316 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
2317 // CHECK2-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
2318 // CHECK2-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
2319 // CHECK2-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
2320 // CHECK2-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
2321 // CHECK2-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
2322 // CHECK2-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
2323 // CHECK2-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
2324 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
2325 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2326 // CHECK2-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
2327 // CHECK2-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
2328 // CHECK2-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
2329 // CHECK2-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
2330 // CHECK2-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
2331 // CHECK2-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
2332 // CHECK2-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
2333 // CHECK2-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
2334 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2335 // CHECK2-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
2336 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
2337 // CHECK2-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
2338 // CHECK2-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
2339 // CHECK2-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
2340 // CHECK2-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
2341 // CHECK2-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
2342 // CHECK2-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
2343 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
2344 // CHECK2-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
2345 // CHECK2-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
2346 // CHECK2-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
2347 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
2348 // CHECK2-NEXT:    ret void
2349 //
2350 //
2351 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..9
2352 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
2353 // CHECK2-NEXT:  entry:
2354 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2355 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2356 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2357 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
2358 // CHECK2-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
2359 // CHECK2-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
2360 // CHECK2-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
2361 // CHECK2-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
2362 // CHECK2-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
2363 // CHECK2-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
2364 // CHECK2-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
2365 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2366 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2367 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2368 // CHECK2-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
2369 // CHECK2-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
2370 // CHECK2-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
2371 // CHECK2-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
2372 // CHECK2-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
2373 // CHECK2-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
2374 // CHECK2-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
2375 // CHECK2-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
2376 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2377 // CHECK2-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
2378 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
2379 // CHECK2-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
2380 // CHECK2-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
2381 // CHECK2-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
2382 // CHECK2-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
2383 // CHECK2-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
2384 // CHECK2-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
2385 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
2386 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
2387 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
2388 // CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
2389 // CHECK2-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
2390 // CHECK2-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
2391 // CHECK2-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
2392 // CHECK2-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
2393 // CHECK2-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
2394 // CHECK2-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
2395 // CHECK2-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
2396 // CHECK2-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
2397 // CHECK2-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
2398 // CHECK2-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
2399 // CHECK2-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
2400 // CHECK2-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
2401 // CHECK2-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
2402 // CHECK2-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
2403 // CHECK2-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
2404 // CHECK2-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
2405 // CHECK2-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
2406 // CHECK2-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
2407 // CHECK2-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
2408 // CHECK2-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
2409 // CHECK2-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
2410 // CHECK2-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
2411 // CHECK2-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
2412 // CHECK2-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
2413 // CHECK2-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
2414 // CHECK2-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
2415 // CHECK2-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
2416 // CHECK2-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
2417 // CHECK2-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
2418 // CHECK2-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
2419 // CHECK2-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
2420 // CHECK2-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
2421 // CHECK2-NEXT:    ret void
2422 //
2423 //
2424 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
2425 // CHECK2-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
2426 // CHECK2-NEXT:  entry:
2427 // CHECK2-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
2428 // CHECK2-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
2429 // CHECK2-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
2430 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
2431 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2432 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
2433 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
2434 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
2435 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP1]])
2436 // CHECK2-NEXT:    ret void
2437 //
2438 //
2439 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..11
2440 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
2441 // CHECK2-NEXT:  entry:
2442 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2443 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2444 // CHECK2-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
2445 // CHECK2-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
2446 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2447 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2448 // CHECK2-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
2449 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
2450 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2451 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
2452 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
2453 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
2454 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i64 [[TMP1]])
2455 // CHECK2-NEXT:    ret void
2456 //
2457 //
2458 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..12
2459 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
2460 // CHECK2-NEXT:  entry:
2461 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2462 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2463 // CHECK2-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
2464 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2465 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2466 // CHECK2-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
2467 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
2468 // CHECK2-NEXT:    ret void
2469 //
2470 //
2471 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
2472 // CHECK2-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
2473 // CHECK2-NEXT:  entry:
2474 // CHECK2-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
2475 // CHECK2-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
2476 // CHECK2-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
2477 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
2478 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
2479 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
2480 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
2481 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
2482 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP1]])
2483 // CHECK2-NEXT:    ret void
2484 //
2485 //
2486 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..15
2487 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
2488 // CHECK2-NEXT:  entry:
2489 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2490 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2491 // CHECK2-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
2492 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2493 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2494 // CHECK2-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
2495 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
2496 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[CONV]])
2497 // CHECK2-NEXT:    ret void
2498 //
2499 //
2500 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..16
2501 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
2502 // CHECK2-NEXT:  entry:
2503 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2504 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2505 // CHECK2-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
2506 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2507 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2508 // CHECK2-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
2509 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
2510 // CHECK2-NEXT:    ret void
2511 //
2512 //
2513 // CHECK2-LABEL: define {{[^@]+}}@_Z6bazzzziPi
2514 // CHECK2-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
2515 // CHECK2-NEXT:  entry:
2516 // CHECK2-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
2517 // CHECK2-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
2518 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
2519 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
2520 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
2521 // CHECK2-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
2522 // CHECK2-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
2523 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
2524 // CHECK2-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
2525 // CHECK2-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2526 // CHECK2-NEXT:    [[TMP3:%.*]] = bitcast i8** [[TMP2]] to i64*
2527 // CHECK2-NEXT:    store i64 [[TMP1]], i64* [[TMP3]], align 8
2528 // CHECK2-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2529 // CHECK2-NEXT:    [[TMP5:%.*]] = bitcast i8** [[TMP4]] to i64*
2530 // CHECK2-NEXT:    store i64 [[TMP1]], i64* [[TMP5]], align 8
2531 // CHECK2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
2532 // CHECK2-NEXT:    store i8* null, i8** [[TMP6]], align 8
2533 // CHECK2-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2534 // CHECK2-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2535 // CHECK2-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2536 // CHECK2-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
2537 // CHECK2-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
2538 // CHECK2:       omp_offload.failed:
2539 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i64 [[TMP1]]) #[[ATTR3]]
2540 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
2541 // CHECK2:       omp_offload.cont:
2542 // CHECK2-NEXT:    ret void
2543 //
2544 //
2545 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
2546 // CHECK2-SAME: (i64 [[VLA:%.*]]) #[[ATTR2]] {
2547 // CHECK2-NEXT:  entry:
2548 // CHECK2-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
2549 // CHECK2-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
2550 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
2551 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i64 [[TMP0]])
2552 // CHECK2-NEXT:    ret void
2553 //
2554 //
2555 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..19
2556 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR2]] {
2557 // CHECK2-NEXT:  entry:
2558 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2559 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2560 // CHECK2-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
2561 // CHECK2-NEXT:    [[F:%.*]] = alloca i32*, align 8
2562 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2563 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2564 // CHECK2-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
2565 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
2566 // CHECK2-NEXT:    ret void
2567 //
2568 //
2569 // CHECK2-LABEL: define {{[^@]+}}@_Z3bari
2570 // CHECK2-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
2571 // CHECK2-NEXT:  entry:
2572 // CHECK2-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
2573 // CHECK2-NEXT:    [[A:%.*]] = alloca i32, align 4
2574 // CHECK2-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
2575 // CHECK2-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
2576 // CHECK2-NEXT:    store i32 0, i32* [[A]], align 4
2577 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
2578 // CHECK2-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
2579 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
2580 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
2581 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
2582 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
2583 // CHECK2-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
2584 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
2585 // CHECK2-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
2586 // CHECK2-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
2587 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
2588 // CHECK2-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
2589 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
2590 // CHECK2-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
2591 // CHECK2-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
2592 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
2593 // CHECK2-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
2594 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
2595 // CHECK2-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
2596 // CHECK2-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
2597 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
2598 // CHECK2-NEXT:    ret i32 [[TMP8]]
2599 //
2600 //
2601 // CHECK2-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
2602 // CHECK2-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
2603 // CHECK2-NEXT:  entry:
2604 // CHECK2-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
2605 // CHECK2-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
2606 // CHECK2-NEXT:    [[B:%.*]] = alloca i32, align 4
2607 // CHECK2-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
2608 // CHECK2-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
2609 // CHECK2-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
2610 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 8
2611 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 8
2612 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 8
2613 // CHECK2-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 8
2614 // CHECK2-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
2615 // CHECK2-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
2616 // CHECK2-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
2617 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
2618 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
2619 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
2620 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
2621 // CHECK2-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
2622 // CHECK2-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
2623 // CHECK2-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
2624 // CHECK2-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
2625 // CHECK2-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
2626 // CHECK2-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
2627 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
2628 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_CASTED]] to i32*
2629 // CHECK2-NEXT:    store i32 [[TMP5]], i32* [[CONV]], align 4
2630 // CHECK2-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
2631 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[N_ADDR]], align 4
2632 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP7]], 60
2633 // CHECK2-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
2634 // CHECK2:       omp_if.then:
2635 // CHECK2-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
2636 // CHECK2-NEXT:    [[TMP8:%.*]] = mul nuw i64 2, [[TMP2]]
2637 // CHECK2-NEXT:    [[TMP9:%.*]] = mul nuw i64 [[TMP8]], 2
2638 // CHECK2-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2639 // CHECK2-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
2640 // CHECK2-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 8
2641 // CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2642 // CHECK2-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
2643 // CHECK2-NEXT:    store double* [[A]], double** [[TMP13]], align 8
2644 // CHECK2-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
2645 // CHECK2-NEXT:    store i64 8, i64* [[TMP14]], align 8
2646 // CHECK2-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
2647 // CHECK2-NEXT:    store i8* null, i8** [[TMP15]], align 8
2648 // CHECK2-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
2649 // CHECK2-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
2650 // CHECK2-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
2651 // CHECK2-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
2652 // CHECK2-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i64*
2653 // CHECK2-NEXT:    store i64 [[TMP6]], i64* [[TMP19]], align 8
2654 // CHECK2-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
2655 // CHECK2-NEXT:    store i64 4, i64* [[TMP20]], align 8
2656 // CHECK2-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
2657 // CHECK2-NEXT:    store i8* null, i8** [[TMP21]], align 8
2658 // CHECK2-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
2659 // CHECK2-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
2660 // CHECK2-NEXT:    store i64 2, i64* [[TMP23]], align 8
2661 // CHECK2-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
2662 // CHECK2-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
2663 // CHECK2-NEXT:    store i64 2, i64* [[TMP25]], align 8
2664 // CHECK2-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
2665 // CHECK2-NEXT:    store i64 8, i64* [[TMP26]], align 8
2666 // CHECK2-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
2667 // CHECK2-NEXT:    store i8* null, i8** [[TMP27]], align 8
2668 // CHECK2-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
2669 // CHECK2-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
2670 // CHECK2-NEXT:    store i64 [[TMP2]], i64* [[TMP29]], align 8
2671 // CHECK2-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
2672 // CHECK2-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i64*
2673 // CHECK2-NEXT:    store i64 [[TMP2]], i64* [[TMP31]], align 8
2674 // CHECK2-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
2675 // CHECK2-NEXT:    store i64 8, i64* [[TMP32]], align 8
2676 // CHECK2-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
2677 // CHECK2-NEXT:    store i8* null, i8** [[TMP33]], align 8
2678 // CHECK2-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
2679 // CHECK2-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
2680 // CHECK2-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 8
2681 // CHECK2-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
2682 // CHECK2-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
2683 // CHECK2-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 8
2684 // CHECK2-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
2685 // CHECK2-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 8
2686 // CHECK2-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 4
2687 // CHECK2-NEXT:    store i8* null, i8** [[TMP39]], align 8
2688 // CHECK2-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2689 // CHECK2-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2690 // CHECK2-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
2691 // CHECK2-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2692 // CHECK2-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
2693 // CHECK2-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
2694 // CHECK2:       omp_offload.failed:
2695 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
2696 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
2697 // CHECK2:       omp_offload.cont:
2698 // CHECK2-NEXT:    br label [[OMP_IF_END:%.*]]
2699 // CHECK2:       omp_if.else:
2700 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
2701 // CHECK2-NEXT:    br label [[OMP_IF_END]]
2702 // CHECK2:       omp_if.end:
2703 // CHECK2-NEXT:    [[TMP45:%.*]] = mul nsw i64 1, [[TMP2]]
2704 // CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP45]]
2705 // CHECK2-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
2706 // CHECK2-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
2707 // CHECK2-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP46]] to i32
2708 // CHECK2-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
2709 // CHECK2-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP47]]
2710 // CHECK2-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
2711 // CHECK2-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
2712 // CHECK2-NEXT:    ret i32 [[ADD4]]
2713 //
2714 //
2715 // CHECK2-LABEL: define {{[^@]+}}@_ZL7fstatici
2716 // CHECK2-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
2717 // CHECK2-NEXT:  entry:
2718 // CHECK2-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
2719 // CHECK2-NEXT:    [[A:%.*]] = alloca i32, align 4
2720 // CHECK2-NEXT:    [[AA:%.*]] = alloca i16, align 2
2721 // CHECK2-NEXT:    [[AAA:%.*]] = alloca i8, align 1
2722 // CHECK2-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
2723 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
2724 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
2725 // CHECK2-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
2726 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 8
2727 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 8
2728 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 8
2729 // CHECK2-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
2730 // CHECK2-NEXT:    store i32 0, i32* [[A]], align 4
2731 // CHECK2-NEXT:    store i16 0, i16* [[AA]], align 2
2732 // CHECK2-NEXT:    store i8 0, i8* [[AAA]], align 1
2733 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
2734 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
2735 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
2736 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
2737 // CHECK2-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
2738 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
2739 // CHECK2-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
2740 // CHECK2-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
2741 // CHECK2-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
2742 // CHECK2-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
2743 // CHECK2-NEXT:    store i8 [[TMP4]], i8* [[CONV2]], align 1
2744 // CHECK2-NEXT:    [[TMP5:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
2745 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
2746 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
2747 // CHECK2-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
2748 // CHECK2:       omp_if.then:
2749 // CHECK2-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2750 // CHECK2-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
2751 // CHECK2-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
2752 // CHECK2-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2753 // CHECK2-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
2754 // CHECK2-NEXT:    store i64 [[TMP1]], i64* [[TMP10]], align 8
2755 // CHECK2-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
2756 // CHECK2-NEXT:    store i8* null, i8** [[TMP11]], align 8
2757 // CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
2758 // CHECK2-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
2759 // CHECK2-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
2760 // CHECK2-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
2761 // CHECK2-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
2762 // CHECK2-NEXT:    store i64 [[TMP3]], i64* [[TMP15]], align 8
2763 // CHECK2-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
2764 // CHECK2-NEXT:    store i8* null, i8** [[TMP16]], align 8
2765 // CHECK2-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
2766 // CHECK2-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
2767 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP18]], align 8
2768 // CHECK2-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
2769 // CHECK2-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
2770 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP20]], align 8
2771 // CHECK2-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
2772 // CHECK2-NEXT:    store i8* null, i8** [[TMP21]], align 8
2773 // CHECK2-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
2774 // CHECK2-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
2775 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 8
2776 // CHECK2-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
2777 // CHECK2-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
2778 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 8
2779 // CHECK2-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
2780 // CHECK2-NEXT:    store i8* null, i8** [[TMP26]], align 8
2781 // CHECK2-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2782 // CHECK2-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2783 // CHECK2-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2784 // CHECK2-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
2785 // CHECK2-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
2786 // CHECK2:       omp_offload.failed:
2787 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
2788 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
2789 // CHECK2:       omp_offload.cont:
2790 // CHECK2-NEXT:    br label [[OMP_IF_END:%.*]]
2791 // CHECK2:       omp_if.else:
2792 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
2793 // CHECK2-NEXT:    br label [[OMP_IF_END]]
2794 // CHECK2:       omp_if.end:
2795 // CHECK2-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
2796 // CHECK2-NEXT:    ret i32 [[TMP31]]
2797 //
2798 //
2799 // CHECK2-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
2800 // CHECK2-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
2801 // CHECK2-NEXT:  entry:
2802 // CHECK2-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
2803 // CHECK2-NEXT:    [[A:%.*]] = alloca i32, align 4
2804 // CHECK2-NEXT:    [[AA:%.*]] = alloca i16, align 2
2805 // CHECK2-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
2806 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
2807 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
2808 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
2809 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
2810 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
2811 // CHECK2-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
2812 // CHECK2-NEXT:    store i32 0, i32* [[A]], align 4
2813 // CHECK2-NEXT:    store i16 0, i16* [[AA]], align 2
2814 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
2815 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
2816 // CHECK2-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
2817 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
2818 // CHECK2-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
2819 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
2820 // CHECK2-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
2821 // CHECK2-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
2822 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
2823 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
2824 // CHECK2-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
2825 // CHECK2:       omp_if.then:
2826 // CHECK2-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2827 // CHECK2-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
2828 // CHECK2-NEXT:    store i64 [[TMP1]], i64* [[TMP6]], align 8
2829 // CHECK2-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2830 // CHECK2-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
2831 // CHECK2-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
2832 // CHECK2-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
2833 // CHECK2-NEXT:    store i8* null, i8** [[TMP9]], align 8
2834 // CHECK2-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
2835 // CHECK2-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i64*
2836 // CHECK2-NEXT:    store i64 [[TMP3]], i64* [[TMP11]], align 8
2837 // CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
2838 // CHECK2-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
2839 // CHECK2-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
2840 // CHECK2-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
2841 // CHECK2-NEXT:    store i8* null, i8** [[TMP14]], align 8
2842 // CHECK2-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
2843 // CHECK2-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
2844 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 8
2845 // CHECK2-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
2846 // CHECK2-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
2847 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 8
2848 // CHECK2-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
2849 // CHECK2-NEXT:    store i8* null, i8** [[TMP19]], align 8
2850 // CHECK2-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2851 // CHECK2-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2852 // CHECK2-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2853 // CHECK2-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
2854 // CHECK2-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
2855 // CHECK2:       omp_offload.failed:
2856 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
2857 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
2858 // CHECK2:       omp_offload.cont:
2859 // CHECK2-NEXT:    br label [[OMP_IF_END:%.*]]
2860 // CHECK2:       omp_if.else:
2861 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
2862 // CHECK2-NEXT:    br label [[OMP_IF_END]]
2863 // CHECK2:       omp_if.end:
2864 // CHECK2-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
2865 // CHECK2-NEXT:    ret i32 [[TMP24]]
2866 //
2867 //
2868 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
2869 // CHECK2-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
2870 // CHECK2-NEXT:  entry:
2871 // CHECK2-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
2872 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
2873 // CHECK2-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
2874 // CHECK2-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
2875 // CHECK2-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
2876 // CHECK2-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
2877 // CHECK2-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
2878 // CHECK2-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
2879 // CHECK2-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
2880 // CHECK2-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
2881 // CHECK2-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
2882 // CHECK2-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
2883 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
2884 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
2885 // CHECK2-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
2886 // CHECK2-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
2887 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
2888 // CHECK2-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
2889 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
2890 // CHECK2-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
2891 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
2892 // CHECK2-NEXT:    ret void
2893 //
2894 //
2895 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..22
2896 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
2897 // CHECK2-NEXT:  entry:
2898 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2899 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2900 // CHECK2-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
2901 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
2902 // CHECK2-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
2903 // CHECK2-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
2904 // CHECK2-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
2905 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2906 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2907 // CHECK2-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
2908 // CHECK2-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
2909 // CHECK2-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
2910 // CHECK2-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
2911 // CHECK2-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
2912 // CHECK2-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
2913 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
2914 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
2915 // CHECK2-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
2916 // CHECK2-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
2917 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
2918 // CHECK2-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
2919 // CHECK2-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
2920 // CHECK2-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
2921 // CHECK2-NEXT:    store double [[ADD]], double* [[A]], align 8
2922 // CHECK2-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
2923 // CHECK2-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
2924 // CHECK2-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
2925 // CHECK2-NEXT:    store double [[INC]], double* [[A4]], align 8
2926 // CHECK2-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
2927 // CHECK2-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
2928 // CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
2929 // CHECK2-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
2930 // CHECK2-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
2931 // CHECK2-NEXT:    ret void
2932 //
2933 //
2934 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
2935 // CHECK2-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
2936 // CHECK2-NEXT:  entry:
2937 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2938 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2939 // CHECK2-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
2940 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
2941 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
2942 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
2943 // CHECK2-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
2944 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2945 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2946 // CHECK2-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
2947 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
2948 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2949 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2950 // CHECK2-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
2951 // CHECK2-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
2952 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
2953 // CHECK2-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
2954 // CHECK2-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
2955 // CHECK2-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
2956 // CHECK2-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
2957 // CHECK2-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
2958 // CHECK2-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
2959 // CHECK2-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
2960 // CHECK2-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
2961 // CHECK2-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
2962 // CHECK2-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
2963 // CHECK2-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
2964 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
2965 // CHECK2-NEXT:    ret void
2966 //
2967 //
2968 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..24
2969 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
2970 // CHECK2-NEXT:  entry:
2971 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2972 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2973 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
2974 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
2975 // CHECK2-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
2976 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
2977 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2978 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2979 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
2980 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
2981 // CHECK2-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
2982 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
2983 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
2984 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
2985 // CHECK2-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
2986 // CHECK2-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
2987 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
2988 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
2989 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
2990 // CHECK2-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
2991 // CHECK2-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
2992 // CHECK2-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
2993 // CHECK2-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
2994 // CHECK2-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
2995 // CHECK2-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
2996 // CHECK2-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
2997 // CHECK2-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
2998 // CHECK2-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
2999 // CHECK2-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
3000 // CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
3001 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
3002 // CHECK2-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
3003 // CHECK2-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
3004 // CHECK2-NEXT:    ret void
3005 //
3006 //
3007 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
3008 // CHECK2-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
3009 // CHECK2-NEXT:  entry:
3010 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
3011 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
3012 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
3013 // CHECK2-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
3014 // CHECK2-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
3015 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
3016 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
3017 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
3018 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
3019 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
3020 // CHECK2-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
3021 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
3022 // CHECK2-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
3023 // CHECK2-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
3024 // CHECK2-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
3025 // CHECK2-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
3026 // CHECK2-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
3027 // CHECK2-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
3028 // CHECK2-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
3029 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
3030 // CHECK2-NEXT:    ret void
3031 //
3032 //
3033 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..27
3034 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
3035 // CHECK2-NEXT:  entry:
3036 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3037 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3038 // CHECK2-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
3039 // CHECK2-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
3040 // CHECK2-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
3041 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3042 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3043 // CHECK2-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
3044 // CHECK2-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
3045 // CHECK2-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
3046 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
3047 // CHECK2-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
3048 // CHECK2-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
3049 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
3050 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
3051 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
3052 // CHECK2-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
3053 // CHECK2-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
3054 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
3055 // CHECK2-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
3056 // CHECK2-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
3057 // CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
3058 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
3059 // CHECK2-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
3060 // CHECK2-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
3061 // CHECK2-NEXT:    ret void
3062 //
3063 //
3064 // CHECK2-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
3065 // CHECK2-SAME: () #[[ATTR4]] {
3066 // CHECK2-NEXT:  entry:
3067 // CHECK2-NEXT:    call void @__tgt_register_requires(i64 1)
3068 // CHECK2-NEXT:    ret void
3069 //
3070 //
3071 // CHECK3-LABEL: define {{[^@]+}}@_Z3fooi
3072 // CHECK3-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
3073 // CHECK3-NEXT:  entry:
3074 // CHECK3-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
3075 // CHECK3-NEXT:    [[A:%.*]] = alloca i32, align 4
3076 // CHECK3-NEXT:    [[AA:%.*]] = alloca i16, align 2
3077 // CHECK3-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
3078 // CHECK3-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
3079 // CHECK3-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
3080 // CHECK3-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
3081 // CHECK3-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
3082 // CHECK3-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
3083 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
3084 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
3085 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
3086 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
3087 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED3:%.*]] = alloca i32, align 4
3088 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
3089 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
3090 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
3091 // CHECK3-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
3092 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
3093 // CHECK3-NEXT:    [[AA_CASTED4:%.*]] = alloca i32, align 4
3094 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS6:%.*]] = alloca [1 x i8*], align 4
3095 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [1 x i8*], align 4
3096 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [1 x i8*], align 4
3097 // CHECK3-NEXT:    [[A_CASTED9:%.*]] = alloca i32, align 4
3098 // CHECK3-NEXT:    [[AA_CASTED10:%.*]] = alloca i32, align 4
3099 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS12:%.*]] = alloca [2 x i8*], align 4
3100 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
3101 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
3102 // CHECK3-NEXT:    [[A_CASTED17:%.*]] = alloca i32, align 4
3103 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [9 x i8*], align 4
3104 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [9 x i8*], align 4
3105 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [9 x i8*], align 4
3106 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
3107 // CHECK3-NEXT:    [[NN:%.*]] = alloca i32, align 4
3108 // CHECK3-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
3109 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
3110 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
3111 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
3112 // CHECK3-NEXT:    [[NN_CASTED32:%.*]] = alloca i32, align 4
3113 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 4
3114 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 4
3115 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 4
3116 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
3117 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
3118 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
3119 // CHECK3-NEXT:    store i16 0, i16* [[AA]], align 2
3120 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
3121 // CHECK3-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
3122 // CHECK3-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
3123 // CHECK3-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP1]], align 4
3124 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
3125 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
3126 // CHECK3-NEXT:    [[TMP4:%.*]] = mul nuw i32 5, [[TMP3]]
3127 // CHECK3-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP4]], align 8
3128 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[__VLA_EXPR1]], align 4
3129 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
3130 // CHECK3-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4
3131 // CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
3132 // CHECK3-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_2]], align 4
3133 // CHECK3-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
3134 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
3135 // CHECK3-NEXT:    store i16 [[TMP7]], i16* [[CONV]], align 2
3136 // CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[AA_CASTED]], align 4
3137 // CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
3138 // CHECK3-NEXT:    store i32 [[TMP9]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
3139 // CHECK3-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
3140 // CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
3141 // CHECK3-NEXT:    store i32 [[TMP11]], i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
3142 // CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
3143 // CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3144 // CHECK3-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i32*
3145 // CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP14]], align 4
3146 // CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3147 // CHECK3-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
3148 // CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP16]], align 4
3149 // CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
3150 // CHECK3-NEXT:    store i8* null, i8** [[TMP17]], align 4
3151 // CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
3152 // CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
3153 // CHECK3-NEXT:    store i32 [[TMP10]], i32* [[TMP19]], align 4
3154 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
3155 // CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
3156 // CHECK3-NEXT:    store i32 [[TMP10]], i32* [[TMP21]], align 4
3157 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
3158 // CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
3159 // CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
3160 // CHECK3-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
3161 // CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP24]], align 4
3162 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
3163 // CHECK3-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
3164 // CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP26]], align 4
3165 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
3166 // CHECK3-NEXT:    store i8* null, i8** [[TMP27]], align 4
3167 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3168 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3169 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
3170 // CHECK3-NEXT:    [[TMP31:%.*]] = load i16, i16* [[AA]], align 2
3171 // CHECK3-NEXT:    store i16 [[TMP31]], i16* [[TMP30]], align 4
3172 // CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
3173 // CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
3174 // CHECK3-NEXT:    store i32 [[TMP33]], i32* [[TMP32]], align 4
3175 // CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
3176 // CHECK3-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
3177 // CHECK3-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
3178 // CHECK3-NEXT:    [[TMP36:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
3179 // CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8* [[TMP36]] to %struct.kmp_task_t_with_privates*
3180 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 0
3181 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP38]], i32 0, i32 0
3182 // CHECK3-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[TMP39]], align 4
3183 // CHECK3-NEXT:    [[TMP41:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
3184 // CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP40]], i8* align 4 [[TMP41]], i32 12, i1 false)
3185 // CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 1
3186 // CHECK3-NEXT:    [[TMP43:%.*]] = bitcast i8* [[TMP40]] to %struct.anon*
3187 // CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 0
3188 // CHECK3-NEXT:    [[TMP45:%.*]] = bitcast [3 x i64]* [[TMP44]] to i8*
3189 // CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP45]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
3190 // CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 1
3191 // CHECK3-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
3192 // CHECK3-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP28]] to i8*
3193 // CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP47]], i8* align 4 [[TMP48]], i32 12, i1 false)
3194 // CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 2
3195 // CHECK3-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
3196 // CHECK3-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP29]] to i8*
3197 // CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i32 12, i1 false)
3198 // CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 3
3199 // CHECK3-NEXT:    [[TMP53:%.*]] = load i16, i16* [[AA]], align 2
3200 // CHECK3-NEXT:    store i16 [[TMP53]], i16* [[TMP52]], align 4
3201 // CHECK3-NEXT:    [[TMP54:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP36]])
3202 // CHECK3-NEXT:    [[TMP55:%.*]] = load i32, i32* [[A]], align 4
3203 // CHECK3-NEXT:    store i32 [[TMP55]], i32* [[A_CASTED]], align 4
3204 // CHECK3-NEXT:    [[TMP56:%.*]] = load i32, i32* [[A_CASTED]], align 4
3205 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i32 [[TMP56]]) #[[ATTR3:[0-9]+]]
3206 // CHECK3-NEXT:    [[TMP57:%.*]] = load i16, i16* [[AA]], align 2
3207 // CHECK3-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
3208 // CHECK3-NEXT:    store i16 [[TMP57]], i16* [[CONV5]], align 2
3209 // CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
3210 // CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
3211 // CHECK3-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32*
3212 // CHECK3-NEXT:    store i32 [[TMP58]], i32* [[TMP60]], align 4
3213 // CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
3214 // CHECK3-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32*
3215 // CHECK3-NEXT:    store i32 [[TMP58]], i32* [[TMP62]], align 4
3216 // CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
3217 // CHECK3-NEXT:    store i8* null, i8** [[TMP63]], align 4
3218 // CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
3219 // CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
3220 // CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3221 // CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
3222 // CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
3223 // CHECK3:       omp_offload.failed:
3224 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR3]]
3225 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
3226 // CHECK3:       omp_offload.cont:
3227 // CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
3228 // CHECK3-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
3229 // CHECK3-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
3230 // CHECK3-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
3231 // CHECK3-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
3232 // CHECK3-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
3233 // CHECK3-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
3234 // CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
3235 // CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
3236 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
3237 // CHECK3:       omp_if.then:
3238 // CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
3239 // CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
3240 // CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
3241 // CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
3242 // CHECK3-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
3243 // CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
3244 // CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
3245 // CHECK3-NEXT:    store i8* null, i8** [[TMP77]], align 4
3246 // CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
3247 // CHECK3-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
3248 // CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
3249 // CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
3250 // CHECK3-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
3251 // CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
3252 // CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
3253 // CHECK3-NEXT:    store i8* null, i8** [[TMP82]], align 4
3254 // CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
3255 // CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
3256 // CHECK3-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3257 // CHECK3-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
3258 // CHECK3-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
3259 // CHECK3:       omp_offload.failed15:
3260 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
3261 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
3262 // CHECK3:       omp_offload.cont16:
3263 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
3264 // CHECK3:       omp_if.else:
3265 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
3266 // CHECK3-NEXT:    br label [[OMP_IF_END]]
3267 // CHECK3:       omp_if.end:
3268 // CHECK3-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
3269 // CHECK3-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED17]], align 4
3270 // CHECK3-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED17]], align 4
3271 // CHECK3-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
3272 // CHECK3-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP89]], 20
3273 // CHECK3-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
3274 // CHECK3:       omp_if.then19:
3275 // CHECK3-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
3276 // CHECK3-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
3277 // CHECK3-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
3278 // CHECK3-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
3279 // CHECK3-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
3280 // CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
3281 // CHECK3-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
3282 // CHECK3-NEXT:    store i32 [[TMP88]], i32* [[TMP96]], align 4
3283 // CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
3284 // CHECK3-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
3285 // CHECK3-NEXT:    store i32 [[TMP88]], i32* [[TMP98]], align 4
3286 // CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
3287 // CHECK3-NEXT:    store i64 4, i64* [[TMP99]], align 4
3288 // CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
3289 // CHECK3-NEXT:    store i8* null, i8** [[TMP100]], align 4
3290 // CHECK3-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
3291 // CHECK3-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
3292 // CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
3293 // CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
3294 // CHECK3-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
3295 // CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
3296 // CHECK3-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
3297 // CHECK3-NEXT:    store i64 40, i64* [[TMP105]], align 4
3298 // CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
3299 // CHECK3-NEXT:    store i8* null, i8** [[TMP106]], align 4
3300 // CHECK3-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
3301 // CHECK3-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
3302 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP108]], align 4
3303 // CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
3304 // CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
3305 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP110]], align 4
3306 // CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
3307 // CHECK3-NEXT:    store i64 4, i64* [[TMP111]], align 4
3308 // CHECK3-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
3309 // CHECK3-NEXT:    store i8* null, i8** [[TMP112]], align 4
3310 // CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
3311 // CHECK3-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
3312 // CHECK3-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
3313 // CHECK3-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
3314 // CHECK3-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
3315 // CHECK3-NEXT:    store float* [[VLA]], float** [[TMP116]], align 4
3316 // CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
3317 // CHECK3-NEXT:    store i64 [[TMP91]], i64* [[TMP117]], align 4
3318 // CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
3319 // CHECK3-NEXT:    store i8* null, i8** [[TMP118]], align 4
3320 // CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
3321 // CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
3322 // CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
3323 // CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
3324 // CHECK3-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
3325 // CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 4
3326 // CHECK3-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
3327 // CHECK3-NEXT:    store i64 400, i64* [[TMP123]], align 4
3328 // CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 4
3329 // CHECK3-NEXT:    store i8* null, i8** [[TMP124]], align 4
3330 // CHECK3-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
3331 // CHECK3-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i32*
3332 // CHECK3-NEXT:    store i32 5, i32* [[TMP126]], align 4
3333 // CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
3334 // CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
3335 // CHECK3-NEXT:    store i32 5, i32* [[TMP128]], align 4
3336 // CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
3337 // CHECK3-NEXT:    store i64 4, i64* [[TMP129]], align 4
3338 // CHECK3-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 5
3339 // CHECK3-NEXT:    store i8* null, i8** [[TMP130]], align 4
3340 // CHECK3-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
3341 // CHECK3-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
3342 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP132]], align 4
3343 // CHECK3-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
3344 // CHECK3-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
3345 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP134]], align 4
3346 // CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
3347 // CHECK3-NEXT:    store i64 4, i64* [[TMP135]], align 4
3348 // CHECK3-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 6
3349 // CHECK3-NEXT:    store i8* null, i8** [[TMP136]], align 4
3350 // CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
3351 // CHECK3-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
3352 // CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 4
3353 // CHECK3-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
3354 // CHECK3-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
3355 // CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 4
3356 // CHECK3-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
3357 // CHECK3-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 4
3358 // CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 7
3359 // CHECK3-NEXT:    store i8* null, i8** [[TMP142]], align 4
3360 // CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
3361 // CHECK3-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
3362 // CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 4
3363 // CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
3364 // CHECK3-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
3365 // CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 4
3366 // CHECK3-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
3367 // CHECK3-NEXT:    store i64 12, i64* [[TMP147]], align 4
3368 // CHECK3-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 8
3369 // CHECK3-NEXT:    store i8* null, i8** [[TMP148]], align 4
3370 // CHECK3-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
3371 // CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
3372 // CHECK3-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
3373 // CHECK3-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3374 // CHECK3-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
3375 // CHECK3-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
3376 // CHECK3:       omp_offload.failed23:
3377 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
3378 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
3379 // CHECK3:       omp_offload.cont24:
3380 // CHECK3-NEXT:    br label [[OMP_IF_END26:%.*]]
3381 // CHECK3:       omp_if.else25:
3382 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
3383 // CHECK3-NEXT:    br label [[OMP_IF_END26]]
3384 // CHECK3:       omp_if.end26:
3385 // CHECK3-NEXT:    store i32 0, i32* [[NN]], align 4
3386 // CHECK3-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
3387 // CHECK3-NEXT:    store i32 [[TMP154]], i32* [[NN_CASTED]], align 4
3388 // CHECK3-NEXT:    [[TMP155:%.*]] = load i32, i32* [[NN_CASTED]], align 4
3389 // CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
3390 // CHECK3-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
3391 // CHECK3-NEXT:    store i32 [[TMP155]], i32* [[TMP157]], align 4
3392 // CHECK3-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
3393 // CHECK3-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32*
3394 // CHECK3-NEXT:    store i32 [[TMP155]], i32* [[TMP159]], align 4
3395 // CHECK3-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
3396 // CHECK3-NEXT:    store i8* null, i8** [[TMP160]], align 4
3397 // CHECK3-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
3398 // CHECK3-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
3399 // CHECK3-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3400 // CHECK3-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
3401 // CHECK3-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
3402 // CHECK3:       omp_offload.failed30:
3403 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i32 [[TMP155]]) #[[ATTR3]]
3404 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
3405 // CHECK3:       omp_offload.cont31:
3406 // CHECK3-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
3407 // CHECK3-NEXT:    store i32 [[TMP165]], i32* [[NN_CASTED32]], align 4
3408 // CHECK3-NEXT:    [[TMP166:%.*]] = load i32, i32* [[NN_CASTED32]], align 4
3409 // CHECK3-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
3410 // CHECK3-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i32*
3411 // CHECK3-NEXT:    store i32 [[TMP166]], i32* [[TMP168]], align 4
3412 // CHECK3-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
3413 // CHECK3-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i32*
3414 // CHECK3-NEXT:    store i32 [[TMP166]], i32* [[TMP170]], align 4
3415 // CHECK3-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
3416 // CHECK3-NEXT:    store i8* null, i8** [[TMP171]], align 4
3417 // CHECK3-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
3418 // CHECK3-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
3419 // CHECK3-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3420 // CHECK3-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
3421 // CHECK3-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
3422 // CHECK3:       omp_offload.failed36:
3423 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i32 [[TMP166]]) #[[ATTR3]]
3424 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
3425 // CHECK3:       omp_offload.cont37:
3426 // CHECK3-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
3427 // CHECK3-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
3428 // CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
3429 // CHECK3-NEXT:    ret i32 [[TMP176]]
3430 //
3431 //
3432 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
3433 // CHECK3-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
3434 // CHECK3-NEXT:  entry:
3435 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
3436 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
3437 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
3438 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
3439 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
3440 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
3441 // CHECK3-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
3442 // CHECK3-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
3443 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
3444 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
3445 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
3446 // CHECK3-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
3447 // CHECK3-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
3448 // CHECK3-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
3449 // CHECK3-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
3450 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
3451 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
3452 // CHECK3-NEXT:    ret void
3453 //
3454 //
3455 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined.
3456 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
3457 // CHECK3-NEXT:  entry:
3458 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3459 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3460 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
3461 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3462 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3463 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
3464 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
3465 // CHECK3-NEXT:    ret void
3466 //
3467 //
3468 // CHECK3-LABEL: define {{[^@]+}}@.omp_task_privates_map.
3469 // CHECK3-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
3470 // CHECK3-NEXT:  entry:
3471 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 4
3472 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 4
3473 // CHECK3-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 4
3474 // CHECK3-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 4
3475 // CHECK3-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 4
3476 // CHECK3-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 4
3477 // CHECK3-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 4
3478 // CHECK3-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 4
3479 // CHECK3-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 4
3480 // CHECK3-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 4
3481 // CHECK3-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 4
3482 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
3483 // CHECK3-NEXT:    [[TMP7:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 4
3484 // CHECK3-NEXT:    store [3 x i64]* [[TMP6]], [3 x i64]** [[TMP7]], align 4
3485 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
3486 // CHECK3-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 4
3487 // CHECK3-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 4
3488 // CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
3489 // CHECK3-NEXT:    [[TMP11:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 4
3490 // CHECK3-NEXT:    store [3 x i8*]* [[TMP10]], [3 x i8*]** [[TMP11]], align 4
3491 // CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
3492 // CHECK3-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 4
3493 // CHECK3-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 4
3494 // CHECK3-NEXT:    ret void
3495 //
3496 //
3497 // CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry.
3498 // CHECK3-SAME: (i32 [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
3499 // CHECK3-NEXT:  entry:
3500 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
3501 // CHECK3-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 4
3502 // CHECK3-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 4
3503 // CHECK3-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
3504 // CHECK3-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
3505 // CHECK3-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
3506 // CHECK3-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 4
3507 // CHECK3-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 4
3508 // CHECK3-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 4
3509 // CHECK3-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 4
3510 // CHECK3-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
3511 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
3512 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
3513 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
3514 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
3515 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
3516 // CHECK3-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
3517 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
3518 // CHECK3-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
3519 // CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
3520 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
3521 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
3522 // CHECK3-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 4
3523 // CHECK3-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
3524 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
3525 // CHECK3-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
3526 // CHECK3-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
3527 // CHECK3-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META15:![0-9]+]])
3528 // CHECK3-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
3529 // CHECK3-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
3530 // CHECK3-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META22:![0-9]+]])
3531 // CHECK3-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !24
3532 // CHECK3-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 4, !noalias !24
3533 // CHECK3-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
3534 // CHECK3-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
3535 // CHECK3-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !24
3536 // CHECK3-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
3537 // CHECK3-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
3538 // CHECK3-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
3539 // CHECK3-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
3540 // CHECK3-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
3541 // CHECK3-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
3542 // CHECK3-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 4, !noalias !24
3543 // CHECK3-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 4, !noalias !24
3544 // CHECK3-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 4, !noalias !24
3545 // CHECK3-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 4, !noalias !24
3546 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i32 0, i32 0
3547 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i32 0, i32 0
3548 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i32 0, i32 0
3549 // CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
3550 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
3551 // CHECK3-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
3552 // CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
3553 // CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
3554 // CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
3555 // CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
3556 // CHECK3:       omp_offload.failed.i:
3557 // CHECK3-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
3558 // CHECK3-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
3559 // CHECK3-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !24
3560 // CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !24
3561 // CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
3562 // CHECK3-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
3563 // CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
3564 // CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
3565 // CHECK3-NEXT:    store i32 [[TMP33]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
3566 // CHECK3-NEXT:    [[TMP34:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
3567 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i32 [[TMP30]], i32 [[TMP32]], i32 [[TMP34]]) #[[ATTR3]]
3568 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
3569 // CHECK3:       .omp_outlined..1.exit:
3570 // CHECK3-NEXT:    ret i32 0
3571 //
3572 //
3573 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
3574 // CHECK3-SAME: (i32 [[A:%.*]]) #[[ATTR2]] {
3575 // CHECK3-NEXT:  entry:
3576 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
3577 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
3578 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
3579 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
3580 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
3581 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
3582 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]])
3583 // CHECK3-NEXT:    ret void
3584 //
3585 //
3586 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..2
3587 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]]) #[[ATTR2]] {
3588 // CHECK3-NEXT:  entry:
3589 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3590 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3591 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
3592 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3593 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3594 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
3595 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
3596 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
3597 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
3598 // CHECK3-NEXT:    ret void
3599 //
3600 //
3601 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
3602 // CHECK3-SAME: (i32 [[AA:%.*]]) #[[ATTR2]] {
3603 // CHECK3-NEXT:  entry:
3604 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
3605 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
3606 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
3607 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
3608 // CHECK3-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
3609 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
3610 // CHECK3-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
3611 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
3612 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP1]])
3613 // CHECK3-NEXT:    ret void
3614 //
3615 //
3616 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..3
3617 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
3618 // CHECK3-NEXT:  entry:
3619 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3620 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3621 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
3622 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3623 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3624 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
3625 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
3626 // CHECK3-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
3627 // CHECK3-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
3628 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
3629 // CHECK3-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
3630 // CHECK3-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
3631 // CHECK3-NEXT:    ret void
3632 //
3633 //
3634 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
3635 // CHECK3-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
3636 // CHECK3-NEXT:  entry:
3637 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
3638 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
3639 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
3640 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
3641 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
3642 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
3643 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
3644 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
3645 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
3646 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
3647 // CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
3648 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
3649 // CHECK3-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
3650 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
3651 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
3652 // CHECK3-NEXT:    ret void
3653 //
3654 //
3655 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..6
3656 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
3657 // CHECK3-NEXT:  entry:
3658 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3659 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3660 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
3661 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
3662 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3663 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3664 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
3665 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
3666 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
3667 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
3668 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
3669 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
3670 // CHECK3-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
3671 // CHECK3-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
3672 // CHECK3-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
3673 // CHECK3-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
3674 // CHECK3-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
3675 // CHECK3-NEXT:    ret void
3676 //
3677 //
3678 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
3679 // CHECK3-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
3680 // CHECK3-NEXT:  entry:
3681 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
3682 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
3683 // CHECK3-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
3684 // CHECK3-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
3685 // CHECK3-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
3686 // CHECK3-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
3687 // CHECK3-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
3688 // CHECK3-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
3689 // CHECK3-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
3690 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
3691 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
3692 // CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
3693 // CHECK3-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
3694 // CHECK3-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
3695 // CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
3696 // CHECK3-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
3697 // CHECK3-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
3698 // CHECK3-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
3699 // CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
3700 // CHECK3-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
3701 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
3702 // CHECK3-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
3703 // CHECK3-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
3704 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
3705 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
3706 // CHECK3-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
3707 // CHECK3-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
3708 // CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
3709 // CHECK3-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
3710 // CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
3711 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
3712 // CHECK3-NEXT:    ret void
3713 //
3714 //
3715 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..9
3716 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
3717 // CHECK3-NEXT:  entry:
3718 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3719 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3720 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
3721 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
3722 // CHECK3-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
3723 // CHECK3-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
3724 // CHECK3-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
3725 // CHECK3-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
3726 // CHECK3-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
3727 // CHECK3-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
3728 // CHECK3-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
3729 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3730 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3731 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
3732 // CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
3733 // CHECK3-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
3734 // CHECK3-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
3735 // CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
3736 // CHECK3-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
3737 // CHECK3-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
3738 // CHECK3-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
3739 // CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
3740 // CHECK3-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
3741 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
3742 // CHECK3-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
3743 // CHECK3-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
3744 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
3745 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
3746 // CHECK3-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
3747 // CHECK3-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
3748 // CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
3749 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
3750 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
3751 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
3752 // CHECK3-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
3753 // CHECK3-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
3754 // CHECK3-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
3755 // CHECK3-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
3756 // CHECK3-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
3757 // CHECK3-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
3758 // CHECK3-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
3759 // CHECK3-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
3760 // CHECK3-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
3761 // CHECK3-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
3762 // CHECK3-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
3763 // CHECK3-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
3764 // CHECK3-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
3765 // CHECK3-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
3766 // CHECK3-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
3767 // CHECK3-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
3768 // CHECK3-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
3769 // CHECK3-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
3770 // CHECK3-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
3771 // CHECK3-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
3772 // CHECK3-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
3773 // CHECK3-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
3774 // CHECK3-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
3775 // CHECK3-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
3776 // CHECK3-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
3777 // CHECK3-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
3778 // CHECK3-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
3779 // CHECK3-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
3780 // CHECK3-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
3781 // CHECK3-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
3782 // CHECK3-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
3783 // CHECK3-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
3784 // CHECK3-NEXT:    ret void
3785 //
3786 //
3787 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
3788 // CHECK3-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
3789 // CHECK3-NEXT:  entry:
3790 // CHECK3-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
3791 // CHECK3-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
3792 // CHECK3-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
3793 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
3794 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
3795 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
3796 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP1]])
3797 // CHECK3-NEXT:    ret void
3798 //
3799 //
3800 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..11
3801 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
3802 // CHECK3-NEXT:  entry:
3803 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3804 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3805 // CHECK3-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
3806 // CHECK3-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
3807 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3808 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3809 // CHECK3-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
3810 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
3811 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
3812 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
3813 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i32 [[TMP1]])
3814 // CHECK3-NEXT:    ret void
3815 //
3816 //
3817 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..12
3818 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
3819 // CHECK3-NEXT:  entry:
3820 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3821 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3822 // CHECK3-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
3823 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3824 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3825 // CHECK3-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
3826 // CHECK3-NEXT:    ret void
3827 //
3828 //
3829 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
3830 // CHECK3-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
3831 // CHECK3-NEXT:  entry:
3832 // CHECK3-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
3833 // CHECK3-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
3834 // CHECK3-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
3835 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
3836 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
3837 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
3838 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i32 [[TMP1]])
3839 // CHECK3-NEXT:    ret void
3840 //
3841 //
3842 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..15
3843 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
3844 // CHECK3-NEXT:  entry:
3845 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3846 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3847 // CHECK3-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
3848 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3849 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3850 // CHECK3-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
3851 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
3852 // CHECK3-NEXT:    ret void
3853 //
3854 //
3855 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..16
3856 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
3857 // CHECK3-NEXT:  entry:
3858 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3859 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3860 // CHECK3-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
3861 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3862 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3863 // CHECK3-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
3864 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
3865 // CHECK3-NEXT:    ret void
3866 //
3867 //
3868 // CHECK3-LABEL: define {{[^@]+}}@_Z6bazzzziPi
3869 // CHECK3-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
3870 // CHECK3-NEXT:  entry:
3871 // CHECK3-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
3872 // CHECK3-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
3873 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
3874 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
3875 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
3876 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
3877 // CHECK3-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
3878 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
3879 // CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3880 // CHECK3-NEXT:    [[TMP2:%.*]] = bitcast i8** [[TMP1]] to i32*
3881 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[TMP2]], align 4
3882 // CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3883 // CHECK3-NEXT:    [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32*
3884 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[TMP4]], align 4
3885 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
3886 // CHECK3-NEXT:    store i8* null, i8** [[TMP5]], align 4
3887 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3888 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3889 // CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3890 // CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
3891 // CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
3892 // CHECK3:       omp_offload.failed:
3893 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i32 [[TMP0]]) #[[ATTR3]]
3894 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
3895 // CHECK3:       omp_offload.cont:
3896 // CHECK3-NEXT:    ret void
3897 //
3898 //
3899 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
3900 // CHECK3-SAME: (i32 [[VLA:%.*]]) #[[ATTR2]] {
3901 // CHECK3-NEXT:  entry:
3902 // CHECK3-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
3903 // CHECK3-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
3904 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
3905 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i32 [[TMP0]])
3906 // CHECK3-NEXT:    ret void
3907 //
3908 //
3909 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..19
3910 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR2]] {
3911 // CHECK3-NEXT:  entry:
3912 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
3913 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
3914 // CHECK3-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
3915 // CHECK3-NEXT:    [[F:%.*]] = alloca i32*, align 4
3916 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
3917 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
3918 // CHECK3-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
3919 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
3920 // CHECK3-NEXT:    ret void
3921 //
3922 //
3923 // CHECK3-LABEL: define {{[^@]+}}@_Z3bari
3924 // CHECK3-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
3925 // CHECK3-NEXT:  entry:
3926 // CHECK3-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
3927 // CHECK3-NEXT:    [[A:%.*]] = alloca i32, align 4
3928 // CHECK3-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
3929 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
3930 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
3931 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
3932 // CHECK3-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
3933 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
3934 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
3935 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
3936 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
3937 // CHECK3-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
3938 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
3939 // CHECK3-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
3940 // CHECK3-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
3941 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
3942 // CHECK3-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
3943 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
3944 // CHECK3-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
3945 // CHECK3-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
3946 // CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
3947 // CHECK3-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
3948 // CHECK3-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
3949 // CHECK3-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
3950 // CHECK3-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
3951 // CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
3952 // CHECK3-NEXT:    ret i32 [[TMP8]]
3953 //
3954 //
3955 // CHECK3-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
3956 // CHECK3-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
3957 // CHECK3-NEXT:  entry:
3958 // CHECK3-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
3959 // CHECK3-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
3960 // CHECK3-NEXT:    [[B:%.*]] = alloca i32, align 4
3961 // CHECK3-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
3962 // CHECK3-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
3963 // CHECK3-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
3964 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 4
3965 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 4
3966 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 4
3967 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 4
3968 // CHECK3-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
3969 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
3970 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
3971 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
3972 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
3973 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
3974 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
3975 // CHECK3-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
3976 // CHECK3-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
3977 // CHECK3-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
3978 // CHECK3-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
3979 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
3980 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
3981 // CHECK3-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
3982 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
3983 // CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
3984 // CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 60
3985 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
3986 // CHECK3:       omp_if.then:
3987 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
3988 // CHECK3-NEXT:    [[TMP7:%.*]] = mul nuw i32 2, [[TMP1]]
3989 // CHECK3-NEXT:    [[TMP8:%.*]] = mul nuw i32 [[TMP7]], 2
3990 // CHECK3-NEXT:    [[TMP9:%.*]] = sext i32 [[TMP8]] to i64
3991 // CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3992 // CHECK3-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
3993 // CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 4
3994 // CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3995 // CHECK3-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
3996 // CHECK3-NEXT:    store double* [[A]], double** [[TMP13]], align 4
3997 // CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
3998 // CHECK3-NEXT:    store i64 8, i64* [[TMP14]], align 4
3999 // CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
4000 // CHECK3-NEXT:    store i8* null, i8** [[TMP15]], align 4
4001 // CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
4002 // CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
4003 // CHECK3-NEXT:    store i32 [[TMP5]], i32* [[TMP17]], align 4
4004 // CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
4005 // CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
4006 // CHECK3-NEXT:    store i32 [[TMP5]], i32* [[TMP19]], align 4
4007 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
4008 // CHECK3-NEXT:    store i64 4, i64* [[TMP20]], align 4
4009 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
4010 // CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
4011 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
4012 // CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
4013 // CHECK3-NEXT:    store i32 2, i32* [[TMP23]], align 4
4014 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
4015 // CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
4016 // CHECK3-NEXT:    store i32 2, i32* [[TMP25]], align 4
4017 // CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
4018 // CHECK3-NEXT:    store i64 4, i64* [[TMP26]], align 4
4019 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
4020 // CHECK3-NEXT:    store i8* null, i8** [[TMP27]], align 4
4021 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
4022 // CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
4023 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP29]], align 4
4024 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
4025 // CHECK3-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
4026 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP31]], align 4
4027 // CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
4028 // CHECK3-NEXT:    store i64 4, i64* [[TMP32]], align 4
4029 // CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
4030 // CHECK3-NEXT:    store i8* null, i8** [[TMP33]], align 4
4031 // CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
4032 // CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
4033 // CHECK3-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 4
4034 // CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
4035 // CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
4036 // CHECK3-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 4
4037 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
4038 // CHECK3-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 4
4039 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 4
4040 // CHECK3-NEXT:    store i8* null, i8** [[TMP39]], align 4
4041 // CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4042 // CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4043 // CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
4044 // CHECK3-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4045 // CHECK3-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
4046 // CHECK3-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
4047 // CHECK3:       omp_offload.failed:
4048 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
4049 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
4050 // CHECK3:       omp_offload.cont:
4051 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
4052 // CHECK3:       omp_if.else:
4053 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
4054 // CHECK3-NEXT:    br label [[OMP_IF_END]]
4055 // CHECK3:       omp_if.end:
4056 // CHECK3-NEXT:    [[TMP45:%.*]] = mul nsw i32 1, [[TMP1]]
4057 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP45]]
4058 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
4059 // CHECK3-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
4060 // CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP46]] to i32
4061 // CHECK3-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
4062 // CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP47]]
4063 // CHECK3-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
4064 // CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
4065 // CHECK3-NEXT:    ret i32 [[ADD3]]
4066 //
4067 //
4068 // CHECK3-LABEL: define {{[^@]+}}@_ZL7fstatici
4069 // CHECK3-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
4070 // CHECK3-NEXT:  entry:
4071 // CHECK3-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
4072 // CHECK3-NEXT:    [[A:%.*]] = alloca i32, align 4
4073 // CHECK3-NEXT:    [[AA:%.*]] = alloca i16, align 2
4074 // CHECK3-NEXT:    [[AAA:%.*]] = alloca i8, align 1
4075 // CHECK3-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
4076 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4077 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4078 // CHECK3-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
4079 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 4
4080 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 4
4081 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 4
4082 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
4083 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
4084 // CHECK3-NEXT:    store i16 0, i16* [[AA]], align 2
4085 // CHECK3-NEXT:    store i8 0, i8* [[AAA]], align 1
4086 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
4087 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
4088 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
4089 // CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
4090 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4091 // CHECK3-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
4092 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4093 // CHECK3-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
4094 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
4095 // CHECK3-NEXT:    store i8 [[TMP4]], i8* [[CONV1]], align 1
4096 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
4097 // CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
4098 // CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
4099 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
4100 // CHECK3:       omp_if.then:
4101 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4102 // CHECK3-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
4103 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
4104 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4105 // CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
4106 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP10]], align 4
4107 // CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
4108 // CHECK3-NEXT:    store i8* null, i8** [[TMP11]], align 4
4109 // CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
4110 // CHECK3-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
4111 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
4112 // CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
4113 // CHECK3-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
4114 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP15]], align 4
4115 // CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
4116 // CHECK3-NEXT:    store i8* null, i8** [[TMP16]], align 4
4117 // CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
4118 // CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
4119 // CHECK3-NEXT:    store i32 [[TMP5]], i32* [[TMP18]], align 4
4120 // CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
4121 // CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
4122 // CHECK3-NEXT:    store i32 [[TMP5]], i32* [[TMP20]], align 4
4123 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
4124 // CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
4125 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
4126 // CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
4127 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 4
4128 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
4129 // CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
4130 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 4
4131 // CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
4132 // CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
4133 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4134 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4135 // CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4136 // CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
4137 // CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
4138 // CHECK3:       omp_offload.failed:
4139 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
4140 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
4141 // CHECK3:       omp_offload.cont:
4142 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
4143 // CHECK3:       omp_if.else:
4144 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
4145 // CHECK3-NEXT:    br label [[OMP_IF_END]]
4146 // CHECK3:       omp_if.end:
4147 // CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
4148 // CHECK3-NEXT:    ret i32 [[TMP31]]
4149 //
4150 //
4151 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
4152 // CHECK3-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
4153 // CHECK3-NEXT:  entry:
4154 // CHECK3-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
4155 // CHECK3-NEXT:    [[A:%.*]] = alloca i32, align 4
4156 // CHECK3-NEXT:    [[AA:%.*]] = alloca i16, align 2
4157 // CHECK3-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
4158 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4159 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4160 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
4161 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
4162 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
4163 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
4164 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
4165 // CHECK3-NEXT:    store i16 0, i16* [[AA]], align 2
4166 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
4167 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
4168 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
4169 // CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
4170 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4171 // CHECK3-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
4172 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4173 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
4174 // CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
4175 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
4176 // CHECK3:       omp_if.then:
4177 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4178 // CHECK3-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
4179 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP6]], align 4
4180 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4181 // CHECK3-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
4182 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
4183 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
4184 // CHECK3-NEXT:    store i8* null, i8** [[TMP9]], align 4
4185 // CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
4186 // CHECK3-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i32*
4187 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP11]], align 4
4188 // CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
4189 // CHECK3-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
4190 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
4191 // CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
4192 // CHECK3-NEXT:    store i8* null, i8** [[TMP14]], align 4
4193 // CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
4194 // CHECK3-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
4195 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 4
4196 // CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
4197 // CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
4198 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 4
4199 // CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
4200 // CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
4201 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4202 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4203 // CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4204 // CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
4205 // CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
4206 // CHECK3:       omp_offload.failed:
4207 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
4208 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
4209 // CHECK3:       omp_offload.cont:
4210 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
4211 // CHECK3:       omp_if.else:
4212 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
4213 // CHECK3-NEXT:    br label [[OMP_IF_END]]
4214 // CHECK3:       omp_if.end:
4215 // CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
4216 // CHECK3-NEXT:    ret i32 [[TMP24]]
4217 //
4218 //
4219 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
4220 // CHECK3-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
4221 // CHECK3-NEXT:  entry:
4222 // CHECK3-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
4223 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
4224 // CHECK3-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
4225 // CHECK3-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
4226 // CHECK3-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
4227 // CHECK3-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
4228 // CHECK3-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
4229 // CHECK3-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
4230 // CHECK3-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
4231 // CHECK3-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
4232 // CHECK3-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
4233 // CHECK3-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
4234 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
4235 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
4236 // CHECK3-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
4237 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
4238 // CHECK3-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
4239 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
4240 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
4241 // CHECK3-NEXT:    ret void
4242 //
4243 //
4244 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..22
4245 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
4246 // CHECK3-NEXT:  entry:
4247 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
4248 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
4249 // CHECK3-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
4250 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
4251 // CHECK3-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
4252 // CHECK3-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
4253 // CHECK3-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
4254 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
4255 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
4256 // CHECK3-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
4257 // CHECK3-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
4258 // CHECK3-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
4259 // CHECK3-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
4260 // CHECK3-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
4261 // CHECK3-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
4262 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
4263 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
4264 // CHECK3-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
4265 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
4266 // CHECK3-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
4267 // CHECK3-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
4268 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
4269 // CHECK3-NEXT:    store double [[ADD]], double* [[A]], align 4
4270 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
4271 // CHECK3-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
4272 // CHECK3-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
4273 // CHECK3-NEXT:    store double [[INC]], double* [[A3]], align 4
4274 // CHECK3-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
4275 // CHECK3-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
4276 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
4277 // CHECK3-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
4278 // CHECK3-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
4279 // CHECK3-NEXT:    ret void
4280 //
4281 //
4282 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
4283 // CHECK3-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
4284 // CHECK3-NEXT:  entry:
4285 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4286 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4287 // CHECK3-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
4288 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
4289 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4290 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4291 // CHECK3-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
4292 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4293 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4294 // CHECK3-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
4295 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
4296 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4297 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
4298 // CHECK3-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
4299 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
4300 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
4301 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
4302 // CHECK3-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
4303 // CHECK3-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4304 // CHECK3-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
4305 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4306 // CHECK3-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
4307 // CHECK3-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
4308 // CHECK3-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
4309 // CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
4310 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
4311 // CHECK3-NEXT:    ret void
4312 //
4313 //
4314 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..24
4315 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
4316 // CHECK3-NEXT:  entry:
4317 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
4318 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
4319 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4320 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4321 // CHECK3-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
4322 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
4323 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
4324 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
4325 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4326 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4327 // CHECK3-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
4328 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
4329 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4330 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
4331 // CHECK3-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
4332 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
4333 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
4334 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
4335 // CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
4336 // CHECK3-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
4337 // CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
4338 // CHECK3-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
4339 // CHECK3-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
4340 // CHECK3-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
4341 // CHECK3-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
4342 // CHECK3-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
4343 // CHECK3-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
4344 // CHECK3-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
4345 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
4346 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
4347 // CHECK3-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
4348 // CHECK3-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
4349 // CHECK3-NEXT:    ret void
4350 //
4351 //
4352 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
4353 // CHECK3-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
4354 // CHECK3-NEXT:  entry:
4355 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4356 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4357 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
4358 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4359 // CHECK3-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4360 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4361 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4362 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
4363 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4364 // CHECK3-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
4365 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
4366 // CHECK3-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
4367 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
4368 // CHECK3-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
4369 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4370 // CHECK3-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
4371 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4372 // CHECK3-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
4373 // CHECK3-NEXT:    ret void
4374 //
4375 //
4376 // CHECK3-LABEL: define {{[^@]+}}@.omp_outlined..27
4377 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
4378 // CHECK3-NEXT:  entry:
4379 // CHECK3-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
4380 // CHECK3-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
4381 // CHECK3-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4382 // CHECK3-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4383 // CHECK3-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
4384 // CHECK3-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
4385 // CHECK3-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
4386 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4387 // CHECK3-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4388 // CHECK3-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
4389 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4390 // CHECK3-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
4391 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
4392 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
4393 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
4394 // CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
4395 // CHECK3-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
4396 // CHECK3-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
4397 // CHECK3-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
4398 // CHECK3-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
4399 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
4400 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
4401 // CHECK3-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
4402 // CHECK3-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
4403 // CHECK3-NEXT:    ret void
4404 //
4405 //
4406 // CHECK3-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
4407 // CHECK3-SAME: () #[[ATTR4]] {
4408 // CHECK3-NEXT:  entry:
4409 // CHECK3-NEXT:    call void @__tgt_register_requires(i64 1)
4410 // CHECK3-NEXT:    ret void
4411 //
4412 //
4413 // CHECK4-LABEL: define {{[^@]+}}@_Z3fooi
4414 // CHECK4-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
4415 // CHECK4-NEXT:  entry:
4416 // CHECK4-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
4417 // CHECK4-NEXT:    [[A:%.*]] = alloca i32, align 4
4418 // CHECK4-NEXT:    [[AA:%.*]] = alloca i16, align 2
4419 // CHECK4-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
4420 // CHECK4-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
4421 // CHECK4-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
4422 // CHECK4-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
4423 // CHECK4-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
4424 // CHECK4-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
4425 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
4426 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
4427 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4428 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
4429 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR__CASTED3:%.*]] = alloca i32, align 4
4430 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
4431 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
4432 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
4433 // CHECK4-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
4434 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4435 // CHECK4-NEXT:    [[AA_CASTED4:%.*]] = alloca i32, align 4
4436 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS6:%.*]] = alloca [1 x i8*], align 4
4437 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [1 x i8*], align 4
4438 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [1 x i8*], align 4
4439 // CHECK4-NEXT:    [[A_CASTED9:%.*]] = alloca i32, align 4
4440 // CHECK4-NEXT:    [[AA_CASTED10:%.*]] = alloca i32, align 4
4441 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS12:%.*]] = alloca [2 x i8*], align 4
4442 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
4443 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
4444 // CHECK4-NEXT:    [[A_CASTED17:%.*]] = alloca i32, align 4
4445 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [9 x i8*], align 4
4446 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [9 x i8*], align 4
4447 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [9 x i8*], align 4
4448 // CHECK4-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
4449 // CHECK4-NEXT:    [[NN:%.*]] = alloca i32, align 4
4450 // CHECK4-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
4451 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
4452 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
4453 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
4454 // CHECK4-NEXT:    [[NN_CASTED32:%.*]] = alloca i32, align 4
4455 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 4
4456 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 4
4457 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 4
4458 // CHECK4-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
4459 // CHECK4-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
4460 // CHECK4-NEXT:    store i32 0, i32* [[A]], align 4
4461 // CHECK4-NEXT:    store i16 0, i16* [[AA]], align 2
4462 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
4463 // CHECK4-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
4464 // CHECK4-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
4465 // CHECK4-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP1]], align 4
4466 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
4467 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
4468 // CHECK4-NEXT:    [[TMP4:%.*]] = mul nuw i32 5, [[TMP3]]
4469 // CHECK4-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP4]], align 8
4470 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[__VLA_EXPR1]], align 4
4471 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
4472 // CHECK4-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4
4473 // CHECK4-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
4474 // CHECK4-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_2]], align 4
4475 // CHECK4-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
4476 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4477 // CHECK4-NEXT:    store i16 [[TMP7]], i16* [[CONV]], align 2
4478 // CHECK4-NEXT:    [[TMP8:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4479 // CHECK4-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
4480 // CHECK4-NEXT:    store i32 [[TMP9]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
4481 // CHECK4-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
4482 // CHECK4-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
4483 // CHECK4-NEXT:    store i32 [[TMP11]], i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
4484 // CHECK4-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
4485 // CHECK4-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4486 // CHECK4-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i32*
4487 // CHECK4-NEXT:    store i32 [[TMP8]], i32* [[TMP14]], align 4
4488 // CHECK4-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4489 // CHECK4-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
4490 // CHECK4-NEXT:    store i32 [[TMP8]], i32* [[TMP16]], align 4
4491 // CHECK4-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
4492 // CHECK4-NEXT:    store i8* null, i8** [[TMP17]], align 4
4493 // CHECK4-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
4494 // CHECK4-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
4495 // CHECK4-NEXT:    store i32 [[TMP10]], i32* [[TMP19]], align 4
4496 // CHECK4-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
4497 // CHECK4-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
4498 // CHECK4-NEXT:    store i32 [[TMP10]], i32* [[TMP21]], align 4
4499 // CHECK4-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
4500 // CHECK4-NEXT:    store i8* null, i8** [[TMP22]], align 4
4501 // CHECK4-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
4502 // CHECK4-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
4503 // CHECK4-NEXT:    store i32 [[TMP12]], i32* [[TMP24]], align 4
4504 // CHECK4-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
4505 // CHECK4-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
4506 // CHECK4-NEXT:    store i32 [[TMP12]], i32* [[TMP26]], align 4
4507 // CHECK4-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
4508 // CHECK4-NEXT:    store i8* null, i8** [[TMP27]], align 4
4509 // CHECK4-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4510 // CHECK4-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4511 // CHECK4-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
4512 // CHECK4-NEXT:    [[TMP31:%.*]] = load i16, i16* [[AA]], align 2
4513 // CHECK4-NEXT:    store i16 [[TMP31]], i16* [[TMP30]], align 4
4514 // CHECK4-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
4515 // CHECK4-NEXT:    [[TMP33:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
4516 // CHECK4-NEXT:    store i32 [[TMP33]], i32* [[TMP32]], align 4
4517 // CHECK4-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
4518 // CHECK4-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
4519 // CHECK4-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
4520 // CHECK4-NEXT:    [[TMP36:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
4521 // CHECK4-NEXT:    [[TMP37:%.*]] = bitcast i8* [[TMP36]] to %struct.kmp_task_t_with_privates*
4522 // CHECK4-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 0
4523 // CHECK4-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP38]], i32 0, i32 0
4524 // CHECK4-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[TMP39]], align 4
4525 // CHECK4-NEXT:    [[TMP41:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
4526 // CHECK4-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP40]], i8* align 4 [[TMP41]], i32 12, i1 false)
4527 // CHECK4-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 1
4528 // CHECK4-NEXT:    [[TMP43:%.*]] = bitcast i8* [[TMP40]] to %struct.anon*
4529 // CHECK4-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 0
4530 // CHECK4-NEXT:    [[TMP45:%.*]] = bitcast [3 x i64]* [[TMP44]] to i8*
4531 // CHECK4-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP45]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
4532 // CHECK4-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 1
4533 // CHECK4-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
4534 // CHECK4-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP28]] to i8*
4535 // CHECK4-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP47]], i8* align 4 [[TMP48]], i32 12, i1 false)
4536 // CHECK4-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 2
4537 // CHECK4-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
4538 // CHECK4-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP29]] to i8*
4539 // CHECK4-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i32 12, i1 false)
4540 // CHECK4-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 3
4541 // CHECK4-NEXT:    [[TMP53:%.*]] = load i16, i16* [[AA]], align 2
4542 // CHECK4-NEXT:    store i16 [[TMP53]], i16* [[TMP52]], align 4
4543 // CHECK4-NEXT:    [[TMP54:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP36]])
4544 // CHECK4-NEXT:    [[TMP55:%.*]] = load i32, i32* [[A]], align 4
4545 // CHECK4-NEXT:    store i32 [[TMP55]], i32* [[A_CASTED]], align 4
4546 // CHECK4-NEXT:    [[TMP56:%.*]] = load i32, i32* [[A_CASTED]], align 4
4547 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i32 [[TMP56]]) #[[ATTR3:[0-9]+]]
4548 // CHECK4-NEXT:    [[TMP57:%.*]] = load i16, i16* [[AA]], align 2
4549 // CHECK4-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
4550 // CHECK4-NEXT:    store i16 [[TMP57]], i16* [[CONV5]], align 2
4551 // CHECK4-NEXT:    [[TMP58:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
4552 // CHECK4-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
4553 // CHECK4-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32*
4554 // CHECK4-NEXT:    store i32 [[TMP58]], i32* [[TMP60]], align 4
4555 // CHECK4-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
4556 // CHECK4-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32*
4557 // CHECK4-NEXT:    store i32 [[TMP58]], i32* [[TMP62]], align 4
4558 // CHECK4-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
4559 // CHECK4-NEXT:    store i8* null, i8** [[TMP63]], align 4
4560 // CHECK4-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
4561 // CHECK4-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
4562 // CHECK4-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4563 // CHECK4-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
4564 // CHECK4-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
4565 // CHECK4:       omp_offload.failed:
4566 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR3]]
4567 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT]]
4568 // CHECK4:       omp_offload.cont:
4569 // CHECK4-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
4570 // CHECK4-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
4571 // CHECK4-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
4572 // CHECK4-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
4573 // CHECK4-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
4574 // CHECK4-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
4575 // CHECK4-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
4576 // CHECK4-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
4577 // CHECK4-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
4578 // CHECK4-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
4579 // CHECK4:       omp_if.then:
4580 // CHECK4-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
4581 // CHECK4-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
4582 // CHECK4-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
4583 // CHECK4-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
4584 // CHECK4-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
4585 // CHECK4-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
4586 // CHECK4-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
4587 // CHECK4-NEXT:    store i8* null, i8** [[TMP77]], align 4
4588 // CHECK4-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
4589 // CHECK4-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
4590 // CHECK4-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
4591 // CHECK4-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
4592 // CHECK4-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
4593 // CHECK4-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
4594 // CHECK4-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
4595 // CHECK4-NEXT:    store i8* null, i8** [[TMP82]], align 4
4596 // CHECK4-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
4597 // CHECK4-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
4598 // CHECK4-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4599 // CHECK4-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
4600 // CHECK4-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
4601 // CHECK4:       omp_offload.failed15:
4602 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
4603 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
4604 // CHECK4:       omp_offload.cont16:
4605 // CHECK4-NEXT:    br label [[OMP_IF_END:%.*]]
4606 // CHECK4:       omp_if.else:
4607 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
4608 // CHECK4-NEXT:    br label [[OMP_IF_END]]
4609 // CHECK4:       omp_if.end:
4610 // CHECK4-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
4611 // CHECK4-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED17]], align 4
4612 // CHECK4-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED17]], align 4
4613 // CHECK4-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
4614 // CHECK4-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP89]], 20
4615 // CHECK4-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
4616 // CHECK4:       omp_if.then19:
4617 // CHECK4-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
4618 // CHECK4-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
4619 // CHECK4-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
4620 // CHECK4-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
4621 // CHECK4-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
4622 // CHECK4-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
4623 // CHECK4-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
4624 // CHECK4-NEXT:    store i32 [[TMP88]], i32* [[TMP96]], align 4
4625 // CHECK4-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
4626 // CHECK4-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
4627 // CHECK4-NEXT:    store i32 [[TMP88]], i32* [[TMP98]], align 4
4628 // CHECK4-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
4629 // CHECK4-NEXT:    store i64 4, i64* [[TMP99]], align 4
4630 // CHECK4-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
4631 // CHECK4-NEXT:    store i8* null, i8** [[TMP100]], align 4
4632 // CHECK4-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
4633 // CHECK4-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
4634 // CHECK4-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
4635 // CHECK4-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
4636 // CHECK4-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
4637 // CHECK4-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
4638 // CHECK4-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
4639 // CHECK4-NEXT:    store i64 40, i64* [[TMP105]], align 4
4640 // CHECK4-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
4641 // CHECK4-NEXT:    store i8* null, i8** [[TMP106]], align 4
4642 // CHECK4-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
4643 // CHECK4-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
4644 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP108]], align 4
4645 // CHECK4-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
4646 // CHECK4-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
4647 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP110]], align 4
4648 // CHECK4-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
4649 // CHECK4-NEXT:    store i64 4, i64* [[TMP111]], align 4
4650 // CHECK4-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
4651 // CHECK4-NEXT:    store i8* null, i8** [[TMP112]], align 4
4652 // CHECK4-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
4653 // CHECK4-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
4654 // CHECK4-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
4655 // CHECK4-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
4656 // CHECK4-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
4657 // CHECK4-NEXT:    store float* [[VLA]], float** [[TMP116]], align 4
4658 // CHECK4-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
4659 // CHECK4-NEXT:    store i64 [[TMP91]], i64* [[TMP117]], align 4
4660 // CHECK4-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
4661 // CHECK4-NEXT:    store i8* null, i8** [[TMP118]], align 4
4662 // CHECK4-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
4663 // CHECK4-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
4664 // CHECK4-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
4665 // CHECK4-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
4666 // CHECK4-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
4667 // CHECK4-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 4
4668 // CHECK4-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
4669 // CHECK4-NEXT:    store i64 400, i64* [[TMP123]], align 4
4670 // CHECK4-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 4
4671 // CHECK4-NEXT:    store i8* null, i8** [[TMP124]], align 4
4672 // CHECK4-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
4673 // CHECK4-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i32*
4674 // CHECK4-NEXT:    store i32 5, i32* [[TMP126]], align 4
4675 // CHECK4-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
4676 // CHECK4-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
4677 // CHECK4-NEXT:    store i32 5, i32* [[TMP128]], align 4
4678 // CHECK4-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
4679 // CHECK4-NEXT:    store i64 4, i64* [[TMP129]], align 4
4680 // CHECK4-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 5
4681 // CHECK4-NEXT:    store i8* null, i8** [[TMP130]], align 4
4682 // CHECK4-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
4683 // CHECK4-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
4684 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[TMP132]], align 4
4685 // CHECK4-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
4686 // CHECK4-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
4687 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[TMP134]], align 4
4688 // CHECK4-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
4689 // CHECK4-NEXT:    store i64 4, i64* [[TMP135]], align 4
4690 // CHECK4-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 6
4691 // CHECK4-NEXT:    store i8* null, i8** [[TMP136]], align 4
4692 // CHECK4-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
4693 // CHECK4-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
4694 // CHECK4-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 4
4695 // CHECK4-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
4696 // CHECK4-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
4697 // CHECK4-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 4
4698 // CHECK4-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
4699 // CHECK4-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 4
4700 // CHECK4-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 7
4701 // CHECK4-NEXT:    store i8* null, i8** [[TMP142]], align 4
4702 // CHECK4-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
4703 // CHECK4-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
4704 // CHECK4-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 4
4705 // CHECK4-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
4706 // CHECK4-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
4707 // CHECK4-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 4
4708 // CHECK4-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
4709 // CHECK4-NEXT:    store i64 12, i64* [[TMP147]], align 4
4710 // CHECK4-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 8
4711 // CHECK4-NEXT:    store i8* null, i8** [[TMP148]], align 4
4712 // CHECK4-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
4713 // CHECK4-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
4714 // CHECK4-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
4715 // CHECK4-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4716 // CHECK4-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
4717 // CHECK4-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
4718 // CHECK4:       omp_offload.failed23:
4719 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
4720 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
4721 // CHECK4:       omp_offload.cont24:
4722 // CHECK4-NEXT:    br label [[OMP_IF_END26:%.*]]
4723 // CHECK4:       omp_if.else25:
4724 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
4725 // CHECK4-NEXT:    br label [[OMP_IF_END26]]
4726 // CHECK4:       omp_if.end26:
4727 // CHECK4-NEXT:    store i32 0, i32* [[NN]], align 4
4728 // CHECK4-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
4729 // CHECK4-NEXT:    store i32 [[TMP154]], i32* [[NN_CASTED]], align 4
4730 // CHECK4-NEXT:    [[TMP155:%.*]] = load i32, i32* [[NN_CASTED]], align 4
4731 // CHECK4-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
4732 // CHECK4-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
4733 // CHECK4-NEXT:    store i32 [[TMP155]], i32* [[TMP157]], align 4
4734 // CHECK4-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
4735 // CHECK4-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32*
4736 // CHECK4-NEXT:    store i32 [[TMP155]], i32* [[TMP159]], align 4
4737 // CHECK4-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
4738 // CHECK4-NEXT:    store i8* null, i8** [[TMP160]], align 4
4739 // CHECK4-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
4740 // CHECK4-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
4741 // CHECK4-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4742 // CHECK4-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
4743 // CHECK4-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
4744 // CHECK4:       omp_offload.failed30:
4745 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i32 [[TMP155]]) #[[ATTR3]]
4746 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
4747 // CHECK4:       omp_offload.cont31:
4748 // CHECK4-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
4749 // CHECK4-NEXT:    store i32 [[TMP165]], i32* [[NN_CASTED32]], align 4
4750 // CHECK4-NEXT:    [[TMP166:%.*]] = load i32, i32* [[NN_CASTED32]], align 4
4751 // CHECK4-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
4752 // CHECK4-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i32*
4753 // CHECK4-NEXT:    store i32 [[TMP166]], i32* [[TMP168]], align 4
4754 // CHECK4-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
4755 // CHECK4-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i32*
4756 // CHECK4-NEXT:    store i32 [[TMP166]], i32* [[TMP170]], align 4
4757 // CHECK4-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
4758 // CHECK4-NEXT:    store i8* null, i8** [[TMP171]], align 4
4759 // CHECK4-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
4760 // CHECK4-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
4761 // CHECK4-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4762 // CHECK4-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
4763 // CHECK4-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
4764 // CHECK4:       omp_offload.failed36:
4765 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i32 [[TMP166]]) #[[ATTR3]]
4766 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
4767 // CHECK4:       omp_offload.cont37:
4768 // CHECK4-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
4769 // CHECK4-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
4770 // CHECK4-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
4771 // CHECK4-NEXT:    ret i32 [[TMP176]]
4772 //
4773 //
4774 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
4775 // CHECK4-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
4776 // CHECK4-NEXT:  entry:
4777 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4778 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
4779 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
4780 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4781 // CHECK4-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
4782 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4783 // CHECK4-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
4784 // CHECK4-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
4785 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4786 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
4787 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
4788 // CHECK4-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
4789 // CHECK4-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
4790 // CHECK4-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4791 // CHECK4-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
4792 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4793 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
4794 // CHECK4-NEXT:    ret void
4795 //
4796 //
4797 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined.
4798 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
4799 // CHECK4-NEXT:  entry:
4800 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
4801 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
4802 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4803 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
4804 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
4805 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4806 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4807 // CHECK4-NEXT:    ret void
4808 //
4809 //
4810 // CHECK4-LABEL: define {{[^@]+}}@.omp_task_privates_map.
4811 // CHECK4-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
4812 // CHECK4-NEXT:  entry:
4813 // CHECK4-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 4
4814 // CHECK4-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 4
4815 // CHECK4-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 4
4816 // CHECK4-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 4
4817 // CHECK4-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 4
4818 // CHECK4-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 4
4819 // CHECK4-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 4
4820 // CHECK4-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 4
4821 // CHECK4-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 4
4822 // CHECK4-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 4
4823 // CHECK4-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 4
4824 // CHECK4-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
4825 // CHECK4-NEXT:    [[TMP7:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 4
4826 // CHECK4-NEXT:    store [3 x i64]* [[TMP6]], [3 x i64]** [[TMP7]], align 4
4827 // CHECK4-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
4828 // CHECK4-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 4
4829 // CHECK4-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 4
4830 // CHECK4-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
4831 // CHECK4-NEXT:    [[TMP11:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 4
4832 // CHECK4-NEXT:    store [3 x i8*]* [[TMP10]], [3 x i8*]** [[TMP11]], align 4
4833 // CHECK4-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
4834 // CHECK4-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 4
4835 // CHECK4-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 4
4836 // CHECK4-NEXT:    ret void
4837 //
4838 //
4839 // CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry.
4840 // CHECK4-SAME: (i32 [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
4841 // CHECK4-NEXT:  entry:
4842 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
4843 // CHECK4-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 4
4844 // CHECK4-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 4
4845 // CHECK4-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
4846 // CHECK4-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
4847 // CHECK4-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
4848 // CHECK4-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 4
4849 // CHECK4-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 4
4850 // CHECK4-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 4
4851 // CHECK4-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 4
4852 // CHECK4-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
4853 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
4854 // CHECK4-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
4855 // CHECK4-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
4856 // CHECK4-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
4857 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
4858 // CHECK4-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
4859 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
4860 // CHECK4-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
4861 // CHECK4-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
4862 // CHECK4-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
4863 // CHECK4-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
4864 // CHECK4-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 4
4865 // CHECK4-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
4866 // CHECK4-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
4867 // CHECK4-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
4868 // CHECK4-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
4869 // CHECK4-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META15:![0-9]+]])
4870 // CHECK4-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
4871 // CHECK4-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
4872 // CHECK4-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META22:![0-9]+]])
4873 // CHECK4-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !24
4874 // CHECK4-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 4, !noalias !24
4875 // CHECK4-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
4876 // CHECK4-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
4877 // CHECK4-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !24
4878 // CHECK4-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
4879 // CHECK4-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
4880 // CHECK4-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
4881 // CHECK4-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
4882 // CHECK4-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
4883 // CHECK4-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
4884 // CHECK4-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 4, !noalias !24
4885 // CHECK4-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 4, !noalias !24
4886 // CHECK4-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 4, !noalias !24
4887 // CHECK4-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 4, !noalias !24
4888 // CHECK4-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i32 0, i32 0
4889 // CHECK4-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i32 0, i32 0
4890 // CHECK4-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i32 0, i32 0
4891 // CHECK4-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
4892 // CHECK4-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
4893 // CHECK4-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
4894 // CHECK4-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
4895 // CHECK4-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
4896 // CHECK4-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
4897 // CHECK4-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
4898 // CHECK4:       omp_offload.failed.i:
4899 // CHECK4-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
4900 // CHECK4-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
4901 // CHECK4-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !24
4902 // CHECK4-NEXT:    [[TMP30:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !24
4903 // CHECK4-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
4904 // CHECK4-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
4905 // CHECK4-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
4906 // CHECK4-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
4907 // CHECK4-NEXT:    store i32 [[TMP33]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
4908 // CHECK4-NEXT:    [[TMP34:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
4909 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i32 [[TMP30]], i32 [[TMP32]], i32 [[TMP34]]) #[[ATTR3]]
4910 // CHECK4-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
4911 // CHECK4:       .omp_outlined..1.exit:
4912 // CHECK4-NEXT:    ret i32 0
4913 //
4914 //
4915 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
4916 // CHECK4-SAME: (i32 [[A:%.*]]) #[[ATTR2]] {
4917 // CHECK4-NEXT:  entry:
4918 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4919 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4920 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4921 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
4922 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
4923 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
4924 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]])
4925 // CHECK4-NEXT:    ret void
4926 //
4927 //
4928 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..2
4929 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]]) #[[ATTR2]] {
4930 // CHECK4-NEXT:  entry:
4931 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
4932 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
4933 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4934 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
4935 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
4936 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4937 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
4938 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
4939 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
4940 // CHECK4-NEXT:    ret void
4941 //
4942 //
4943 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
4944 // CHECK4-SAME: (i32 [[AA:%.*]]) #[[ATTR2]] {
4945 // CHECK4-NEXT:  entry:
4946 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4947 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4948 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4949 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4950 // CHECK4-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
4951 // CHECK4-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4952 // CHECK4-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
4953 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4954 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP1]])
4955 // CHECK4-NEXT:    ret void
4956 //
4957 //
4958 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..3
4959 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
4960 // CHECK4-NEXT:  entry:
4961 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
4962 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
4963 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4964 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
4965 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
4966 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4967 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4968 // CHECK4-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
4969 // CHECK4-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
4970 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
4971 // CHECK4-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
4972 // CHECK4-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
4973 // CHECK4-NEXT:    ret void
4974 //
4975 //
4976 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
4977 // CHECK4-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
4978 // CHECK4-NEXT:  entry:
4979 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
4980 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
4981 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
4982 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
4983 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
4984 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
4985 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
4986 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
4987 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
4988 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
4989 // CHECK4-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
4990 // CHECK4-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
4991 // CHECK4-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
4992 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
4993 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
4994 // CHECK4-NEXT:    ret void
4995 //
4996 //
4997 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..6
4998 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
4999 // CHECK4-NEXT:  entry:
5000 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5001 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5002 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5003 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
5004 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5005 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5006 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5007 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
5008 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
5009 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
5010 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
5011 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
5012 // CHECK4-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
5013 // CHECK4-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
5014 // CHECK4-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
5015 // CHECK4-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
5016 // CHECK4-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
5017 // CHECK4-NEXT:    ret void
5018 //
5019 //
5020 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
5021 // CHECK4-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
5022 // CHECK4-NEXT:  entry:
5023 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5024 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
5025 // CHECK4-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
5026 // CHECK4-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
5027 // CHECK4-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
5028 // CHECK4-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
5029 // CHECK4-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
5030 // CHECK4-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
5031 // CHECK4-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
5032 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
5033 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5034 // CHECK4-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
5035 // CHECK4-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
5036 // CHECK4-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
5037 // CHECK4-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
5038 // CHECK4-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
5039 // CHECK4-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
5040 // CHECK4-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
5041 // CHECK4-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
5042 // CHECK4-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
5043 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
5044 // CHECK4-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
5045 // CHECK4-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
5046 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
5047 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
5048 // CHECK4-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
5049 // CHECK4-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
5050 // CHECK4-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
5051 // CHECK4-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
5052 // CHECK4-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
5053 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
5054 // CHECK4-NEXT:    ret void
5055 //
5056 //
5057 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..9
5058 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
5059 // CHECK4-NEXT:  entry:
5060 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5061 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5062 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5063 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
5064 // CHECK4-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
5065 // CHECK4-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
5066 // CHECK4-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
5067 // CHECK4-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
5068 // CHECK4-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
5069 // CHECK4-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
5070 // CHECK4-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
5071 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5072 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5073 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5074 // CHECK4-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
5075 // CHECK4-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
5076 // CHECK4-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
5077 // CHECK4-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
5078 // CHECK4-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
5079 // CHECK4-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
5080 // CHECK4-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
5081 // CHECK4-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
5082 // CHECK4-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
5083 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
5084 // CHECK4-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
5085 // CHECK4-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
5086 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
5087 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
5088 // CHECK4-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
5089 // CHECK4-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
5090 // CHECK4-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
5091 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
5092 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
5093 // CHECK4-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
5094 // CHECK4-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
5095 // CHECK4-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
5096 // CHECK4-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
5097 // CHECK4-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
5098 // CHECK4-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
5099 // CHECK4-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
5100 // CHECK4-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
5101 // CHECK4-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
5102 // CHECK4-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
5103 // CHECK4-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
5104 // CHECK4-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
5105 // CHECK4-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
5106 // CHECK4-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
5107 // CHECK4-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
5108 // CHECK4-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
5109 // CHECK4-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
5110 // CHECK4-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
5111 // CHECK4-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
5112 // CHECK4-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
5113 // CHECK4-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
5114 // CHECK4-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
5115 // CHECK4-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
5116 // CHECK4-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
5117 // CHECK4-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
5118 // CHECK4-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
5119 // CHECK4-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
5120 // CHECK4-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
5121 // CHECK4-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
5122 // CHECK4-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
5123 // CHECK4-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
5124 // CHECK4-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
5125 // CHECK4-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
5126 // CHECK4-NEXT:    ret void
5127 //
5128 //
5129 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
5130 // CHECK4-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
5131 // CHECK4-NEXT:  entry:
5132 // CHECK4-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
5133 // CHECK4-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
5134 // CHECK4-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
5135 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
5136 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
5137 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
5138 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP1]])
5139 // CHECK4-NEXT:    ret void
5140 //
5141 //
5142 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..11
5143 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
5144 // CHECK4-NEXT:  entry:
5145 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5146 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5147 // CHECK4-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
5148 // CHECK4-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
5149 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5150 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5151 // CHECK4-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
5152 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
5153 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
5154 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
5155 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i32 [[TMP1]])
5156 // CHECK4-NEXT:    ret void
5157 //
5158 //
5159 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..12
5160 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
5161 // CHECK4-NEXT:  entry:
5162 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5163 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5164 // CHECK4-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
5165 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5166 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5167 // CHECK4-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
5168 // CHECK4-NEXT:    ret void
5169 //
5170 //
5171 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
5172 // CHECK4-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
5173 // CHECK4-NEXT:  entry:
5174 // CHECK4-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
5175 // CHECK4-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
5176 // CHECK4-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
5177 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
5178 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
5179 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
5180 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i32 [[TMP1]])
5181 // CHECK4-NEXT:    ret void
5182 //
5183 //
5184 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..15
5185 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
5186 // CHECK4-NEXT:  entry:
5187 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5188 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5189 // CHECK4-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
5190 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5191 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5192 // CHECK4-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
5193 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
5194 // CHECK4-NEXT:    ret void
5195 //
5196 //
5197 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..16
5198 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
5199 // CHECK4-NEXT:  entry:
5200 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5201 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5202 // CHECK4-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
5203 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5204 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5205 // CHECK4-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
5206 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
5207 // CHECK4-NEXT:    ret void
5208 //
5209 //
5210 // CHECK4-LABEL: define {{[^@]+}}@_Z6bazzzziPi
5211 // CHECK4-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
5212 // CHECK4-NEXT:  entry:
5213 // CHECK4-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5214 // CHECK4-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
5215 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
5216 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
5217 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
5218 // CHECK4-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5219 // CHECK4-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
5220 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5221 // CHECK4-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5222 // CHECK4-NEXT:    [[TMP2:%.*]] = bitcast i8** [[TMP1]] to i32*
5223 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[TMP2]], align 4
5224 // CHECK4-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5225 // CHECK4-NEXT:    [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32*
5226 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[TMP4]], align 4
5227 // CHECK4-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
5228 // CHECK4-NEXT:    store i8* null, i8** [[TMP5]], align 4
5229 // CHECK4-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5230 // CHECK4-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5231 // CHECK4-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
5232 // CHECK4-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
5233 // CHECK4-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
5234 // CHECK4:       omp_offload.failed:
5235 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i32 [[TMP0]]) #[[ATTR3]]
5236 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT]]
5237 // CHECK4:       omp_offload.cont:
5238 // CHECK4-NEXT:    ret void
5239 //
5240 //
5241 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
5242 // CHECK4-SAME: (i32 [[VLA:%.*]]) #[[ATTR2]] {
5243 // CHECK4-NEXT:  entry:
5244 // CHECK4-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
5245 // CHECK4-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
5246 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
5247 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i32 [[TMP0]])
5248 // CHECK4-NEXT:    ret void
5249 //
5250 //
5251 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..19
5252 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR2]] {
5253 // CHECK4-NEXT:  entry:
5254 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5255 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5256 // CHECK4-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
5257 // CHECK4-NEXT:    [[F:%.*]] = alloca i32*, align 4
5258 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5259 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5260 // CHECK4-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
5261 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
5262 // CHECK4-NEXT:    ret void
5263 //
5264 //
5265 // CHECK4-LABEL: define {{[^@]+}}@_Z3bari
5266 // CHECK4-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
5267 // CHECK4-NEXT:  entry:
5268 // CHECK4-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5269 // CHECK4-NEXT:    [[A:%.*]] = alloca i32, align 4
5270 // CHECK4-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
5271 // CHECK4-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5272 // CHECK4-NEXT:    store i32 0, i32* [[A]], align 4
5273 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5274 // CHECK4-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
5275 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
5276 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
5277 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
5278 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
5279 // CHECK4-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
5280 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
5281 // CHECK4-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
5282 // CHECK4-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
5283 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
5284 // CHECK4-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
5285 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
5286 // CHECK4-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
5287 // CHECK4-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
5288 // CHECK4-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
5289 // CHECK4-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
5290 // CHECK4-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
5291 // CHECK4-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
5292 // CHECK4-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
5293 // CHECK4-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
5294 // CHECK4-NEXT:    ret i32 [[TMP8]]
5295 //
5296 //
5297 // CHECK4-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
5298 // CHECK4-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
5299 // CHECK4-NEXT:  entry:
5300 // CHECK4-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
5301 // CHECK4-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5302 // CHECK4-NEXT:    [[B:%.*]] = alloca i32, align 4
5303 // CHECK4-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
5304 // CHECK4-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
5305 // CHECK4-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
5306 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 4
5307 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 4
5308 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 4
5309 // CHECK4-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 4
5310 // CHECK4-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
5311 // CHECK4-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5312 // CHECK4-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
5313 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5314 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
5315 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
5316 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
5317 // CHECK4-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
5318 // CHECK4-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
5319 // CHECK4-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
5320 // CHECK4-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
5321 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
5322 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
5323 // CHECK4-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
5324 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
5325 // CHECK4-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
5326 // CHECK4-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 60
5327 // CHECK4-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
5328 // CHECK4:       omp_if.then:
5329 // CHECK4-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
5330 // CHECK4-NEXT:    [[TMP7:%.*]] = mul nuw i32 2, [[TMP1]]
5331 // CHECK4-NEXT:    [[TMP8:%.*]] = mul nuw i32 [[TMP7]], 2
5332 // CHECK4-NEXT:    [[TMP9:%.*]] = sext i32 [[TMP8]] to i64
5333 // CHECK4-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5334 // CHECK4-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
5335 // CHECK4-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 4
5336 // CHECK4-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5337 // CHECK4-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
5338 // CHECK4-NEXT:    store double* [[A]], double** [[TMP13]], align 4
5339 // CHECK4-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
5340 // CHECK4-NEXT:    store i64 8, i64* [[TMP14]], align 4
5341 // CHECK4-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
5342 // CHECK4-NEXT:    store i8* null, i8** [[TMP15]], align 4
5343 // CHECK4-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
5344 // CHECK4-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
5345 // CHECK4-NEXT:    store i32 [[TMP5]], i32* [[TMP17]], align 4
5346 // CHECK4-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
5347 // CHECK4-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
5348 // CHECK4-NEXT:    store i32 [[TMP5]], i32* [[TMP19]], align 4
5349 // CHECK4-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
5350 // CHECK4-NEXT:    store i64 4, i64* [[TMP20]], align 4
5351 // CHECK4-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
5352 // CHECK4-NEXT:    store i8* null, i8** [[TMP21]], align 4
5353 // CHECK4-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
5354 // CHECK4-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
5355 // CHECK4-NEXT:    store i32 2, i32* [[TMP23]], align 4
5356 // CHECK4-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
5357 // CHECK4-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
5358 // CHECK4-NEXT:    store i32 2, i32* [[TMP25]], align 4
5359 // CHECK4-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
5360 // CHECK4-NEXT:    store i64 4, i64* [[TMP26]], align 4
5361 // CHECK4-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
5362 // CHECK4-NEXT:    store i8* null, i8** [[TMP27]], align 4
5363 // CHECK4-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
5364 // CHECK4-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
5365 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP29]], align 4
5366 // CHECK4-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
5367 // CHECK4-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
5368 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP31]], align 4
5369 // CHECK4-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
5370 // CHECK4-NEXT:    store i64 4, i64* [[TMP32]], align 4
5371 // CHECK4-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
5372 // CHECK4-NEXT:    store i8* null, i8** [[TMP33]], align 4
5373 // CHECK4-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
5374 // CHECK4-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
5375 // CHECK4-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 4
5376 // CHECK4-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
5377 // CHECK4-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
5378 // CHECK4-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 4
5379 // CHECK4-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
5380 // CHECK4-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 4
5381 // CHECK4-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 4
5382 // CHECK4-NEXT:    store i8* null, i8** [[TMP39]], align 4
5383 // CHECK4-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5384 // CHECK4-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5385 // CHECK4-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
5386 // CHECK4-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
5387 // CHECK4-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
5388 // CHECK4-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
5389 // CHECK4:       omp_offload.failed:
5390 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
5391 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT]]
5392 // CHECK4:       omp_offload.cont:
5393 // CHECK4-NEXT:    br label [[OMP_IF_END:%.*]]
5394 // CHECK4:       omp_if.else:
5395 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
5396 // CHECK4-NEXT:    br label [[OMP_IF_END]]
5397 // CHECK4:       omp_if.end:
5398 // CHECK4-NEXT:    [[TMP45:%.*]] = mul nsw i32 1, [[TMP1]]
5399 // CHECK4-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP45]]
5400 // CHECK4-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
5401 // CHECK4-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
5402 // CHECK4-NEXT:    [[CONV:%.*]] = sext i16 [[TMP46]] to i32
5403 // CHECK4-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
5404 // CHECK4-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP47]]
5405 // CHECK4-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
5406 // CHECK4-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
5407 // CHECK4-NEXT:    ret i32 [[ADD3]]
5408 //
5409 //
5410 // CHECK4-LABEL: define {{[^@]+}}@_ZL7fstatici
5411 // CHECK4-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
5412 // CHECK4-NEXT:  entry:
5413 // CHECK4-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5414 // CHECK4-NEXT:    [[A:%.*]] = alloca i32, align 4
5415 // CHECK4-NEXT:    [[AA:%.*]] = alloca i16, align 2
5416 // CHECK4-NEXT:    [[AAA:%.*]] = alloca i8, align 1
5417 // CHECK4-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
5418 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
5419 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
5420 // CHECK4-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
5421 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 4
5422 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 4
5423 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 4
5424 // CHECK4-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5425 // CHECK4-NEXT:    store i32 0, i32* [[A]], align 4
5426 // CHECK4-NEXT:    store i16 0, i16* [[AA]], align 2
5427 // CHECK4-NEXT:    store i8 0, i8* [[AAA]], align 1
5428 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
5429 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
5430 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
5431 // CHECK4-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
5432 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
5433 // CHECK4-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
5434 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
5435 // CHECK4-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
5436 // CHECK4-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
5437 // CHECK4-NEXT:    store i8 [[TMP4]], i8* [[CONV1]], align 1
5438 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
5439 // CHECK4-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
5440 // CHECK4-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
5441 // CHECK4-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
5442 // CHECK4:       omp_if.then:
5443 // CHECK4-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5444 // CHECK4-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
5445 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
5446 // CHECK4-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5447 // CHECK4-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
5448 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP10]], align 4
5449 // CHECK4-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
5450 // CHECK4-NEXT:    store i8* null, i8** [[TMP11]], align 4
5451 // CHECK4-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
5452 // CHECK4-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
5453 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
5454 // CHECK4-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
5455 // CHECK4-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
5456 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[TMP15]], align 4
5457 // CHECK4-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
5458 // CHECK4-NEXT:    store i8* null, i8** [[TMP16]], align 4
5459 // CHECK4-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
5460 // CHECK4-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
5461 // CHECK4-NEXT:    store i32 [[TMP5]], i32* [[TMP18]], align 4
5462 // CHECK4-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
5463 // CHECK4-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
5464 // CHECK4-NEXT:    store i32 [[TMP5]], i32* [[TMP20]], align 4
5465 // CHECK4-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
5466 // CHECK4-NEXT:    store i8* null, i8** [[TMP21]], align 4
5467 // CHECK4-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
5468 // CHECK4-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
5469 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 4
5470 // CHECK4-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
5471 // CHECK4-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
5472 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 4
5473 // CHECK4-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
5474 // CHECK4-NEXT:    store i8* null, i8** [[TMP26]], align 4
5475 // CHECK4-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5476 // CHECK4-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5477 // CHECK4-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
5478 // CHECK4-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
5479 // CHECK4-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
5480 // CHECK4:       omp_offload.failed:
5481 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
5482 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT]]
5483 // CHECK4:       omp_offload.cont:
5484 // CHECK4-NEXT:    br label [[OMP_IF_END:%.*]]
5485 // CHECK4:       omp_if.else:
5486 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
5487 // CHECK4-NEXT:    br label [[OMP_IF_END]]
5488 // CHECK4:       omp_if.end:
5489 // CHECK4-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
5490 // CHECK4-NEXT:    ret i32 [[TMP31]]
5491 //
5492 //
5493 // CHECK4-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
5494 // CHECK4-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
5495 // CHECK4-NEXT:  entry:
5496 // CHECK4-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5497 // CHECK4-NEXT:    [[A:%.*]] = alloca i32, align 4
5498 // CHECK4-NEXT:    [[AA:%.*]] = alloca i16, align 2
5499 // CHECK4-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
5500 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
5501 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
5502 // CHECK4-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
5503 // CHECK4-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
5504 // CHECK4-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
5505 // CHECK4-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5506 // CHECK4-NEXT:    store i32 0, i32* [[A]], align 4
5507 // CHECK4-NEXT:    store i16 0, i16* [[AA]], align 2
5508 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
5509 // CHECK4-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
5510 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
5511 // CHECK4-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
5512 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
5513 // CHECK4-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
5514 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
5515 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
5516 // CHECK4-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
5517 // CHECK4-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
5518 // CHECK4:       omp_if.then:
5519 // CHECK4-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5520 // CHECK4-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
5521 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP6]], align 4
5522 // CHECK4-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5523 // CHECK4-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
5524 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
5525 // CHECK4-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
5526 // CHECK4-NEXT:    store i8* null, i8** [[TMP9]], align 4
5527 // CHECK4-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
5528 // CHECK4-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i32*
5529 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[TMP11]], align 4
5530 // CHECK4-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
5531 // CHECK4-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
5532 // CHECK4-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
5533 // CHECK4-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
5534 // CHECK4-NEXT:    store i8* null, i8** [[TMP14]], align 4
5535 // CHECK4-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
5536 // CHECK4-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
5537 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 4
5538 // CHECK4-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
5539 // CHECK4-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
5540 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 4
5541 // CHECK4-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
5542 // CHECK4-NEXT:    store i8* null, i8** [[TMP19]], align 4
5543 // CHECK4-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5544 // CHECK4-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5545 // CHECK4-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
5546 // CHECK4-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
5547 // CHECK4-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
5548 // CHECK4:       omp_offload.failed:
5549 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
5550 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT]]
5551 // CHECK4:       omp_offload.cont:
5552 // CHECK4-NEXT:    br label [[OMP_IF_END:%.*]]
5553 // CHECK4:       omp_if.else:
5554 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
5555 // CHECK4-NEXT:    br label [[OMP_IF_END]]
5556 // CHECK4:       omp_if.end:
5557 // CHECK4-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
5558 // CHECK4-NEXT:    ret i32 [[TMP24]]
5559 //
5560 //
5561 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
5562 // CHECK4-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
5563 // CHECK4-NEXT:  entry:
5564 // CHECK4-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
5565 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
5566 // CHECK4-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
5567 // CHECK4-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
5568 // CHECK4-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
5569 // CHECK4-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
5570 // CHECK4-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
5571 // CHECK4-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
5572 // CHECK4-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
5573 // CHECK4-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
5574 // CHECK4-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
5575 // CHECK4-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
5576 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
5577 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
5578 // CHECK4-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
5579 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
5580 // CHECK4-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
5581 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
5582 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
5583 // CHECK4-NEXT:    ret void
5584 //
5585 //
5586 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..22
5587 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
5588 // CHECK4-NEXT:  entry:
5589 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5590 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5591 // CHECK4-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
5592 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
5593 // CHECK4-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
5594 // CHECK4-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
5595 // CHECK4-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
5596 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5597 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5598 // CHECK4-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
5599 // CHECK4-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
5600 // CHECK4-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
5601 // CHECK4-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
5602 // CHECK4-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
5603 // CHECK4-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
5604 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
5605 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
5606 // CHECK4-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
5607 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
5608 // CHECK4-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
5609 // CHECK4-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
5610 // CHECK4-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
5611 // CHECK4-NEXT:    store double [[ADD]], double* [[A]], align 4
5612 // CHECK4-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
5613 // CHECK4-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
5614 // CHECK4-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
5615 // CHECK4-NEXT:    store double [[INC]], double* [[A3]], align 4
5616 // CHECK4-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
5617 // CHECK4-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
5618 // CHECK4-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
5619 // CHECK4-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
5620 // CHECK4-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
5621 // CHECK4-NEXT:    ret void
5622 //
5623 //
5624 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
5625 // CHECK4-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
5626 // CHECK4-NEXT:  entry:
5627 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5628 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
5629 // CHECK4-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
5630 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
5631 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
5632 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
5633 // CHECK4-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
5634 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5635 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
5636 // CHECK4-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
5637 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
5638 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
5639 // CHECK4-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
5640 // CHECK4-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
5641 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
5642 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
5643 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
5644 // CHECK4-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
5645 // CHECK4-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
5646 // CHECK4-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
5647 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
5648 // CHECK4-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
5649 // CHECK4-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
5650 // CHECK4-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
5651 // CHECK4-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
5652 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
5653 // CHECK4-NEXT:    ret void
5654 //
5655 //
5656 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..24
5657 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
5658 // CHECK4-NEXT:  entry:
5659 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5660 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5661 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5662 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
5663 // CHECK4-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
5664 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
5665 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5666 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5667 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5668 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
5669 // CHECK4-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
5670 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
5671 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
5672 // CHECK4-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
5673 // CHECK4-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
5674 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
5675 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
5676 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
5677 // CHECK4-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
5678 // CHECK4-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
5679 // CHECK4-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
5680 // CHECK4-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
5681 // CHECK4-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
5682 // CHECK4-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
5683 // CHECK4-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
5684 // CHECK4-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
5685 // CHECK4-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
5686 // CHECK4-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
5687 // CHECK4-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
5688 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
5689 // CHECK4-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
5690 // CHECK4-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
5691 // CHECK4-NEXT:    ret void
5692 //
5693 //
5694 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
5695 // CHECK4-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
5696 // CHECK4-NEXT:  entry:
5697 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5698 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
5699 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
5700 // CHECK4-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
5701 // CHECK4-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
5702 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5703 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
5704 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
5705 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
5706 // CHECK4-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
5707 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
5708 // CHECK4-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
5709 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
5710 // CHECK4-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
5711 // CHECK4-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
5712 // CHECK4-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
5713 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
5714 // CHECK4-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
5715 // CHECK4-NEXT:    ret void
5716 //
5717 //
5718 // CHECK4-LABEL: define {{[^@]+}}@.omp_outlined..27
5719 // CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
5720 // CHECK4-NEXT:  entry:
5721 // CHECK4-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
5722 // CHECK4-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
5723 // CHECK4-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
5724 // CHECK4-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
5725 // CHECK4-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
5726 // CHECK4-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
5727 // CHECK4-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
5728 // CHECK4-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
5729 // CHECK4-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
5730 // CHECK4-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
5731 // CHECK4-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
5732 // CHECK4-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
5733 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
5734 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
5735 // CHECK4-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
5736 // CHECK4-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
5737 // CHECK4-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
5738 // CHECK4-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
5739 // CHECK4-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
5740 // CHECK4-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
5741 // CHECK4-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
5742 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
5743 // CHECK4-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
5744 // CHECK4-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
5745 // CHECK4-NEXT:    ret void
5746 //
5747 //
5748 // CHECK4-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
5749 // CHECK4-SAME: () #[[ATTR4]] {
5750 // CHECK4-NEXT:  entry:
5751 // CHECK4-NEXT:    call void @__tgt_register_requires(i64 1)
5752 // CHECK4-NEXT:    ret void
5753 //
5754 //
5755 // CHECK5-LABEL: define {{[^@]+}}@_Z3fooi
5756 // CHECK5-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
5757 // CHECK5-NEXT:  entry:
5758 // CHECK5-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5759 // CHECK5-NEXT:    [[A:%.*]] = alloca i32, align 4
5760 // CHECK5-NEXT:    [[AA:%.*]] = alloca i16, align 2
5761 // CHECK5-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
5762 // CHECK5-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
5763 // CHECK5-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
5764 // CHECK5-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
5765 // CHECK5-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
5766 // CHECK5-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
5767 // CHECK5-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
5768 // CHECK5-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
5769 // CHECK5-NEXT:    [[NN:%.*]] = alloca i32, align 4
5770 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5771 // CHECK5-NEXT:    store i32 0, i32* [[A]], align 4
5772 // CHECK5-NEXT:    store i16 0, i16* [[AA]], align 2
5773 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5774 // CHECK5-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
5775 // CHECK5-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
5776 // CHECK5-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
5777 // CHECK5-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
5778 // CHECK5-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
5779 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
5780 // CHECK5-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
5781 // CHECK5-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
5782 // CHECK5-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
5783 // CHECK5-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
5784 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
5785 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
5786 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
5787 // CHECK5-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
5788 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
5789 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
5790 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
5791 // CHECK5-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
5792 // CHECK5-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
5793 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
5794 // CHECK5-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
5795 // CHECK5-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
5796 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
5797 // CHECK5-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
5798 // CHECK5-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
5799 // CHECK5-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
5800 // CHECK5-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
5801 // CHECK5-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
5802 // CHECK5-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
5803 // CHECK5-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
5804 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
5805 // CHECK5-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
5806 // CHECK5-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
5807 // CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
5808 // CHECK5-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
5809 // CHECK5-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
5810 // CHECK5-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
5811 // CHECK5-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
5812 // CHECK5-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
5813 // CHECK5-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
5814 // CHECK5-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
5815 // CHECK5-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
5816 // CHECK5-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
5817 // CHECK5-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
5818 // CHECK5-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
5819 // CHECK5-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
5820 // CHECK5-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
5821 // CHECK5-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
5822 // CHECK5-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
5823 // CHECK5-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
5824 // CHECK5-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
5825 // CHECK5-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
5826 // CHECK5-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
5827 // CHECK5-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
5828 // CHECK5-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
5829 // CHECK5-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
5830 // CHECK5-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
5831 // CHECK5-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
5832 // CHECK5-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
5833 // CHECK5-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
5834 // CHECK5-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
5835 // CHECK5-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
5836 // CHECK5-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
5837 // CHECK5-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
5838 // CHECK5-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
5839 // CHECK5-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
5840 // CHECK5-NEXT:    store i32 0, i32* [[NN]], align 4
5841 // CHECK5-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
5842 // CHECK5-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
5843 // CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
5844 // CHECK5-NEXT:    ret i32 [[TMP20]]
5845 //
5846 //
5847 // CHECK5-LABEL: define {{[^@]+}}@_Z6bazzzziPi
5848 // CHECK5-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
5849 // CHECK5-NEXT:  entry:
5850 // CHECK5-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5851 // CHECK5-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
5852 // CHECK5-NEXT:    [[F1:%.*]] = alloca i32*, align 8
5853 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5854 // CHECK5-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
5855 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5856 // CHECK5-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
5857 // CHECK5-NEXT:    ret void
5858 //
5859 //
5860 // CHECK5-LABEL: define {{[^@]+}}@_Z3bari
5861 // CHECK5-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
5862 // CHECK5-NEXT:  entry:
5863 // CHECK5-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5864 // CHECK5-NEXT:    [[A:%.*]] = alloca i32, align 4
5865 // CHECK5-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
5866 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5867 // CHECK5-NEXT:    store i32 0, i32* [[A]], align 4
5868 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5869 // CHECK5-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
5870 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
5871 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
5872 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
5873 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
5874 // CHECK5-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
5875 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
5876 // CHECK5-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
5877 // CHECK5-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
5878 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
5879 // CHECK5-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
5880 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
5881 // CHECK5-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
5882 // CHECK5-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
5883 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
5884 // CHECK5-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
5885 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
5886 // CHECK5-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
5887 // CHECK5-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
5888 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
5889 // CHECK5-NEXT:    ret i32 [[TMP8]]
5890 //
5891 //
5892 // CHECK5-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
5893 // CHECK5-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
5894 // CHECK5-NEXT:  entry:
5895 // CHECK5-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
5896 // CHECK5-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5897 // CHECK5-NEXT:    [[B:%.*]] = alloca i32, align 4
5898 // CHECK5-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
5899 // CHECK5-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
5900 // CHECK5-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
5901 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5902 // CHECK5-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
5903 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
5904 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
5905 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
5906 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
5907 // CHECK5-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
5908 // CHECK5-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
5909 // CHECK5-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
5910 // CHECK5-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
5911 // CHECK5-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
5912 // CHECK5-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
5913 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
5914 // CHECK5-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
5915 // CHECK5-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
5916 // CHECK5-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
5917 // CHECK5-NEXT:    store double [[ADD2]], double* [[A]], align 8
5918 // CHECK5-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
5919 // CHECK5-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
5920 // CHECK5-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
5921 // CHECK5-NEXT:    store double [[INC]], double* [[A3]], align 8
5922 // CHECK5-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
5923 // CHECK5-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
5924 // CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
5925 // CHECK5-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
5926 // CHECK5-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
5927 // CHECK5-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
5928 // CHECK5-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
5929 // CHECK5-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
5930 // CHECK5-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
5931 // CHECK5-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
5932 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
5933 // CHECK5-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
5934 // CHECK5-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
5935 // CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
5936 // CHECK5-NEXT:    ret i32 [[ADD9]]
5937 //
5938 //
5939 // CHECK5-LABEL: define {{[^@]+}}@_ZL7fstatici
5940 // CHECK5-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
5941 // CHECK5-NEXT:  entry:
5942 // CHECK5-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5943 // CHECK5-NEXT:    [[A:%.*]] = alloca i32, align 4
5944 // CHECK5-NEXT:    [[AA:%.*]] = alloca i16, align 2
5945 // CHECK5-NEXT:    [[AAA:%.*]] = alloca i8, align 1
5946 // CHECK5-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
5947 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5948 // CHECK5-NEXT:    store i32 0, i32* [[A]], align 4
5949 // CHECK5-NEXT:    store i16 0, i16* [[AA]], align 2
5950 // CHECK5-NEXT:    store i8 0, i8* [[AAA]], align 1
5951 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
5952 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
5953 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
5954 // CHECK5-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
5955 // CHECK5-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
5956 // CHECK5-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
5957 // CHECK5-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
5958 // CHECK5-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
5959 // CHECK5-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
5960 // CHECK5-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
5961 // CHECK5-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
5962 // CHECK5-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
5963 // CHECK5-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
5964 // CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
5965 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
5966 // CHECK5-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
5967 // CHECK5-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
5968 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
5969 // CHECK5-NEXT:    ret i32 [[TMP4]]
5970 //
5971 //
5972 // CHECK5-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
5973 // CHECK5-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
5974 // CHECK5-NEXT:  entry:
5975 // CHECK5-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
5976 // CHECK5-NEXT:    [[A:%.*]] = alloca i32, align 4
5977 // CHECK5-NEXT:    [[AA:%.*]] = alloca i16, align 2
5978 // CHECK5-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
5979 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
5980 // CHECK5-NEXT:    store i32 0, i32* [[A]], align 4
5981 // CHECK5-NEXT:    store i16 0, i16* [[AA]], align 2
5982 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
5983 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
5984 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
5985 // CHECK5-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
5986 // CHECK5-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
5987 // CHECK5-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
5988 // CHECK5-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
5989 // CHECK5-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
5990 // CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
5991 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
5992 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
5993 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
5994 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
5995 // CHECK5-NEXT:    ret i32 [[TMP3]]
5996 //
5997 //
5998 // CHECK6-LABEL: define {{[^@]+}}@_Z3fooi
5999 // CHECK6-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
6000 // CHECK6-NEXT:  entry:
6001 // CHECK6-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6002 // CHECK6-NEXT:    [[A:%.*]] = alloca i32, align 4
6003 // CHECK6-NEXT:    [[AA:%.*]] = alloca i16, align 2
6004 // CHECK6-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
6005 // CHECK6-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
6006 // CHECK6-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
6007 // CHECK6-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
6008 // CHECK6-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
6009 // CHECK6-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
6010 // CHECK6-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
6011 // CHECK6-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
6012 // CHECK6-NEXT:    [[NN:%.*]] = alloca i32, align 4
6013 // CHECK6-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6014 // CHECK6-NEXT:    store i32 0, i32* [[A]], align 4
6015 // CHECK6-NEXT:    store i16 0, i16* [[AA]], align 2
6016 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6017 // CHECK6-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
6018 // CHECK6-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
6019 // CHECK6-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
6020 // CHECK6-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
6021 // CHECK6-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
6022 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
6023 // CHECK6-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
6024 // CHECK6-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
6025 // CHECK6-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
6026 // CHECK6-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
6027 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
6028 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
6029 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
6030 // CHECK6-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
6031 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
6032 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
6033 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6034 // CHECK6-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
6035 // CHECK6-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
6036 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
6037 // CHECK6-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
6038 // CHECK6-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
6039 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
6040 // CHECK6-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
6041 // CHECK6-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
6042 // CHECK6-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
6043 // CHECK6-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
6044 // CHECK6-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
6045 // CHECK6-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
6046 // CHECK6-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
6047 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
6048 // CHECK6-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
6049 // CHECK6-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
6050 // CHECK6-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
6051 // CHECK6-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
6052 // CHECK6-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
6053 // CHECK6-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
6054 // CHECK6-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
6055 // CHECK6-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
6056 // CHECK6-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
6057 // CHECK6-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
6058 // CHECK6-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
6059 // CHECK6-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
6060 // CHECK6-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
6061 // CHECK6-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
6062 // CHECK6-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
6063 // CHECK6-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
6064 // CHECK6-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
6065 // CHECK6-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
6066 // CHECK6-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
6067 // CHECK6-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
6068 // CHECK6-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
6069 // CHECK6-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
6070 // CHECK6-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
6071 // CHECK6-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
6072 // CHECK6-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
6073 // CHECK6-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
6074 // CHECK6-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
6075 // CHECK6-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
6076 // CHECK6-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
6077 // CHECK6-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
6078 // CHECK6-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
6079 // CHECK6-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
6080 // CHECK6-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
6081 // CHECK6-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
6082 // CHECK6-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
6083 // CHECK6-NEXT:    store i32 0, i32* [[NN]], align 4
6084 // CHECK6-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
6085 // CHECK6-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
6086 // CHECK6-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
6087 // CHECK6-NEXT:    ret i32 [[TMP20]]
6088 //
6089 //
6090 // CHECK6-LABEL: define {{[^@]+}}@_Z6bazzzziPi
6091 // CHECK6-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
6092 // CHECK6-NEXT:  entry:
6093 // CHECK6-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6094 // CHECK6-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
6095 // CHECK6-NEXT:    [[F1:%.*]] = alloca i32*, align 8
6096 // CHECK6-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6097 // CHECK6-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
6098 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6099 // CHECK6-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
6100 // CHECK6-NEXT:    ret void
6101 //
6102 //
6103 // CHECK6-LABEL: define {{[^@]+}}@_Z3bari
6104 // CHECK6-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
6105 // CHECK6-NEXT:  entry:
6106 // CHECK6-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6107 // CHECK6-NEXT:    [[A:%.*]] = alloca i32, align 4
6108 // CHECK6-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
6109 // CHECK6-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6110 // CHECK6-NEXT:    store i32 0, i32* [[A]], align 4
6111 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6112 // CHECK6-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
6113 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
6114 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
6115 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6116 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
6117 // CHECK6-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
6118 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
6119 // CHECK6-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
6120 // CHECK6-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
6121 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
6122 // CHECK6-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
6123 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
6124 // CHECK6-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
6125 // CHECK6-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
6126 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
6127 // CHECK6-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
6128 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
6129 // CHECK6-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
6130 // CHECK6-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
6131 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
6132 // CHECK6-NEXT:    ret i32 [[TMP8]]
6133 //
6134 //
6135 // CHECK6-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
6136 // CHECK6-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
6137 // CHECK6-NEXT:  entry:
6138 // CHECK6-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
6139 // CHECK6-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6140 // CHECK6-NEXT:    [[B:%.*]] = alloca i32, align 4
6141 // CHECK6-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
6142 // CHECK6-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
6143 // CHECK6-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
6144 // CHECK6-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6145 // CHECK6-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
6146 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6147 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6148 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
6149 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
6150 // CHECK6-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
6151 // CHECK6-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
6152 // CHECK6-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
6153 // CHECK6-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
6154 // CHECK6-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
6155 // CHECK6-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
6156 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
6157 // CHECK6-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
6158 // CHECK6-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
6159 // CHECK6-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
6160 // CHECK6-NEXT:    store double [[ADD2]], double* [[A]], align 8
6161 // CHECK6-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
6162 // CHECK6-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
6163 // CHECK6-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
6164 // CHECK6-NEXT:    store double [[INC]], double* [[A3]], align 8
6165 // CHECK6-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
6166 // CHECK6-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
6167 // CHECK6-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
6168 // CHECK6-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
6169 // CHECK6-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
6170 // CHECK6-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
6171 // CHECK6-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
6172 // CHECK6-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
6173 // CHECK6-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
6174 // CHECK6-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
6175 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
6176 // CHECK6-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
6177 // CHECK6-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
6178 // CHECK6-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
6179 // CHECK6-NEXT:    ret i32 [[ADD9]]
6180 //
6181 //
6182 // CHECK6-LABEL: define {{[^@]+}}@_ZL7fstatici
6183 // CHECK6-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
6184 // CHECK6-NEXT:  entry:
6185 // CHECK6-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6186 // CHECK6-NEXT:    [[A:%.*]] = alloca i32, align 4
6187 // CHECK6-NEXT:    [[AA:%.*]] = alloca i16, align 2
6188 // CHECK6-NEXT:    [[AAA:%.*]] = alloca i8, align 1
6189 // CHECK6-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
6190 // CHECK6-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6191 // CHECK6-NEXT:    store i32 0, i32* [[A]], align 4
6192 // CHECK6-NEXT:    store i16 0, i16* [[AA]], align 2
6193 // CHECK6-NEXT:    store i8 0, i8* [[AAA]], align 1
6194 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
6195 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6196 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6197 // CHECK6-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
6198 // CHECK6-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
6199 // CHECK6-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
6200 // CHECK6-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
6201 // CHECK6-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
6202 // CHECK6-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
6203 // CHECK6-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
6204 // CHECK6-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
6205 // CHECK6-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
6206 // CHECK6-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
6207 // CHECK6-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
6208 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
6209 // CHECK6-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
6210 // CHECK6-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
6211 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
6212 // CHECK6-NEXT:    ret i32 [[TMP4]]
6213 //
6214 //
6215 // CHECK6-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
6216 // CHECK6-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
6217 // CHECK6-NEXT:  entry:
6218 // CHECK6-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6219 // CHECK6-NEXT:    [[A:%.*]] = alloca i32, align 4
6220 // CHECK6-NEXT:    [[AA:%.*]] = alloca i16, align 2
6221 // CHECK6-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
6222 // CHECK6-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6223 // CHECK6-NEXT:    store i32 0, i32* [[A]], align 4
6224 // CHECK6-NEXT:    store i16 0, i16* [[AA]], align 2
6225 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
6226 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6227 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6228 // CHECK6-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
6229 // CHECK6-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
6230 // CHECK6-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
6231 // CHECK6-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
6232 // CHECK6-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
6233 // CHECK6-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
6234 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
6235 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
6236 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
6237 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
6238 // CHECK6-NEXT:    ret i32 [[TMP3]]
6239 //
6240 //
6241 // CHECK7-LABEL: define {{[^@]+}}@_Z3fooi
6242 // CHECK7-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
6243 // CHECK7-NEXT:  entry:
6244 // CHECK7-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6245 // CHECK7-NEXT:    [[A:%.*]] = alloca i32, align 4
6246 // CHECK7-NEXT:    [[AA:%.*]] = alloca i16, align 2
6247 // CHECK7-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
6248 // CHECK7-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
6249 // CHECK7-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
6250 // CHECK7-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
6251 // CHECK7-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
6252 // CHECK7-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
6253 // CHECK7-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
6254 // CHECK7-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
6255 // CHECK7-NEXT:    [[NN:%.*]] = alloca i32, align 4
6256 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6257 // CHECK7-NEXT:    store i32 0, i32* [[A]], align 4
6258 // CHECK7-NEXT:    store i16 0, i16* [[AA]], align 2
6259 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6260 // CHECK7-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
6261 // CHECK7-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
6262 // CHECK7-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
6263 // CHECK7-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
6264 // CHECK7-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
6265 // CHECK7-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
6266 // CHECK7-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
6267 // CHECK7-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
6268 // CHECK7-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
6269 // CHECK7-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
6270 // CHECK7-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
6271 // CHECK7-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
6272 // CHECK7-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
6273 // CHECK7-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
6274 // CHECK7-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6275 // CHECK7-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
6276 // CHECK7-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
6277 // CHECK7-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
6278 // CHECK7-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
6279 // CHECK7-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
6280 // CHECK7-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
6281 // CHECK7-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
6282 // CHECK7-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
6283 // CHECK7-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
6284 // CHECK7-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
6285 // CHECK7-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
6286 // CHECK7-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
6287 // CHECK7-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
6288 // CHECK7-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
6289 // CHECK7-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
6290 // CHECK7-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
6291 // CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
6292 // CHECK7-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
6293 // CHECK7-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
6294 // CHECK7-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
6295 // CHECK7-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
6296 // CHECK7-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
6297 // CHECK7-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
6298 // CHECK7-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
6299 // CHECK7-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
6300 // CHECK7-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
6301 // CHECK7-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
6302 // CHECK7-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
6303 // CHECK7-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
6304 // CHECK7-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
6305 // CHECK7-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
6306 // CHECK7-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
6307 // CHECK7-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
6308 // CHECK7-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
6309 // CHECK7-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
6310 // CHECK7-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
6311 // CHECK7-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
6312 // CHECK7-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
6313 // CHECK7-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
6314 // CHECK7-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
6315 // CHECK7-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
6316 // CHECK7-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
6317 // CHECK7-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
6318 // CHECK7-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
6319 // CHECK7-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
6320 // CHECK7-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
6321 // CHECK7-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
6322 // CHECK7-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
6323 // CHECK7-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
6324 // CHECK7-NEXT:    store i32 0, i32* [[NN]], align 4
6325 // CHECK7-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
6326 // CHECK7-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
6327 // CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
6328 // CHECK7-NEXT:    ret i32 [[TMP18]]
6329 //
6330 //
6331 // CHECK7-LABEL: define {{[^@]+}}@_Z6bazzzziPi
6332 // CHECK7-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
6333 // CHECK7-NEXT:  entry:
6334 // CHECK7-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6335 // CHECK7-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
6336 // CHECK7-NEXT:    [[F1:%.*]] = alloca i32*, align 4
6337 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6338 // CHECK7-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
6339 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6340 // CHECK7-NEXT:    ret void
6341 //
6342 //
6343 // CHECK7-LABEL: define {{[^@]+}}@_Z3bari
6344 // CHECK7-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
6345 // CHECK7-NEXT:  entry:
6346 // CHECK7-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6347 // CHECK7-NEXT:    [[A:%.*]] = alloca i32, align 4
6348 // CHECK7-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
6349 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6350 // CHECK7-NEXT:    store i32 0, i32* [[A]], align 4
6351 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6352 // CHECK7-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
6353 // CHECK7-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
6354 // CHECK7-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
6355 // CHECK7-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6356 // CHECK7-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
6357 // CHECK7-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
6358 // CHECK7-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
6359 // CHECK7-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
6360 // CHECK7-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
6361 // CHECK7-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
6362 // CHECK7-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
6363 // CHECK7-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
6364 // CHECK7-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
6365 // CHECK7-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
6366 // CHECK7-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
6367 // CHECK7-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
6368 // CHECK7-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
6369 // CHECK7-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
6370 // CHECK7-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
6371 // CHECK7-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
6372 // CHECK7-NEXT:    ret i32 [[TMP8]]
6373 //
6374 //
6375 // CHECK7-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
6376 // CHECK7-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
6377 // CHECK7-NEXT:  entry:
6378 // CHECK7-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
6379 // CHECK7-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6380 // CHECK7-NEXT:    [[B:%.*]] = alloca i32, align 4
6381 // CHECK7-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
6382 // CHECK7-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
6383 // CHECK7-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
6384 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6385 // CHECK7-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
6386 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6387 // CHECK7-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6388 // CHECK7-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
6389 // CHECK7-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
6390 // CHECK7-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
6391 // CHECK7-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
6392 // CHECK7-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
6393 // CHECK7-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
6394 // CHECK7-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
6395 // CHECK7-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
6396 // CHECK7-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
6397 // CHECK7-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
6398 // CHECK7-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
6399 // CHECK7-NEXT:    store double [[ADD2]], double* [[A]], align 4
6400 // CHECK7-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
6401 // CHECK7-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
6402 // CHECK7-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
6403 // CHECK7-NEXT:    store double [[INC]], double* [[A3]], align 4
6404 // CHECK7-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
6405 // CHECK7-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
6406 // CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
6407 // CHECK7-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
6408 // CHECK7-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
6409 // CHECK7-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
6410 // CHECK7-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
6411 // CHECK7-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
6412 // CHECK7-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
6413 // CHECK7-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
6414 // CHECK7-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
6415 // CHECK7-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
6416 // CHECK7-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
6417 // CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
6418 // CHECK7-NEXT:    ret i32 [[ADD9]]
6419 //
6420 //
6421 // CHECK7-LABEL: define {{[^@]+}}@_ZL7fstatici
6422 // CHECK7-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
6423 // CHECK7-NEXT:  entry:
6424 // CHECK7-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6425 // CHECK7-NEXT:    [[A:%.*]] = alloca i32, align 4
6426 // CHECK7-NEXT:    [[AA:%.*]] = alloca i16, align 2
6427 // CHECK7-NEXT:    [[AAA:%.*]] = alloca i8, align 1
6428 // CHECK7-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
6429 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6430 // CHECK7-NEXT:    store i32 0, i32* [[A]], align 4
6431 // CHECK7-NEXT:    store i16 0, i16* [[AA]], align 2
6432 // CHECK7-NEXT:    store i8 0, i8* [[AAA]], align 1
6433 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
6434 // CHECK7-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6435 // CHECK7-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6436 // CHECK7-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
6437 // CHECK7-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
6438 // CHECK7-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
6439 // CHECK7-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
6440 // CHECK7-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
6441 // CHECK7-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
6442 // CHECK7-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
6443 // CHECK7-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
6444 // CHECK7-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
6445 // CHECK7-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
6446 // CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
6447 // CHECK7-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
6448 // CHECK7-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
6449 // CHECK7-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
6450 // CHECK7-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
6451 // CHECK7-NEXT:    ret i32 [[TMP4]]
6452 //
6453 //
6454 // CHECK7-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
6455 // CHECK7-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
6456 // CHECK7-NEXT:  entry:
6457 // CHECK7-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6458 // CHECK7-NEXT:    [[A:%.*]] = alloca i32, align 4
6459 // CHECK7-NEXT:    [[AA:%.*]] = alloca i16, align 2
6460 // CHECK7-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
6461 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6462 // CHECK7-NEXT:    store i32 0, i32* [[A]], align 4
6463 // CHECK7-NEXT:    store i16 0, i16* [[AA]], align 2
6464 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
6465 // CHECK7-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6466 // CHECK7-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6467 // CHECK7-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
6468 // CHECK7-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
6469 // CHECK7-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
6470 // CHECK7-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
6471 // CHECK7-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
6472 // CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
6473 // CHECK7-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
6474 // CHECK7-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
6475 // CHECK7-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
6476 // CHECK7-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
6477 // CHECK7-NEXT:    ret i32 [[TMP3]]
6478 //
6479 //
6480 // CHECK8-LABEL: define {{[^@]+}}@_Z3fooi
6481 // CHECK8-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
6482 // CHECK8-NEXT:  entry:
6483 // CHECK8-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6484 // CHECK8-NEXT:    [[A:%.*]] = alloca i32, align 4
6485 // CHECK8-NEXT:    [[AA:%.*]] = alloca i16, align 2
6486 // CHECK8-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
6487 // CHECK8-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
6488 // CHECK8-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
6489 // CHECK8-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
6490 // CHECK8-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
6491 // CHECK8-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
6492 // CHECK8-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
6493 // CHECK8-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
6494 // CHECK8-NEXT:    [[NN:%.*]] = alloca i32, align 4
6495 // CHECK8-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6496 // CHECK8-NEXT:    store i32 0, i32* [[A]], align 4
6497 // CHECK8-NEXT:    store i16 0, i16* [[AA]], align 2
6498 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6499 // CHECK8-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
6500 // CHECK8-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
6501 // CHECK8-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
6502 // CHECK8-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
6503 // CHECK8-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
6504 // CHECK8-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
6505 // CHECK8-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
6506 // CHECK8-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
6507 // CHECK8-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
6508 // CHECK8-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
6509 // CHECK8-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
6510 // CHECK8-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
6511 // CHECK8-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
6512 // CHECK8-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
6513 // CHECK8-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6514 // CHECK8-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
6515 // CHECK8-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
6516 // CHECK8-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
6517 // CHECK8-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
6518 // CHECK8-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
6519 // CHECK8-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
6520 // CHECK8-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
6521 // CHECK8-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
6522 // CHECK8-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
6523 // CHECK8-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
6524 // CHECK8-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
6525 // CHECK8-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
6526 // CHECK8-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
6527 // CHECK8-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
6528 // CHECK8-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
6529 // CHECK8-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
6530 // CHECK8-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
6531 // CHECK8-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
6532 // CHECK8-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
6533 // CHECK8-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
6534 // CHECK8-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
6535 // CHECK8-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
6536 // CHECK8-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
6537 // CHECK8-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
6538 // CHECK8-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
6539 // CHECK8-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
6540 // CHECK8-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
6541 // CHECK8-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
6542 // CHECK8-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
6543 // CHECK8-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
6544 // CHECK8-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
6545 // CHECK8-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
6546 // CHECK8-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
6547 // CHECK8-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
6548 // CHECK8-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
6549 // CHECK8-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
6550 // CHECK8-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
6551 // CHECK8-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
6552 // CHECK8-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
6553 // CHECK8-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
6554 // CHECK8-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
6555 // CHECK8-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
6556 // CHECK8-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
6557 // CHECK8-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
6558 // CHECK8-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
6559 // CHECK8-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
6560 // CHECK8-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
6561 // CHECK8-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
6562 // CHECK8-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
6563 // CHECK8-NEXT:    store i32 0, i32* [[NN]], align 4
6564 // CHECK8-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
6565 // CHECK8-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
6566 // CHECK8-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
6567 // CHECK8-NEXT:    ret i32 [[TMP18]]
6568 //
6569 //
6570 // CHECK8-LABEL: define {{[^@]+}}@_Z6bazzzziPi
6571 // CHECK8-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
6572 // CHECK8-NEXT:  entry:
6573 // CHECK8-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6574 // CHECK8-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
6575 // CHECK8-NEXT:    [[F1:%.*]] = alloca i32*, align 4
6576 // CHECK8-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6577 // CHECK8-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
6578 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6579 // CHECK8-NEXT:    ret void
6580 //
6581 //
6582 // CHECK8-LABEL: define {{[^@]+}}@_Z3bari
6583 // CHECK8-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
6584 // CHECK8-NEXT:  entry:
6585 // CHECK8-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6586 // CHECK8-NEXT:    [[A:%.*]] = alloca i32, align 4
6587 // CHECK8-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
6588 // CHECK8-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6589 // CHECK8-NEXT:    store i32 0, i32* [[A]], align 4
6590 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6591 // CHECK8-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
6592 // CHECK8-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
6593 // CHECK8-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
6594 // CHECK8-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6595 // CHECK8-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
6596 // CHECK8-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
6597 // CHECK8-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
6598 // CHECK8-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
6599 // CHECK8-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
6600 // CHECK8-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
6601 // CHECK8-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
6602 // CHECK8-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
6603 // CHECK8-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
6604 // CHECK8-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
6605 // CHECK8-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
6606 // CHECK8-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
6607 // CHECK8-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
6608 // CHECK8-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
6609 // CHECK8-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
6610 // CHECK8-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
6611 // CHECK8-NEXT:    ret i32 [[TMP8]]
6612 //
6613 //
6614 // CHECK8-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
6615 // CHECK8-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
6616 // CHECK8-NEXT:  entry:
6617 // CHECK8-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
6618 // CHECK8-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6619 // CHECK8-NEXT:    [[B:%.*]] = alloca i32, align 4
6620 // CHECK8-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
6621 // CHECK8-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
6622 // CHECK8-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
6623 // CHECK8-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6624 // CHECK8-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
6625 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
6626 // CHECK8-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6627 // CHECK8-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
6628 // CHECK8-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
6629 // CHECK8-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
6630 // CHECK8-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
6631 // CHECK8-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
6632 // CHECK8-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
6633 // CHECK8-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
6634 // CHECK8-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
6635 // CHECK8-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
6636 // CHECK8-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
6637 // CHECK8-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
6638 // CHECK8-NEXT:    store double [[ADD2]], double* [[A]], align 4
6639 // CHECK8-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
6640 // CHECK8-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
6641 // CHECK8-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
6642 // CHECK8-NEXT:    store double [[INC]], double* [[A3]], align 4
6643 // CHECK8-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
6644 // CHECK8-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
6645 // CHECK8-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
6646 // CHECK8-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
6647 // CHECK8-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
6648 // CHECK8-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
6649 // CHECK8-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
6650 // CHECK8-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
6651 // CHECK8-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
6652 // CHECK8-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
6653 // CHECK8-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
6654 // CHECK8-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
6655 // CHECK8-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
6656 // CHECK8-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
6657 // CHECK8-NEXT:    ret i32 [[ADD9]]
6658 //
6659 //
6660 // CHECK8-LABEL: define {{[^@]+}}@_ZL7fstatici
6661 // CHECK8-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
6662 // CHECK8-NEXT:  entry:
6663 // CHECK8-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6664 // CHECK8-NEXT:    [[A:%.*]] = alloca i32, align 4
6665 // CHECK8-NEXT:    [[AA:%.*]] = alloca i16, align 2
6666 // CHECK8-NEXT:    [[AAA:%.*]] = alloca i8, align 1
6667 // CHECK8-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
6668 // CHECK8-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6669 // CHECK8-NEXT:    store i32 0, i32* [[A]], align 4
6670 // CHECK8-NEXT:    store i16 0, i16* [[AA]], align 2
6671 // CHECK8-NEXT:    store i8 0, i8* [[AAA]], align 1
6672 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
6673 // CHECK8-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6674 // CHECK8-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6675 // CHECK8-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
6676 // CHECK8-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
6677 // CHECK8-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
6678 // CHECK8-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
6679 // CHECK8-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
6680 // CHECK8-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
6681 // CHECK8-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
6682 // CHECK8-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
6683 // CHECK8-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
6684 // CHECK8-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
6685 // CHECK8-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
6686 // CHECK8-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
6687 // CHECK8-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
6688 // CHECK8-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
6689 // CHECK8-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
6690 // CHECK8-NEXT:    ret i32 [[TMP4]]
6691 //
6692 //
6693 // CHECK8-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
6694 // CHECK8-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
6695 // CHECK8-NEXT:  entry:
6696 // CHECK8-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
6697 // CHECK8-NEXT:    [[A:%.*]] = alloca i32, align 4
6698 // CHECK8-NEXT:    [[AA:%.*]] = alloca i16, align 2
6699 // CHECK8-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
6700 // CHECK8-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
6701 // CHECK8-NEXT:    store i32 0, i32* [[A]], align 4
6702 // CHECK8-NEXT:    store i16 0, i16* [[AA]], align 2
6703 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
6704 // CHECK8-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6705 // CHECK8-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
6706 // CHECK8-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
6707 // CHECK8-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
6708 // CHECK8-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
6709 // CHECK8-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
6710 // CHECK8-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
6711 // CHECK8-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
6712 // CHECK8-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
6713 // CHECK8-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
6714 // CHECK8-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
6715 // CHECK8-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
6716 // CHECK8-NEXT:    ret i32 [[TMP3]]
6717 //
6718 //
6719 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
6720 // CHECK9-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
6721 // CHECK9-NEXT:  entry:
6722 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
6723 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
6724 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
6725 // CHECK9-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
6726 // CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
6727 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
6728 // CHECK9-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
6729 // CHECK9-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
6730 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
6731 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
6732 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
6733 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
6734 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
6735 // CHECK9-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
6736 // CHECK9-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
6737 // CHECK9-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
6738 // CHECK9-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
6739 // CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
6740 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
6741 // CHECK9-NEXT:    ret void
6742 //
6743 //
6744 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined.
6745 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
6746 // CHECK9-NEXT:  entry:
6747 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6748 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6749 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
6750 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6751 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6752 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
6753 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
6754 // CHECK9-NEXT:    ret void
6755 //
6756 //
6757 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
6758 // CHECK9-SAME: (i64 [[AA:%.*]]) #[[ATTR0]] {
6759 // CHECK9-NEXT:  entry:
6760 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
6761 // CHECK9-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
6762 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
6763 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
6764 // CHECK9-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
6765 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
6766 // CHECK9-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
6767 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
6768 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP1]])
6769 // CHECK9-NEXT:    ret void
6770 //
6771 //
6772 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..1
6773 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
6774 // CHECK9-NEXT:  entry:
6775 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6776 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6777 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
6778 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6779 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6780 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
6781 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
6782 // CHECK9-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
6783 // CHECK9-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
6784 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
6785 // CHECK9-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
6786 // CHECK9-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
6787 // CHECK9-NEXT:    ret void
6788 //
6789 //
6790 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
6791 // CHECK9-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
6792 // CHECK9-NEXT:  entry:
6793 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
6794 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
6795 // CHECK9-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
6796 // CHECK9-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
6797 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
6798 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
6799 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
6800 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
6801 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
6802 // CHECK9-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
6803 // CHECK9-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
6804 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
6805 // CHECK9-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
6806 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
6807 // CHECK9-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
6808 // CHECK9-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
6809 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
6810 // CHECK9-NEXT:    ret void
6811 //
6812 //
6813 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..2
6814 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
6815 // CHECK9-NEXT:  entry:
6816 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6817 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6818 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
6819 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
6820 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6821 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6822 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
6823 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
6824 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
6825 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
6826 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
6827 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
6828 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
6829 // CHECK9-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
6830 // CHECK9-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
6831 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
6832 // CHECK9-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
6833 // CHECK9-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
6834 // CHECK9-NEXT:    ret void
6835 //
6836 //
6837 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
6838 // CHECK9-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
6839 // CHECK9-NEXT:  entry:
6840 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
6841 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
6842 // CHECK9-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
6843 // CHECK9-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
6844 // CHECK9-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
6845 // CHECK9-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
6846 // CHECK9-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
6847 // CHECK9-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
6848 // CHECK9-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
6849 // CHECK9-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
6850 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
6851 // CHECK9-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
6852 // CHECK9-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
6853 // CHECK9-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
6854 // CHECK9-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
6855 // CHECK9-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
6856 // CHECK9-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
6857 // CHECK9-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
6858 // CHECK9-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
6859 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
6860 // CHECK9-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
6861 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
6862 // CHECK9-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
6863 // CHECK9-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
6864 // CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
6865 // CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
6866 // CHECK9-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
6867 // CHECK9-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
6868 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
6869 // CHECK9-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
6870 // CHECK9-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
6871 // CHECK9-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
6872 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
6873 // CHECK9-NEXT:    ret void
6874 //
6875 //
6876 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..3
6877 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
6878 // CHECK9-NEXT:  entry:
6879 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6880 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6881 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
6882 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
6883 // CHECK9-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
6884 // CHECK9-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
6885 // CHECK9-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
6886 // CHECK9-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
6887 // CHECK9-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
6888 // CHECK9-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
6889 // CHECK9-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
6890 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6891 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6892 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
6893 // CHECK9-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
6894 // CHECK9-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
6895 // CHECK9-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
6896 // CHECK9-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
6897 // CHECK9-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
6898 // CHECK9-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
6899 // CHECK9-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
6900 // CHECK9-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
6901 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
6902 // CHECK9-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
6903 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
6904 // CHECK9-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
6905 // CHECK9-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
6906 // CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
6907 // CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
6908 // CHECK9-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
6909 // CHECK9-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
6910 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
6911 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
6912 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
6913 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
6914 // CHECK9-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
6915 // CHECK9-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
6916 // CHECK9-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
6917 // CHECK9-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
6918 // CHECK9-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
6919 // CHECK9-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
6920 // CHECK9-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
6921 // CHECK9-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
6922 // CHECK9-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
6923 // CHECK9-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
6924 // CHECK9-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
6925 // CHECK9-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
6926 // CHECK9-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
6927 // CHECK9-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
6928 // CHECK9-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
6929 // CHECK9-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
6930 // CHECK9-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
6931 // CHECK9-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
6932 // CHECK9-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
6933 // CHECK9-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
6934 // CHECK9-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
6935 // CHECK9-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
6936 // CHECK9-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
6937 // CHECK9-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
6938 // CHECK9-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
6939 // CHECK9-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
6940 // CHECK9-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
6941 // CHECK9-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
6942 // CHECK9-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
6943 // CHECK9-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
6944 // CHECK9-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
6945 // CHECK9-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
6946 // CHECK9-NEXT:    ret void
6947 //
6948 //
6949 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
6950 // CHECK9-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
6951 // CHECK9-NEXT:  entry:
6952 // CHECK9-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
6953 // CHECK9-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
6954 // CHECK9-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
6955 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
6956 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
6957 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
6958 // CHECK9-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
6959 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
6960 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i64 [[TMP1]])
6961 // CHECK9-NEXT:    ret void
6962 //
6963 //
6964 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..4
6965 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
6966 // CHECK9-NEXT:  entry:
6967 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6968 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6969 // CHECK9-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
6970 // CHECK9-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
6971 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6972 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6973 // CHECK9-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
6974 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
6975 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
6976 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
6977 // CHECK9-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
6978 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
6979 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP1]])
6980 // CHECK9-NEXT:    ret void
6981 //
6982 //
6983 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..5
6984 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
6985 // CHECK9-NEXT:  entry:
6986 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6987 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6988 // CHECK9-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
6989 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6990 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6991 // CHECK9-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
6992 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
6993 // CHECK9-NEXT:    ret void
6994 //
6995 //
6996 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
6997 // CHECK9-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
6998 // CHECK9-NEXT:  entry:
6999 // CHECK9-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7000 // CHECK9-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
7001 // CHECK9-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7002 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7003 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
7004 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
7005 // CHECK9-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
7006 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
7007 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]])
7008 // CHECK9-NEXT:    ret void
7009 //
7010 //
7011 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..6
7012 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
7013 // CHECK9-NEXT:  entry:
7014 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7015 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7016 // CHECK9-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7017 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7018 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7019 // CHECK9-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7020 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7021 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[CONV]])
7022 // CHECK9-NEXT:    ret void
7023 //
7024 //
7025 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..7
7026 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
7027 // CHECK9-NEXT:  entry:
7028 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7029 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7030 // CHECK9-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
7031 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7032 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7033 // CHECK9-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
7034 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
7035 // CHECK9-NEXT:    ret void
7036 //
7037 //
7038 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
7039 // CHECK9-SAME: (i64 [[VLA:%.*]]) #[[ATTR0]] {
7040 // CHECK9-NEXT:  entry:
7041 // CHECK9-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7042 // CHECK9-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7043 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7044 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i64 [[TMP0]])
7045 // CHECK9-NEXT:    ret void
7046 //
7047 //
7048 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..8
7049 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR0]] {
7050 // CHECK9-NEXT:  entry:
7051 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7052 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7053 // CHECK9-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7054 // CHECK9-NEXT:    [[F:%.*]] = alloca i32*, align 8
7055 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7056 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7057 // CHECK9-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7058 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7059 // CHECK9-NEXT:    ret void
7060 //
7061 //
7062 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
7063 // CHECK9-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7064 // CHECK9-NEXT:  entry:
7065 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7066 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7067 // CHECK9-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
7068 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7069 // CHECK9-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
7070 // CHECK9-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7071 // CHECK9-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
7072 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7073 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7074 // CHECK9-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
7075 // CHECK9-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7076 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7077 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7078 // CHECK9-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
7079 // CHECK9-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7080 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7081 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
7082 // CHECK9-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
7083 // CHECK9-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
7084 // CHECK9-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
7085 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7086 // CHECK9-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
7087 // CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7088 // CHECK9-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
7089 // CHECK9-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
7090 // CHECK9-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
7091 // CHECK9-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
7092 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
7093 // CHECK9-NEXT:    ret void
7094 //
7095 //
7096 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..9
7097 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7098 // CHECK9-NEXT:  entry:
7099 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7100 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7101 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7102 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7103 // CHECK9-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
7104 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7105 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7106 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7107 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7108 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7109 // CHECK9-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
7110 // CHECK9-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7111 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7112 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7113 // CHECK9-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
7114 // CHECK9-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7115 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7116 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
7117 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
7118 // CHECK9-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
7119 // CHECK9-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
7120 // CHECK9-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
7121 // CHECK9-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
7122 // CHECK9-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
7123 // CHECK9-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
7124 // CHECK9-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
7125 // CHECK9-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
7126 // CHECK9-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
7127 // CHECK9-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
7128 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
7129 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
7130 // CHECK9-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
7131 // CHECK9-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
7132 // CHECK9-NEXT:    ret void
7133 //
7134 //
7135 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
7136 // CHECK9-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
7137 // CHECK9-NEXT:  entry:
7138 // CHECK9-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
7139 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
7140 // CHECK9-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7141 // CHECK9-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
7142 // CHECK9-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
7143 // CHECK9-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
7144 // CHECK9-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
7145 // CHECK9-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
7146 // CHECK9-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7147 // CHECK9-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
7148 // CHECK9-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
7149 // CHECK9-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
7150 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
7151 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7152 // CHECK9-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
7153 // CHECK9-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
7154 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
7155 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
7156 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
7157 // CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
7158 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
7159 // CHECK9-NEXT:    ret void
7160 //
7161 //
7162 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..10
7163 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
7164 // CHECK9-NEXT:  entry:
7165 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7166 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7167 // CHECK9-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
7168 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
7169 // CHECK9-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7170 // CHECK9-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
7171 // CHECK9-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
7172 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7173 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7174 // CHECK9-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
7175 // CHECK9-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
7176 // CHECK9-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7177 // CHECK9-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
7178 // CHECK9-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
7179 // CHECK9-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
7180 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
7181 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7182 // CHECK9-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
7183 // CHECK9-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
7184 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
7185 // CHECK9-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
7186 // CHECK9-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
7187 // CHECK9-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
7188 // CHECK9-NEXT:    store double [[ADD]], double* [[A]], align 8
7189 // CHECK9-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
7190 // CHECK9-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
7191 // CHECK9-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
7192 // CHECK9-NEXT:    store double [[INC]], double* [[A4]], align 8
7193 // CHECK9-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
7194 // CHECK9-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
7195 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
7196 // CHECK9-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
7197 // CHECK9-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
7198 // CHECK9-NEXT:    ret void
7199 //
7200 //
7201 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
7202 // CHECK9-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7203 // CHECK9-NEXT:  entry:
7204 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7205 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7206 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7207 // CHECK9-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
7208 // CHECK9-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7209 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7210 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7211 // CHECK9-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7212 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7213 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7214 // CHECK9-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7215 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7216 // CHECK9-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
7217 // CHECK9-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
7218 // CHECK9-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
7219 // CHECK9-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
7220 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7221 // CHECK9-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
7222 // CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7223 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
7224 // CHECK9-NEXT:    ret void
7225 //
7226 //
7227 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..11
7228 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7229 // CHECK9-NEXT:  entry:
7230 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7231 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7232 // CHECK9-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7233 // CHECK9-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7234 // CHECK9-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7235 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7236 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7237 // CHECK9-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7238 // CHECK9-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7239 // CHECK9-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7240 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7241 // CHECK9-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7242 // CHECK9-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7243 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7244 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
7245 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
7246 // CHECK9-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
7247 // CHECK9-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
7248 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
7249 // CHECK9-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
7250 // CHECK9-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
7251 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
7252 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
7253 // CHECK9-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
7254 // CHECK9-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
7255 // CHECK9-NEXT:    ret void
7256 //
7257 //
7258 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
7259 // CHECK10-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
7260 // CHECK10-NEXT:  entry:
7261 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7262 // CHECK10-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
7263 // CHECK10-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
7264 // CHECK10-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7265 // CHECK10-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
7266 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7267 // CHECK10-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
7268 // CHECK10-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
7269 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7270 // CHECK10-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
7271 // CHECK10-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
7272 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
7273 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
7274 // CHECK10-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
7275 // CHECK10-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
7276 // CHECK10-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7277 // CHECK10-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
7278 // CHECK10-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7279 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
7280 // CHECK10-NEXT:    ret void
7281 //
7282 //
7283 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined.
7284 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
7285 // CHECK10-NEXT:  entry:
7286 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7287 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7288 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7289 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7290 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7291 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7292 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7293 // CHECK10-NEXT:    ret void
7294 //
7295 //
7296 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
7297 // CHECK10-SAME: (i64 [[AA:%.*]]) #[[ATTR0]] {
7298 // CHECK10-NEXT:  entry:
7299 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7300 // CHECK10-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7301 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7302 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7303 // CHECK10-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
7304 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7305 // CHECK10-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
7306 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7307 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP1]])
7308 // CHECK10-NEXT:    ret void
7309 //
7310 //
7311 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..1
7312 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
7313 // CHECK10-NEXT:  entry:
7314 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7315 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7316 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7317 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7318 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7319 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7320 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7321 // CHECK10-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
7322 // CHECK10-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
7323 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
7324 // CHECK10-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
7325 // CHECK10-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
7326 // CHECK10-NEXT:    ret void
7327 //
7328 //
7329 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
7330 // CHECK10-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
7331 // CHECK10-NEXT:  entry:
7332 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7333 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7334 // CHECK10-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
7335 // CHECK10-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7336 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7337 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7338 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7339 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7340 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
7341 // CHECK10-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
7342 // CHECK10-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
7343 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
7344 // CHECK10-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
7345 // CHECK10-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7346 // CHECK10-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
7347 // CHECK10-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7348 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
7349 // CHECK10-NEXT:    ret void
7350 //
7351 //
7352 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..2
7353 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
7354 // CHECK10-NEXT:  entry:
7355 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7356 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7357 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7358 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7359 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7360 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7361 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7362 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7363 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7364 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7365 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
7366 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
7367 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
7368 // CHECK10-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
7369 // CHECK10-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
7370 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
7371 // CHECK10-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
7372 // CHECK10-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
7373 // CHECK10-NEXT:    ret void
7374 //
7375 //
7376 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
7377 // CHECK10-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
7378 // CHECK10-NEXT:  entry:
7379 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7380 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
7381 // CHECK10-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7382 // CHECK10-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
7383 // CHECK10-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
7384 // CHECK10-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
7385 // CHECK10-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
7386 // CHECK10-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
7387 // CHECK10-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
7388 // CHECK10-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
7389 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7390 // CHECK10-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
7391 // CHECK10-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7392 // CHECK10-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
7393 // CHECK10-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
7394 // CHECK10-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
7395 // CHECK10-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
7396 // CHECK10-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
7397 // CHECK10-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
7398 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7399 // CHECK10-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
7400 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7401 // CHECK10-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
7402 // CHECK10-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
7403 // CHECK10-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
7404 // CHECK10-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
7405 // CHECK10-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
7406 // CHECK10-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
7407 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
7408 // CHECK10-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
7409 // CHECK10-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
7410 // CHECK10-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
7411 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
7412 // CHECK10-NEXT:    ret void
7413 //
7414 //
7415 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..3
7416 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
7417 // CHECK10-NEXT:  entry:
7418 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7419 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7420 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7421 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
7422 // CHECK10-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7423 // CHECK10-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
7424 // CHECK10-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
7425 // CHECK10-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
7426 // CHECK10-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
7427 // CHECK10-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
7428 // CHECK10-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
7429 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7430 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7431 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7432 // CHECK10-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
7433 // CHECK10-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7434 // CHECK10-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
7435 // CHECK10-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
7436 // CHECK10-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
7437 // CHECK10-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
7438 // CHECK10-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
7439 // CHECK10-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
7440 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7441 // CHECK10-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
7442 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7443 // CHECK10-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
7444 // CHECK10-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
7445 // CHECK10-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
7446 // CHECK10-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
7447 // CHECK10-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
7448 // CHECK10-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
7449 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
7450 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
7451 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
7452 // CHECK10-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
7453 // CHECK10-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
7454 // CHECK10-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
7455 // CHECK10-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
7456 // CHECK10-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
7457 // CHECK10-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
7458 // CHECK10-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
7459 // CHECK10-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
7460 // CHECK10-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
7461 // CHECK10-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
7462 // CHECK10-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
7463 // CHECK10-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
7464 // CHECK10-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
7465 // CHECK10-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
7466 // CHECK10-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
7467 // CHECK10-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
7468 // CHECK10-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
7469 // CHECK10-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
7470 // CHECK10-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
7471 // CHECK10-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
7472 // CHECK10-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
7473 // CHECK10-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
7474 // CHECK10-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
7475 // CHECK10-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
7476 // CHECK10-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
7477 // CHECK10-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
7478 // CHECK10-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
7479 // CHECK10-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
7480 // CHECK10-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
7481 // CHECK10-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
7482 // CHECK10-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
7483 // CHECK10-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
7484 // CHECK10-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
7485 // CHECK10-NEXT:    ret void
7486 //
7487 //
7488 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
7489 // CHECK10-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
7490 // CHECK10-NEXT:  entry:
7491 // CHECK10-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7492 // CHECK10-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
7493 // CHECK10-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7494 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7495 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
7496 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
7497 // CHECK10-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
7498 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
7499 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i64 [[TMP1]])
7500 // CHECK10-NEXT:    ret void
7501 //
7502 //
7503 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..4
7504 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
7505 // CHECK10-NEXT:  entry:
7506 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7507 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7508 // CHECK10-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7509 // CHECK10-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
7510 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7511 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7512 // CHECK10-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7513 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7514 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
7515 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
7516 // CHECK10-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
7517 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
7518 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP1]])
7519 // CHECK10-NEXT:    ret void
7520 //
7521 //
7522 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..5
7523 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
7524 // CHECK10-NEXT:  entry:
7525 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7526 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7527 // CHECK10-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7528 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7529 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7530 // CHECK10-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7531 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7532 // CHECK10-NEXT:    ret void
7533 //
7534 //
7535 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
7536 // CHECK10-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
7537 // CHECK10-NEXT:  entry:
7538 // CHECK10-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7539 // CHECK10-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
7540 // CHECK10-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7541 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7542 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
7543 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
7544 // CHECK10-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
7545 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
7546 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]])
7547 // CHECK10-NEXT:    ret void
7548 //
7549 //
7550 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..6
7551 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
7552 // CHECK10-NEXT:  entry:
7553 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7554 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7555 // CHECK10-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
7556 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7557 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7558 // CHECK10-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
7559 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
7560 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[CONV]])
7561 // CHECK10-NEXT:    ret void
7562 //
7563 //
7564 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..7
7565 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
7566 // CHECK10-NEXT:  entry:
7567 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7568 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7569 // CHECK10-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
7570 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7571 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7572 // CHECK10-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
7573 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
7574 // CHECK10-NEXT:    ret void
7575 //
7576 //
7577 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
7578 // CHECK10-SAME: (i64 [[VLA:%.*]]) #[[ATTR0]] {
7579 // CHECK10-NEXT:  entry:
7580 // CHECK10-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7581 // CHECK10-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7582 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7583 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i64 [[TMP0]])
7584 // CHECK10-NEXT:    ret void
7585 //
7586 //
7587 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..8
7588 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR0]] {
7589 // CHECK10-NEXT:  entry:
7590 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7591 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7592 // CHECK10-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7593 // CHECK10-NEXT:    [[F:%.*]] = alloca i32*, align 8
7594 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7595 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7596 // CHECK10-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7597 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7598 // CHECK10-NEXT:    ret void
7599 //
7600 //
7601 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
7602 // CHECK10-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7603 // CHECK10-NEXT:  entry:
7604 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7605 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7606 // CHECK10-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
7607 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7608 // CHECK10-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
7609 // CHECK10-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7610 // CHECK10-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
7611 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7612 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7613 // CHECK10-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
7614 // CHECK10-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7615 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7616 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7617 // CHECK10-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
7618 // CHECK10-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7619 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7620 // CHECK10-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
7621 // CHECK10-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
7622 // CHECK10-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
7623 // CHECK10-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
7624 // CHECK10-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7625 // CHECK10-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
7626 // CHECK10-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7627 // CHECK10-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
7628 // CHECK10-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
7629 // CHECK10-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
7630 // CHECK10-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
7631 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
7632 // CHECK10-NEXT:    ret void
7633 //
7634 //
7635 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..9
7636 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7637 // CHECK10-NEXT:  entry:
7638 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7639 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7640 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7641 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7642 // CHECK10-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
7643 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7644 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7645 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7646 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7647 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7648 // CHECK10-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
7649 // CHECK10-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7650 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7651 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7652 // CHECK10-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
7653 // CHECK10-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7654 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7655 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
7656 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
7657 // CHECK10-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
7658 // CHECK10-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
7659 // CHECK10-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
7660 // CHECK10-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
7661 // CHECK10-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
7662 // CHECK10-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
7663 // CHECK10-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
7664 // CHECK10-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
7665 // CHECK10-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
7666 // CHECK10-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
7667 // CHECK10-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
7668 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
7669 // CHECK10-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
7670 // CHECK10-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
7671 // CHECK10-NEXT:    ret void
7672 //
7673 //
7674 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
7675 // CHECK10-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
7676 // CHECK10-NEXT:  entry:
7677 // CHECK10-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
7678 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
7679 // CHECK10-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7680 // CHECK10-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
7681 // CHECK10-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
7682 // CHECK10-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
7683 // CHECK10-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
7684 // CHECK10-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
7685 // CHECK10-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7686 // CHECK10-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
7687 // CHECK10-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
7688 // CHECK10-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
7689 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
7690 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7691 // CHECK10-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
7692 // CHECK10-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
7693 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
7694 // CHECK10-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
7695 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
7696 // CHECK10-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
7697 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
7698 // CHECK10-NEXT:    ret void
7699 //
7700 //
7701 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..10
7702 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
7703 // CHECK10-NEXT:  entry:
7704 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7705 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7706 // CHECK10-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
7707 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
7708 // CHECK10-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
7709 // CHECK10-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
7710 // CHECK10-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
7711 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7712 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7713 // CHECK10-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
7714 // CHECK10-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
7715 // CHECK10-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
7716 // CHECK10-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
7717 // CHECK10-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
7718 // CHECK10-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
7719 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
7720 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
7721 // CHECK10-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
7722 // CHECK10-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
7723 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
7724 // CHECK10-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
7725 // CHECK10-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
7726 // CHECK10-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
7727 // CHECK10-NEXT:    store double [[ADD]], double* [[A]], align 8
7728 // CHECK10-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
7729 // CHECK10-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
7730 // CHECK10-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
7731 // CHECK10-NEXT:    store double [[INC]], double* [[A4]], align 8
7732 // CHECK10-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
7733 // CHECK10-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
7734 // CHECK10-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
7735 // CHECK10-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
7736 // CHECK10-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
7737 // CHECK10-NEXT:    ret void
7738 //
7739 //
7740 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
7741 // CHECK10-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7742 // CHECK10-NEXT:  entry:
7743 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7744 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7745 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7746 // CHECK10-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
7747 // CHECK10-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
7748 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7749 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7750 // CHECK10-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7751 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7752 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7753 // CHECK10-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7754 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7755 // CHECK10-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
7756 // CHECK10-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
7757 // CHECK10-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
7758 // CHECK10-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
7759 // CHECK10-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
7760 // CHECK10-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
7761 // CHECK10-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
7762 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
7763 // CHECK10-NEXT:    ret void
7764 //
7765 //
7766 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..11
7767 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
7768 // CHECK10-NEXT:  entry:
7769 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7770 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7771 // CHECK10-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
7772 // CHECK10-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
7773 // CHECK10-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
7774 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7775 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7776 // CHECK10-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
7777 // CHECK10-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
7778 // CHECK10-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
7779 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
7780 // CHECK10-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
7781 // CHECK10-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
7782 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
7783 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
7784 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
7785 // CHECK10-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
7786 // CHECK10-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
7787 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
7788 // CHECK10-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
7789 // CHECK10-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
7790 // CHECK10-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
7791 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
7792 // CHECK10-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
7793 // CHECK10-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
7794 // CHECK10-NEXT:    ret void
7795 //
7796 //
7797 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
7798 // CHECK11-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
7799 // CHECK11-NEXT:  entry:
7800 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
7801 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
7802 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
7803 // CHECK11-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
7804 // CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
7805 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
7806 // CHECK11-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
7807 // CHECK11-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
7808 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
7809 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
7810 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
7811 // CHECK11-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
7812 // CHECK11-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
7813 // CHECK11-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
7814 // CHECK11-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
7815 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
7816 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
7817 // CHECK11-NEXT:    ret void
7818 //
7819 //
7820 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined.
7821 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
7822 // CHECK11-NEXT:  entry:
7823 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
7824 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
7825 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
7826 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
7827 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
7828 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
7829 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
7830 // CHECK11-NEXT:    ret void
7831 //
7832 //
7833 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
7834 // CHECK11-SAME: (i32 [[AA:%.*]]) #[[ATTR0]] {
7835 // CHECK11-NEXT:  entry:
7836 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
7837 // CHECK11-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
7838 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
7839 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
7840 // CHECK11-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
7841 // CHECK11-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
7842 // CHECK11-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
7843 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
7844 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i32 [[TMP1]])
7845 // CHECK11-NEXT:    ret void
7846 //
7847 //
7848 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..1
7849 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
7850 // CHECK11-NEXT:  entry:
7851 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
7852 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
7853 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
7854 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
7855 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
7856 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
7857 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
7858 // CHECK11-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
7859 // CHECK11-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
7860 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
7861 // CHECK11-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
7862 // CHECK11-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
7863 // CHECK11-NEXT:    ret void
7864 //
7865 //
7866 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
7867 // CHECK11-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
7868 // CHECK11-NEXT:  entry:
7869 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
7870 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
7871 // CHECK11-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
7872 // CHECK11-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
7873 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
7874 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
7875 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
7876 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
7877 // CHECK11-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
7878 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
7879 // CHECK11-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
7880 // CHECK11-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
7881 // CHECK11-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
7882 // CHECK11-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
7883 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
7884 // CHECK11-NEXT:    ret void
7885 //
7886 //
7887 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..2
7888 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
7889 // CHECK11-NEXT:  entry:
7890 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
7891 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
7892 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
7893 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
7894 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
7895 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
7896 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
7897 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
7898 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
7899 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
7900 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
7901 // CHECK11-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
7902 // CHECK11-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
7903 // CHECK11-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
7904 // CHECK11-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
7905 // CHECK11-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
7906 // CHECK11-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
7907 // CHECK11-NEXT:    ret void
7908 //
7909 //
7910 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
7911 // CHECK11-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
7912 // CHECK11-NEXT:  entry:
7913 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
7914 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
7915 // CHECK11-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
7916 // CHECK11-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
7917 // CHECK11-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
7918 // CHECK11-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
7919 // CHECK11-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
7920 // CHECK11-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
7921 // CHECK11-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
7922 // CHECK11-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
7923 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
7924 // CHECK11-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
7925 // CHECK11-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
7926 // CHECK11-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
7927 // CHECK11-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
7928 // CHECK11-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
7929 // CHECK11-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
7930 // CHECK11-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
7931 // CHECK11-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
7932 // CHECK11-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
7933 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
7934 // CHECK11-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
7935 // CHECK11-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
7936 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
7937 // CHECK11-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
7938 // CHECK11-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
7939 // CHECK11-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
7940 // CHECK11-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
7941 // CHECK11-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
7942 // CHECK11-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
7943 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
7944 // CHECK11-NEXT:    ret void
7945 //
7946 //
7947 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..3
7948 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
7949 // CHECK11-NEXT:  entry:
7950 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
7951 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
7952 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
7953 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
7954 // CHECK11-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
7955 // CHECK11-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
7956 // CHECK11-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
7957 // CHECK11-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
7958 // CHECK11-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
7959 // CHECK11-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
7960 // CHECK11-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
7961 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
7962 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
7963 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
7964 // CHECK11-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
7965 // CHECK11-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
7966 // CHECK11-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
7967 // CHECK11-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
7968 // CHECK11-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
7969 // CHECK11-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
7970 // CHECK11-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
7971 // CHECK11-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
7972 // CHECK11-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
7973 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
7974 // CHECK11-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
7975 // CHECK11-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
7976 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
7977 // CHECK11-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
7978 // CHECK11-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
7979 // CHECK11-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
7980 // CHECK11-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
7981 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
7982 // CHECK11-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
7983 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
7984 // CHECK11-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
7985 // CHECK11-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
7986 // CHECK11-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
7987 // CHECK11-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
7988 // CHECK11-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
7989 // CHECK11-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
7990 // CHECK11-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
7991 // CHECK11-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
7992 // CHECK11-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
7993 // CHECK11-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
7994 // CHECK11-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
7995 // CHECK11-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
7996 // CHECK11-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
7997 // CHECK11-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
7998 // CHECK11-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
7999 // CHECK11-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
8000 // CHECK11-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
8001 // CHECK11-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
8002 // CHECK11-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
8003 // CHECK11-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
8004 // CHECK11-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
8005 // CHECK11-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
8006 // CHECK11-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
8007 // CHECK11-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
8008 // CHECK11-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
8009 // CHECK11-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
8010 // CHECK11-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
8011 // CHECK11-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
8012 // CHECK11-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
8013 // CHECK11-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
8014 // CHECK11-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
8015 // CHECK11-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
8016 // CHECK11-NEXT:    ret void
8017 //
8018 //
8019 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
8020 // CHECK11-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
8021 // CHECK11-NEXT:  entry:
8022 // CHECK11-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8023 // CHECK11-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
8024 // CHECK11-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8025 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
8026 // CHECK11-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
8027 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
8028 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8029 // CHECK11-NEXT:    ret void
8030 //
8031 //
8032 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..4
8033 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
8034 // CHECK11-NEXT:  entry:
8035 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8036 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8037 // CHECK11-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8038 // CHECK11-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
8039 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8040 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8041 // CHECK11-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8042 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
8043 // CHECK11-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
8044 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
8045 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8046 // CHECK11-NEXT:    ret void
8047 //
8048 //
8049 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..5
8050 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
8051 // CHECK11-NEXT:  entry:
8052 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8053 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8054 // CHECK11-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8055 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8056 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8057 // CHECK11-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8058 // CHECK11-NEXT:    ret void
8059 //
8060 //
8061 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
8062 // CHECK11-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
8063 // CHECK11-NEXT:  entry:
8064 // CHECK11-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8065 // CHECK11-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
8066 // CHECK11-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8067 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
8068 // CHECK11-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
8069 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
8070 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8071 // CHECK11-NEXT:    ret void
8072 //
8073 //
8074 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..6
8075 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
8076 // CHECK11-NEXT:  entry:
8077 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8078 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8079 // CHECK11-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8080 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8081 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8082 // CHECK11-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8083 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
8084 // CHECK11-NEXT:    ret void
8085 //
8086 //
8087 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..7
8088 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
8089 // CHECK11-NEXT:  entry:
8090 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8091 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8092 // CHECK11-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
8093 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8094 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8095 // CHECK11-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
8096 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
8097 // CHECK11-NEXT:    ret void
8098 //
8099 //
8100 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
8101 // CHECK11-SAME: (i32 [[VLA:%.*]]) #[[ATTR0]] {
8102 // CHECK11-NEXT:  entry:
8103 // CHECK11-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8104 // CHECK11-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8105 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8106 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32 [[TMP0]])
8107 // CHECK11-NEXT:    ret void
8108 //
8109 //
8110 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..8
8111 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR0]] {
8112 // CHECK11-NEXT:  entry:
8113 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8114 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8115 // CHECK11-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8116 // CHECK11-NEXT:    [[F:%.*]] = alloca i32*, align 4
8117 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8118 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8119 // CHECK11-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8120 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8121 // CHECK11-NEXT:    ret void
8122 //
8123 //
8124 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
8125 // CHECK11-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8126 // CHECK11-NEXT:  entry:
8127 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8128 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8129 // CHECK11-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
8130 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8131 // CHECK11-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
8132 // CHECK11-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8133 // CHECK11-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
8134 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8135 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8136 // CHECK11-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
8137 // CHECK11-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8138 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8139 // CHECK11-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
8140 // CHECK11-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8141 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8142 // CHECK11-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
8143 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
8144 // CHECK11-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
8145 // CHECK11-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8146 // CHECK11-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
8147 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8148 // CHECK11-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
8149 // CHECK11-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
8150 // CHECK11-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
8151 // CHECK11-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
8152 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
8153 // CHECK11-NEXT:    ret void
8154 //
8155 //
8156 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..9
8157 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8158 // CHECK11-NEXT:  entry:
8159 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8160 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8161 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8162 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8163 // CHECK11-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
8164 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8165 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8166 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8167 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8168 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8169 // CHECK11-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
8170 // CHECK11-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8171 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8172 // CHECK11-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
8173 // CHECK11-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8174 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8175 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
8176 // CHECK11-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
8177 // CHECK11-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
8178 // CHECK11-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
8179 // CHECK11-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
8180 // CHECK11-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
8181 // CHECK11-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
8182 // CHECK11-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
8183 // CHECK11-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
8184 // CHECK11-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
8185 // CHECK11-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
8186 // CHECK11-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
8187 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
8188 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
8189 // CHECK11-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
8190 // CHECK11-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
8191 // CHECK11-NEXT:    ret void
8192 //
8193 //
8194 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
8195 // CHECK11-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
8196 // CHECK11-NEXT:  entry:
8197 // CHECK11-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
8198 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
8199 // CHECK11-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8200 // CHECK11-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
8201 // CHECK11-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
8202 // CHECK11-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
8203 // CHECK11-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
8204 // CHECK11-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
8205 // CHECK11-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8206 // CHECK11-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
8207 // CHECK11-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
8208 // CHECK11-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
8209 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8210 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
8211 // CHECK11-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
8212 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
8213 // CHECK11-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
8214 // CHECK11-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
8215 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
8216 // CHECK11-NEXT:    ret void
8217 //
8218 //
8219 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..10
8220 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
8221 // CHECK11-NEXT:  entry:
8222 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8223 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8224 // CHECK11-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
8225 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
8226 // CHECK11-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8227 // CHECK11-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
8228 // CHECK11-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
8229 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8230 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8231 // CHECK11-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
8232 // CHECK11-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
8233 // CHECK11-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8234 // CHECK11-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
8235 // CHECK11-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
8236 // CHECK11-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
8237 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8238 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
8239 // CHECK11-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
8240 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
8241 // CHECK11-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
8242 // CHECK11-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
8243 // CHECK11-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
8244 // CHECK11-NEXT:    store double [[ADD]], double* [[A]], align 4
8245 // CHECK11-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
8246 // CHECK11-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
8247 // CHECK11-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
8248 // CHECK11-NEXT:    store double [[INC]], double* [[A3]], align 4
8249 // CHECK11-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
8250 // CHECK11-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
8251 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
8252 // CHECK11-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
8253 // CHECK11-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
8254 // CHECK11-NEXT:    ret void
8255 //
8256 //
8257 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
8258 // CHECK11-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8259 // CHECK11-NEXT:  entry:
8260 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8261 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8262 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8263 // CHECK11-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
8264 // CHECK11-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8265 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8266 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8267 // CHECK11-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8268 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8269 // CHECK11-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8270 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8271 // CHECK11-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
8272 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
8273 // CHECK11-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
8274 // CHECK11-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8275 // CHECK11-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
8276 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8277 // CHECK11-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
8278 // CHECK11-NEXT:    ret void
8279 //
8280 //
8281 // CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..11
8282 // CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8283 // CHECK11-NEXT:  entry:
8284 // CHECK11-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8285 // CHECK11-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8286 // CHECK11-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8287 // CHECK11-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8288 // CHECK11-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8289 // CHECK11-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8290 // CHECK11-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8291 // CHECK11-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8292 // CHECK11-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8293 // CHECK11-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8294 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8295 // CHECK11-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8296 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8297 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
8298 // CHECK11-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
8299 // CHECK11-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
8300 // CHECK11-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
8301 // CHECK11-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
8302 // CHECK11-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
8303 // CHECK11-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
8304 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
8305 // CHECK11-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
8306 // CHECK11-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
8307 // CHECK11-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
8308 // CHECK11-NEXT:    ret void
8309 //
8310 //
8311 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
8312 // CHECK12-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
8313 // CHECK12-NEXT:  entry:
8314 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8315 // CHECK12-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
8316 // CHECK12-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
8317 // CHECK12-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8318 // CHECK12-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
8319 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8320 // CHECK12-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
8321 // CHECK12-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
8322 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8323 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
8324 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
8325 // CHECK12-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
8326 // CHECK12-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
8327 // CHECK12-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8328 // CHECK12-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
8329 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8330 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
8331 // CHECK12-NEXT:    ret void
8332 //
8333 //
8334 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined.
8335 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
8336 // CHECK12-NEXT:  entry:
8337 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8338 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8339 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8340 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8341 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8342 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8343 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8344 // CHECK12-NEXT:    ret void
8345 //
8346 //
8347 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
8348 // CHECK12-SAME: (i32 [[AA:%.*]]) #[[ATTR0]] {
8349 // CHECK12-NEXT:  entry:
8350 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8351 // CHECK12-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8352 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8353 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8354 // CHECK12-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
8355 // CHECK12-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8356 // CHECK12-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
8357 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8358 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8359 // CHECK12-NEXT:    ret void
8360 //
8361 //
8362 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..1
8363 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
8364 // CHECK12-NEXT:  entry:
8365 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8366 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8367 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8368 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8369 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8370 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8371 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8372 // CHECK12-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
8373 // CHECK12-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
8374 // CHECK12-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
8375 // CHECK12-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
8376 // CHECK12-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
8377 // CHECK12-NEXT:    ret void
8378 //
8379 //
8380 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
8381 // CHECK12-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
8382 // CHECK12-NEXT:  entry:
8383 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8384 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8385 // CHECK12-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
8386 // CHECK12-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8387 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8388 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8389 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8390 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
8391 // CHECK12-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
8392 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
8393 // CHECK12-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
8394 // CHECK12-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8395 // CHECK12-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
8396 // CHECK12-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8397 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
8398 // CHECK12-NEXT:    ret void
8399 //
8400 //
8401 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..2
8402 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
8403 // CHECK12-NEXT:  entry:
8404 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8405 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8406 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8407 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8408 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8409 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8410 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8411 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8412 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8413 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
8414 // CHECK12-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
8415 // CHECK12-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
8416 // CHECK12-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
8417 // CHECK12-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
8418 // CHECK12-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
8419 // CHECK12-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
8420 // CHECK12-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
8421 // CHECK12-NEXT:    ret void
8422 //
8423 //
8424 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
8425 // CHECK12-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
8426 // CHECK12-NEXT:  entry:
8427 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8428 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
8429 // CHECK12-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8430 // CHECK12-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
8431 // CHECK12-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
8432 // CHECK12-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
8433 // CHECK12-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
8434 // CHECK12-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
8435 // CHECK12-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
8436 // CHECK12-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
8437 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8438 // CHECK12-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
8439 // CHECK12-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8440 // CHECK12-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
8441 // CHECK12-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
8442 // CHECK12-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
8443 // CHECK12-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
8444 // CHECK12-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
8445 // CHECK12-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
8446 // CHECK12-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
8447 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8448 // CHECK12-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
8449 // CHECK12-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
8450 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
8451 // CHECK12-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
8452 // CHECK12-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
8453 // CHECK12-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
8454 // CHECK12-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
8455 // CHECK12-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
8456 // CHECK12-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
8457 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
8458 // CHECK12-NEXT:    ret void
8459 //
8460 //
8461 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..3
8462 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
8463 // CHECK12-NEXT:  entry:
8464 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8465 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8466 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8467 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
8468 // CHECK12-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8469 // CHECK12-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
8470 // CHECK12-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
8471 // CHECK12-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
8472 // CHECK12-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
8473 // CHECK12-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
8474 // CHECK12-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
8475 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8476 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8477 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8478 // CHECK12-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
8479 // CHECK12-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8480 // CHECK12-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
8481 // CHECK12-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
8482 // CHECK12-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
8483 // CHECK12-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
8484 // CHECK12-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
8485 // CHECK12-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
8486 // CHECK12-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
8487 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8488 // CHECK12-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
8489 // CHECK12-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
8490 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
8491 // CHECK12-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
8492 // CHECK12-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
8493 // CHECK12-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
8494 // CHECK12-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
8495 // CHECK12-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
8496 // CHECK12-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
8497 // CHECK12-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
8498 // CHECK12-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
8499 // CHECK12-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
8500 // CHECK12-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
8501 // CHECK12-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
8502 // CHECK12-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
8503 // CHECK12-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
8504 // CHECK12-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
8505 // CHECK12-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
8506 // CHECK12-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
8507 // CHECK12-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
8508 // CHECK12-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
8509 // CHECK12-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
8510 // CHECK12-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
8511 // CHECK12-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
8512 // CHECK12-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
8513 // CHECK12-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
8514 // CHECK12-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
8515 // CHECK12-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
8516 // CHECK12-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
8517 // CHECK12-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
8518 // CHECK12-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
8519 // CHECK12-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
8520 // CHECK12-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
8521 // CHECK12-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
8522 // CHECK12-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
8523 // CHECK12-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
8524 // CHECK12-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
8525 // CHECK12-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
8526 // CHECK12-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
8527 // CHECK12-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
8528 // CHECK12-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
8529 // CHECK12-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
8530 // CHECK12-NEXT:    ret void
8531 //
8532 //
8533 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
8534 // CHECK12-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
8535 // CHECK12-NEXT:  entry:
8536 // CHECK12-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8537 // CHECK12-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
8538 // CHECK12-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8539 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
8540 // CHECK12-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
8541 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
8542 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8543 // CHECK12-NEXT:    ret void
8544 //
8545 //
8546 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..4
8547 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
8548 // CHECK12-NEXT:  entry:
8549 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8550 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8551 // CHECK12-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8552 // CHECK12-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
8553 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8554 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8555 // CHECK12-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8556 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
8557 // CHECK12-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
8558 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
8559 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8560 // CHECK12-NEXT:    ret void
8561 //
8562 //
8563 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..5
8564 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
8565 // CHECK12-NEXT:  entry:
8566 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8567 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8568 // CHECK12-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8569 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8570 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8571 // CHECK12-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8572 // CHECK12-NEXT:    ret void
8573 //
8574 //
8575 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
8576 // CHECK12-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
8577 // CHECK12-NEXT:  entry:
8578 // CHECK12-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8579 // CHECK12-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
8580 // CHECK12-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8581 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
8582 // CHECK12-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
8583 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
8584 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]])
8585 // CHECK12-NEXT:    ret void
8586 //
8587 //
8588 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..6
8589 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
8590 // CHECK12-NEXT:  entry:
8591 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8592 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8593 // CHECK12-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
8594 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8595 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8596 // CHECK12-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
8597 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
8598 // CHECK12-NEXT:    ret void
8599 //
8600 //
8601 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..7
8602 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
8603 // CHECK12-NEXT:  entry:
8604 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8605 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8606 // CHECK12-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
8607 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8608 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8609 // CHECK12-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
8610 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
8611 // CHECK12-NEXT:    ret void
8612 //
8613 //
8614 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
8615 // CHECK12-SAME: (i32 [[VLA:%.*]]) #[[ATTR0]] {
8616 // CHECK12-NEXT:  entry:
8617 // CHECK12-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8618 // CHECK12-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8619 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8620 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32 [[TMP0]])
8621 // CHECK12-NEXT:    ret void
8622 //
8623 //
8624 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..8
8625 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR0]] {
8626 // CHECK12-NEXT:  entry:
8627 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8628 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8629 // CHECK12-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8630 // CHECK12-NEXT:    [[F:%.*]] = alloca i32*, align 4
8631 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8632 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8633 // CHECK12-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8634 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8635 // CHECK12-NEXT:    ret void
8636 //
8637 //
8638 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
8639 // CHECK12-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8640 // CHECK12-NEXT:  entry:
8641 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8642 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8643 // CHECK12-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
8644 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8645 // CHECK12-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
8646 // CHECK12-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8647 // CHECK12-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
8648 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8649 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8650 // CHECK12-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
8651 // CHECK12-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8652 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8653 // CHECK12-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
8654 // CHECK12-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8655 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8656 // CHECK12-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
8657 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
8658 // CHECK12-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
8659 // CHECK12-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8660 // CHECK12-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
8661 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8662 // CHECK12-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
8663 // CHECK12-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
8664 // CHECK12-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
8665 // CHECK12-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
8666 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
8667 // CHECK12-NEXT:    ret void
8668 //
8669 //
8670 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..9
8671 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8672 // CHECK12-NEXT:  entry:
8673 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8674 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8675 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8676 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8677 // CHECK12-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
8678 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8679 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8680 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8681 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8682 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8683 // CHECK12-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
8684 // CHECK12-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8685 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8686 // CHECK12-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
8687 // CHECK12-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8688 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8689 // CHECK12-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
8690 // CHECK12-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
8691 // CHECK12-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
8692 // CHECK12-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
8693 // CHECK12-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
8694 // CHECK12-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
8695 // CHECK12-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
8696 // CHECK12-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
8697 // CHECK12-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
8698 // CHECK12-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
8699 // CHECK12-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
8700 // CHECK12-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
8701 // CHECK12-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
8702 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
8703 // CHECK12-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
8704 // CHECK12-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
8705 // CHECK12-NEXT:    ret void
8706 //
8707 //
8708 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
8709 // CHECK12-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
8710 // CHECK12-NEXT:  entry:
8711 // CHECK12-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
8712 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
8713 // CHECK12-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8714 // CHECK12-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
8715 // CHECK12-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
8716 // CHECK12-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
8717 // CHECK12-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
8718 // CHECK12-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
8719 // CHECK12-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8720 // CHECK12-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
8721 // CHECK12-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
8722 // CHECK12-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
8723 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8724 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
8725 // CHECK12-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
8726 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
8727 // CHECK12-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
8728 // CHECK12-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
8729 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
8730 // CHECK12-NEXT:    ret void
8731 //
8732 //
8733 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..10
8734 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
8735 // CHECK12-NEXT:  entry:
8736 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8737 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8738 // CHECK12-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
8739 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
8740 // CHECK12-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
8741 // CHECK12-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
8742 // CHECK12-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
8743 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8744 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8745 // CHECK12-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
8746 // CHECK12-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
8747 // CHECK12-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
8748 // CHECK12-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
8749 // CHECK12-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
8750 // CHECK12-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
8751 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
8752 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
8753 // CHECK12-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
8754 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
8755 // CHECK12-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
8756 // CHECK12-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
8757 // CHECK12-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
8758 // CHECK12-NEXT:    store double [[ADD]], double* [[A]], align 4
8759 // CHECK12-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
8760 // CHECK12-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
8761 // CHECK12-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
8762 // CHECK12-NEXT:    store double [[INC]], double* [[A3]], align 4
8763 // CHECK12-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
8764 // CHECK12-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
8765 // CHECK12-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
8766 // CHECK12-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
8767 // CHECK12-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
8768 // CHECK12-NEXT:    ret void
8769 //
8770 //
8771 // CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
8772 // CHECK12-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8773 // CHECK12-NEXT:  entry:
8774 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8775 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8776 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8777 // CHECK12-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
8778 // CHECK12-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
8779 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8780 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8781 // CHECK12-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8782 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8783 // CHECK12-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8784 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8785 // CHECK12-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
8786 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
8787 // CHECK12-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
8788 // CHECK12-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
8789 // CHECK12-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
8790 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
8791 // CHECK12-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
8792 // CHECK12-NEXT:    ret void
8793 //
8794 //
8795 // CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..11
8796 // CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
8797 // CHECK12-NEXT:  entry:
8798 // CHECK12-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
8799 // CHECK12-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
8800 // CHECK12-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
8801 // CHECK12-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
8802 // CHECK12-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
8803 // CHECK12-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
8804 // CHECK12-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
8805 // CHECK12-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
8806 // CHECK12-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
8807 // CHECK12-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
8808 // CHECK12-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
8809 // CHECK12-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
8810 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
8811 // CHECK12-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
8812 // CHECK12-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
8813 // CHECK12-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
8814 // CHECK12-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
8815 // CHECK12-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
8816 // CHECK12-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
8817 // CHECK12-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
8818 // CHECK12-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
8819 // CHECK12-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
8820 // CHECK12-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
8821 // CHECK12-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
8822 // CHECK12-NEXT:    ret void
8823 //
8824 //
8825 // CHECK13-LABEL: define {{[^@]+}}@_Z3fooi
8826 // CHECK13-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
8827 // CHECK13-NEXT:  entry:
8828 // CHECK13-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
8829 // CHECK13-NEXT:    [[A:%.*]] = alloca i32, align 4
8830 // CHECK13-NEXT:    [[AA:%.*]] = alloca i16, align 2
8831 // CHECK13-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
8832 // CHECK13-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
8833 // CHECK13-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
8834 // CHECK13-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
8835 // CHECK13-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
8836 // CHECK13-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
8837 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
8838 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
8839 // CHECK13-NEXT:    [[NN:%.*]] = alloca i32, align 4
8840 // CHECK13-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
8841 // CHECK13-NEXT:    store i32 0, i32* [[A]], align 4
8842 // CHECK13-NEXT:    store i16 0, i16* [[AA]], align 2
8843 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
8844 // CHECK13-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
8845 // CHECK13-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
8846 // CHECK13-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
8847 // CHECK13-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
8848 // CHECK13-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
8849 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
8850 // CHECK13-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
8851 // CHECK13-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
8852 // CHECK13-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
8853 // CHECK13-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
8854 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
8855 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
8856 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
8857 // CHECK13-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
8858 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
8859 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
8860 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
8861 // CHECK13-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
8862 // CHECK13-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
8863 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
8864 // CHECK13-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
8865 // CHECK13-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
8866 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
8867 // CHECK13-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
8868 // CHECK13-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
8869 // CHECK13-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
8870 // CHECK13-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
8871 // CHECK13-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
8872 // CHECK13-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
8873 // CHECK13-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
8874 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
8875 // CHECK13-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
8876 // CHECK13-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
8877 // CHECK13-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
8878 // CHECK13-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
8879 // CHECK13-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
8880 // CHECK13-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
8881 // CHECK13-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
8882 // CHECK13-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
8883 // CHECK13-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
8884 // CHECK13-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
8885 // CHECK13-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
8886 // CHECK13-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
8887 // CHECK13-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
8888 // CHECK13-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
8889 // CHECK13-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
8890 // CHECK13-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
8891 // CHECK13-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
8892 // CHECK13-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
8893 // CHECK13-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
8894 // CHECK13-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
8895 // CHECK13-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
8896 // CHECK13-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
8897 // CHECK13-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
8898 // CHECK13-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
8899 // CHECK13-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
8900 // CHECK13-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
8901 // CHECK13-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
8902 // CHECK13-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
8903 // CHECK13-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
8904 // CHECK13-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
8905 // CHECK13-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
8906 // CHECK13-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
8907 // CHECK13-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
8908 // CHECK13-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
8909 // CHECK13-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
8910 // CHECK13-NEXT:    store i32 0, i32* [[NN]], align 4
8911 // CHECK13-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
8912 // CHECK13-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
8913 // CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
8914 // CHECK13-NEXT:    ret i32 [[TMP20]]
8915 //
8916 //
8917 // CHECK13-LABEL: define {{[^@]+}}@_Z6bazzzziPi
8918 // CHECK13-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
8919 // CHECK13-NEXT:  entry:
8920 // CHECK13-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
8921 // CHECK13-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
8922 // CHECK13-NEXT:    [[F1:%.*]] = alloca i32*, align 8
8923 // CHECK13-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
8924 // CHECK13-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
8925 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
8926 // CHECK13-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
8927 // CHECK13-NEXT:    ret void
8928 //
8929 //
8930 // CHECK13-LABEL: define {{[^@]+}}@_Z3bari
8931 // CHECK13-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
8932 // CHECK13-NEXT:  entry:
8933 // CHECK13-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
8934 // CHECK13-NEXT:    [[A:%.*]] = alloca i32, align 4
8935 // CHECK13-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
8936 // CHECK13-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
8937 // CHECK13-NEXT:    store i32 0, i32* [[A]], align 4
8938 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
8939 // CHECK13-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
8940 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
8941 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
8942 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
8943 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
8944 // CHECK13-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
8945 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
8946 // CHECK13-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
8947 // CHECK13-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
8948 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
8949 // CHECK13-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
8950 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
8951 // CHECK13-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
8952 // CHECK13-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
8953 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
8954 // CHECK13-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
8955 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
8956 // CHECK13-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
8957 // CHECK13-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
8958 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
8959 // CHECK13-NEXT:    ret i32 [[TMP8]]
8960 //
8961 //
8962 // CHECK13-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
8963 // CHECK13-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
8964 // CHECK13-NEXT:  entry:
8965 // CHECK13-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
8966 // CHECK13-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
8967 // CHECK13-NEXT:    [[B:%.*]] = alloca i32, align 4
8968 // CHECK13-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
8969 // CHECK13-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
8970 // CHECK13-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
8971 // CHECK13-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
8972 // CHECK13-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
8973 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
8974 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
8975 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
8976 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
8977 // CHECK13-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
8978 // CHECK13-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
8979 // CHECK13-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
8980 // CHECK13-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
8981 // CHECK13-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
8982 // CHECK13-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
8983 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
8984 // CHECK13-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
8985 // CHECK13-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
8986 // CHECK13-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
8987 // CHECK13-NEXT:    store double [[ADD2]], double* [[A]], align 8
8988 // CHECK13-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
8989 // CHECK13-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
8990 // CHECK13-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
8991 // CHECK13-NEXT:    store double [[INC]], double* [[A3]], align 8
8992 // CHECK13-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
8993 // CHECK13-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
8994 // CHECK13-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
8995 // CHECK13-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
8996 // CHECK13-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
8997 // CHECK13-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
8998 // CHECK13-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
8999 // CHECK13-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
9000 // CHECK13-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
9001 // CHECK13-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
9002 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
9003 // CHECK13-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
9004 // CHECK13-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
9005 // CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
9006 // CHECK13-NEXT:    ret i32 [[ADD9]]
9007 //
9008 //
9009 // CHECK13-LABEL: define {{[^@]+}}@_ZL7fstatici
9010 // CHECK13-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
9011 // CHECK13-NEXT:  entry:
9012 // CHECK13-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9013 // CHECK13-NEXT:    [[A:%.*]] = alloca i32, align 4
9014 // CHECK13-NEXT:    [[AA:%.*]] = alloca i16, align 2
9015 // CHECK13-NEXT:    [[AAA:%.*]] = alloca i8, align 1
9016 // CHECK13-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9017 // CHECK13-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9018 // CHECK13-NEXT:    store i32 0, i32* [[A]], align 4
9019 // CHECK13-NEXT:    store i16 0, i16* [[AA]], align 2
9020 // CHECK13-NEXT:    store i8 0, i8* [[AAA]], align 1
9021 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9022 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9023 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9024 // CHECK13-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9025 // CHECK13-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9026 // CHECK13-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9027 // CHECK13-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9028 // CHECK13-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9029 // CHECK13-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
9030 // CHECK13-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
9031 // CHECK13-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
9032 // CHECK13-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
9033 // CHECK13-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
9034 // CHECK13-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
9035 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9036 // CHECK13-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
9037 // CHECK13-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
9038 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
9039 // CHECK13-NEXT:    ret i32 [[TMP4]]
9040 //
9041 //
9042 // CHECK13-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
9043 // CHECK13-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
9044 // CHECK13-NEXT:  entry:
9045 // CHECK13-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9046 // CHECK13-NEXT:    [[A:%.*]] = alloca i32, align 4
9047 // CHECK13-NEXT:    [[AA:%.*]] = alloca i16, align 2
9048 // CHECK13-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9049 // CHECK13-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9050 // CHECK13-NEXT:    store i32 0, i32* [[A]], align 4
9051 // CHECK13-NEXT:    store i16 0, i16* [[AA]], align 2
9052 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9053 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9054 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9055 // CHECK13-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9056 // CHECK13-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9057 // CHECK13-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9058 // CHECK13-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9059 // CHECK13-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9060 // CHECK13-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
9061 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9062 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
9063 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
9064 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9065 // CHECK13-NEXT:    ret i32 [[TMP3]]
9066 //
9067 //
9068 // CHECK14-LABEL: define {{[^@]+}}@_Z3fooi
9069 // CHECK14-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
9070 // CHECK14-NEXT:  entry:
9071 // CHECK14-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9072 // CHECK14-NEXT:    [[A:%.*]] = alloca i32, align 4
9073 // CHECK14-NEXT:    [[AA:%.*]] = alloca i16, align 2
9074 // CHECK14-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
9075 // CHECK14-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
9076 // CHECK14-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
9077 // CHECK14-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
9078 // CHECK14-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
9079 // CHECK14-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
9080 // CHECK14-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
9081 // CHECK14-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
9082 // CHECK14-NEXT:    [[NN:%.*]] = alloca i32, align 4
9083 // CHECK14-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9084 // CHECK14-NEXT:    store i32 0, i32* [[A]], align 4
9085 // CHECK14-NEXT:    store i16 0, i16* [[AA]], align 2
9086 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9087 // CHECK14-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
9088 // CHECK14-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
9089 // CHECK14-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
9090 // CHECK14-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
9091 // CHECK14-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
9092 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
9093 // CHECK14-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
9094 // CHECK14-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
9095 // CHECK14-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
9096 // CHECK14-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
9097 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
9098 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
9099 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
9100 // CHECK14-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
9101 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9102 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
9103 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9104 // CHECK14-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
9105 // CHECK14-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
9106 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
9107 // CHECK14-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
9108 // CHECK14-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
9109 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
9110 // CHECK14-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
9111 // CHECK14-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
9112 // CHECK14-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
9113 // CHECK14-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
9114 // CHECK14-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
9115 // CHECK14-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
9116 // CHECK14-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
9117 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
9118 // CHECK14-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
9119 // CHECK14-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
9120 // CHECK14-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
9121 // CHECK14-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
9122 // CHECK14-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
9123 // CHECK14-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
9124 // CHECK14-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
9125 // CHECK14-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
9126 // CHECK14-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
9127 // CHECK14-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
9128 // CHECK14-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
9129 // CHECK14-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
9130 // CHECK14-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
9131 // CHECK14-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
9132 // CHECK14-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
9133 // CHECK14-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
9134 // CHECK14-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
9135 // CHECK14-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
9136 // CHECK14-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
9137 // CHECK14-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
9138 // CHECK14-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
9139 // CHECK14-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
9140 // CHECK14-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
9141 // CHECK14-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
9142 // CHECK14-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
9143 // CHECK14-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
9144 // CHECK14-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
9145 // CHECK14-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
9146 // CHECK14-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
9147 // CHECK14-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
9148 // CHECK14-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
9149 // CHECK14-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
9150 // CHECK14-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
9151 // CHECK14-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
9152 // CHECK14-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
9153 // CHECK14-NEXT:    store i32 0, i32* [[NN]], align 4
9154 // CHECK14-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
9155 // CHECK14-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
9156 // CHECK14-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
9157 // CHECK14-NEXT:    ret i32 [[TMP20]]
9158 //
9159 //
9160 // CHECK14-LABEL: define {{[^@]+}}@_Z6bazzzziPi
9161 // CHECK14-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
9162 // CHECK14-NEXT:  entry:
9163 // CHECK14-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9164 // CHECK14-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
9165 // CHECK14-NEXT:    [[F1:%.*]] = alloca i32*, align 8
9166 // CHECK14-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9167 // CHECK14-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
9168 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9169 // CHECK14-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
9170 // CHECK14-NEXT:    ret void
9171 //
9172 //
9173 // CHECK14-LABEL: define {{[^@]+}}@_Z3bari
9174 // CHECK14-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
9175 // CHECK14-NEXT:  entry:
9176 // CHECK14-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9177 // CHECK14-NEXT:    [[A:%.*]] = alloca i32, align 4
9178 // CHECK14-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
9179 // CHECK14-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9180 // CHECK14-NEXT:    store i32 0, i32* [[A]], align 4
9181 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9182 // CHECK14-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
9183 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
9184 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
9185 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9186 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
9187 // CHECK14-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
9188 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9189 // CHECK14-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
9190 // CHECK14-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
9191 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
9192 // CHECK14-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
9193 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
9194 // CHECK14-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
9195 // CHECK14-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
9196 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
9197 // CHECK14-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
9198 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
9199 // CHECK14-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
9200 // CHECK14-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
9201 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9202 // CHECK14-NEXT:    ret i32 [[TMP8]]
9203 //
9204 //
9205 // CHECK14-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
9206 // CHECK14-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
9207 // CHECK14-NEXT:  entry:
9208 // CHECK14-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
9209 // CHECK14-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9210 // CHECK14-NEXT:    [[B:%.*]] = alloca i32, align 4
9211 // CHECK14-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
9212 // CHECK14-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
9213 // CHECK14-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
9214 // CHECK14-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9215 // CHECK14-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
9216 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9217 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9218 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
9219 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
9220 // CHECK14-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
9221 // CHECK14-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
9222 // CHECK14-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
9223 // CHECK14-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
9224 // CHECK14-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
9225 // CHECK14-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
9226 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
9227 // CHECK14-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
9228 // CHECK14-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
9229 // CHECK14-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
9230 // CHECK14-NEXT:    store double [[ADD2]], double* [[A]], align 8
9231 // CHECK14-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
9232 // CHECK14-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
9233 // CHECK14-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
9234 // CHECK14-NEXT:    store double [[INC]], double* [[A3]], align 8
9235 // CHECK14-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
9236 // CHECK14-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
9237 // CHECK14-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
9238 // CHECK14-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
9239 // CHECK14-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
9240 // CHECK14-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
9241 // CHECK14-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
9242 // CHECK14-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
9243 // CHECK14-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
9244 // CHECK14-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
9245 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
9246 // CHECK14-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
9247 // CHECK14-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
9248 // CHECK14-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
9249 // CHECK14-NEXT:    ret i32 [[ADD9]]
9250 //
9251 //
9252 // CHECK14-LABEL: define {{[^@]+}}@_ZL7fstatici
9253 // CHECK14-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
9254 // CHECK14-NEXT:  entry:
9255 // CHECK14-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9256 // CHECK14-NEXT:    [[A:%.*]] = alloca i32, align 4
9257 // CHECK14-NEXT:    [[AA:%.*]] = alloca i16, align 2
9258 // CHECK14-NEXT:    [[AAA:%.*]] = alloca i8, align 1
9259 // CHECK14-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9260 // CHECK14-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9261 // CHECK14-NEXT:    store i32 0, i32* [[A]], align 4
9262 // CHECK14-NEXT:    store i16 0, i16* [[AA]], align 2
9263 // CHECK14-NEXT:    store i8 0, i8* [[AAA]], align 1
9264 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9265 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9266 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9267 // CHECK14-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9268 // CHECK14-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9269 // CHECK14-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9270 // CHECK14-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9271 // CHECK14-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9272 // CHECK14-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
9273 // CHECK14-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
9274 // CHECK14-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
9275 // CHECK14-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
9276 // CHECK14-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
9277 // CHECK14-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
9278 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9279 // CHECK14-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
9280 // CHECK14-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
9281 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
9282 // CHECK14-NEXT:    ret i32 [[TMP4]]
9283 //
9284 //
9285 // CHECK14-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
9286 // CHECK14-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
9287 // CHECK14-NEXT:  entry:
9288 // CHECK14-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9289 // CHECK14-NEXT:    [[A:%.*]] = alloca i32, align 4
9290 // CHECK14-NEXT:    [[AA:%.*]] = alloca i16, align 2
9291 // CHECK14-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9292 // CHECK14-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9293 // CHECK14-NEXT:    store i32 0, i32* [[A]], align 4
9294 // CHECK14-NEXT:    store i16 0, i16* [[AA]], align 2
9295 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9296 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9297 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9298 // CHECK14-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9299 // CHECK14-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9300 // CHECK14-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9301 // CHECK14-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9302 // CHECK14-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9303 // CHECK14-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
9304 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9305 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
9306 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
9307 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9308 // CHECK14-NEXT:    ret i32 [[TMP3]]
9309 //
9310 //
9311 // CHECK15-LABEL: define {{[^@]+}}@_Z3fooi
9312 // CHECK15-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
9313 // CHECK15-NEXT:  entry:
9314 // CHECK15-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9315 // CHECK15-NEXT:    [[A:%.*]] = alloca i32, align 4
9316 // CHECK15-NEXT:    [[AA:%.*]] = alloca i16, align 2
9317 // CHECK15-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
9318 // CHECK15-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
9319 // CHECK15-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
9320 // CHECK15-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
9321 // CHECK15-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
9322 // CHECK15-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
9323 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
9324 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
9325 // CHECK15-NEXT:    [[NN:%.*]] = alloca i32, align 4
9326 // CHECK15-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9327 // CHECK15-NEXT:    store i32 0, i32* [[A]], align 4
9328 // CHECK15-NEXT:    store i16 0, i16* [[AA]], align 2
9329 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9330 // CHECK15-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
9331 // CHECK15-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
9332 // CHECK15-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
9333 // CHECK15-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
9334 // CHECK15-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
9335 // CHECK15-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
9336 // CHECK15-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
9337 // CHECK15-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
9338 // CHECK15-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
9339 // CHECK15-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
9340 // CHECK15-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
9341 // CHECK15-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
9342 // CHECK15-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
9343 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
9344 // CHECK15-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9345 // CHECK15-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
9346 // CHECK15-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
9347 // CHECK15-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
9348 // CHECK15-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
9349 // CHECK15-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
9350 // CHECK15-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9351 // CHECK15-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
9352 // CHECK15-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
9353 // CHECK15-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
9354 // CHECK15-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
9355 // CHECK15-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
9356 // CHECK15-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
9357 // CHECK15-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
9358 // CHECK15-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
9359 // CHECK15-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
9360 // CHECK15-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
9361 // CHECK15-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
9362 // CHECK15-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
9363 // CHECK15-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
9364 // CHECK15-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
9365 // CHECK15-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
9366 // CHECK15-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
9367 // CHECK15-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
9368 // CHECK15-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
9369 // CHECK15-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
9370 // CHECK15-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
9371 // CHECK15-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
9372 // CHECK15-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
9373 // CHECK15-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
9374 // CHECK15-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
9375 // CHECK15-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
9376 // CHECK15-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
9377 // CHECK15-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
9378 // CHECK15-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
9379 // CHECK15-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
9380 // CHECK15-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
9381 // CHECK15-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
9382 // CHECK15-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
9383 // CHECK15-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
9384 // CHECK15-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
9385 // CHECK15-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
9386 // CHECK15-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
9387 // CHECK15-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
9388 // CHECK15-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
9389 // CHECK15-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
9390 // CHECK15-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
9391 // CHECK15-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
9392 // CHECK15-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
9393 // CHECK15-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
9394 // CHECK15-NEXT:    store i32 0, i32* [[NN]], align 4
9395 // CHECK15-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
9396 // CHECK15-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
9397 // CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
9398 // CHECK15-NEXT:    ret i32 [[TMP18]]
9399 //
9400 //
9401 // CHECK15-LABEL: define {{[^@]+}}@_Z6bazzzziPi
9402 // CHECK15-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
9403 // CHECK15-NEXT:  entry:
9404 // CHECK15-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9405 // CHECK15-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
9406 // CHECK15-NEXT:    [[F1:%.*]] = alloca i32*, align 4
9407 // CHECK15-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9408 // CHECK15-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
9409 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9410 // CHECK15-NEXT:    ret void
9411 //
9412 //
9413 // CHECK15-LABEL: define {{[^@]+}}@_Z3bari
9414 // CHECK15-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
9415 // CHECK15-NEXT:  entry:
9416 // CHECK15-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9417 // CHECK15-NEXT:    [[A:%.*]] = alloca i32, align 4
9418 // CHECK15-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
9419 // CHECK15-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9420 // CHECK15-NEXT:    store i32 0, i32* [[A]], align 4
9421 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9422 // CHECK15-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
9423 // CHECK15-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
9424 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
9425 // CHECK15-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9426 // CHECK15-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
9427 // CHECK15-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
9428 // CHECK15-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9429 // CHECK15-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
9430 // CHECK15-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
9431 // CHECK15-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
9432 // CHECK15-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
9433 // CHECK15-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
9434 // CHECK15-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
9435 // CHECK15-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
9436 // CHECK15-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
9437 // CHECK15-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
9438 // CHECK15-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
9439 // CHECK15-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
9440 // CHECK15-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
9441 // CHECK15-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9442 // CHECK15-NEXT:    ret i32 [[TMP8]]
9443 //
9444 //
9445 // CHECK15-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
9446 // CHECK15-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
9447 // CHECK15-NEXT:  entry:
9448 // CHECK15-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
9449 // CHECK15-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9450 // CHECK15-NEXT:    [[B:%.*]] = alloca i32, align 4
9451 // CHECK15-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
9452 // CHECK15-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
9453 // CHECK15-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
9454 // CHECK15-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9455 // CHECK15-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
9456 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9457 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9458 // CHECK15-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
9459 // CHECK15-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
9460 // CHECK15-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
9461 // CHECK15-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
9462 // CHECK15-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
9463 // CHECK15-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
9464 // CHECK15-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
9465 // CHECK15-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
9466 // CHECK15-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
9467 // CHECK15-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
9468 // CHECK15-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
9469 // CHECK15-NEXT:    store double [[ADD2]], double* [[A]], align 4
9470 // CHECK15-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
9471 // CHECK15-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
9472 // CHECK15-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
9473 // CHECK15-NEXT:    store double [[INC]], double* [[A3]], align 4
9474 // CHECK15-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
9475 // CHECK15-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
9476 // CHECK15-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
9477 // CHECK15-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
9478 // CHECK15-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
9479 // CHECK15-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
9480 // CHECK15-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
9481 // CHECK15-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
9482 // CHECK15-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
9483 // CHECK15-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
9484 // CHECK15-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
9485 // CHECK15-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
9486 // CHECK15-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
9487 // CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
9488 // CHECK15-NEXT:    ret i32 [[ADD9]]
9489 //
9490 //
9491 // CHECK15-LABEL: define {{[^@]+}}@_ZL7fstatici
9492 // CHECK15-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
9493 // CHECK15-NEXT:  entry:
9494 // CHECK15-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9495 // CHECK15-NEXT:    [[A:%.*]] = alloca i32, align 4
9496 // CHECK15-NEXT:    [[AA:%.*]] = alloca i16, align 2
9497 // CHECK15-NEXT:    [[AAA:%.*]] = alloca i8, align 1
9498 // CHECK15-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9499 // CHECK15-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9500 // CHECK15-NEXT:    store i32 0, i32* [[A]], align 4
9501 // CHECK15-NEXT:    store i16 0, i16* [[AA]], align 2
9502 // CHECK15-NEXT:    store i8 0, i8* [[AAA]], align 1
9503 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9504 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9505 // CHECK15-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9506 // CHECK15-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9507 // CHECK15-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9508 // CHECK15-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9509 // CHECK15-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9510 // CHECK15-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9511 // CHECK15-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
9512 // CHECK15-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
9513 // CHECK15-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
9514 // CHECK15-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
9515 // CHECK15-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
9516 // CHECK15-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
9517 // CHECK15-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9518 // CHECK15-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
9519 // CHECK15-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
9520 // CHECK15-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
9521 // CHECK15-NEXT:    ret i32 [[TMP4]]
9522 //
9523 //
9524 // CHECK15-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
9525 // CHECK15-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
9526 // CHECK15-NEXT:  entry:
9527 // CHECK15-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9528 // CHECK15-NEXT:    [[A:%.*]] = alloca i32, align 4
9529 // CHECK15-NEXT:    [[AA:%.*]] = alloca i16, align 2
9530 // CHECK15-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9531 // CHECK15-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9532 // CHECK15-NEXT:    store i32 0, i32* [[A]], align 4
9533 // CHECK15-NEXT:    store i16 0, i16* [[AA]], align 2
9534 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9535 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9536 // CHECK15-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9537 // CHECK15-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9538 // CHECK15-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9539 // CHECK15-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9540 // CHECK15-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9541 // CHECK15-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9542 // CHECK15-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
9543 // CHECK15-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9544 // CHECK15-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
9545 // CHECK15-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
9546 // CHECK15-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9547 // CHECK15-NEXT:    ret i32 [[TMP3]]
9548 //
9549 //
9550 // CHECK16-LABEL: define {{[^@]+}}@_Z3fooi
9551 // CHECK16-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
9552 // CHECK16-NEXT:  entry:
9553 // CHECK16-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9554 // CHECK16-NEXT:    [[A:%.*]] = alloca i32, align 4
9555 // CHECK16-NEXT:    [[AA:%.*]] = alloca i16, align 2
9556 // CHECK16-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
9557 // CHECK16-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
9558 // CHECK16-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
9559 // CHECK16-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
9560 // CHECK16-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
9561 // CHECK16-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
9562 // CHECK16-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
9563 // CHECK16-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
9564 // CHECK16-NEXT:    [[NN:%.*]] = alloca i32, align 4
9565 // CHECK16-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9566 // CHECK16-NEXT:    store i32 0, i32* [[A]], align 4
9567 // CHECK16-NEXT:    store i16 0, i16* [[AA]], align 2
9568 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9569 // CHECK16-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
9570 // CHECK16-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
9571 // CHECK16-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
9572 // CHECK16-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
9573 // CHECK16-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
9574 // CHECK16-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
9575 // CHECK16-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
9576 // CHECK16-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
9577 // CHECK16-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
9578 // CHECK16-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
9579 // CHECK16-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
9580 // CHECK16-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
9581 // CHECK16-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
9582 // CHECK16-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
9583 // CHECK16-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9584 // CHECK16-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
9585 // CHECK16-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
9586 // CHECK16-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
9587 // CHECK16-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
9588 // CHECK16-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
9589 // CHECK16-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9590 // CHECK16-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
9591 // CHECK16-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
9592 // CHECK16-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
9593 // CHECK16-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
9594 // CHECK16-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
9595 // CHECK16-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
9596 // CHECK16-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
9597 // CHECK16-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
9598 // CHECK16-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
9599 // CHECK16-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
9600 // CHECK16-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
9601 // CHECK16-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
9602 // CHECK16-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
9603 // CHECK16-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
9604 // CHECK16-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
9605 // CHECK16-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
9606 // CHECK16-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
9607 // CHECK16-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
9608 // CHECK16-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
9609 // CHECK16-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
9610 // CHECK16-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
9611 // CHECK16-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
9612 // CHECK16-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
9613 // CHECK16-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
9614 // CHECK16-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
9615 // CHECK16-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
9616 // CHECK16-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
9617 // CHECK16-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
9618 // CHECK16-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
9619 // CHECK16-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
9620 // CHECK16-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
9621 // CHECK16-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
9622 // CHECK16-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
9623 // CHECK16-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
9624 // CHECK16-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
9625 // CHECK16-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
9626 // CHECK16-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
9627 // CHECK16-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
9628 // CHECK16-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
9629 // CHECK16-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
9630 // CHECK16-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
9631 // CHECK16-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
9632 // CHECK16-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
9633 // CHECK16-NEXT:    store i32 0, i32* [[NN]], align 4
9634 // CHECK16-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
9635 // CHECK16-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
9636 // CHECK16-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
9637 // CHECK16-NEXT:    ret i32 [[TMP18]]
9638 //
9639 //
9640 // CHECK16-LABEL: define {{[^@]+}}@_Z6bazzzziPi
9641 // CHECK16-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
9642 // CHECK16-NEXT:  entry:
9643 // CHECK16-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9644 // CHECK16-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
9645 // CHECK16-NEXT:    [[F1:%.*]] = alloca i32*, align 4
9646 // CHECK16-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9647 // CHECK16-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
9648 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9649 // CHECK16-NEXT:    ret void
9650 //
9651 //
9652 // CHECK16-LABEL: define {{[^@]+}}@_Z3bari
9653 // CHECK16-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
9654 // CHECK16-NEXT:  entry:
9655 // CHECK16-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9656 // CHECK16-NEXT:    [[A:%.*]] = alloca i32, align 4
9657 // CHECK16-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
9658 // CHECK16-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9659 // CHECK16-NEXT:    store i32 0, i32* [[A]], align 4
9660 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9661 // CHECK16-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
9662 // CHECK16-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
9663 // CHECK16-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
9664 // CHECK16-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9665 // CHECK16-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
9666 // CHECK16-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
9667 // CHECK16-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9668 // CHECK16-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
9669 // CHECK16-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
9670 // CHECK16-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
9671 // CHECK16-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
9672 // CHECK16-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
9673 // CHECK16-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
9674 // CHECK16-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
9675 // CHECK16-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
9676 // CHECK16-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
9677 // CHECK16-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
9678 // CHECK16-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
9679 // CHECK16-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
9680 // CHECK16-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9681 // CHECK16-NEXT:    ret i32 [[TMP8]]
9682 //
9683 //
9684 // CHECK16-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
9685 // CHECK16-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
9686 // CHECK16-NEXT:  entry:
9687 // CHECK16-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
9688 // CHECK16-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9689 // CHECK16-NEXT:    [[B:%.*]] = alloca i32, align 4
9690 // CHECK16-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
9691 // CHECK16-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
9692 // CHECK16-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
9693 // CHECK16-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9694 // CHECK16-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
9695 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
9696 // CHECK16-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9697 // CHECK16-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
9698 // CHECK16-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
9699 // CHECK16-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
9700 // CHECK16-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
9701 // CHECK16-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
9702 // CHECK16-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
9703 // CHECK16-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
9704 // CHECK16-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
9705 // CHECK16-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
9706 // CHECK16-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
9707 // CHECK16-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
9708 // CHECK16-NEXT:    store double [[ADD2]], double* [[A]], align 4
9709 // CHECK16-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
9710 // CHECK16-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
9711 // CHECK16-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
9712 // CHECK16-NEXT:    store double [[INC]], double* [[A3]], align 4
9713 // CHECK16-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
9714 // CHECK16-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
9715 // CHECK16-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
9716 // CHECK16-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
9717 // CHECK16-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
9718 // CHECK16-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
9719 // CHECK16-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
9720 // CHECK16-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
9721 // CHECK16-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
9722 // CHECK16-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
9723 // CHECK16-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
9724 // CHECK16-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
9725 // CHECK16-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
9726 // CHECK16-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
9727 // CHECK16-NEXT:    ret i32 [[ADD9]]
9728 //
9729 //
9730 // CHECK16-LABEL: define {{[^@]+}}@_ZL7fstatici
9731 // CHECK16-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
9732 // CHECK16-NEXT:  entry:
9733 // CHECK16-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9734 // CHECK16-NEXT:    [[A:%.*]] = alloca i32, align 4
9735 // CHECK16-NEXT:    [[AA:%.*]] = alloca i16, align 2
9736 // CHECK16-NEXT:    [[AAA:%.*]] = alloca i8, align 1
9737 // CHECK16-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9738 // CHECK16-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9739 // CHECK16-NEXT:    store i32 0, i32* [[A]], align 4
9740 // CHECK16-NEXT:    store i16 0, i16* [[AA]], align 2
9741 // CHECK16-NEXT:    store i8 0, i8* [[AAA]], align 1
9742 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9743 // CHECK16-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9744 // CHECK16-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9745 // CHECK16-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9746 // CHECK16-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9747 // CHECK16-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9748 // CHECK16-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9749 // CHECK16-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9750 // CHECK16-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
9751 // CHECK16-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
9752 // CHECK16-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
9753 // CHECK16-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
9754 // CHECK16-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
9755 // CHECK16-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
9756 // CHECK16-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9757 // CHECK16-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
9758 // CHECK16-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
9759 // CHECK16-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
9760 // CHECK16-NEXT:    ret i32 [[TMP4]]
9761 //
9762 //
9763 // CHECK16-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
9764 // CHECK16-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
9765 // CHECK16-NEXT:  entry:
9766 // CHECK16-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9767 // CHECK16-NEXT:    [[A:%.*]] = alloca i32, align 4
9768 // CHECK16-NEXT:    [[AA:%.*]] = alloca i16, align 2
9769 // CHECK16-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
9770 // CHECK16-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9771 // CHECK16-NEXT:    store i32 0, i32* [[A]], align 4
9772 // CHECK16-NEXT:    store i16 0, i16* [[AA]], align 2
9773 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
9774 // CHECK16-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
9775 // CHECK16-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
9776 // CHECK16-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
9777 // CHECK16-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
9778 // CHECK16-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
9779 // CHECK16-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
9780 // CHECK16-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
9781 // CHECK16-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
9782 // CHECK16-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
9783 // CHECK16-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
9784 // CHECK16-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
9785 // CHECK16-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
9786 // CHECK16-NEXT:    ret i32 [[TMP3]]
9787 //
9788 //
9789 // CHECK17-LABEL: define {{[^@]+}}@_Z3fooi
9790 // CHECK17-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
9791 // CHECK17-NEXT:  entry:
9792 // CHECK17-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
9793 // CHECK17-NEXT:    [[A:%.*]] = alloca i32, align 4
9794 // CHECK17-NEXT:    [[AA:%.*]] = alloca i16, align 2
9795 // CHECK17-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
9796 // CHECK17-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
9797 // CHECK17-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
9798 // CHECK17-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
9799 // CHECK17-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
9800 // CHECK17-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
9801 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
9802 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
9803 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
9804 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
9805 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED4:%.*]] = alloca i64, align 8
9806 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
9807 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
9808 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
9809 // CHECK17-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
9810 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
9811 // CHECK17-NEXT:    [[AA_CASTED7:%.*]] = alloca i64, align 8
9812 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS9:%.*]] = alloca [1 x i8*], align 8
9813 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS10:%.*]] = alloca [1 x i8*], align 8
9814 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS11:%.*]] = alloca [1 x i8*], align 8
9815 // CHECK17-NEXT:    [[A_CASTED12:%.*]] = alloca i64, align 8
9816 // CHECK17-NEXT:    [[AA_CASTED14:%.*]] = alloca i64, align 8
9817 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS16:%.*]] = alloca [2 x i8*], align 8
9818 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
9819 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
9820 // CHECK17-NEXT:    [[A_CASTED21:%.*]] = alloca i64, align 8
9821 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [9 x i8*], align 8
9822 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [9 x i8*], align 8
9823 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [9 x i8*], align 8
9824 // CHECK17-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
9825 // CHECK17-NEXT:    [[NN:%.*]] = alloca i32, align 4
9826 // CHECK17-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
9827 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 8
9828 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 8
9829 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 8
9830 // CHECK17-NEXT:    [[NN_CASTED38:%.*]] = alloca i64, align 8
9831 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS40:%.*]] = alloca [1 x i8*], align 8
9832 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS41:%.*]] = alloca [1 x i8*], align 8
9833 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS42:%.*]] = alloca [1 x i8*], align 8
9834 // CHECK17-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
9835 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
9836 // CHECK17-NEXT:    store i32 0, i32* [[A]], align 4
9837 // CHECK17-NEXT:    store i16 0, i16* [[AA]], align 2
9838 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
9839 // CHECK17-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
9840 // CHECK17-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
9841 // CHECK17-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
9842 // CHECK17-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP2]], align 4
9843 // CHECK17-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
9844 // CHECK17-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
9845 // CHECK17-NEXT:    [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
9846 // CHECK17-NEXT:    [[TMP6:%.*]] = mul nuw i64 5, [[TMP5]]
9847 // CHECK17-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP6]], align 8
9848 // CHECK17-NEXT:    store i64 [[TMP5]], i64* [[__VLA_EXPR1]], align 8
9849 // CHECK17-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
9850 // CHECK17-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_]], align 4
9851 // CHECK17-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
9852 // CHECK17-NEXT:    store i32 [[TMP8]], i32* [[DOTCAPTURE_EXPR_2]], align 4
9853 // CHECK17-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
9854 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
9855 // CHECK17-NEXT:    store i16 [[TMP9]], i16* [[CONV]], align 2
9856 // CHECK17-NEXT:    [[TMP10:%.*]] = load i64, i64* [[AA_CASTED]], align 8
9857 // CHECK17-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
9858 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
9859 // CHECK17-NEXT:    store i32 [[TMP11]], i32* [[CONV3]], align 4
9860 // CHECK17-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
9861 // CHECK17-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
9862 // CHECK17-NEXT:    [[CONV5:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED4]] to i32*
9863 // CHECK17-NEXT:    store i32 [[TMP13]], i32* [[CONV5]], align 4
9864 // CHECK17-NEXT:    [[TMP14:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED4]], align 8
9865 // CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
9866 // CHECK17-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
9867 // CHECK17-NEXT:    store i64 [[TMP10]], i64* [[TMP16]], align 8
9868 // CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
9869 // CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
9870 // CHECK17-NEXT:    store i64 [[TMP10]], i64* [[TMP18]], align 8
9871 // CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
9872 // CHECK17-NEXT:    store i8* null, i8** [[TMP19]], align 8
9873 // CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
9874 // CHECK17-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
9875 // CHECK17-NEXT:    store i64 [[TMP12]], i64* [[TMP21]], align 8
9876 // CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
9877 // CHECK17-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
9878 // CHECK17-NEXT:    store i64 [[TMP12]], i64* [[TMP23]], align 8
9879 // CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
9880 // CHECK17-NEXT:    store i8* null, i8** [[TMP24]], align 8
9881 // CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
9882 // CHECK17-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i64*
9883 // CHECK17-NEXT:    store i64 [[TMP14]], i64* [[TMP26]], align 8
9884 // CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
9885 // CHECK17-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
9886 // CHECK17-NEXT:    store i64 [[TMP14]], i64* [[TMP28]], align 8
9887 // CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
9888 // CHECK17-NEXT:    store i8* null, i8** [[TMP29]], align 8
9889 // CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
9890 // CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
9891 // CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
9892 // CHECK17-NEXT:    [[TMP33:%.*]] = load i16, i16* [[AA]], align 2
9893 // CHECK17-NEXT:    store i16 [[TMP33]], i16* [[TMP32]], align 4
9894 // CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
9895 // CHECK17-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
9896 // CHECK17-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
9897 // CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
9898 // CHECK17-NEXT:    [[TMP37:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
9899 // CHECK17-NEXT:    store i32 [[TMP37]], i32* [[TMP36]], align 4
9900 // CHECK17-NEXT:    [[TMP38:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
9901 // CHECK17-NEXT:    [[TMP39:%.*]] = bitcast i8* [[TMP38]] to %struct.kmp_task_t_with_privates*
9902 // CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 0
9903 // CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP40]], i32 0, i32 0
9904 // CHECK17-NEXT:    [[TMP42:%.*]] = load i8*, i8** [[TMP41]], align 8
9905 // CHECK17-NEXT:    [[TMP43:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
9906 // CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP42]], i8* align 4 [[TMP43]], i64 12, i1 false)
9907 // CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 1
9908 // CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8* [[TMP42]] to %struct.anon*
9909 // CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 0
9910 // CHECK17-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
9911 // CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP30]] to i8*
9912 // CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP47]], i8* align 8 [[TMP48]], i64 24, i1 false)
9913 // CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 1
9914 // CHECK17-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
9915 // CHECK17-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP31]] to i8*
9916 // CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP50]], i8* align 8 [[TMP51]], i64 24, i1 false)
9917 // CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 2
9918 // CHECK17-NEXT:    [[TMP53:%.*]] = bitcast [3 x i64]* [[TMP52]] to i8*
9919 // CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP53]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
9920 // CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 3
9921 // CHECK17-NEXT:    [[TMP55:%.*]] = load i16, i16* [[AA]], align 2
9922 // CHECK17-NEXT:    store i16 [[TMP55]], i16* [[TMP54]], align 8
9923 // CHECK17-NEXT:    [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP38]])
9924 // CHECK17-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
9925 // CHECK17-NEXT:    [[CONV6:%.*]] = bitcast i64* [[A_CASTED]] to i32*
9926 // CHECK17-NEXT:    store i32 [[TMP57]], i32* [[CONV6]], align 4
9927 // CHECK17-NEXT:    [[TMP58:%.*]] = load i64, i64* [[A_CASTED]], align 8
9928 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i64 [[TMP58]]) #[[ATTR3:[0-9]+]]
9929 // CHECK17-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
9930 // CHECK17-NEXT:    [[CONV8:%.*]] = bitcast i64* [[AA_CASTED7]] to i16*
9931 // CHECK17-NEXT:    store i16 [[TMP59]], i16* [[CONV8]], align 2
9932 // CHECK17-NEXT:    [[TMP60:%.*]] = load i64, i64* [[AA_CASTED7]], align 8
9933 // CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
9934 // CHECK17-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i64*
9935 // CHECK17-NEXT:    store i64 [[TMP60]], i64* [[TMP62]], align 8
9936 // CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
9937 // CHECK17-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i64*
9938 // CHECK17-NEXT:    store i64 [[TMP60]], i64* [[TMP64]], align 8
9939 // CHECK17-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
9940 // CHECK17-NEXT:    store i8* null, i8** [[TMP65]], align 8
9941 // CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
9942 // CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
9943 // CHECK17-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
9944 // CHECK17-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
9945 // CHECK17-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
9946 // CHECK17:       omp_offload.failed:
9947 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR3]]
9948 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
9949 // CHECK17:       omp_offload.cont:
9950 // CHECK17-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
9951 // CHECK17-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
9952 // CHECK17-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
9953 // CHECK17-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
9954 // CHECK17-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
9955 // CHECK17-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
9956 // CHECK17-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
9957 // CHECK17-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
9958 // CHECK17-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
9959 // CHECK17-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
9960 // CHECK17-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
9961 // CHECK17:       omp_if.then:
9962 // CHECK17-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
9963 // CHECK17-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
9964 // CHECK17-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
9965 // CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
9966 // CHECK17-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
9967 // CHECK17-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
9968 // CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
9969 // CHECK17-NEXT:    store i8* null, i8** [[TMP79]], align 8
9970 // CHECK17-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
9971 // CHECK17-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
9972 // CHECK17-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
9973 // CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
9974 // CHECK17-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
9975 // CHECK17-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
9976 // CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
9977 // CHECK17-NEXT:    store i8* null, i8** [[TMP84]], align 8
9978 // CHECK17-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
9979 // CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
9980 // CHECK17-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
9981 // CHECK17-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
9982 // CHECK17-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
9983 // CHECK17:       omp_offload.failed19:
9984 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
9985 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
9986 // CHECK17:       omp_offload.cont20:
9987 // CHECK17-NEXT:    br label [[OMP_IF_END:%.*]]
9988 // CHECK17:       omp_if.else:
9989 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
9990 // CHECK17-NEXT:    br label [[OMP_IF_END]]
9991 // CHECK17:       omp_if.end:
9992 // CHECK17-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
9993 // CHECK17-NEXT:    [[CONV22:%.*]] = bitcast i64* [[A_CASTED21]] to i32*
9994 // CHECK17-NEXT:    store i32 [[TMP89]], i32* [[CONV22]], align 4
9995 // CHECK17-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED21]], align 8
9996 // CHECK17-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
9997 // CHECK17-NEXT:    [[CMP23:%.*]] = icmp sgt i32 [[TMP91]], 20
9998 // CHECK17-NEXT:    br i1 [[CMP23]], label [[OMP_IF_THEN24:%.*]], label [[OMP_IF_ELSE30:%.*]]
9999 // CHECK17:       omp_if.then24:
10000 // CHECK17-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
10001 // CHECK17-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
10002 // CHECK17-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
10003 // CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
10004 // CHECK17-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
10005 // CHECK17-NEXT:    store i64 [[TMP90]], i64* [[TMP96]], align 8
10006 // CHECK17-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
10007 // CHECK17-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
10008 // CHECK17-NEXT:    store i64 [[TMP90]], i64* [[TMP98]], align 8
10009 // CHECK17-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
10010 // CHECK17-NEXT:    store i64 4, i64* [[TMP99]], align 8
10011 // CHECK17-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
10012 // CHECK17-NEXT:    store i8* null, i8** [[TMP100]], align 8
10013 // CHECK17-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
10014 // CHECK17-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
10015 // CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
10016 // CHECK17-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
10017 // CHECK17-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
10018 // CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
10019 // CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
10020 // CHECK17-NEXT:    store i64 40, i64* [[TMP105]], align 8
10021 // CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
10022 // CHECK17-NEXT:    store i8* null, i8** [[TMP106]], align 8
10023 // CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 2
10024 // CHECK17-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
10025 // CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP108]], align 8
10026 // CHECK17-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 2
10027 // CHECK17-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
10028 // CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP110]], align 8
10029 // CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
10030 // CHECK17-NEXT:    store i64 8, i64* [[TMP111]], align 8
10031 // CHECK17-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 2
10032 // CHECK17-NEXT:    store i8* null, i8** [[TMP112]], align 8
10033 // CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 3
10034 // CHECK17-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
10035 // CHECK17-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
10036 // CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 3
10037 // CHECK17-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
10038 // CHECK17-NEXT:    store float* [[VLA]], float** [[TMP116]], align 8
10039 // CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
10040 // CHECK17-NEXT:    store i64 [[TMP92]], i64* [[TMP117]], align 8
10041 // CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 3
10042 // CHECK17-NEXT:    store i8* null, i8** [[TMP118]], align 8
10043 // CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 4
10044 // CHECK17-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
10045 // CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
10046 // CHECK17-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 4
10047 // CHECK17-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
10048 // CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 8
10049 // CHECK17-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
10050 // CHECK17-NEXT:    store i64 400, i64* [[TMP123]], align 8
10051 // CHECK17-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 4
10052 // CHECK17-NEXT:    store i8* null, i8** [[TMP124]], align 8
10053 // CHECK17-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 5
10054 // CHECK17-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
10055 // CHECK17-NEXT:    store i64 5, i64* [[TMP126]], align 8
10056 // CHECK17-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 5
10057 // CHECK17-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
10058 // CHECK17-NEXT:    store i64 5, i64* [[TMP128]], align 8
10059 // CHECK17-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
10060 // CHECK17-NEXT:    store i64 8, i64* [[TMP129]], align 8
10061 // CHECK17-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 5
10062 // CHECK17-NEXT:    store i8* null, i8** [[TMP130]], align 8
10063 // CHECK17-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 6
10064 // CHECK17-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
10065 // CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP132]], align 8
10066 // CHECK17-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 6
10067 // CHECK17-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
10068 // CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP134]], align 8
10069 // CHECK17-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
10070 // CHECK17-NEXT:    store i64 8, i64* [[TMP135]], align 8
10071 // CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 6
10072 // CHECK17-NEXT:    store i8* null, i8** [[TMP136]], align 8
10073 // CHECK17-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 7
10074 // CHECK17-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
10075 // CHECK17-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 8
10076 // CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 7
10077 // CHECK17-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
10078 // CHECK17-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 8
10079 // CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
10080 // CHECK17-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 8
10081 // CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 7
10082 // CHECK17-NEXT:    store i8* null, i8** [[TMP142]], align 8
10083 // CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 8
10084 // CHECK17-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
10085 // CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 8
10086 // CHECK17-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 8
10087 // CHECK17-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
10088 // CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 8
10089 // CHECK17-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
10090 // CHECK17-NEXT:    store i64 16, i64* [[TMP147]], align 8
10091 // CHECK17-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 8
10092 // CHECK17-NEXT:    store i8* null, i8** [[TMP148]], align 8
10093 // CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
10094 // CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
10095 // CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
10096 // CHECK17-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10097 // CHECK17-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
10098 // CHECK17-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
10099 // CHECK17:       omp_offload.failed28:
10100 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
10101 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
10102 // CHECK17:       omp_offload.cont29:
10103 // CHECK17-NEXT:    br label [[OMP_IF_END31:%.*]]
10104 // CHECK17:       omp_if.else30:
10105 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
10106 // CHECK17-NEXT:    br label [[OMP_IF_END31]]
10107 // CHECK17:       omp_if.end31:
10108 // CHECK17-NEXT:    store i32 0, i32* [[NN]], align 4
10109 // CHECK17-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
10110 // CHECK17-NEXT:    [[CONV32:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
10111 // CHECK17-NEXT:    store i32 [[TMP154]], i32* [[CONV32]], align 4
10112 // CHECK17-NEXT:    [[TMP155:%.*]] = load i64, i64* [[NN_CASTED]], align 8
10113 // CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
10114 // CHECK17-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i64*
10115 // CHECK17-NEXT:    store i64 [[TMP155]], i64* [[TMP157]], align 8
10116 // CHECK17-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
10117 // CHECK17-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i64*
10118 // CHECK17-NEXT:    store i64 [[TMP155]], i64* [[TMP159]], align 8
10119 // CHECK17-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 0
10120 // CHECK17-NEXT:    store i8* null, i8** [[TMP160]], align 8
10121 // CHECK17-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
10122 // CHECK17-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
10123 // CHECK17-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10124 // CHECK17-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
10125 // CHECK17-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
10126 // CHECK17:       omp_offload.failed36:
10127 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i64 [[TMP155]]) #[[ATTR3]]
10128 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
10129 // CHECK17:       omp_offload.cont37:
10130 // CHECK17-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
10131 // CHECK17-NEXT:    [[CONV39:%.*]] = bitcast i64* [[NN_CASTED38]] to i32*
10132 // CHECK17-NEXT:    store i32 [[TMP165]], i32* [[CONV39]], align 4
10133 // CHECK17-NEXT:    [[TMP166:%.*]] = load i64, i64* [[NN_CASTED38]], align 8
10134 // CHECK17-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
10135 // CHECK17-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i64*
10136 // CHECK17-NEXT:    store i64 [[TMP166]], i64* [[TMP168]], align 8
10137 // CHECK17-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
10138 // CHECK17-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
10139 // CHECK17-NEXT:    store i64 [[TMP166]], i64* [[TMP170]], align 8
10140 // CHECK17-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 0
10141 // CHECK17-NEXT:    store i8* null, i8** [[TMP171]], align 8
10142 // CHECK17-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
10143 // CHECK17-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
10144 // CHECK17-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10145 // CHECK17-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
10146 // CHECK17-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
10147 // CHECK17:       omp_offload.failed43:
10148 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i64 [[TMP166]]) #[[ATTR3]]
10149 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
10150 // CHECK17:       omp_offload.cont44:
10151 // CHECK17-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
10152 // CHECK17-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
10153 // CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
10154 // CHECK17-NEXT:    ret i32 [[TMP176]]
10155 //
10156 //
10157 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
10158 // CHECK17-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
10159 // CHECK17-NEXT:  entry:
10160 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
10161 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
10162 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
10163 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
10164 // CHECK17-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
10165 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
10166 // CHECK17-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
10167 // CHECK17-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
10168 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
10169 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
10170 // CHECK17-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
10171 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
10172 // CHECK17-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
10173 // CHECK17-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
10174 // CHECK17-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
10175 // CHECK17-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
10176 // CHECK17-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
10177 // CHECK17-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
10178 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
10179 // CHECK17-NEXT:    ret void
10180 //
10181 //
10182 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined.
10183 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
10184 // CHECK17-NEXT:  entry:
10185 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10186 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10187 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
10188 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10189 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10190 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
10191 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
10192 // CHECK17-NEXT:    ret void
10193 //
10194 //
10195 // CHECK17-LABEL: define {{[^@]+}}@.omp_task_privates_map.
10196 // CHECK17-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
10197 // CHECK17-NEXT:  entry:
10198 // CHECK17-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
10199 // CHECK17-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 8
10200 // CHECK17-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 8
10201 // CHECK17-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 8
10202 // CHECK17-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 8
10203 // CHECK17-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
10204 // CHECK17-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 8
10205 // CHECK17-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 8
10206 // CHECK17-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 8
10207 // CHECK17-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 8
10208 // CHECK17-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
10209 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
10210 // CHECK17-NEXT:    [[TMP7:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 8
10211 // CHECK17-NEXT:    store [3 x i8*]* [[TMP6]], [3 x i8*]** [[TMP7]], align 8
10212 // CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
10213 // CHECK17-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 8
10214 // CHECK17-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 8
10215 // CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
10216 // CHECK17-NEXT:    [[TMP11:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 8
10217 // CHECK17-NEXT:    store [3 x i64]* [[TMP10]], [3 x i64]** [[TMP11]], align 8
10218 // CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
10219 // CHECK17-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 8
10220 // CHECK17-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 8
10221 // CHECK17-NEXT:    ret void
10222 //
10223 //
10224 // CHECK17-LABEL: define {{[^@]+}}@.omp_task_entry.
10225 // CHECK17-SAME: (i32 signext [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
10226 // CHECK17-NEXT:  entry:
10227 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
10228 // CHECK17-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
10229 // CHECK17-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
10230 // CHECK17-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
10231 // CHECK17-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
10232 // CHECK17-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
10233 // CHECK17-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 8
10234 // CHECK17-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 8
10235 // CHECK17-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 8
10236 // CHECK17-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 8
10237 // CHECK17-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
10238 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
10239 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
10240 // CHECK17-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
10241 // CHECK17-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
10242 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
10243 // CHECK17-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
10244 // CHECK17-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
10245 // CHECK17-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
10246 // CHECK17-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
10247 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
10248 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
10249 // CHECK17-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
10250 // CHECK17-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
10251 // CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
10252 // CHECK17-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
10253 // CHECK17-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
10254 // CHECK17-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META14:![0-9]+]])
10255 // CHECK17-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META17:![0-9]+]])
10256 // CHECK17-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META19:![0-9]+]])
10257 // CHECK17-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META21:![0-9]+]])
10258 // CHECK17-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !23
10259 // CHECK17-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !23
10260 // CHECK17-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
10261 // CHECK17-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
10262 // CHECK17-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !23
10263 // CHECK17-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
10264 // CHECK17-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
10265 // CHECK17-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
10266 // CHECK17-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
10267 // CHECK17-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
10268 // CHECK17-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
10269 // CHECK17-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !23
10270 // CHECK17-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 8, !noalias !23
10271 // CHECK17-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 8, !noalias !23
10272 // CHECK17-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 8, !noalias !23
10273 // CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i64 0, i64 0
10274 // CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i64 0, i64 0
10275 // CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i64 0, i64 0
10276 // CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
10277 // CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
10278 // CHECK17-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
10279 // CHECK17-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
10280 // CHECK17-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
10281 // CHECK17-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
10282 // CHECK17-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
10283 // CHECK17:       omp_offload.failed.i:
10284 // CHECK17-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
10285 // CHECK17-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
10286 // CHECK17-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !23
10287 // CHECK17-NEXT:    [[TMP30:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !23
10288 // CHECK17-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
10289 // CHECK17-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
10290 // CHECK17-NEXT:    store i32 [[TMP31]], i32* [[CONV4_I]], align 4, !noalias !23
10291 // CHECK17-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !23
10292 // CHECK17-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
10293 // CHECK17-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
10294 // CHECK17-NEXT:    store i32 [[TMP33]], i32* [[CONV6_I]], align 4, !noalias !23
10295 // CHECK17-NEXT:    [[TMP34:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !23
10296 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i64 [[TMP30]], i64 [[TMP32]], i64 [[TMP34]]) #[[ATTR3]]
10297 // CHECK17-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
10298 // CHECK17:       .omp_outlined..1.exit:
10299 // CHECK17-NEXT:    ret i32 0
10300 //
10301 //
10302 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
10303 // CHECK17-SAME: (i64 [[A:%.*]]) #[[ATTR2]] {
10304 // CHECK17-NEXT:  entry:
10305 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
10306 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
10307 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
10308 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
10309 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10310 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[A_CASTED]] to i32*
10311 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
10312 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
10313 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]])
10314 // CHECK17-NEXT:    ret void
10315 //
10316 //
10317 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..2
10318 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]]) #[[ATTR2]] {
10319 // CHECK17-NEXT:  entry:
10320 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10321 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10322 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
10323 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10324 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10325 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
10326 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
10327 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10328 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
10329 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
10330 // CHECK17-NEXT:    ret void
10331 //
10332 //
10333 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
10334 // CHECK17-SAME: (i64 [[AA:%.*]]) #[[ATTR2]] {
10335 // CHECK17-NEXT:  entry:
10336 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
10337 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
10338 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
10339 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
10340 // CHECK17-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
10341 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
10342 // CHECK17-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
10343 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
10344 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP1]])
10345 // CHECK17-NEXT:    ret void
10346 //
10347 //
10348 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..3
10349 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
10350 // CHECK17-NEXT:  entry:
10351 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10352 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10353 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
10354 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10355 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10356 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
10357 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
10358 // CHECK17-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
10359 // CHECK17-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
10360 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
10361 // CHECK17-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
10362 // CHECK17-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
10363 // CHECK17-NEXT:    ret void
10364 //
10365 //
10366 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
10367 // CHECK17-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
10368 // CHECK17-NEXT:  entry:
10369 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
10370 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
10371 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
10372 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
10373 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
10374 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
10375 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
10376 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
10377 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10378 // CHECK17-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
10379 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
10380 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
10381 // CHECK17-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
10382 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
10383 // CHECK17-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
10384 // CHECK17-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
10385 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
10386 // CHECK17-NEXT:    ret void
10387 //
10388 //
10389 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..6
10390 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
10391 // CHECK17-NEXT:  entry:
10392 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10393 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10394 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
10395 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
10396 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10397 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10398 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
10399 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
10400 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
10401 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
10402 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10403 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
10404 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
10405 // CHECK17-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
10406 // CHECK17-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
10407 // CHECK17-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
10408 // CHECK17-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
10409 // CHECK17-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
10410 // CHECK17-NEXT:    ret void
10411 //
10412 //
10413 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
10414 // CHECK17-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
10415 // CHECK17-NEXT:  entry:
10416 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
10417 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
10418 // CHECK17-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
10419 // CHECK17-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
10420 // CHECK17-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
10421 // CHECK17-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
10422 // CHECK17-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
10423 // CHECK17-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
10424 // CHECK17-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
10425 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
10426 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
10427 // CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
10428 // CHECK17-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
10429 // CHECK17-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
10430 // CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
10431 // CHECK17-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
10432 // CHECK17-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
10433 // CHECK17-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
10434 // CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
10435 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
10436 // CHECK17-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
10437 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
10438 // CHECK17-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
10439 // CHECK17-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
10440 // CHECK17-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
10441 // CHECK17-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
10442 // CHECK17-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
10443 // CHECK17-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
10444 // CHECK17-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
10445 // CHECK17-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
10446 // CHECK17-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
10447 // CHECK17-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
10448 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
10449 // CHECK17-NEXT:    ret void
10450 //
10451 //
10452 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..9
10453 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
10454 // CHECK17-NEXT:  entry:
10455 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10456 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10457 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
10458 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
10459 // CHECK17-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
10460 // CHECK17-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
10461 // CHECK17-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
10462 // CHECK17-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
10463 // CHECK17-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
10464 // CHECK17-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
10465 // CHECK17-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
10466 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10467 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10468 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
10469 // CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
10470 // CHECK17-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
10471 // CHECK17-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
10472 // CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
10473 // CHECK17-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
10474 // CHECK17-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
10475 // CHECK17-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
10476 // CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
10477 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
10478 // CHECK17-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
10479 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
10480 // CHECK17-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
10481 // CHECK17-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
10482 // CHECK17-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
10483 // CHECK17-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
10484 // CHECK17-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
10485 // CHECK17-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
10486 // CHECK17-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
10487 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
10488 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
10489 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
10490 // CHECK17-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
10491 // CHECK17-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
10492 // CHECK17-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
10493 // CHECK17-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
10494 // CHECK17-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
10495 // CHECK17-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
10496 // CHECK17-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
10497 // CHECK17-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
10498 // CHECK17-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
10499 // CHECK17-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
10500 // CHECK17-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
10501 // CHECK17-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
10502 // CHECK17-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
10503 // CHECK17-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
10504 // CHECK17-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
10505 // CHECK17-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
10506 // CHECK17-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
10507 // CHECK17-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
10508 // CHECK17-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
10509 // CHECK17-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
10510 // CHECK17-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
10511 // CHECK17-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
10512 // CHECK17-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
10513 // CHECK17-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
10514 // CHECK17-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
10515 // CHECK17-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
10516 // CHECK17-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
10517 // CHECK17-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
10518 // CHECK17-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
10519 // CHECK17-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
10520 // CHECK17-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
10521 // CHECK17-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
10522 // CHECK17-NEXT:    ret void
10523 //
10524 //
10525 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
10526 // CHECK17-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
10527 // CHECK17-NEXT:  entry:
10528 // CHECK17-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
10529 // CHECK17-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
10530 // CHECK17-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
10531 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
10532 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10533 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
10534 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
10535 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
10536 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP1]])
10537 // CHECK17-NEXT:    ret void
10538 //
10539 //
10540 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..11
10541 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
10542 // CHECK17-NEXT:  entry:
10543 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10544 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10545 // CHECK17-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
10546 // CHECK17-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
10547 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10548 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10549 // CHECK17-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
10550 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
10551 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10552 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
10553 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
10554 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
10555 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i64 [[TMP1]])
10556 // CHECK17-NEXT:    ret void
10557 //
10558 //
10559 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..12
10560 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
10561 // CHECK17-NEXT:  entry:
10562 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10563 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10564 // CHECK17-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
10565 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10566 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10567 // CHECK17-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
10568 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
10569 // CHECK17-NEXT:    ret void
10570 //
10571 //
10572 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
10573 // CHECK17-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
10574 // CHECK17-NEXT:  entry:
10575 // CHECK17-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
10576 // CHECK17-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
10577 // CHECK17-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
10578 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
10579 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
10580 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
10581 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
10582 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
10583 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP1]])
10584 // CHECK17-NEXT:    ret void
10585 //
10586 //
10587 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..15
10588 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
10589 // CHECK17-NEXT:  entry:
10590 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10591 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10592 // CHECK17-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
10593 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10594 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10595 // CHECK17-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
10596 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
10597 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[CONV]])
10598 // CHECK17-NEXT:    ret void
10599 //
10600 //
10601 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..16
10602 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
10603 // CHECK17-NEXT:  entry:
10604 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10605 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10606 // CHECK17-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
10607 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10608 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10609 // CHECK17-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
10610 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
10611 // CHECK17-NEXT:    ret void
10612 //
10613 //
10614 // CHECK17-LABEL: define {{[^@]+}}@_Z6bazzzziPi
10615 // CHECK17-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
10616 // CHECK17-NEXT:  entry:
10617 // CHECK17-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
10618 // CHECK17-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
10619 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
10620 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
10621 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
10622 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
10623 // CHECK17-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
10624 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
10625 // CHECK17-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
10626 // CHECK17-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10627 // CHECK17-NEXT:    [[TMP3:%.*]] = bitcast i8** [[TMP2]] to i64*
10628 // CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP3]], align 8
10629 // CHECK17-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10630 // CHECK17-NEXT:    [[TMP5:%.*]] = bitcast i8** [[TMP4]] to i64*
10631 // CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP5]], align 8
10632 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
10633 // CHECK17-NEXT:    store i8* null, i8** [[TMP6]], align 8
10634 // CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10635 // CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10636 // CHECK17-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10637 // CHECK17-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
10638 // CHECK17-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10639 // CHECK17:       omp_offload.failed:
10640 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i64 [[TMP1]]) #[[ATTR3]]
10641 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10642 // CHECK17:       omp_offload.cont:
10643 // CHECK17-NEXT:    ret void
10644 //
10645 //
10646 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
10647 // CHECK17-SAME: (i64 [[VLA:%.*]]) #[[ATTR2]] {
10648 // CHECK17-NEXT:  entry:
10649 // CHECK17-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
10650 // CHECK17-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
10651 // CHECK17-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
10652 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i64 [[TMP0]])
10653 // CHECK17-NEXT:    ret void
10654 //
10655 //
10656 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..19
10657 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR2]] {
10658 // CHECK17-NEXT:  entry:
10659 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10660 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10661 // CHECK17-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
10662 // CHECK17-NEXT:    [[F:%.*]] = alloca i32*, align 8
10663 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10664 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10665 // CHECK17-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
10666 // CHECK17-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
10667 // CHECK17-NEXT:    ret void
10668 //
10669 //
10670 // CHECK17-LABEL: define {{[^@]+}}@_Z3bari
10671 // CHECK17-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
10672 // CHECK17-NEXT:  entry:
10673 // CHECK17-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
10674 // CHECK17-NEXT:    [[A:%.*]] = alloca i32, align 4
10675 // CHECK17-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
10676 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
10677 // CHECK17-NEXT:    store i32 0, i32* [[A]], align 4
10678 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
10679 // CHECK17-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
10680 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
10681 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
10682 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
10683 // CHECK17-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
10684 // CHECK17-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
10685 // CHECK17-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
10686 // CHECK17-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
10687 // CHECK17-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
10688 // CHECK17-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
10689 // CHECK17-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
10690 // CHECK17-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
10691 // CHECK17-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
10692 // CHECK17-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
10693 // CHECK17-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
10694 // CHECK17-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
10695 // CHECK17-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
10696 // CHECK17-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
10697 // CHECK17-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
10698 // CHECK17-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
10699 // CHECK17-NEXT:    ret i32 [[TMP8]]
10700 //
10701 //
10702 // CHECK17-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
10703 // CHECK17-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
10704 // CHECK17-NEXT:  entry:
10705 // CHECK17-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
10706 // CHECK17-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
10707 // CHECK17-NEXT:    [[B:%.*]] = alloca i32, align 4
10708 // CHECK17-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
10709 // CHECK17-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
10710 // CHECK17-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
10711 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 8
10712 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 8
10713 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 8
10714 // CHECK17-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 8
10715 // CHECK17-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
10716 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
10717 // CHECK17-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
10718 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
10719 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
10720 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
10721 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
10722 // CHECK17-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
10723 // CHECK17-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
10724 // CHECK17-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
10725 // CHECK17-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
10726 // CHECK17-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
10727 // CHECK17-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
10728 // CHECK17-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
10729 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_CASTED]] to i32*
10730 // CHECK17-NEXT:    store i32 [[TMP5]], i32* [[CONV]], align 4
10731 // CHECK17-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
10732 // CHECK17-NEXT:    [[TMP7:%.*]] = load i32, i32* [[N_ADDR]], align 4
10733 // CHECK17-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP7]], 60
10734 // CHECK17-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
10735 // CHECK17:       omp_if.then:
10736 // CHECK17-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
10737 // CHECK17-NEXT:    [[TMP8:%.*]] = mul nuw i64 2, [[TMP2]]
10738 // CHECK17-NEXT:    [[TMP9:%.*]] = mul nuw i64 [[TMP8]], 2
10739 // CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10740 // CHECK17-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
10741 // CHECK17-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 8
10742 // CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10743 // CHECK17-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
10744 // CHECK17-NEXT:    store double* [[A]], double** [[TMP13]], align 8
10745 // CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
10746 // CHECK17-NEXT:    store i64 8, i64* [[TMP14]], align 8
10747 // CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
10748 // CHECK17-NEXT:    store i8* null, i8** [[TMP15]], align 8
10749 // CHECK17-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
10750 // CHECK17-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
10751 // CHECK17-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
10752 // CHECK17-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
10753 // CHECK17-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i64*
10754 // CHECK17-NEXT:    store i64 [[TMP6]], i64* [[TMP19]], align 8
10755 // CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
10756 // CHECK17-NEXT:    store i64 4, i64* [[TMP20]], align 8
10757 // CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
10758 // CHECK17-NEXT:    store i8* null, i8** [[TMP21]], align 8
10759 // CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
10760 // CHECK17-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
10761 // CHECK17-NEXT:    store i64 2, i64* [[TMP23]], align 8
10762 // CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
10763 // CHECK17-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
10764 // CHECK17-NEXT:    store i64 2, i64* [[TMP25]], align 8
10765 // CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
10766 // CHECK17-NEXT:    store i64 8, i64* [[TMP26]], align 8
10767 // CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
10768 // CHECK17-NEXT:    store i8* null, i8** [[TMP27]], align 8
10769 // CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
10770 // CHECK17-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
10771 // CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP29]], align 8
10772 // CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
10773 // CHECK17-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i64*
10774 // CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP31]], align 8
10775 // CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
10776 // CHECK17-NEXT:    store i64 8, i64* [[TMP32]], align 8
10777 // CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
10778 // CHECK17-NEXT:    store i8* null, i8** [[TMP33]], align 8
10779 // CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
10780 // CHECK17-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
10781 // CHECK17-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 8
10782 // CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
10783 // CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
10784 // CHECK17-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 8
10785 // CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
10786 // CHECK17-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 8
10787 // CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 4
10788 // CHECK17-NEXT:    store i8* null, i8** [[TMP39]], align 8
10789 // CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10790 // CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10791 // CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
10792 // CHECK17-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10793 // CHECK17-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
10794 // CHECK17-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10795 // CHECK17:       omp_offload.failed:
10796 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
10797 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10798 // CHECK17:       omp_offload.cont:
10799 // CHECK17-NEXT:    br label [[OMP_IF_END:%.*]]
10800 // CHECK17:       omp_if.else:
10801 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
10802 // CHECK17-NEXT:    br label [[OMP_IF_END]]
10803 // CHECK17:       omp_if.end:
10804 // CHECK17-NEXT:    [[TMP45:%.*]] = mul nsw i64 1, [[TMP2]]
10805 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP45]]
10806 // CHECK17-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
10807 // CHECK17-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
10808 // CHECK17-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP46]] to i32
10809 // CHECK17-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
10810 // CHECK17-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP47]]
10811 // CHECK17-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
10812 // CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
10813 // CHECK17-NEXT:    ret i32 [[ADD4]]
10814 //
10815 //
10816 // CHECK17-LABEL: define {{[^@]+}}@_ZL7fstatici
10817 // CHECK17-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
10818 // CHECK17-NEXT:  entry:
10819 // CHECK17-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
10820 // CHECK17-NEXT:    [[A:%.*]] = alloca i32, align 4
10821 // CHECK17-NEXT:    [[AA:%.*]] = alloca i16, align 2
10822 // CHECK17-NEXT:    [[AAA:%.*]] = alloca i8, align 1
10823 // CHECK17-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
10824 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
10825 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
10826 // CHECK17-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
10827 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 8
10828 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 8
10829 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 8
10830 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
10831 // CHECK17-NEXT:    store i32 0, i32* [[A]], align 4
10832 // CHECK17-NEXT:    store i16 0, i16* [[AA]], align 2
10833 // CHECK17-NEXT:    store i8 0, i8* [[AAA]], align 1
10834 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
10835 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
10836 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
10837 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
10838 // CHECK17-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
10839 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
10840 // CHECK17-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
10841 // CHECK17-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
10842 // CHECK17-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
10843 // CHECK17-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
10844 // CHECK17-NEXT:    store i8 [[TMP4]], i8* [[CONV2]], align 1
10845 // CHECK17-NEXT:    [[TMP5:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
10846 // CHECK17-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
10847 // CHECK17-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
10848 // CHECK17-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
10849 // CHECK17:       omp_if.then:
10850 // CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10851 // CHECK17-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
10852 // CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
10853 // CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10854 // CHECK17-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
10855 // CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP10]], align 8
10856 // CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
10857 // CHECK17-NEXT:    store i8* null, i8** [[TMP11]], align 8
10858 // CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
10859 // CHECK17-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
10860 // CHECK17-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
10861 // CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
10862 // CHECK17-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
10863 // CHECK17-NEXT:    store i64 [[TMP3]], i64* [[TMP15]], align 8
10864 // CHECK17-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
10865 // CHECK17-NEXT:    store i8* null, i8** [[TMP16]], align 8
10866 // CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
10867 // CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
10868 // CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP18]], align 8
10869 // CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
10870 // CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
10871 // CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP20]], align 8
10872 // CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
10873 // CHECK17-NEXT:    store i8* null, i8** [[TMP21]], align 8
10874 // CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
10875 // CHECK17-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
10876 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 8
10877 // CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
10878 // CHECK17-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
10879 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 8
10880 // CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
10881 // CHECK17-NEXT:    store i8* null, i8** [[TMP26]], align 8
10882 // CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10883 // CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10884 // CHECK17-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10885 // CHECK17-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
10886 // CHECK17-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10887 // CHECK17:       omp_offload.failed:
10888 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
10889 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10890 // CHECK17:       omp_offload.cont:
10891 // CHECK17-NEXT:    br label [[OMP_IF_END:%.*]]
10892 // CHECK17:       omp_if.else:
10893 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
10894 // CHECK17-NEXT:    br label [[OMP_IF_END]]
10895 // CHECK17:       omp_if.end:
10896 // CHECK17-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
10897 // CHECK17-NEXT:    ret i32 [[TMP31]]
10898 //
10899 //
10900 // CHECK17-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
10901 // CHECK17-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
10902 // CHECK17-NEXT:  entry:
10903 // CHECK17-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
10904 // CHECK17-NEXT:    [[A:%.*]] = alloca i32, align 4
10905 // CHECK17-NEXT:    [[AA:%.*]] = alloca i16, align 2
10906 // CHECK17-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
10907 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
10908 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
10909 // CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
10910 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
10911 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
10912 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
10913 // CHECK17-NEXT:    store i32 0, i32* [[A]], align 4
10914 // CHECK17-NEXT:    store i16 0, i16* [[AA]], align 2
10915 // CHECK17-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
10916 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
10917 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
10918 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
10919 // CHECK17-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
10920 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
10921 // CHECK17-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
10922 // CHECK17-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
10923 // CHECK17-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
10924 // CHECK17-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
10925 // CHECK17-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
10926 // CHECK17:       omp_if.then:
10927 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10928 // CHECK17-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
10929 // CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP6]], align 8
10930 // CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10931 // CHECK17-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
10932 // CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
10933 // CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
10934 // CHECK17-NEXT:    store i8* null, i8** [[TMP9]], align 8
10935 // CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
10936 // CHECK17-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i64*
10937 // CHECK17-NEXT:    store i64 [[TMP3]], i64* [[TMP11]], align 8
10938 // CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
10939 // CHECK17-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
10940 // CHECK17-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
10941 // CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
10942 // CHECK17-NEXT:    store i8* null, i8** [[TMP14]], align 8
10943 // CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
10944 // CHECK17-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
10945 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 8
10946 // CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
10947 // CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
10948 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 8
10949 // CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
10950 // CHECK17-NEXT:    store i8* null, i8** [[TMP19]], align 8
10951 // CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10952 // CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10953 // CHECK17-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10954 // CHECK17-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
10955 // CHECK17-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10956 // CHECK17:       omp_offload.failed:
10957 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
10958 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10959 // CHECK17:       omp_offload.cont:
10960 // CHECK17-NEXT:    br label [[OMP_IF_END:%.*]]
10961 // CHECK17:       omp_if.else:
10962 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
10963 // CHECK17-NEXT:    br label [[OMP_IF_END]]
10964 // CHECK17:       omp_if.end:
10965 // CHECK17-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
10966 // CHECK17-NEXT:    ret i32 [[TMP24]]
10967 //
10968 //
10969 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
10970 // CHECK17-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
10971 // CHECK17-NEXT:  entry:
10972 // CHECK17-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
10973 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
10974 // CHECK17-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
10975 // CHECK17-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
10976 // CHECK17-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
10977 // CHECK17-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
10978 // CHECK17-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
10979 // CHECK17-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
10980 // CHECK17-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
10981 // CHECK17-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
10982 // CHECK17-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
10983 // CHECK17-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
10984 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
10985 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
10986 // CHECK17-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
10987 // CHECK17-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
10988 // CHECK17-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
10989 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
10990 // CHECK17-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
10991 // CHECK17-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
10992 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
10993 // CHECK17-NEXT:    ret void
10994 //
10995 //
10996 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..22
10997 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
10998 // CHECK17-NEXT:  entry:
10999 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11000 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11001 // CHECK17-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
11002 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
11003 // CHECK17-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
11004 // CHECK17-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
11005 // CHECK17-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
11006 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11007 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11008 // CHECK17-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
11009 // CHECK17-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
11010 // CHECK17-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
11011 // CHECK17-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
11012 // CHECK17-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
11013 // CHECK17-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
11014 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
11015 // CHECK17-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
11016 // CHECK17-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
11017 // CHECK17-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
11018 // CHECK17-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
11019 // CHECK17-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
11020 // CHECK17-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
11021 // CHECK17-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
11022 // CHECK17-NEXT:    store double [[ADD]], double* [[A]], align 8
11023 // CHECK17-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
11024 // CHECK17-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
11025 // CHECK17-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
11026 // CHECK17-NEXT:    store double [[INC]], double* [[A4]], align 8
11027 // CHECK17-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
11028 // CHECK17-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
11029 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
11030 // CHECK17-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
11031 // CHECK17-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
11032 // CHECK17-NEXT:    ret void
11033 //
11034 //
11035 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
11036 // CHECK17-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
11037 // CHECK17-NEXT:  entry:
11038 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11039 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11040 // CHECK17-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
11041 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
11042 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
11043 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
11044 // CHECK17-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
11045 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11046 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11047 // CHECK17-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
11048 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
11049 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11050 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11051 // CHECK17-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
11052 // CHECK17-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
11053 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
11054 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
11055 // CHECK17-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
11056 // CHECK17-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
11057 // CHECK17-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
11058 // CHECK17-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
11059 // CHECK17-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
11060 // CHECK17-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
11061 // CHECK17-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
11062 // CHECK17-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
11063 // CHECK17-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
11064 // CHECK17-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
11065 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
11066 // CHECK17-NEXT:    ret void
11067 //
11068 //
11069 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..24
11070 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
11071 // CHECK17-NEXT:  entry:
11072 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11073 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11074 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11075 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11076 // CHECK17-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
11077 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
11078 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11079 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11080 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11081 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11082 // CHECK17-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
11083 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
11084 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11085 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11086 // CHECK17-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
11087 // CHECK17-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
11088 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
11089 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
11090 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
11091 // CHECK17-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
11092 // CHECK17-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
11093 // CHECK17-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
11094 // CHECK17-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
11095 // CHECK17-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
11096 // CHECK17-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
11097 // CHECK17-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
11098 // CHECK17-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
11099 // CHECK17-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
11100 // CHECK17-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
11101 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
11102 // CHECK17-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
11103 // CHECK17-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
11104 // CHECK17-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
11105 // CHECK17-NEXT:    ret void
11106 //
11107 //
11108 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
11109 // CHECK17-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
11110 // CHECK17-NEXT:  entry:
11111 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11112 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11113 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
11114 // CHECK17-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
11115 // CHECK17-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
11116 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11117 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11118 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
11119 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11120 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11121 // CHECK17-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
11122 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
11123 // CHECK17-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
11124 // CHECK17-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
11125 // CHECK17-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
11126 // CHECK17-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
11127 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
11128 // CHECK17-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
11129 // CHECK17-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
11130 // CHECK17-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
11131 // CHECK17-NEXT:    ret void
11132 //
11133 //
11134 // CHECK17-LABEL: define {{[^@]+}}@.omp_outlined..27
11135 // CHECK17-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
11136 // CHECK17-NEXT:  entry:
11137 // CHECK17-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11138 // CHECK17-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11139 // CHECK17-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11140 // CHECK17-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11141 // CHECK17-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
11142 // CHECK17-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11143 // CHECK17-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11144 // CHECK17-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11145 // CHECK17-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11146 // CHECK17-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
11147 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11148 // CHECK17-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11149 // CHECK17-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
11150 // CHECK17-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
11151 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
11152 // CHECK17-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
11153 // CHECK17-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
11154 // CHECK17-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
11155 // CHECK17-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
11156 // CHECK17-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
11157 // CHECK17-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
11158 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
11159 // CHECK17-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
11160 // CHECK17-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
11161 // CHECK17-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
11162 // CHECK17-NEXT:    ret void
11163 //
11164 //
11165 // CHECK17-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
11166 // CHECK17-SAME: () #[[ATTR4]] {
11167 // CHECK17-NEXT:  entry:
11168 // CHECK17-NEXT:    call void @__tgt_register_requires(i64 1)
11169 // CHECK17-NEXT:    ret void
11170 //
11171 //
11172 // CHECK18-LABEL: define {{[^@]+}}@_Z3fooi
11173 // CHECK18-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
11174 // CHECK18-NEXT:  entry:
11175 // CHECK18-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
11176 // CHECK18-NEXT:    [[A:%.*]] = alloca i32, align 4
11177 // CHECK18-NEXT:    [[AA:%.*]] = alloca i16, align 2
11178 // CHECK18-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
11179 // CHECK18-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
11180 // CHECK18-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
11181 // CHECK18-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
11182 // CHECK18-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
11183 // CHECK18-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
11184 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
11185 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
11186 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
11187 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
11188 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR__CASTED4:%.*]] = alloca i64, align 8
11189 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
11190 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
11191 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
11192 // CHECK18-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
11193 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
11194 // CHECK18-NEXT:    [[AA_CASTED7:%.*]] = alloca i64, align 8
11195 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS9:%.*]] = alloca [1 x i8*], align 8
11196 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS10:%.*]] = alloca [1 x i8*], align 8
11197 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS11:%.*]] = alloca [1 x i8*], align 8
11198 // CHECK18-NEXT:    [[A_CASTED12:%.*]] = alloca i64, align 8
11199 // CHECK18-NEXT:    [[AA_CASTED14:%.*]] = alloca i64, align 8
11200 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS16:%.*]] = alloca [2 x i8*], align 8
11201 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
11202 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
11203 // CHECK18-NEXT:    [[A_CASTED21:%.*]] = alloca i64, align 8
11204 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [9 x i8*], align 8
11205 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [9 x i8*], align 8
11206 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [9 x i8*], align 8
11207 // CHECK18-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
11208 // CHECK18-NEXT:    [[NN:%.*]] = alloca i32, align 4
11209 // CHECK18-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
11210 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 8
11211 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 8
11212 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 8
11213 // CHECK18-NEXT:    [[NN_CASTED38:%.*]] = alloca i64, align 8
11214 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS40:%.*]] = alloca [1 x i8*], align 8
11215 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS41:%.*]] = alloca [1 x i8*], align 8
11216 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS42:%.*]] = alloca [1 x i8*], align 8
11217 // CHECK18-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
11218 // CHECK18-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
11219 // CHECK18-NEXT:    store i32 0, i32* [[A]], align 4
11220 // CHECK18-NEXT:    store i16 0, i16* [[AA]], align 2
11221 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
11222 // CHECK18-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
11223 // CHECK18-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
11224 // CHECK18-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
11225 // CHECK18-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP2]], align 4
11226 // CHECK18-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
11227 // CHECK18-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
11228 // CHECK18-NEXT:    [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
11229 // CHECK18-NEXT:    [[TMP6:%.*]] = mul nuw i64 5, [[TMP5]]
11230 // CHECK18-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP6]], align 8
11231 // CHECK18-NEXT:    store i64 [[TMP5]], i64* [[__VLA_EXPR1]], align 8
11232 // CHECK18-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
11233 // CHECK18-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_]], align 4
11234 // CHECK18-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
11235 // CHECK18-NEXT:    store i32 [[TMP8]], i32* [[DOTCAPTURE_EXPR_2]], align 4
11236 // CHECK18-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
11237 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
11238 // CHECK18-NEXT:    store i16 [[TMP9]], i16* [[CONV]], align 2
11239 // CHECK18-NEXT:    [[TMP10:%.*]] = load i64, i64* [[AA_CASTED]], align 8
11240 // CHECK18-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
11241 // CHECK18-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
11242 // CHECK18-NEXT:    store i32 [[TMP11]], i32* [[CONV3]], align 4
11243 // CHECK18-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
11244 // CHECK18-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
11245 // CHECK18-NEXT:    [[CONV5:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED4]] to i32*
11246 // CHECK18-NEXT:    store i32 [[TMP13]], i32* [[CONV5]], align 4
11247 // CHECK18-NEXT:    [[TMP14:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED4]], align 8
11248 // CHECK18-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
11249 // CHECK18-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
11250 // CHECK18-NEXT:    store i64 [[TMP10]], i64* [[TMP16]], align 8
11251 // CHECK18-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
11252 // CHECK18-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
11253 // CHECK18-NEXT:    store i64 [[TMP10]], i64* [[TMP18]], align 8
11254 // CHECK18-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
11255 // CHECK18-NEXT:    store i8* null, i8** [[TMP19]], align 8
11256 // CHECK18-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
11257 // CHECK18-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
11258 // CHECK18-NEXT:    store i64 [[TMP12]], i64* [[TMP21]], align 8
11259 // CHECK18-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
11260 // CHECK18-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
11261 // CHECK18-NEXT:    store i64 [[TMP12]], i64* [[TMP23]], align 8
11262 // CHECK18-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
11263 // CHECK18-NEXT:    store i8* null, i8** [[TMP24]], align 8
11264 // CHECK18-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
11265 // CHECK18-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i64*
11266 // CHECK18-NEXT:    store i64 [[TMP14]], i64* [[TMP26]], align 8
11267 // CHECK18-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
11268 // CHECK18-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
11269 // CHECK18-NEXT:    store i64 [[TMP14]], i64* [[TMP28]], align 8
11270 // CHECK18-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
11271 // CHECK18-NEXT:    store i8* null, i8** [[TMP29]], align 8
11272 // CHECK18-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
11273 // CHECK18-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
11274 // CHECK18-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
11275 // CHECK18-NEXT:    [[TMP33:%.*]] = load i16, i16* [[AA]], align 2
11276 // CHECK18-NEXT:    store i16 [[TMP33]], i16* [[TMP32]], align 4
11277 // CHECK18-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
11278 // CHECK18-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
11279 // CHECK18-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
11280 // CHECK18-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
11281 // CHECK18-NEXT:    [[TMP37:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
11282 // CHECK18-NEXT:    store i32 [[TMP37]], i32* [[TMP36]], align 4
11283 // CHECK18-NEXT:    [[TMP38:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
11284 // CHECK18-NEXT:    [[TMP39:%.*]] = bitcast i8* [[TMP38]] to %struct.kmp_task_t_with_privates*
11285 // CHECK18-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 0
11286 // CHECK18-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP40]], i32 0, i32 0
11287 // CHECK18-NEXT:    [[TMP42:%.*]] = load i8*, i8** [[TMP41]], align 8
11288 // CHECK18-NEXT:    [[TMP43:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
11289 // CHECK18-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP42]], i8* align 4 [[TMP43]], i64 12, i1 false)
11290 // CHECK18-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 1
11291 // CHECK18-NEXT:    [[TMP45:%.*]] = bitcast i8* [[TMP42]] to %struct.anon*
11292 // CHECK18-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 0
11293 // CHECK18-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
11294 // CHECK18-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP30]] to i8*
11295 // CHECK18-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP47]], i8* align 8 [[TMP48]], i64 24, i1 false)
11296 // CHECK18-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 1
11297 // CHECK18-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
11298 // CHECK18-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP31]] to i8*
11299 // CHECK18-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP50]], i8* align 8 [[TMP51]], i64 24, i1 false)
11300 // CHECK18-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 2
11301 // CHECK18-NEXT:    [[TMP53:%.*]] = bitcast [3 x i64]* [[TMP52]] to i8*
11302 // CHECK18-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP53]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
11303 // CHECK18-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 3
11304 // CHECK18-NEXT:    [[TMP55:%.*]] = load i16, i16* [[AA]], align 2
11305 // CHECK18-NEXT:    store i16 [[TMP55]], i16* [[TMP54]], align 8
11306 // CHECK18-NEXT:    [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP38]])
11307 // CHECK18-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
11308 // CHECK18-NEXT:    [[CONV6:%.*]] = bitcast i64* [[A_CASTED]] to i32*
11309 // CHECK18-NEXT:    store i32 [[TMP57]], i32* [[CONV6]], align 4
11310 // CHECK18-NEXT:    [[TMP58:%.*]] = load i64, i64* [[A_CASTED]], align 8
11311 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i64 [[TMP58]]) #[[ATTR3:[0-9]+]]
11312 // CHECK18-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
11313 // CHECK18-NEXT:    [[CONV8:%.*]] = bitcast i64* [[AA_CASTED7]] to i16*
11314 // CHECK18-NEXT:    store i16 [[TMP59]], i16* [[CONV8]], align 2
11315 // CHECK18-NEXT:    [[TMP60:%.*]] = load i64, i64* [[AA_CASTED7]], align 8
11316 // CHECK18-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
11317 // CHECK18-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i64*
11318 // CHECK18-NEXT:    store i64 [[TMP60]], i64* [[TMP62]], align 8
11319 // CHECK18-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
11320 // CHECK18-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i64*
11321 // CHECK18-NEXT:    store i64 [[TMP60]], i64* [[TMP64]], align 8
11322 // CHECK18-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
11323 // CHECK18-NEXT:    store i8* null, i8** [[TMP65]], align 8
11324 // CHECK18-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
11325 // CHECK18-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
11326 // CHECK18-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
11327 // CHECK18-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
11328 // CHECK18-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
11329 // CHECK18:       omp_offload.failed:
11330 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR3]]
11331 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT]]
11332 // CHECK18:       omp_offload.cont:
11333 // CHECK18-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
11334 // CHECK18-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
11335 // CHECK18-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
11336 // CHECK18-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
11337 // CHECK18-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
11338 // CHECK18-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
11339 // CHECK18-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
11340 // CHECK18-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
11341 // CHECK18-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
11342 // CHECK18-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
11343 // CHECK18-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
11344 // CHECK18:       omp_if.then:
11345 // CHECK18-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
11346 // CHECK18-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
11347 // CHECK18-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
11348 // CHECK18-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
11349 // CHECK18-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
11350 // CHECK18-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
11351 // CHECK18-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
11352 // CHECK18-NEXT:    store i8* null, i8** [[TMP79]], align 8
11353 // CHECK18-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
11354 // CHECK18-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
11355 // CHECK18-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
11356 // CHECK18-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
11357 // CHECK18-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
11358 // CHECK18-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
11359 // CHECK18-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
11360 // CHECK18-NEXT:    store i8* null, i8** [[TMP84]], align 8
11361 // CHECK18-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
11362 // CHECK18-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
11363 // CHECK18-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
11364 // CHECK18-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
11365 // CHECK18-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
11366 // CHECK18:       omp_offload.failed19:
11367 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
11368 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
11369 // CHECK18:       omp_offload.cont20:
11370 // CHECK18-NEXT:    br label [[OMP_IF_END:%.*]]
11371 // CHECK18:       omp_if.else:
11372 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
11373 // CHECK18-NEXT:    br label [[OMP_IF_END]]
11374 // CHECK18:       omp_if.end:
11375 // CHECK18-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
11376 // CHECK18-NEXT:    [[CONV22:%.*]] = bitcast i64* [[A_CASTED21]] to i32*
11377 // CHECK18-NEXT:    store i32 [[TMP89]], i32* [[CONV22]], align 4
11378 // CHECK18-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED21]], align 8
11379 // CHECK18-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
11380 // CHECK18-NEXT:    [[CMP23:%.*]] = icmp sgt i32 [[TMP91]], 20
11381 // CHECK18-NEXT:    br i1 [[CMP23]], label [[OMP_IF_THEN24:%.*]], label [[OMP_IF_ELSE30:%.*]]
11382 // CHECK18:       omp_if.then24:
11383 // CHECK18-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
11384 // CHECK18-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
11385 // CHECK18-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
11386 // CHECK18-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
11387 // CHECK18-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
11388 // CHECK18-NEXT:    store i64 [[TMP90]], i64* [[TMP96]], align 8
11389 // CHECK18-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
11390 // CHECK18-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
11391 // CHECK18-NEXT:    store i64 [[TMP90]], i64* [[TMP98]], align 8
11392 // CHECK18-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
11393 // CHECK18-NEXT:    store i64 4, i64* [[TMP99]], align 8
11394 // CHECK18-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
11395 // CHECK18-NEXT:    store i8* null, i8** [[TMP100]], align 8
11396 // CHECK18-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
11397 // CHECK18-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
11398 // CHECK18-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
11399 // CHECK18-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
11400 // CHECK18-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
11401 // CHECK18-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
11402 // CHECK18-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
11403 // CHECK18-NEXT:    store i64 40, i64* [[TMP105]], align 8
11404 // CHECK18-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
11405 // CHECK18-NEXT:    store i8* null, i8** [[TMP106]], align 8
11406 // CHECK18-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 2
11407 // CHECK18-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
11408 // CHECK18-NEXT:    store i64 [[TMP2]], i64* [[TMP108]], align 8
11409 // CHECK18-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 2
11410 // CHECK18-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
11411 // CHECK18-NEXT:    store i64 [[TMP2]], i64* [[TMP110]], align 8
11412 // CHECK18-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
11413 // CHECK18-NEXT:    store i64 8, i64* [[TMP111]], align 8
11414 // CHECK18-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 2
11415 // CHECK18-NEXT:    store i8* null, i8** [[TMP112]], align 8
11416 // CHECK18-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 3
11417 // CHECK18-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
11418 // CHECK18-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
11419 // CHECK18-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 3
11420 // CHECK18-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
11421 // CHECK18-NEXT:    store float* [[VLA]], float** [[TMP116]], align 8
11422 // CHECK18-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
11423 // CHECK18-NEXT:    store i64 [[TMP92]], i64* [[TMP117]], align 8
11424 // CHECK18-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 3
11425 // CHECK18-NEXT:    store i8* null, i8** [[TMP118]], align 8
11426 // CHECK18-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 4
11427 // CHECK18-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
11428 // CHECK18-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
11429 // CHECK18-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 4
11430 // CHECK18-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
11431 // CHECK18-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 8
11432 // CHECK18-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
11433 // CHECK18-NEXT:    store i64 400, i64* [[TMP123]], align 8
11434 // CHECK18-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 4
11435 // CHECK18-NEXT:    store i8* null, i8** [[TMP124]], align 8
11436 // CHECK18-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 5
11437 // CHECK18-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
11438 // CHECK18-NEXT:    store i64 5, i64* [[TMP126]], align 8
11439 // CHECK18-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 5
11440 // CHECK18-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
11441 // CHECK18-NEXT:    store i64 5, i64* [[TMP128]], align 8
11442 // CHECK18-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
11443 // CHECK18-NEXT:    store i64 8, i64* [[TMP129]], align 8
11444 // CHECK18-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 5
11445 // CHECK18-NEXT:    store i8* null, i8** [[TMP130]], align 8
11446 // CHECK18-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 6
11447 // CHECK18-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
11448 // CHECK18-NEXT:    store i64 [[TMP5]], i64* [[TMP132]], align 8
11449 // CHECK18-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 6
11450 // CHECK18-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
11451 // CHECK18-NEXT:    store i64 [[TMP5]], i64* [[TMP134]], align 8
11452 // CHECK18-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
11453 // CHECK18-NEXT:    store i64 8, i64* [[TMP135]], align 8
11454 // CHECK18-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 6
11455 // CHECK18-NEXT:    store i8* null, i8** [[TMP136]], align 8
11456 // CHECK18-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 7
11457 // CHECK18-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
11458 // CHECK18-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 8
11459 // CHECK18-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 7
11460 // CHECK18-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
11461 // CHECK18-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 8
11462 // CHECK18-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
11463 // CHECK18-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 8
11464 // CHECK18-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 7
11465 // CHECK18-NEXT:    store i8* null, i8** [[TMP142]], align 8
11466 // CHECK18-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 8
11467 // CHECK18-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
11468 // CHECK18-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 8
11469 // CHECK18-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 8
11470 // CHECK18-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
11471 // CHECK18-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 8
11472 // CHECK18-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
11473 // CHECK18-NEXT:    store i64 16, i64* [[TMP147]], align 8
11474 // CHECK18-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 8
11475 // CHECK18-NEXT:    store i8* null, i8** [[TMP148]], align 8
11476 // CHECK18-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
11477 // CHECK18-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
11478 // CHECK18-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
11479 // CHECK18-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
11480 // CHECK18-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
11481 // CHECK18-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
11482 // CHECK18:       omp_offload.failed28:
11483 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
11484 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
11485 // CHECK18:       omp_offload.cont29:
11486 // CHECK18-NEXT:    br label [[OMP_IF_END31:%.*]]
11487 // CHECK18:       omp_if.else30:
11488 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
11489 // CHECK18-NEXT:    br label [[OMP_IF_END31]]
11490 // CHECK18:       omp_if.end31:
11491 // CHECK18-NEXT:    store i32 0, i32* [[NN]], align 4
11492 // CHECK18-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
11493 // CHECK18-NEXT:    [[CONV32:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
11494 // CHECK18-NEXT:    store i32 [[TMP154]], i32* [[CONV32]], align 4
11495 // CHECK18-NEXT:    [[TMP155:%.*]] = load i64, i64* [[NN_CASTED]], align 8
11496 // CHECK18-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
11497 // CHECK18-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i64*
11498 // CHECK18-NEXT:    store i64 [[TMP155]], i64* [[TMP157]], align 8
11499 // CHECK18-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
11500 // CHECK18-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i64*
11501 // CHECK18-NEXT:    store i64 [[TMP155]], i64* [[TMP159]], align 8
11502 // CHECK18-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 0
11503 // CHECK18-NEXT:    store i8* null, i8** [[TMP160]], align 8
11504 // CHECK18-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
11505 // CHECK18-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
11506 // CHECK18-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
11507 // CHECK18-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
11508 // CHECK18-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
11509 // CHECK18:       omp_offload.failed36:
11510 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i64 [[TMP155]]) #[[ATTR3]]
11511 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
11512 // CHECK18:       omp_offload.cont37:
11513 // CHECK18-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
11514 // CHECK18-NEXT:    [[CONV39:%.*]] = bitcast i64* [[NN_CASTED38]] to i32*
11515 // CHECK18-NEXT:    store i32 [[TMP165]], i32* [[CONV39]], align 4
11516 // CHECK18-NEXT:    [[TMP166:%.*]] = load i64, i64* [[NN_CASTED38]], align 8
11517 // CHECK18-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
11518 // CHECK18-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i64*
11519 // CHECK18-NEXT:    store i64 [[TMP166]], i64* [[TMP168]], align 8
11520 // CHECK18-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
11521 // CHECK18-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
11522 // CHECK18-NEXT:    store i64 [[TMP166]], i64* [[TMP170]], align 8
11523 // CHECK18-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 0
11524 // CHECK18-NEXT:    store i8* null, i8** [[TMP171]], align 8
11525 // CHECK18-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
11526 // CHECK18-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
11527 // CHECK18-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
11528 // CHECK18-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
11529 // CHECK18-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
11530 // CHECK18:       omp_offload.failed43:
11531 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i64 [[TMP166]]) #[[ATTR3]]
11532 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
11533 // CHECK18:       omp_offload.cont44:
11534 // CHECK18-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
11535 // CHECK18-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
11536 // CHECK18-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
11537 // CHECK18-NEXT:    ret i32 [[TMP176]]
11538 //
11539 //
11540 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
11541 // CHECK18-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
11542 // CHECK18-NEXT:  entry:
11543 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11544 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
11545 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
11546 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
11547 // CHECK18-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
11548 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11549 // CHECK18-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
11550 // CHECK18-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
11551 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11552 // CHECK18-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
11553 // CHECK18-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
11554 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
11555 // CHECK18-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
11556 // CHECK18-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
11557 // CHECK18-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
11558 // CHECK18-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
11559 // CHECK18-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
11560 // CHECK18-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
11561 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
11562 // CHECK18-NEXT:    ret void
11563 //
11564 //
11565 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined.
11566 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
11567 // CHECK18-NEXT:  entry:
11568 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11569 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11570 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11571 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11572 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11573 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11574 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11575 // CHECK18-NEXT:    ret void
11576 //
11577 //
11578 // CHECK18-LABEL: define {{[^@]+}}@.omp_task_privates_map.
11579 // CHECK18-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
11580 // CHECK18-NEXT:  entry:
11581 // CHECK18-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
11582 // CHECK18-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 8
11583 // CHECK18-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 8
11584 // CHECK18-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 8
11585 // CHECK18-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 8
11586 // CHECK18-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
11587 // CHECK18-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 8
11588 // CHECK18-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 8
11589 // CHECK18-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 8
11590 // CHECK18-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 8
11591 // CHECK18-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
11592 // CHECK18-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
11593 // CHECK18-NEXT:    [[TMP7:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 8
11594 // CHECK18-NEXT:    store [3 x i8*]* [[TMP6]], [3 x i8*]** [[TMP7]], align 8
11595 // CHECK18-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
11596 // CHECK18-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 8
11597 // CHECK18-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 8
11598 // CHECK18-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
11599 // CHECK18-NEXT:    [[TMP11:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 8
11600 // CHECK18-NEXT:    store [3 x i64]* [[TMP10]], [3 x i64]** [[TMP11]], align 8
11601 // CHECK18-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
11602 // CHECK18-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 8
11603 // CHECK18-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 8
11604 // CHECK18-NEXT:    ret void
11605 //
11606 //
11607 // CHECK18-LABEL: define {{[^@]+}}@.omp_task_entry.
11608 // CHECK18-SAME: (i32 signext [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
11609 // CHECK18-NEXT:  entry:
11610 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
11611 // CHECK18-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
11612 // CHECK18-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
11613 // CHECK18-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
11614 // CHECK18-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
11615 // CHECK18-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
11616 // CHECK18-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 8
11617 // CHECK18-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 8
11618 // CHECK18-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 8
11619 // CHECK18-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 8
11620 // CHECK18-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
11621 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
11622 // CHECK18-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
11623 // CHECK18-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
11624 // CHECK18-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
11625 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
11626 // CHECK18-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
11627 // CHECK18-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
11628 // CHECK18-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
11629 // CHECK18-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
11630 // CHECK18-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
11631 // CHECK18-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
11632 // CHECK18-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
11633 // CHECK18-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
11634 // CHECK18-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
11635 // CHECK18-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
11636 // CHECK18-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
11637 // CHECK18-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META14:![0-9]+]])
11638 // CHECK18-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META17:![0-9]+]])
11639 // CHECK18-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META19:![0-9]+]])
11640 // CHECK18-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META21:![0-9]+]])
11641 // CHECK18-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !23
11642 // CHECK18-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !23
11643 // CHECK18-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
11644 // CHECK18-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
11645 // CHECK18-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !23
11646 // CHECK18-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
11647 // CHECK18-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !23
11648 // CHECK18-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !23
11649 // CHECK18-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !23
11650 // CHECK18-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
11651 // CHECK18-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
11652 // CHECK18-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !23
11653 // CHECK18-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 8, !noalias !23
11654 // CHECK18-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 8, !noalias !23
11655 // CHECK18-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 8, !noalias !23
11656 // CHECK18-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i64 0, i64 0
11657 // CHECK18-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i64 0, i64 0
11658 // CHECK18-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i64 0, i64 0
11659 // CHECK18-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
11660 // CHECK18-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
11661 // CHECK18-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
11662 // CHECK18-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
11663 // CHECK18-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
11664 // CHECK18-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
11665 // CHECK18-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
11666 // CHECK18:       omp_offload.failed.i:
11667 // CHECK18-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
11668 // CHECK18-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
11669 // CHECK18-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !23
11670 // CHECK18-NEXT:    [[TMP30:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !23
11671 // CHECK18-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
11672 // CHECK18-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
11673 // CHECK18-NEXT:    store i32 [[TMP31]], i32* [[CONV4_I]], align 4, !noalias !23
11674 // CHECK18-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !23
11675 // CHECK18-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
11676 // CHECK18-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
11677 // CHECK18-NEXT:    store i32 [[TMP33]], i32* [[CONV6_I]], align 4, !noalias !23
11678 // CHECK18-NEXT:    [[TMP34:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !23
11679 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i64 [[TMP30]], i64 [[TMP32]], i64 [[TMP34]]) #[[ATTR3]]
11680 // CHECK18-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
11681 // CHECK18:       .omp_outlined..1.exit:
11682 // CHECK18-NEXT:    ret i32 0
11683 //
11684 //
11685 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
11686 // CHECK18-SAME: (i64 [[A:%.*]]) #[[ATTR2]] {
11687 // CHECK18-NEXT:  entry:
11688 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11689 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
11690 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11691 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11692 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11693 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[A_CASTED]] to i32*
11694 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
11695 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
11696 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]])
11697 // CHECK18-NEXT:    ret void
11698 //
11699 //
11700 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..2
11701 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]]) #[[ATTR2]] {
11702 // CHECK18-NEXT:  entry:
11703 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11704 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11705 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11706 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11707 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11708 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11709 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11710 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11711 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
11712 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
11713 // CHECK18-NEXT:    ret void
11714 //
11715 //
11716 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
11717 // CHECK18-SAME: (i64 [[AA:%.*]]) #[[ATTR2]] {
11718 // CHECK18-NEXT:  entry:
11719 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11720 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
11721 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11722 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11723 // CHECK18-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
11724 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
11725 // CHECK18-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
11726 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
11727 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP1]])
11728 // CHECK18-NEXT:    ret void
11729 //
11730 //
11731 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..3
11732 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
11733 // CHECK18-NEXT:  entry:
11734 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11735 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11736 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11737 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11738 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11739 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11740 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11741 // CHECK18-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
11742 // CHECK18-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
11743 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
11744 // CHECK18-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
11745 // CHECK18-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
11746 // CHECK18-NEXT:    ret void
11747 //
11748 //
11749 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
11750 // CHECK18-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
11751 // CHECK18-NEXT:  entry:
11752 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11753 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11754 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
11755 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
11756 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11757 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11758 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11759 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11760 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11761 // CHECK18-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
11762 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
11763 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
11764 // CHECK18-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
11765 // CHECK18-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
11766 // CHECK18-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
11767 // CHECK18-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
11768 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
11769 // CHECK18-NEXT:    ret void
11770 //
11771 //
11772 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..6
11773 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR2]] {
11774 // CHECK18-NEXT:  entry:
11775 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11776 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11777 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11778 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
11779 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11780 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11781 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11782 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
11783 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11784 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
11785 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11786 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
11787 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
11788 // CHECK18-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
11789 // CHECK18-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
11790 // CHECK18-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
11791 // CHECK18-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
11792 // CHECK18-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
11793 // CHECK18-NEXT:    ret void
11794 //
11795 //
11796 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
11797 // CHECK18-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
11798 // CHECK18-NEXT:  entry:
11799 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11800 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
11801 // CHECK18-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
11802 // CHECK18-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
11803 // CHECK18-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
11804 // CHECK18-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
11805 // CHECK18-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
11806 // CHECK18-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
11807 // CHECK18-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
11808 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
11809 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11810 // CHECK18-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
11811 // CHECK18-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
11812 // CHECK18-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
11813 // CHECK18-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
11814 // CHECK18-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
11815 // CHECK18-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
11816 // CHECK18-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
11817 // CHECK18-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
11818 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11819 // CHECK18-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
11820 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
11821 // CHECK18-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
11822 // CHECK18-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
11823 // CHECK18-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
11824 // CHECK18-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
11825 // CHECK18-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
11826 // CHECK18-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
11827 // CHECK18-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
11828 // CHECK18-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
11829 // CHECK18-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
11830 // CHECK18-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
11831 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
11832 // CHECK18-NEXT:    ret void
11833 //
11834 //
11835 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..9
11836 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR2]] {
11837 // CHECK18-NEXT:  entry:
11838 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11839 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11840 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
11841 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
11842 // CHECK18-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
11843 // CHECK18-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
11844 // CHECK18-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
11845 // CHECK18-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
11846 // CHECK18-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
11847 // CHECK18-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
11848 // CHECK18-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
11849 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11850 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11851 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
11852 // CHECK18-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
11853 // CHECK18-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
11854 // CHECK18-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
11855 // CHECK18-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
11856 // CHECK18-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
11857 // CHECK18-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
11858 // CHECK18-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
11859 // CHECK18-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
11860 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
11861 // CHECK18-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
11862 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
11863 // CHECK18-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
11864 // CHECK18-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
11865 // CHECK18-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
11866 // CHECK18-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
11867 // CHECK18-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
11868 // CHECK18-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
11869 // CHECK18-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
11870 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
11871 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
11872 // CHECK18-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
11873 // CHECK18-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
11874 // CHECK18-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
11875 // CHECK18-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
11876 // CHECK18-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
11877 // CHECK18-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
11878 // CHECK18-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
11879 // CHECK18-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
11880 // CHECK18-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
11881 // CHECK18-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
11882 // CHECK18-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
11883 // CHECK18-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
11884 // CHECK18-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
11885 // CHECK18-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
11886 // CHECK18-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
11887 // CHECK18-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
11888 // CHECK18-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
11889 // CHECK18-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
11890 // CHECK18-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
11891 // CHECK18-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
11892 // CHECK18-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
11893 // CHECK18-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
11894 // CHECK18-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
11895 // CHECK18-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
11896 // CHECK18-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
11897 // CHECK18-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
11898 // CHECK18-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
11899 // CHECK18-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
11900 // CHECK18-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
11901 // CHECK18-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
11902 // CHECK18-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
11903 // CHECK18-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
11904 // CHECK18-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
11905 // CHECK18-NEXT:    ret void
11906 //
11907 //
11908 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
11909 // CHECK18-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
11910 // CHECK18-NEXT:  entry:
11911 // CHECK18-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
11912 // CHECK18-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
11913 // CHECK18-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
11914 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
11915 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11916 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
11917 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
11918 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
11919 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP1]])
11920 // CHECK18-NEXT:    ret void
11921 //
11922 //
11923 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..11
11924 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
11925 // CHECK18-NEXT:  entry:
11926 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11927 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11928 // CHECK18-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
11929 // CHECK18-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
11930 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11931 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11932 // CHECK18-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
11933 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
11934 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11935 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
11936 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
11937 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
11938 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i64 [[TMP1]])
11939 // CHECK18-NEXT:    ret void
11940 //
11941 //
11942 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..12
11943 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
11944 // CHECK18-NEXT:  entry:
11945 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11946 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11947 // CHECK18-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
11948 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11949 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11950 // CHECK18-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
11951 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
11952 // CHECK18-NEXT:    ret void
11953 //
11954 //
11955 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
11956 // CHECK18-SAME: (i64 [[NN:%.*]]) #[[ATTR2]] {
11957 // CHECK18-NEXT:  entry:
11958 // CHECK18-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
11959 // CHECK18-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
11960 // CHECK18-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
11961 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
11962 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
11963 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
11964 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
11965 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
11966 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP1]])
11967 // CHECK18-NEXT:    ret void
11968 //
11969 //
11970 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..15
11971 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR2]] {
11972 // CHECK18-NEXT:  entry:
11973 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11974 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11975 // CHECK18-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
11976 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11977 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11978 // CHECK18-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
11979 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
11980 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[CONV]])
11981 // CHECK18-NEXT:    ret void
11982 //
11983 //
11984 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..16
11985 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
11986 // CHECK18-NEXT:  entry:
11987 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11988 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11989 // CHECK18-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
11990 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11991 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11992 // CHECK18-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
11993 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
11994 // CHECK18-NEXT:    ret void
11995 //
11996 //
11997 // CHECK18-LABEL: define {{[^@]+}}@_Z6bazzzziPi
11998 // CHECK18-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
11999 // CHECK18-NEXT:  entry:
12000 // CHECK18-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
12001 // CHECK18-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
12002 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
12003 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
12004 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
12005 // CHECK18-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
12006 // CHECK18-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
12007 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
12008 // CHECK18-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
12009 // CHECK18-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12010 // CHECK18-NEXT:    [[TMP3:%.*]] = bitcast i8** [[TMP2]] to i64*
12011 // CHECK18-NEXT:    store i64 [[TMP1]], i64* [[TMP3]], align 8
12012 // CHECK18-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12013 // CHECK18-NEXT:    [[TMP5:%.*]] = bitcast i8** [[TMP4]] to i64*
12014 // CHECK18-NEXT:    store i64 [[TMP1]], i64* [[TMP5]], align 8
12015 // CHECK18-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
12016 // CHECK18-NEXT:    store i8* null, i8** [[TMP6]], align 8
12017 // CHECK18-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12018 // CHECK18-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12019 // CHECK18-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12020 // CHECK18-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
12021 // CHECK18-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
12022 // CHECK18:       omp_offload.failed:
12023 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i64 [[TMP1]]) #[[ATTR3]]
12024 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT]]
12025 // CHECK18:       omp_offload.cont:
12026 // CHECK18-NEXT:    ret void
12027 //
12028 //
12029 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
12030 // CHECK18-SAME: (i64 [[VLA:%.*]]) #[[ATTR2]] {
12031 // CHECK18-NEXT:  entry:
12032 // CHECK18-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
12033 // CHECK18-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
12034 // CHECK18-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
12035 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i64 [[TMP0]])
12036 // CHECK18-NEXT:    ret void
12037 //
12038 //
12039 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..19
12040 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR2]] {
12041 // CHECK18-NEXT:  entry:
12042 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
12043 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
12044 // CHECK18-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
12045 // CHECK18-NEXT:    [[F:%.*]] = alloca i32*, align 8
12046 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
12047 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
12048 // CHECK18-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
12049 // CHECK18-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
12050 // CHECK18-NEXT:    ret void
12051 //
12052 //
12053 // CHECK18-LABEL: define {{[^@]+}}@_Z3bari
12054 // CHECK18-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
12055 // CHECK18-NEXT:  entry:
12056 // CHECK18-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
12057 // CHECK18-NEXT:    [[A:%.*]] = alloca i32, align 4
12058 // CHECK18-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
12059 // CHECK18-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
12060 // CHECK18-NEXT:    store i32 0, i32* [[A]], align 4
12061 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
12062 // CHECK18-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
12063 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
12064 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
12065 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
12066 // CHECK18-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
12067 // CHECK18-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
12068 // CHECK18-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
12069 // CHECK18-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
12070 // CHECK18-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
12071 // CHECK18-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
12072 // CHECK18-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
12073 // CHECK18-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
12074 // CHECK18-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
12075 // CHECK18-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
12076 // CHECK18-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
12077 // CHECK18-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
12078 // CHECK18-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
12079 // CHECK18-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
12080 // CHECK18-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
12081 // CHECK18-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
12082 // CHECK18-NEXT:    ret i32 [[TMP8]]
12083 //
12084 //
12085 // CHECK18-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
12086 // CHECK18-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
12087 // CHECK18-NEXT:  entry:
12088 // CHECK18-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
12089 // CHECK18-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
12090 // CHECK18-NEXT:    [[B:%.*]] = alloca i32, align 4
12091 // CHECK18-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
12092 // CHECK18-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
12093 // CHECK18-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
12094 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 8
12095 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 8
12096 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 8
12097 // CHECK18-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 8
12098 // CHECK18-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
12099 // CHECK18-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
12100 // CHECK18-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
12101 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
12102 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
12103 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
12104 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
12105 // CHECK18-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
12106 // CHECK18-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
12107 // CHECK18-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
12108 // CHECK18-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
12109 // CHECK18-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
12110 // CHECK18-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
12111 // CHECK18-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
12112 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_CASTED]] to i32*
12113 // CHECK18-NEXT:    store i32 [[TMP5]], i32* [[CONV]], align 4
12114 // CHECK18-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
12115 // CHECK18-NEXT:    [[TMP7:%.*]] = load i32, i32* [[N_ADDR]], align 4
12116 // CHECK18-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP7]], 60
12117 // CHECK18-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
12118 // CHECK18:       omp_if.then:
12119 // CHECK18-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
12120 // CHECK18-NEXT:    [[TMP8:%.*]] = mul nuw i64 2, [[TMP2]]
12121 // CHECK18-NEXT:    [[TMP9:%.*]] = mul nuw i64 [[TMP8]], 2
12122 // CHECK18-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12123 // CHECK18-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
12124 // CHECK18-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 8
12125 // CHECK18-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12126 // CHECK18-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
12127 // CHECK18-NEXT:    store double* [[A]], double** [[TMP13]], align 8
12128 // CHECK18-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
12129 // CHECK18-NEXT:    store i64 8, i64* [[TMP14]], align 8
12130 // CHECK18-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
12131 // CHECK18-NEXT:    store i8* null, i8** [[TMP15]], align 8
12132 // CHECK18-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
12133 // CHECK18-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
12134 // CHECK18-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
12135 // CHECK18-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
12136 // CHECK18-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i64*
12137 // CHECK18-NEXT:    store i64 [[TMP6]], i64* [[TMP19]], align 8
12138 // CHECK18-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
12139 // CHECK18-NEXT:    store i64 4, i64* [[TMP20]], align 8
12140 // CHECK18-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
12141 // CHECK18-NEXT:    store i8* null, i8** [[TMP21]], align 8
12142 // CHECK18-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
12143 // CHECK18-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
12144 // CHECK18-NEXT:    store i64 2, i64* [[TMP23]], align 8
12145 // CHECK18-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
12146 // CHECK18-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
12147 // CHECK18-NEXT:    store i64 2, i64* [[TMP25]], align 8
12148 // CHECK18-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
12149 // CHECK18-NEXT:    store i64 8, i64* [[TMP26]], align 8
12150 // CHECK18-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
12151 // CHECK18-NEXT:    store i8* null, i8** [[TMP27]], align 8
12152 // CHECK18-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
12153 // CHECK18-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
12154 // CHECK18-NEXT:    store i64 [[TMP2]], i64* [[TMP29]], align 8
12155 // CHECK18-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
12156 // CHECK18-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i64*
12157 // CHECK18-NEXT:    store i64 [[TMP2]], i64* [[TMP31]], align 8
12158 // CHECK18-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
12159 // CHECK18-NEXT:    store i64 8, i64* [[TMP32]], align 8
12160 // CHECK18-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
12161 // CHECK18-NEXT:    store i8* null, i8** [[TMP33]], align 8
12162 // CHECK18-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
12163 // CHECK18-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
12164 // CHECK18-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 8
12165 // CHECK18-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
12166 // CHECK18-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
12167 // CHECK18-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 8
12168 // CHECK18-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
12169 // CHECK18-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 8
12170 // CHECK18-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 4
12171 // CHECK18-NEXT:    store i8* null, i8** [[TMP39]], align 8
12172 // CHECK18-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12173 // CHECK18-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12174 // CHECK18-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
12175 // CHECK18-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12176 // CHECK18-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
12177 // CHECK18-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
12178 // CHECK18:       omp_offload.failed:
12179 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
12180 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT]]
12181 // CHECK18:       omp_offload.cont:
12182 // CHECK18-NEXT:    br label [[OMP_IF_END:%.*]]
12183 // CHECK18:       omp_if.else:
12184 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
12185 // CHECK18-NEXT:    br label [[OMP_IF_END]]
12186 // CHECK18:       omp_if.end:
12187 // CHECK18-NEXT:    [[TMP45:%.*]] = mul nsw i64 1, [[TMP2]]
12188 // CHECK18-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP45]]
12189 // CHECK18-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
12190 // CHECK18-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
12191 // CHECK18-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP46]] to i32
12192 // CHECK18-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
12193 // CHECK18-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP47]]
12194 // CHECK18-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
12195 // CHECK18-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
12196 // CHECK18-NEXT:    ret i32 [[ADD4]]
12197 //
12198 //
12199 // CHECK18-LABEL: define {{[^@]+}}@_ZL7fstatici
12200 // CHECK18-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
12201 // CHECK18-NEXT:  entry:
12202 // CHECK18-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
12203 // CHECK18-NEXT:    [[A:%.*]] = alloca i32, align 4
12204 // CHECK18-NEXT:    [[AA:%.*]] = alloca i16, align 2
12205 // CHECK18-NEXT:    [[AAA:%.*]] = alloca i8, align 1
12206 // CHECK18-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
12207 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
12208 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
12209 // CHECK18-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
12210 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 8
12211 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 8
12212 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 8
12213 // CHECK18-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
12214 // CHECK18-NEXT:    store i32 0, i32* [[A]], align 4
12215 // CHECK18-NEXT:    store i16 0, i16* [[AA]], align 2
12216 // CHECK18-NEXT:    store i8 0, i8* [[AAA]], align 1
12217 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
12218 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
12219 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
12220 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
12221 // CHECK18-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
12222 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
12223 // CHECK18-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
12224 // CHECK18-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
12225 // CHECK18-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
12226 // CHECK18-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
12227 // CHECK18-NEXT:    store i8 [[TMP4]], i8* [[CONV2]], align 1
12228 // CHECK18-NEXT:    [[TMP5:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
12229 // CHECK18-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
12230 // CHECK18-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
12231 // CHECK18-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
12232 // CHECK18:       omp_if.then:
12233 // CHECK18-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12234 // CHECK18-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
12235 // CHECK18-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
12236 // CHECK18-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12237 // CHECK18-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
12238 // CHECK18-NEXT:    store i64 [[TMP1]], i64* [[TMP10]], align 8
12239 // CHECK18-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
12240 // CHECK18-NEXT:    store i8* null, i8** [[TMP11]], align 8
12241 // CHECK18-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
12242 // CHECK18-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
12243 // CHECK18-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
12244 // CHECK18-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
12245 // CHECK18-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
12246 // CHECK18-NEXT:    store i64 [[TMP3]], i64* [[TMP15]], align 8
12247 // CHECK18-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
12248 // CHECK18-NEXT:    store i8* null, i8** [[TMP16]], align 8
12249 // CHECK18-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
12250 // CHECK18-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
12251 // CHECK18-NEXT:    store i64 [[TMP5]], i64* [[TMP18]], align 8
12252 // CHECK18-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
12253 // CHECK18-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
12254 // CHECK18-NEXT:    store i64 [[TMP5]], i64* [[TMP20]], align 8
12255 // CHECK18-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
12256 // CHECK18-NEXT:    store i8* null, i8** [[TMP21]], align 8
12257 // CHECK18-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
12258 // CHECK18-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
12259 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 8
12260 // CHECK18-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
12261 // CHECK18-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
12262 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 8
12263 // CHECK18-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 3
12264 // CHECK18-NEXT:    store i8* null, i8** [[TMP26]], align 8
12265 // CHECK18-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12266 // CHECK18-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12267 // CHECK18-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12268 // CHECK18-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
12269 // CHECK18-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
12270 // CHECK18:       omp_offload.failed:
12271 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
12272 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT]]
12273 // CHECK18:       omp_offload.cont:
12274 // CHECK18-NEXT:    br label [[OMP_IF_END:%.*]]
12275 // CHECK18:       omp_if.else:
12276 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
12277 // CHECK18-NEXT:    br label [[OMP_IF_END]]
12278 // CHECK18:       omp_if.end:
12279 // CHECK18-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
12280 // CHECK18-NEXT:    ret i32 [[TMP31]]
12281 //
12282 //
12283 // CHECK18-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
12284 // CHECK18-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
12285 // CHECK18-NEXT:  entry:
12286 // CHECK18-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
12287 // CHECK18-NEXT:    [[A:%.*]] = alloca i32, align 4
12288 // CHECK18-NEXT:    [[AA:%.*]] = alloca i16, align 2
12289 // CHECK18-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
12290 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
12291 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
12292 // CHECK18-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 8
12293 // CHECK18-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 8
12294 // CHECK18-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
12295 // CHECK18-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
12296 // CHECK18-NEXT:    store i32 0, i32* [[A]], align 4
12297 // CHECK18-NEXT:    store i16 0, i16* [[AA]], align 2
12298 // CHECK18-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
12299 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
12300 // CHECK18-NEXT:    store i32 [[TMP0]], i32* [[CONV]], align 4
12301 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
12302 // CHECK18-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
12303 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
12304 // CHECK18-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
12305 // CHECK18-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
12306 // CHECK18-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
12307 // CHECK18-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
12308 // CHECK18-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
12309 // CHECK18:       omp_if.then:
12310 // CHECK18-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12311 // CHECK18-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
12312 // CHECK18-NEXT:    store i64 [[TMP1]], i64* [[TMP6]], align 8
12313 // CHECK18-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12314 // CHECK18-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
12315 // CHECK18-NEXT:    store i64 [[TMP1]], i64* [[TMP8]], align 8
12316 // CHECK18-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
12317 // CHECK18-NEXT:    store i8* null, i8** [[TMP9]], align 8
12318 // CHECK18-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
12319 // CHECK18-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i64*
12320 // CHECK18-NEXT:    store i64 [[TMP3]], i64* [[TMP11]], align 8
12321 // CHECK18-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
12322 // CHECK18-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i64*
12323 // CHECK18-NEXT:    store i64 [[TMP3]], i64* [[TMP13]], align 8
12324 // CHECK18-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
12325 // CHECK18-NEXT:    store i8* null, i8** [[TMP14]], align 8
12326 // CHECK18-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
12327 // CHECK18-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
12328 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 8
12329 // CHECK18-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
12330 // CHECK18-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
12331 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 8
12332 // CHECK18-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
12333 // CHECK18-NEXT:    store i8* null, i8** [[TMP19]], align 8
12334 // CHECK18-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12335 // CHECK18-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12336 // CHECK18-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12337 // CHECK18-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
12338 // CHECK18-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
12339 // CHECK18:       omp_offload.failed:
12340 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
12341 // CHECK18-NEXT:    br label [[OMP_OFFLOAD_CONT]]
12342 // CHECK18:       omp_offload.cont:
12343 // CHECK18-NEXT:    br label [[OMP_IF_END:%.*]]
12344 // CHECK18:       omp_if.else:
12345 // CHECK18-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
12346 // CHECK18-NEXT:    br label [[OMP_IF_END]]
12347 // CHECK18:       omp_if.end:
12348 // CHECK18-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
12349 // CHECK18-NEXT:    ret i32 [[TMP24]]
12350 //
12351 //
12352 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
12353 // CHECK18-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
12354 // CHECK18-NEXT:  entry:
12355 // CHECK18-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
12356 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
12357 // CHECK18-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
12358 // CHECK18-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
12359 // CHECK18-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
12360 // CHECK18-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
12361 // CHECK18-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
12362 // CHECK18-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
12363 // CHECK18-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
12364 // CHECK18-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
12365 // CHECK18-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
12366 // CHECK18-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
12367 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
12368 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
12369 // CHECK18-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
12370 // CHECK18-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
12371 // CHECK18-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
12372 // CHECK18-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
12373 // CHECK18-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
12374 // CHECK18-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
12375 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
12376 // CHECK18-NEXT:    ret void
12377 //
12378 //
12379 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..22
12380 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
12381 // CHECK18-NEXT:  entry:
12382 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
12383 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
12384 // CHECK18-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
12385 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
12386 // CHECK18-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
12387 // CHECK18-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
12388 // CHECK18-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
12389 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
12390 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
12391 // CHECK18-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
12392 // CHECK18-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
12393 // CHECK18-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
12394 // CHECK18-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
12395 // CHECK18-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
12396 // CHECK18-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
12397 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
12398 // CHECK18-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
12399 // CHECK18-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
12400 // CHECK18-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
12401 // CHECK18-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
12402 // CHECK18-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
12403 // CHECK18-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
12404 // CHECK18-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
12405 // CHECK18-NEXT:    store double [[ADD]], double* [[A]], align 8
12406 // CHECK18-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
12407 // CHECK18-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
12408 // CHECK18-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
12409 // CHECK18-NEXT:    store double [[INC]], double* [[A4]], align 8
12410 // CHECK18-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
12411 // CHECK18-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
12412 // CHECK18-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
12413 // CHECK18-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
12414 // CHECK18-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
12415 // CHECK18-NEXT:    ret void
12416 //
12417 //
12418 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
12419 // CHECK18-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
12420 // CHECK18-NEXT:  entry:
12421 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
12422 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
12423 // CHECK18-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
12424 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
12425 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
12426 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
12427 // CHECK18-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
12428 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
12429 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
12430 // CHECK18-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
12431 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
12432 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
12433 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
12434 // CHECK18-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
12435 // CHECK18-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
12436 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
12437 // CHECK18-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
12438 // CHECK18-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
12439 // CHECK18-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
12440 // CHECK18-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
12441 // CHECK18-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
12442 // CHECK18-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
12443 // CHECK18-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
12444 // CHECK18-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
12445 // CHECK18-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
12446 // CHECK18-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
12447 // CHECK18-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
12448 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
12449 // CHECK18-NEXT:    ret void
12450 //
12451 //
12452 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..24
12453 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
12454 // CHECK18-NEXT:  entry:
12455 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
12456 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
12457 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
12458 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
12459 // CHECK18-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
12460 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
12461 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
12462 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
12463 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
12464 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
12465 // CHECK18-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
12466 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
12467 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
12468 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
12469 // CHECK18-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
12470 // CHECK18-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
12471 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
12472 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
12473 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
12474 // CHECK18-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
12475 // CHECK18-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
12476 // CHECK18-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
12477 // CHECK18-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
12478 // CHECK18-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
12479 // CHECK18-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
12480 // CHECK18-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
12481 // CHECK18-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
12482 // CHECK18-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
12483 // CHECK18-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
12484 // CHECK18-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
12485 // CHECK18-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
12486 // CHECK18-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
12487 // CHECK18-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
12488 // CHECK18-NEXT:    ret void
12489 //
12490 //
12491 // CHECK18-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
12492 // CHECK18-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
12493 // CHECK18-NEXT:  entry:
12494 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
12495 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
12496 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
12497 // CHECK18-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
12498 // CHECK18-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
12499 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
12500 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
12501 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
12502 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
12503 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
12504 // CHECK18-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
12505 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
12506 // CHECK18-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
12507 // CHECK18-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
12508 // CHECK18-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
12509 // CHECK18-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
12510 // CHECK18-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
12511 // CHECK18-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
12512 // CHECK18-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
12513 // CHECK18-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
12514 // CHECK18-NEXT:    ret void
12515 //
12516 //
12517 // CHECK18-LABEL: define {{[^@]+}}@.omp_outlined..27
12518 // CHECK18-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
12519 // CHECK18-NEXT:  entry:
12520 // CHECK18-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
12521 // CHECK18-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
12522 // CHECK18-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
12523 // CHECK18-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
12524 // CHECK18-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
12525 // CHECK18-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
12526 // CHECK18-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
12527 // CHECK18-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
12528 // CHECK18-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
12529 // CHECK18-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
12530 // CHECK18-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
12531 // CHECK18-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
12532 // CHECK18-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
12533 // CHECK18-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
12534 // CHECK18-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
12535 // CHECK18-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
12536 // CHECK18-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
12537 // CHECK18-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
12538 // CHECK18-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
12539 // CHECK18-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
12540 // CHECK18-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
12541 // CHECK18-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
12542 // CHECK18-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
12543 // CHECK18-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
12544 // CHECK18-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
12545 // CHECK18-NEXT:    ret void
12546 //
12547 //
12548 // CHECK18-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
12549 // CHECK18-SAME: () #[[ATTR4]] {
12550 // CHECK18-NEXT:  entry:
12551 // CHECK18-NEXT:    call void @__tgt_register_requires(i64 1)
12552 // CHECK18-NEXT:    ret void
12553 //
12554 //
12555 // CHECK19-LABEL: define {{[^@]+}}@_Z3fooi
12556 // CHECK19-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
12557 // CHECK19-NEXT:  entry:
12558 // CHECK19-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
12559 // CHECK19-NEXT:    [[A:%.*]] = alloca i32, align 4
12560 // CHECK19-NEXT:    [[AA:%.*]] = alloca i16, align 2
12561 // CHECK19-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
12562 // CHECK19-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
12563 // CHECK19-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
12564 // CHECK19-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
12565 // CHECK19-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
12566 // CHECK19-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
12567 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
12568 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
12569 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
12570 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
12571 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED3:%.*]] = alloca i32, align 4
12572 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
12573 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
12574 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
12575 // CHECK19-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
12576 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
12577 // CHECK19-NEXT:    [[AA_CASTED4:%.*]] = alloca i32, align 4
12578 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS6:%.*]] = alloca [1 x i8*], align 4
12579 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [1 x i8*], align 4
12580 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [1 x i8*], align 4
12581 // CHECK19-NEXT:    [[A_CASTED9:%.*]] = alloca i32, align 4
12582 // CHECK19-NEXT:    [[AA_CASTED10:%.*]] = alloca i32, align 4
12583 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS12:%.*]] = alloca [2 x i8*], align 4
12584 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
12585 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
12586 // CHECK19-NEXT:    [[A_CASTED17:%.*]] = alloca i32, align 4
12587 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [9 x i8*], align 4
12588 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [9 x i8*], align 4
12589 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [9 x i8*], align 4
12590 // CHECK19-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
12591 // CHECK19-NEXT:    [[NN:%.*]] = alloca i32, align 4
12592 // CHECK19-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
12593 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
12594 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
12595 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
12596 // CHECK19-NEXT:    [[NN_CASTED32:%.*]] = alloca i32, align 4
12597 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 4
12598 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 4
12599 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 4
12600 // CHECK19-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
12601 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
12602 // CHECK19-NEXT:    store i32 0, i32* [[A]], align 4
12603 // CHECK19-NEXT:    store i16 0, i16* [[AA]], align 2
12604 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
12605 // CHECK19-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
12606 // CHECK19-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
12607 // CHECK19-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP1]], align 4
12608 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
12609 // CHECK19-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
12610 // CHECK19-NEXT:    [[TMP4:%.*]] = mul nuw i32 5, [[TMP3]]
12611 // CHECK19-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP4]], align 8
12612 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[__VLA_EXPR1]], align 4
12613 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
12614 // CHECK19-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4
12615 // CHECK19-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
12616 // CHECK19-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_2]], align 4
12617 // CHECK19-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
12618 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
12619 // CHECK19-NEXT:    store i16 [[TMP7]], i16* [[CONV]], align 2
12620 // CHECK19-NEXT:    [[TMP8:%.*]] = load i32, i32* [[AA_CASTED]], align 4
12621 // CHECK19-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
12622 // CHECK19-NEXT:    store i32 [[TMP9]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
12623 // CHECK19-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
12624 // CHECK19-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
12625 // CHECK19-NEXT:    store i32 [[TMP11]], i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
12626 // CHECK19-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
12627 // CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12628 // CHECK19-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i32*
12629 // CHECK19-NEXT:    store i32 [[TMP8]], i32* [[TMP14]], align 4
12630 // CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12631 // CHECK19-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
12632 // CHECK19-NEXT:    store i32 [[TMP8]], i32* [[TMP16]], align 4
12633 // CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
12634 // CHECK19-NEXT:    store i8* null, i8** [[TMP17]], align 4
12635 // CHECK19-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
12636 // CHECK19-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
12637 // CHECK19-NEXT:    store i32 [[TMP10]], i32* [[TMP19]], align 4
12638 // CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
12639 // CHECK19-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
12640 // CHECK19-NEXT:    store i32 [[TMP10]], i32* [[TMP21]], align 4
12641 // CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
12642 // CHECK19-NEXT:    store i8* null, i8** [[TMP22]], align 4
12643 // CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
12644 // CHECK19-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
12645 // CHECK19-NEXT:    store i32 [[TMP12]], i32* [[TMP24]], align 4
12646 // CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
12647 // CHECK19-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
12648 // CHECK19-NEXT:    store i32 [[TMP12]], i32* [[TMP26]], align 4
12649 // CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
12650 // CHECK19-NEXT:    store i8* null, i8** [[TMP27]], align 4
12651 // CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
12652 // CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
12653 // CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
12654 // CHECK19-NEXT:    [[TMP31:%.*]] = load i16, i16* [[AA]], align 2
12655 // CHECK19-NEXT:    store i16 [[TMP31]], i16* [[TMP30]], align 4
12656 // CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
12657 // CHECK19-NEXT:    [[TMP33:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
12658 // CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP32]], align 4
12659 // CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
12660 // CHECK19-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
12661 // CHECK19-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
12662 // CHECK19-NEXT:    [[TMP36:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
12663 // CHECK19-NEXT:    [[TMP37:%.*]] = bitcast i8* [[TMP36]] to %struct.kmp_task_t_with_privates*
12664 // CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 0
12665 // CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP38]], i32 0, i32 0
12666 // CHECK19-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[TMP39]], align 4
12667 // CHECK19-NEXT:    [[TMP41:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
12668 // CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP40]], i8* align 4 [[TMP41]], i32 12, i1 false)
12669 // CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 1
12670 // CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8* [[TMP40]] to %struct.anon*
12671 // CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 0
12672 // CHECK19-NEXT:    [[TMP45:%.*]] = bitcast [3 x i64]* [[TMP44]] to i8*
12673 // CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP45]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
12674 // CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 1
12675 // CHECK19-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
12676 // CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP28]] to i8*
12677 // CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP47]], i8* align 4 [[TMP48]], i32 12, i1 false)
12678 // CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 2
12679 // CHECK19-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
12680 // CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP29]] to i8*
12681 // CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i32 12, i1 false)
12682 // CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 3
12683 // CHECK19-NEXT:    [[TMP53:%.*]] = load i16, i16* [[AA]], align 2
12684 // CHECK19-NEXT:    store i16 [[TMP53]], i16* [[TMP52]], align 4
12685 // CHECK19-NEXT:    [[TMP54:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP36]])
12686 // CHECK19-NEXT:    [[TMP55:%.*]] = load i32, i32* [[A]], align 4
12687 // CHECK19-NEXT:    store i32 [[TMP55]], i32* [[A_CASTED]], align 4
12688 // CHECK19-NEXT:    [[TMP56:%.*]] = load i32, i32* [[A_CASTED]], align 4
12689 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i32 [[TMP56]]) #[[ATTR3:[0-9]+]]
12690 // CHECK19-NEXT:    [[TMP57:%.*]] = load i16, i16* [[AA]], align 2
12691 // CHECK19-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
12692 // CHECK19-NEXT:    store i16 [[TMP57]], i16* [[CONV5]], align 2
12693 // CHECK19-NEXT:    [[TMP58:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
12694 // CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
12695 // CHECK19-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32*
12696 // CHECK19-NEXT:    store i32 [[TMP58]], i32* [[TMP60]], align 4
12697 // CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
12698 // CHECK19-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32*
12699 // CHECK19-NEXT:    store i32 [[TMP58]], i32* [[TMP62]], align 4
12700 // CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
12701 // CHECK19-NEXT:    store i8* null, i8** [[TMP63]], align 4
12702 // CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
12703 // CHECK19-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
12704 // CHECK19-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12705 // CHECK19-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
12706 // CHECK19-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
12707 // CHECK19:       omp_offload.failed:
12708 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR3]]
12709 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
12710 // CHECK19:       omp_offload.cont:
12711 // CHECK19-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
12712 // CHECK19-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
12713 // CHECK19-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
12714 // CHECK19-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
12715 // CHECK19-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
12716 // CHECK19-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
12717 // CHECK19-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
12718 // CHECK19-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
12719 // CHECK19-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
12720 // CHECK19-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
12721 // CHECK19:       omp_if.then:
12722 // CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
12723 // CHECK19-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
12724 // CHECK19-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
12725 // CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
12726 // CHECK19-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
12727 // CHECK19-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
12728 // CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
12729 // CHECK19-NEXT:    store i8* null, i8** [[TMP77]], align 4
12730 // CHECK19-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
12731 // CHECK19-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
12732 // CHECK19-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
12733 // CHECK19-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
12734 // CHECK19-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
12735 // CHECK19-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
12736 // CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
12737 // CHECK19-NEXT:    store i8* null, i8** [[TMP82]], align 4
12738 // CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
12739 // CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
12740 // CHECK19-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12741 // CHECK19-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
12742 // CHECK19-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
12743 // CHECK19:       omp_offload.failed15:
12744 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
12745 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
12746 // CHECK19:       omp_offload.cont16:
12747 // CHECK19-NEXT:    br label [[OMP_IF_END:%.*]]
12748 // CHECK19:       omp_if.else:
12749 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
12750 // CHECK19-NEXT:    br label [[OMP_IF_END]]
12751 // CHECK19:       omp_if.end:
12752 // CHECK19-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
12753 // CHECK19-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED17]], align 4
12754 // CHECK19-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED17]], align 4
12755 // CHECK19-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
12756 // CHECK19-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP89]], 20
12757 // CHECK19-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
12758 // CHECK19:       omp_if.then19:
12759 // CHECK19-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
12760 // CHECK19-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
12761 // CHECK19-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
12762 // CHECK19-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
12763 // CHECK19-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
12764 // CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
12765 // CHECK19-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
12766 // CHECK19-NEXT:    store i32 [[TMP88]], i32* [[TMP96]], align 4
12767 // CHECK19-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
12768 // CHECK19-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
12769 // CHECK19-NEXT:    store i32 [[TMP88]], i32* [[TMP98]], align 4
12770 // CHECK19-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
12771 // CHECK19-NEXT:    store i64 4, i64* [[TMP99]], align 4
12772 // CHECK19-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
12773 // CHECK19-NEXT:    store i8* null, i8** [[TMP100]], align 4
12774 // CHECK19-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
12775 // CHECK19-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
12776 // CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
12777 // CHECK19-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
12778 // CHECK19-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
12779 // CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
12780 // CHECK19-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
12781 // CHECK19-NEXT:    store i64 40, i64* [[TMP105]], align 4
12782 // CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
12783 // CHECK19-NEXT:    store i8* null, i8** [[TMP106]], align 4
12784 // CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
12785 // CHECK19-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
12786 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP108]], align 4
12787 // CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
12788 // CHECK19-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
12789 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP110]], align 4
12790 // CHECK19-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
12791 // CHECK19-NEXT:    store i64 4, i64* [[TMP111]], align 4
12792 // CHECK19-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
12793 // CHECK19-NEXT:    store i8* null, i8** [[TMP112]], align 4
12794 // CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
12795 // CHECK19-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
12796 // CHECK19-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
12797 // CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
12798 // CHECK19-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
12799 // CHECK19-NEXT:    store float* [[VLA]], float** [[TMP116]], align 4
12800 // CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
12801 // CHECK19-NEXT:    store i64 [[TMP91]], i64* [[TMP117]], align 4
12802 // CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
12803 // CHECK19-NEXT:    store i8* null, i8** [[TMP118]], align 4
12804 // CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
12805 // CHECK19-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
12806 // CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
12807 // CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
12808 // CHECK19-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
12809 // CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 4
12810 // CHECK19-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
12811 // CHECK19-NEXT:    store i64 400, i64* [[TMP123]], align 4
12812 // CHECK19-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 4
12813 // CHECK19-NEXT:    store i8* null, i8** [[TMP124]], align 4
12814 // CHECK19-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
12815 // CHECK19-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i32*
12816 // CHECK19-NEXT:    store i32 5, i32* [[TMP126]], align 4
12817 // CHECK19-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
12818 // CHECK19-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
12819 // CHECK19-NEXT:    store i32 5, i32* [[TMP128]], align 4
12820 // CHECK19-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
12821 // CHECK19-NEXT:    store i64 4, i64* [[TMP129]], align 4
12822 // CHECK19-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 5
12823 // CHECK19-NEXT:    store i8* null, i8** [[TMP130]], align 4
12824 // CHECK19-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
12825 // CHECK19-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
12826 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP132]], align 4
12827 // CHECK19-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
12828 // CHECK19-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
12829 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP134]], align 4
12830 // CHECK19-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
12831 // CHECK19-NEXT:    store i64 4, i64* [[TMP135]], align 4
12832 // CHECK19-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 6
12833 // CHECK19-NEXT:    store i8* null, i8** [[TMP136]], align 4
12834 // CHECK19-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
12835 // CHECK19-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
12836 // CHECK19-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 4
12837 // CHECK19-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
12838 // CHECK19-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
12839 // CHECK19-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 4
12840 // CHECK19-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
12841 // CHECK19-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 4
12842 // CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 7
12843 // CHECK19-NEXT:    store i8* null, i8** [[TMP142]], align 4
12844 // CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
12845 // CHECK19-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
12846 // CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 4
12847 // CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
12848 // CHECK19-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
12849 // CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 4
12850 // CHECK19-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
12851 // CHECK19-NEXT:    store i64 12, i64* [[TMP147]], align 4
12852 // CHECK19-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 8
12853 // CHECK19-NEXT:    store i8* null, i8** [[TMP148]], align 4
12854 // CHECK19-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
12855 // CHECK19-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
12856 // CHECK19-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
12857 // CHECK19-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12858 // CHECK19-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
12859 // CHECK19-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
12860 // CHECK19:       omp_offload.failed23:
12861 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
12862 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
12863 // CHECK19:       omp_offload.cont24:
12864 // CHECK19-NEXT:    br label [[OMP_IF_END26:%.*]]
12865 // CHECK19:       omp_if.else25:
12866 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
12867 // CHECK19-NEXT:    br label [[OMP_IF_END26]]
12868 // CHECK19:       omp_if.end26:
12869 // CHECK19-NEXT:    store i32 0, i32* [[NN]], align 4
12870 // CHECK19-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
12871 // CHECK19-NEXT:    store i32 [[TMP154]], i32* [[NN_CASTED]], align 4
12872 // CHECK19-NEXT:    [[TMP155:%.*]] = load i32, i32* [[NN_CASTED]], align 4
12873 // CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
12874 // CHECK19-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
12875 // CHECK19-NEXT:    store i32 [[TMP155]], i32* [[TMP157]], align 4
12876 // CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
12877 // CHECK19-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32*
12878 // CHECK19-NEXT:    store i32 [[TMP155]], i32* [[TMP159]], align 4
12879 // CHECK19-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
12880 // CHECK19-NEXT:    store i8* null, i8** [[TMP160]], align 4
12881 // CHECK19-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
12882 // CHECK19-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
12883 // CHECK19-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12884 // CHECK19-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
12885 // CHECK19-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
12886 // CHECK19:       omp_offload.failed30:
12887 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i32 [[TMP155]]) #[[ATTR3]]
12888 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
12889 // CHECK19:       omp_offload.cont31:
12890 // CHECK19-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
12891 // CHECK19-NEXT:    store i32 [[TMP165]], i32* [[NN_CASTED32]], align 4
12892 // CHECK19-NEXT:    [[TMP166:%.*]] = load i32, i32* [[NN_CASTED32]], align 4
12893 // CHECK19-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
12894 // CHECK19-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i32*
12895 // CHECK19-NEXT:    store i32 [[TMP166]], i32* [[TMP168]], align 4
12896 // CHECK19-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
12897 // CHECK19-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i32*
12898 // CHECK19-NEXT:    store i32 [[TMP166]], i32* [[TMP170]], align 4
12899 // CHECK19-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
12900 // CHECK19-NEXT:    store i8* null, i8** [[TMP171]], align 4
12901 // CHECK19-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
12902 // CHECK19-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
12903 // CHECK19-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
12904 // CHECK19-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
12905 // CHECK19-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
12906 // CHECK19:       omp_offload.failed36:
12907 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i32 [[TMP166]]) #[[ATTR3]]
12908 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
12909 // CHECK19:       omp_offload.cont37:
12910 // CHECK19-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
12911 // CHECK19-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
12912 // CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
12913 // CHECK19-NEXT:    ret i32 [[TMP176]]
12914 //
12915 //
12916 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
12917 // CHECK19-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
12918 // CHECK19-NEXT:  entry:
12919 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
12920 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
12921 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
12922 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
12923 // CHECK19-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
12924 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
12925 // CHECK19-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
12926 // CHECK19-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
12927 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
12928 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
12929 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
12930 // CHECK19-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
12931 // CHECK19-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
12932 // CHECK19-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
12933 // CHECK19-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
12934 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
12935 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
12936 // CHECK19-NEXT:    ret void
12937 //
12938 //
12939 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined.
12940 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
12941 // CHECK19-NEXT:  entry:
12942 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
12943 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
12944 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
12945 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
12946 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
12947 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
12948 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
12949 // CHECK19-NEXT:    ret void
12950 //
12951 //
12952 // CHECK19-LABEL: define {{[^@]+}}@.omp_task_privates_map.
12953 // CHECK19-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
12954 // CHECK19-NEXT:  entry:
12955 // CHECK19-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 4
12956 // CHECK19-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 4
12957 // CHECK19-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 4
12958 // CHECK19-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 4
12959 // CHECK19-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 4
12960 // CHECK19-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 4
12961 // CHECK19-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 4
12962 // CHECK19-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 4
12963 // CHECK19-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 4
12964 // CHECK19-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 4
12965 // CHECK19-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 4
12966 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
12967 // CHECK19-NEXT:    [[TMP7:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 4
12968 // CHECK19-NEXT:    store [3 x i64]* [[TMP6]], [3 x i64]** [[TMP7]], align 4
12969 // CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
12970 // CHECK19-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 4
12971 // CHECK19-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 4
12972 // CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
12973 // CHECK19-NEXT:    [[TMP11:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 4
12974 // CHECK19-NEXT:    store [3 x i8*]* [[TMP10]], [3 x i8*]** [[TMP11]], align 4
12975 // CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
12976 // CHECK19-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 4
12977 // CHECK19-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 4
12978 // CHECK19-NEXT:    ret void
12979 //
12980 //
12981 // CHECK19-LABEL: define {{[^@]+}}@.omp_task_entry.
12982 // CHECK19-SAME: (i32 [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
12983 // CHECK19-NEXT:  entry:
12984 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
12985 // CHECK19-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 4
12986 // CHECK19-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 4
12987 // CHECK19-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
12988 // CHECK19-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
12989 // CHECK19-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
12990 // CHECK19-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 4
12991 // CHECK19-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 4
12992 // CHECK19-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 4
12993 // CHECK19-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 4
12994 // CHECK19-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
12995 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
12996 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
12997 // CHECK19-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
12998 // CHECK19-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
12999 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
13000 // CHECK19-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
13001 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
13002 // CHECK19-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
13003 // CHECK19-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
13004 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
13005 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
13006 // CHECK19-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 4
13007 // CHECK19-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
13008 // CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
13009 // CHECK19-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
13010 // CHECK19-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
13011 // CHECK19-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META15:![0-9]+]])
13012 // CHECK19-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
13013 // CHECK19-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
13014 // CHECK19-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META22:![0-9]+]])
13015 // CHECK19-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !24
13016 // CHECK19-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 4, !noalias !24
13017 // CHECK19-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
13018 // CHECK19-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
13019 // CHECK19-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !24
13020 // CHECK19-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
13021 // CHECK19-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
13022 // CHECK19-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
13023 // CHECK19-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
13024 // CHECK19-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
13025 // CHECK19-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
13026 // CHECK19-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 4, !noalias !24
13027 // CHECK19-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 4, !noalias !24
13028 // CHECK19-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 4, !noalias !24
13029 // CHECK19-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 4, !noalias !24
13030 // CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i32 0, i32 0
13031 // CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i32 0, i32 0
13032 // CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i32 0, i32 0
13033 // CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
13034 // CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
13035 // CHECK19-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
13036 // CHECK19-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
13037 // CHECK19-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
13038 // CHECK19-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
13039 // CHECK19-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
13040 // CHECK19:       omp_offload.failed.i:
13041 // CHECK19-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
13042 // CHECK19-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
13043 // CHECK19-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !24
13044 // CHECK19-NEXT:    [[TMP30:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !24
13045 // CHECK19-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
13046 // CHECK19-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
13047 // CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
13048 // CHECK19-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
13049 // CHECK19-NEXT:    store i32 [[TMP33]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
13050 // CHECK19-NEXT:    [[TMP34:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
13051 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i32 [[TMP30]], i32 [[TMP32]], i32 [[TMP34]]) #[[ATTR3]]
13052 // CHECK19-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
13053 // CHECK19:       .omp_outlined..1.exit:
13054 // CHECK19-NEXT:    ret i32 0
13055 //
13056 //
13057 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
13058 // CHECK19-SAME: (i32 [[A:%.*]]) #[[ATTR2]] {
13059 // CHECK19-NEXT:  entry:
13060 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13061 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13062 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13063 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
13064 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
13065 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
13066 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]])
13067 // CHECK19-NEXT:    ret void
13068 //
13069 //
13070 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..2
13071 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]]) #[[ATTR2]] {
13072 // CHECK19-NEXT:  entry:
13073 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13074 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13075 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13076 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13077 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13078 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13079 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
13080 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
13081 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
13082 // CHECK19-NEXT:    ret void
13083 //
13084 //
13085 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
13086 // CHECK19-SAME: (i32 [[AA:%.*]]) #[[ATTR2]] {
13087 // CHECK19-NEXT:  entry:
13088 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13089 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13090 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13091 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13092 // CHECK19-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
13093 // CHECK19-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13094 // CHECK19-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
13095 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13096 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP1]])
13097 // CHECK19-NEXT:    ret void
13098 //
13099 //
13100 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..3
13101 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
13102 // CHECK19-NEXT:  entry:
13103 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13104 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13105 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13106 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13107 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13108 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13109 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13110 // CHECK19-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
13111 // CHECK19-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
13112 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
13113 // CHECK19-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
13114 // CHECK19-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
13115 // CHECK19-NEXT:    ret void
13116 //
13117 //
13118 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
13119 // CHECK19-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
13120 // CHECK19-NEXT:  entry:
13121 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13122 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13123 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13124 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13125 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13126 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13127 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13128 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
13129 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
13130 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
13131 // CHECK19-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
13132 // CHECK19-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13133 // CHECK19-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
13134 // CHECK19-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13135 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
13136 // CHECK19-NEXT:    ret void
13137 //
13138 //
13139 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..6
13140 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
13141 // CHECK19-NEXT:  entry:
13142 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13143 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13144 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13145 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13146 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13147 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13148 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13149 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13150 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13151 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
13152 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
13153 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
13154 // CHECK19-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
13155 // CHECK19-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
13156 // CHECK19-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
13157 // CHECK19-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
13158 // CHECK19-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
13159 // CHECK19-NEXT:    ret void
13160 //
13161 //
13162 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
13163 // CHECK19-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
13164 // CHECK19-NEXT:  entry:
13165 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13166 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
13167 // CHECK19-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
13168 // CHECK19-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
13169 // CHECK19-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
13170 // CHECK19-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
13171 // CHECK19-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
13172 // CHECK19-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
13173 // CHECK19-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
13174 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13175 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13176 // CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
13177 // CHECK19-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
13178 // CHECK19-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
13179 // CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
13180 // CHECK19-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
13181 // CHECK19-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
13182 // CHECK19-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
13183 // CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
13184 // CHECK19-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
13185 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
13186 // CHECK19-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
13187 // CHECK19-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
13188 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
13189 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
13190 // CHECK19-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
13191 // CHECK19-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
13192 // CHECK19-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
13193 // CHECK19-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
13194 // CHECK19-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
13195 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
13196 // CHECK19-NEXT:    ret void
13197 //
13198 //
13199 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..9
13200 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
13201 // CHECK19-NEXT:  entry:
13202 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13203 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13204 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13205 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
13206 // CHECK19-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
13207 // CHECK19-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
13208 // CHECK19-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
13209 // CHECK19-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
13210 // CHECK19-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
13211 // CHECK19-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
13212 // CHECK19-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
13213 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13214 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13215 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13216 // CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
13217 // CHECK19-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
13218 // CHECK19-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
13219 // CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
13220 // CHECK19-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
13221 // CHECK19-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
13222 // CHECK19-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
13223 // CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
13224 // CHECK19-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
13225 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
13226 // CHECK19-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
13227 // CHECK19-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
13228 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
13229 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
13230 // CHECK19-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
13231 // CHECK19-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
13232 // CHECK19-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
13233 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
13234 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
13235 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
13236 // CHECK19-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
13237 // CHECK19-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
13238 // CHECK19-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
13239 // CHECK19-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
13240 // CHECK19-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
13241 // CHECK19-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
13242 // CHECK19-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
13243 // CHECK19-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
13244 // CHECK19-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
13245 // CHECK19-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
13246 // CHECK19-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
13247 // CHECK19-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
13248 // CHECK19-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
13249 // CHECK19-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
13250 // CHECK19-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
13251 // CHECK19-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
13252 // CHECK19-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
13253 // CHECK19-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
13254 // CHECK19-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
13255 // CHECK19-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
13256 // CHECK19-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
13257 // CHECK19-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
13258 // CHECK19-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
13259 // CHECK19-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
13260 // CHECK19-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
13261 // CHECK19-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
13262 // CHECK19-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
13263 // CHECK19-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
13264 // CHECK19-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
13265 // CHECK19-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
13266 // CHECK19-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
13267 // CHECK19-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
13268 // CHECK19-NEXT:    ret void
13269 //
13270 //
13271 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
13272 // CHECK19-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
13273 // CHECK19-NEXT:  entry:
13274 // CHECK19-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
13275 // CHECK19-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
13276 // CHECK19-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
13277 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
13278 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
13279 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
13280 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP1]])
13281 // CHECK19-NEXT:    ret void
13282 //
13283 //
13284 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..11
13285 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
13286 // CHECK19-NEXT:  entry:
13287 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13288 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13289 // CHECK19-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
13290 // CHECK19-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
13291 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13292 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13293 // CHECK19-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
13294 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
13295 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
13296 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
13297 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i32 [[TMP1]])
13298 // CHECK19-NEXT:    ret void
13299 //
13300 //
13301 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..12
13302 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
13303 // CHECK19-NEXT:  entry:
13304 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13305 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13306 // CHECK19-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
13307 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13308 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13309 // CHECK19-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
13310 // CHECK19-NEXT:    ret void
13311 //
13312 //
13313 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
13314 // CHECK19-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
13315 // CHECK19-NEXT:  entry:
13316 // CHECK19-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
13317 // CHECK19-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
13318 // CHECK19-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
13319 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
13320 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
13321 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
13322 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i32 [[TMP1]])
13323 // CHECK19-NEXT:    ret void
13324 //
13325 //
13326 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..15
13327 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
13328 // CHECK19-NEXT:  entry:
13329 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13330 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13331 // CHECK19-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
13332 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13333 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13334 // CHECK19-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
13335 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
13336 // CHECK19-NEXT:    ret void
13337 //
13338 //
13339 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..16
13340 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
13341 // CHECK19-NEXT:  entry:
13342 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13343 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13344 // CHECK19-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
13345 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13346 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13347 // CHECK19-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
13348 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
13349 // CHECK19-NEXT:    ret void
13350 //
13351 //
13352 // CHECK19-LABEL: define {{[^@]+}}@_Z6bazzzziPi
13353 // CHECK19-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
13354 // CHECK19-NEXT:  entry:
13355 // CHECK19-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
13356 // CHECK19-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
13357 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
13358 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
13359 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
13360 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
13361 // CHECK19-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
13362 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
13363 // CHECK19-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13364 // CHECK19-NEXT:    [[TMP2:%.*]] = bitcast i8** [[TMP1]] to i32*
13365 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP2]], align 4
13366 // CHECK19-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13367 // CHECK19-NEXT:    [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32*
13368 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP4]], align 4
13369 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
13370 // CHECK19-NEXT:    store i8* null, i8** [[TMP5]], align 4
13371 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13372 // CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13373 // CHECK19-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
13374 // CHECK19-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
13375 // CHECK19-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
13376 // CHECK19:       omp_offload.failed:
13377 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i32 [[TMP0]]) #[[ATTR3]]
13378 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
13379 // CHECK19:       omp_offload.cont:
13380 // CHECK19-NEXT:    ret void
13381 //
13382 //
13383 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
13384 // CHECK19-SAME: (i32 [[VLA:%.*]]) #[[ATTR2]] {
13385 // CHECK19-NEXT:  entry:
13386 // CHECK19-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
13387 // CHECK19-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
13388 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
13389 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i32 [[TMP0]])
13390 // CHECK19-NEXT:    ret void
13391 //
13392 //
13393 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..19
13394 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR2]] {
13395 // CHECK19-NEXT:  entry:
13396 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13397 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13398 // CHECK19-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
13399 // CHECK19-NEXT:    [[F:%.*]] = alloca i32*, align 4
13400 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13401 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13402 // CHECK19-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
13403 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
13404 // CHECK19-NEXT:    ret void
13405 //
13406 //
13407 // CHECK19-LABEL: define {{[^@]+}}@_Z3bari
13408 // CHECK19-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
13409 // CHECK19-NEXT:  entry:
13410 // CHECK19-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
13411 // CHECK19-NEXT:    [[A:%.*]] = alloca i32, align 4
13412 // CHECK19-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
13413 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
13414 // CHECK19-NEXT:    store i32 0, i32* [[A]], align 4
13415 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
13416 // CHECK19-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
13417 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
13418 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
13419 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
13420 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
13421 // CHECK19-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
13422 // CHECK19-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
13423 // CHECK19-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
13424 // CHECK19-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
13425 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
13426 // CHECK19-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
13427 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
13428 // CHECK19-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
13429 // CHECK19-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
13430 // CHECK19-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
13431 // CHECK19-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
13432 // CHECK19-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
13433 // CHECK19-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
13434 // CHECK19-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
13435 // CHECK19-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
13436 // CHECK19-NEXT:    ret i32 [[TMP8]]
13437 //
13438 //
13439 // CHECK19-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
13440 // CHECK19-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
13441 // CHECK19-NEXT:  entry:
13442 // CHECK19-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
13443 // CHECK19-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
13444 // CHECK19-NEXT:    [[B:%.*]] = alloca i32, align 4
13445 // CHECK19-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
13446 // CHECK19-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
13447 // CHECK19-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
13448 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 4
13449 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 4
13450 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 4
13451 // CHECK19-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 4
13452 // CHECK19-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
13453 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
13454 // CHECK19-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
13455 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
13456 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
13457 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
13458 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
13459 // CHECK19-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
13460 // CHECK19-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
13461 // CHECK19-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
13462 // CHECK19-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
13463 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
13464 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
13465 // CHECK19-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
13466 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
13467 // CHECK19-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
13468 // CHECK19-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 60
13469 // CHECK19-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
13470 // CHECK19:       omp_if.then:
13471 // CHECK19-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
13472 // CHECK19-NEXT:    [[TMP7:%.*]] = mul nuw i32 2, [[TMP1]]
13473 // CHECK19-NEXT:    [[TMP8:%.*]] = mul nuw i32 [[TMP7]], 2
13474 // CHECK19-NEXT:    [[TMP9:%.*]] = sext i32 [[TMP8]] to i64
13475 // CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13476 // CHECK19-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
13477 // CHECK19-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 4
13478 // CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13479 // CHECK19-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
13480 // CHECK19-NEXT:    store double* [[A]], double** [[TMP13]], align 4
13481 // CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
13482 // CHECK19-NEXT:    store i64 8, i64* [[TMP14]], align 4
13483 // CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
13484 // CHECK19-NEXT:    store i8* null, i8** [[TMP15]], align 4
13485 // CHECK19-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
13486 // CHECK19-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
13487 // CHECK19-NEXT:    store i32 [[TMP5]], i32* [[TMP17]], align 4
13488 // CHECK19-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
13489 // CHECK19-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
13490 // CHECK19-NEXT:    store i32 [[TMP5]], i32* [[TMP19]], align 4
13491 // CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
13492 // CHECK19-NEXT:    store i64 4, i64* [[TMP20]], align 4
13493 // CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
13494 // CHECK19-NEXT:    store i8* null, i8** [[TMP21]], align 4
13495 // CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
13496 // CHECK19-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
13497 // CHECK19-NEXT:    store i32 2, i32* [[TMP23]], align 4
13498 // CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
13499 // CHECK19-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
13500 // CHECK19-NEXT:    store i32 2, i32* [[TMP25]], align 4
13501 // CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
13502 // CHECK19-NEXT:    store i64 4, i64* [[TMP26]], align 4
13503 // CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
13504 // CHECK19-NEXT:    store i8* null, i8** [[TMP27]], align 4
13505 // CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
13506 // CHECK19-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
13507 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP29]], align 4
13508 // CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
13509 // CHECK19-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
13510 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP31]], align 4
13511 // CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
13512 // CHECK19-NEXT:    store i64 4, i64* [[TMP32]], align 4
13513 // CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
13514 // CHECK19-NEXT:    store i8* null, i8** [[TMP33]], align 4
13515 // CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
13516 // CHECK19-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
13517 // CHECK19-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 4
13518 // CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
13519 // CHECK19-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
13520 // CHECK19-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 4
13521 // CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
13522 // CHECK19-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 4
13523 // CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 4
13524 // CHECK19-NEXT:    store i8* null, i8** [[TMP39]], align 4
13525 // CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13526 // CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13527 // CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
13528 // CHECK19-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
13529 // CHECK19-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
13530 // CHECK19-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
13531 // CHECK19:       omp_offload.failed:
13532 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
13533 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
13534 // CHECK19:       omp_offload.cont:
13535 // CHECK19-NEXT:    br label [[OMP_IF_END:%.*]]
13536 // CHECK19:       omp_if.else:
13537 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
13538 // CHECK19-NEXT:    br label [[OMP_IF_END]]
13539 // CHECK19:       omp_if.end:
13540 // CHECK19-NEXT:    [[TMP45:%.*]] = mul nsw i32 1, [[TMP1]]
13541 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP45]]
13542 // CHECK19-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
13543 // CHECK19-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
13544 // CHECK19-NEXT:    [[CONV:%.*]] = sext i16 [[TMP46]] to i32
13545 // CHECK19-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
13546 // CHECK19-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP47]]
13547 // CHECK19-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
13548 // CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
13549 // CHECK19-NEXT:    ret i32 [[ADD3]]
13550 //
13551 //
13552 // CHECK19-LABEL: define {{[^@]+}}@_ZL7fstatici
13553 // CHECK19-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
13554 // CHECK19-NEXT:  entry:
13555 // CHECK19-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
13556 // CHECK19-NEXT:    [[A:%.*]] = alloca i32, align 4
13557 // CHECK19-NEXT:    [[AA:%.*]] = alloca i16, align 2
13558 // CHECK19-NEXT:    [[AAA:%.*]] = alloca i8, align 1
13559 // CHECK19-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
13560 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13561 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13562 // CHECK19-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
13563 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 4
13564 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 4
13565 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 4
13566 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
13567 // CHECK19-NEXT:    store i32 0, i32* [[A]], align 4
13568 // CHECK19-NEXT:    store i16 0, i16* [[AA]], align 2
13569 // CHECK19-NEXT:    store i8 0, i8* [[AAA]], align 1
13570 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
13571 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
13572 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
13573 // CHECK19-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
13574 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13575 // CHECK19-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
13576 // CHECK19-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13577 // CHECK19-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
13578 // CHECK19-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
13579 // CHECK19-NEXT:    store i8 [[TMP4]], i8* [[CONV1]], align 1
13580 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
13581 // CHECK19-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
13582 // CHECK19-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
13583 // CHECK19-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
13584 // CHECK19:       omp_if.then:
13585 // CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13586 // CHECK19-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
13587 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
13588 // CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13589 // CHECK19-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
13590 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP10]], align 4
13591 // CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
13592 // CHECK19-NEXT:    store i8* null, i8** [[TMP11]], align 4
13593 // CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
13594 // CHECK19-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
13595 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
13596 // CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
13597 // CHECK19-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
13598 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP15]], align 4
13599 // CHECK19-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
13600 // CHECK19-NEXT:    store i8* null, i8** [[TMP16]], align 4
13601 // CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
13602 // CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
13603 // CHECK19-NEXT:    store i32 [[TMP5]], i32* [[TMP18]], align 4
13604 // CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
13605 // CHECK19-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
13606 // CHECK19-NEXT:    store i32 [[TMP5]], i32* [[TMP20]], align 4
13607 // CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
13608 // CHECK19-NEXT:    store i8* null, i8** [[TMP21]], align 4
13609 // CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
13610 // CHECK19-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
13611 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 4
13612 // CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
13613 // CHECK19-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
13614 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 4
13615 // CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
13616 // CHECK19-NEXT:    store i8* null, i8** [[TMP26]], align 4
13617 // CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13618 // CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13619 // CHECK19-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
13620 // CHECK19-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
13621 // CHECK19-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
13622 // CHECK19:       omp_offload.failed:
13623 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
13624 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
13625 // CHECK19:       omp_offload.cont:
13626 // CHECK19-NEXT:    br label [[OMP_IF_END:%.*]]
13627 // CHECK19:       omp_if.else:
13628 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
13629 // CHECK19-NEXT:    br label [[OMP_IF_END]]
13630 // CHECK19:       omp_if.end:
13631 // CHECK19-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
13632 // CHECK19-NEXT:    ret i32 [[TMP31]]
13633 //
13634 //
13635 // CHECK19-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
13636 // CHECK19-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
13637 // CHECK19-NEXT:  entry:
13638 // CHECK19-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
13639 // CHECK19-NEXT:    [[A:%.*]] = alloca i32, align 4
13640 // CHECK19-NEXT:    [[AA:%.*]] = alloca i16, align 2
13641 // CHECK19-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
13642 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13643 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13644 // CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
13645 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
13646 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
13647 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
13648 // CHECK19-NEXT:    store i32 0, i32* [[A]], align 4
13649 // CHECK19-NEXT:    store i16 0, i16* [[AA]], align 2
13650 // CHECK19-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
13651 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
13652 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
13653 // CHECK19-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
13654 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13655 // CHECK19-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
13656 // CHECK19-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13657 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
13658 // CHECK19-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
13659 // CHECK19-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
13660 // CHECK19:       omp_if.then:
13661 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13662 // CHECK19-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
13663 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP6]], align 4
13664 // CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13665 // CHECK19-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
13666 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
13667 // CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
13668 // CHECK19-NEXT:    store i8* null, i8** [[TMP9]], align 4
13669 // CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
13670 // CHECK19-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i32*
13671 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP11]], align 4
13672 // CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
13673 // CHECK19-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
13674 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
13675 // CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
13676 // CHECK19-NEXT:    store i8* null, i8** [[TMP14]], align 4
13677 // CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
13678 // CHECK19-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
13679 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 4
13680 // CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
13681 // CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
13682 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 4
13683 // CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
13684 // CHECK19-NEXT:    store i8* null, i8** [[TMP19]], align 4
13685 // CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13686 // CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13687 // CHECK19-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
13688 // CHECK19-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
13689 // CHECK19-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
13690 // CHECK19:       omp_offload.failed:
13691 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
13692 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
13693 // CHECK19:       omp_offload.cont:
13694 // CHECK19-NEXT:    br label [[OMP_IF_END:%.*]]
13695 // CHECK19:       omp_if.else:
13696 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
13697 // CHECK19-NEXT:    br label [[OMP_IF_END]]
13698 // CHECK19:       omp_if.end:
13699 // CHECK19-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
13700 // CHECK19-NEXT:    ret i32 [[TMP24]]
13701 //
13702 //
13703 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
13704 // CHECK19-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
13705 // CHECK19-NEXT:  entry:
13706 // CHECK19-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
13707 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
13708 // CHECK19-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
13709 // CHECK19-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
13710 // CHECK19-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
13711 // CHECK19-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
13712 // CHECK19-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
13713 // CHECK19-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
13714 // CHECK19-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
13715 // CHECK19-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
13716 // CHECK19-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
13717 // CHECK19-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
13718 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
13719 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
13720 // CHECK19-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
13721 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
13722 // CHECK19-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
13723 // CHECK19-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
13724 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
13725 // CHECK19-NEXT:    ret void
13726 //
13727 //
13728 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..22
13729 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
13730 // CHECK19-NEXT:  entry:
13731 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13732 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13733 // CHECK19-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
13734 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
13735 // CHECK19-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
13736 // CHECK19-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
13737 // CHECK19-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
13738 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13739 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13740 // CHECK19-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
13741 // CHECK19-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
13742 // CHECK19-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
13743 // CHECK19-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
13744 // CHECK19-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
13745 // CHECK19-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
13746 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
13747 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
13748 // CHECK19-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
13749 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
13750 // CHECK19-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
13751 // CHECK19-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
13752 // CHECK19-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
13753 // CHECK19-NEXT:    store double [[ADD]], double* [[A]], align 4
13754 // CHECK19-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
13755 // CHECK19-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
13756 // CHECK19-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
13757 // CHECK19-NEXT:    store double [[INC]], double* [[A3]], align 4
13758 // CHECK19-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
13759 // CHECK19-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
13760 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
13761 // CHECK19-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
13762 // CHECK19-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
13763 // CHECK19-NEXT:    ret void
13764 //
13765 //
13766 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
13767 // CHECK19-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
13768 // CHECK19-NEXT:  entry:
13769 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13770 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13771 // CHECK19-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
13772 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
13773 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13774 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13775 // CHECK19-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
13776 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13777 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13778 // CHECK19-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
13779 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
13780 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13781 // CHECK19-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
13782 // CHECK19-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
13783 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
13784 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
13785 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
13786 // CHECK19-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
13787 // CHECK19-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13788 // CHECK19-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
13789 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13790 // CHECK19-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
13791 // CHECK19-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
13792 // CHECK19-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
13793 // CHECK19-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
13794 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
13795 // CHECK19-NEXT:    ret void
13796 //
13797 //
13798 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..24
13799 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
13800 // CHECK19-NEXT:  entry:
13801 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13802 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13803 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13804 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13805 // CHECK19-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
13806 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
13807 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13808 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13809 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13810 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13811 // CHECK19-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
13812 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
13813 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13814 // CHECK19-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
13815 // CHECK19-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
13816 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
13817 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
13818 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
13819 // CHECK19-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
13820 // CHECK19-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
13821 // CHECK19-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
13822 // CHECK19-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
13823 // CHECK19-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
13824 // CHECK19-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
13825 // CHECK19-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
13826 // CHECK19-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
13827 // CHECK19-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
13828 // CHECK19-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
13829 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
13830 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
13831 // CHECK19-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
13832 // CHECK19-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
13833 // CHECK19-NEXT:    ret void
13834 //
13835 //
13836 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
13837 // CHECK19-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
13838 // CHECK19-NEXT:  entry:
13839 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13840 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13841 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
13842 // CHECK19-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13843 // CHECK19-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13844 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13845 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13846 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
13847 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13848 // CHECK19-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
13849 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
13850 // CHECK19-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
13851 // CHECK19-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
13852 // CHECK19-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
13853 // CHECK19-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13854 // CHECK19-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
13855 // CHECK19-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13856 // CHECK19-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
13857 // CHECK19-NEXT:    ret void
13858 //
13859 //
13860 // CHECK19-LABEL: define {{[^@]+}}@.omp_outlined..27
13861 // CHECK19-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
13862 // CHECK19-NEXT:  entry:
13863 // CHECK19-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
13864 // CHECK19-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
13865 // CHECK19-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
13866 // CHECK19-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
13867 // CHECK19-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
13868 // CHECK19-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
13869 // CHECK19-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
13870 // CHECK19-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
13871 // CHECK19-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
13872 // CHECK19-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
13873 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
13874 // CHECK19-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
13875 // CHECK19-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
13876 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
13877 // CHECK19-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
13878 // CHECK19-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
13879 // CHECK19-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
13880 // CHECK19-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
13881 // CHECK19-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
13882 // CHECK19-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
13883 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
13884 // CHECK19-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
13885 // CHECK19-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
13886 // CHECK19-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
13887 // CHECK19-NEXT:    ret void
13888 //
13889 //
13890 // CHECK19-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
13891 // CHECK19-SAME: () #[[ATTR4]] {
13892 // CHECK19-NEXT:  entry:
13893 // CHECK19-NEXT:    call void @__tgt_register_requires(i64 1)
13894 // CHECK19-NEXT:    ret void
13895 //
13896 //
13897 // CHECK20-LABEL: define {{[^@]+}}@_Z3fooi
13898 // CHECK20-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
13899 // CHECK20-NEXT:  entry:
13900 // CHECK20-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
13901 // CHECK20-NEXT:    [[A:%.*]] = alloca i32, align 4
13902 // CHECK20-NEXT:    [[AA:%.*]] = alloca i16, align 2
13903 // CHECK20-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
13904 // CHECK20-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
13905 // CHECK20-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
13906 // CHECK20-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
13907 // CHECK20-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
13908 // CHECK20-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
13909 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
13910 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
13911 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
13912 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
13913 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR__CASTED3:%.*]] = alloca i32, align 4
13914 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
13915 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
13916 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
13917 // CHECK20-NEXT:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
13918 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
13919 // CHECK20-NEXT:    [[AA_CASTED4:%.*]] = alloca i32, align 4
13920 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS6:%.*]] = alloca [1 x i8*], align 4
13921 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [1 x i8*], align 4
13922 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [1 x i8*], align 4
13923 // CHECK20-NEXT:    [[A_CASTED9:%.*]] = alloca i32, align 4
13924 // CHECK20-NEXT:    [[AA_CASTED10:%.*]] = alloca i32, align 4
13925 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS12:%.*]] = alloca [2 x i8*], align 4
13926 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
13927 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
13928 // CHECK20-NEXT:    [[A_CASTED17:%.*]] = alloca i32, align 4
13929 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [9 x i8*], align 4
13930 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [9 x i8*], align 4
13931 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [9 x i8*], align 4
13932 // CHECK20-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
13933 // CHECK20-NEXT:    [[NN:%.*]] = alloca i32, align 4
13934 // CHECK20-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
13935 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
13936 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
13937 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
13938 // CHECK20-NEXT:    [[NN_CASTED32:%.*]] = alloca i32, align 4
13939 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 4
13940 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 4
13941 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 4
13942 // CHECK20-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
13943 // CHECK20-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
13944 // CHECK20-NEXT:    store i32 0, i32* [[A]], align 4
13945 // CHECK20-NEXT:    store i16 0, i16* [[AA]], align 2
13946 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
13947 // CHECK20-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
13948 // CHECK20-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
13949 // CHECK20-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP1]], align 4
13950 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
13951 // CHECK20-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
13952 // CHECK20-NEXT:    [[TMP4:%.*]] = mul nuw i32 5, [[TMP3]]
13953 // CHECK20-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP4]], align 8
13954 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[__VLA_EXPR1]], align 4
13955 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
13956 // CHECK20-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4
13957 // CHECK20-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
13958 // CHECK20-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_2]], align 4
13959 // CHECK20-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
13960 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
13961 // CHECK20-NEXT:    store i16 [[TMP7]], i16* [[CONV]], align 2
13962 // CHECK20-NEXT:    [[TMP8:%.*]] = load i32, i32* [[AA_CASTED]], align 4
13963 // CHECK20-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
13964 // CHECK20-NEXT:    store i32 [[TMP9]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
13965 // CHECK20-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
13966 // CHECK20-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
13967 // CHECK20-NEXT:    store i32 [[TMP11]], i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
13968 // CHECK20-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED3]], align 4
13969 // CHECK20-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13970 // CHECK20-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i32*
13971 // CHECK20-NEXT:    store i32 [[TMP8]], i32* [[TMP14]], align 4
13972 // CHECK20-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13973 // CHECK20-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
13974 // CHECK20-NEXT:    store i32 [[TMP8]], i32* [[TMP16]], align 4
13975 // CHECK20-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
13976 // CHECK20-NEXT:    store i8* null, i8** [[TMP17]], align 4
13977 // CHECK20-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
13978 // CHECK20-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
13979 // CHECK20-NEXT:    store i32 [[TMP10]], i32* [[TMP19]], align 4
13980 // CHECK20-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
13981 // CHECK20-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
13982 // CHECK20-NEXT:    store i32 [[TMP10]], i32* [[TMP21]], align 4
13983 // CHECK20-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
13984 // CHECK20-NEXT:    store i8* null, i8** [[TMP22]], align 4
13985 // CHECK20-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
13986 // CHECK20-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
13987 // CHECK20-NEXT:    store i32 [[TMP12]], i32* [[TMP24]], align 4
13988 // CHECK20-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
13989 // CHECK20-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
13990 // CHECK20-NEXT:    store i32 [[TMP12]], i32* [[TMP26]], align 4
13991 // CHECK20-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
13992 // CHECK20-NEXT:    store i8* null, i8** [[TMP27]], align 4
13993 // CHECK20-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
13994 // CHECK20-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
13995 // CHECK20-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
13996 // CHECK20-NEXT:    [[TMP31:%.*]] = load i16, i16* [[AA]], align 2
13997 // CHECK20-NEXT:    store i16 [[TMP31]], i16* [[TMP30]], align 4
13998 // CHECK20-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
13999 // CHECK20-NEXT:    [[TMP33:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
14000 // CHECK20-NEXT:    store i32 [[TMP33]], i32* [[TMP32]], align 4
14001 // CHECK20-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
14002 // CHECK20-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4
14003 // CHECK20-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
14004 // CHECK20-NEXT:    [[TMP36:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
14005 // CHECK20-NEXT:    [[TMP37:%.*]] = bitcast i8* [[TMP36]] to %struct.kmp_task_t_with_privates*
14006 // CHECK20-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 0
14007 // CHECK20-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP38]], i32 0, i32 0
14008 // CHECK20-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[TMP39]], align 4
14009 // CHECK20-NEXT:    [[TMP41:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
14010 // CHECK20-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP40]], i8* align 4 [[TMP41]], i32 12, i1 false)
14011 // CHECK20-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP37]], i32 0, i32 1
14012 // CHECK20-NEXT:    [[TMP43:%.*]] = bitcast i8* [[TMP40]] to %struct.anon*
14013 // CHECK20-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 0
14014 // CHECK20-NEXT:    [[TMP45:%.*]] = bitcast [3 x i64]* [[TMP44]] to i8*
14015 // CHECK20-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP45]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
14016 // CHECK20-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 1
14017 // CHECK20-NEXT:    [[TMP47:%.*]] = bitcast [3 x i8*]* [[TMP46]] to i8*
14018 // CHECK20-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP28]] to i8*
14019 // CHECK20-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP47]], i8* align 4 [[TMP48]], i32 12, i1 false)
14020 // CHECK20-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 2
14021 // CHECK20-NEXT:    [[TMP50:%.*]] = bitcast [3 x i8*]* [[TMP49]] to i8*
14022 // CHECK20-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP29]] to i8*
14023 // CHECK20-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i32 12, i1 false)
14024 // CHECK20-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP42]], i32 0, i32 3
14025 // CHECK20-NEXT:    [[TMP53:%.*]] = load i16, i16* [[AA]], align 2
14026 // CHECK20-NEXT:    store i16 [[TMP53]], i16* [[TMP52]], align 4
14027 // CHECK20-NEXT:    [[TMP54:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP36]])
14028 // CHECK20-NEXT:    [[TMP55:%.*]] = load i32, i32* [[A]], align 4
14029 // CHECK20-NEXT:    store i32 [[TMP55]], i32* [[A_CASTED]], align 4
14030 // CHECK20-NEXT:    [[TMP56:%.*]] = load i32, i32* [[A_CASTED]], align 4
14031 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105(i32 [[TMP56]]) #[[ATTR3:[0-9]+]]
14032 // CHECK20-NEXT:    [[TMP57:%.*]] = load i16, i16* [[AA]], align 2
14033 // CHECK20-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
14034 // CHECK20-NEXT:    store i16 [[TMP57]], i16* [[CONV5]], align 2
14035 // CHECK20-NEXT:    [[TMP58:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
14036 // CHECK20-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
14037 // CHECK20-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32*
14038 // CHECK20-NEXT:    store i32 [[TMP58]], i32* [[TMP60]], align 4
14039 // CHECK20-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
14040 // CHECK20-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32*
14041 // CHECK20-NEXT:    store i32 [[TMP58]], i32* [[TMP62]], align 4
14042 // CHECK20-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
14043 // CHECK20-NEXT:    store i8* null, i8** [[TMP63]], align 4
14044 // CHECK20-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
14045 // CHECK20-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
14046 // CHECK20-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14047 // CHECK20-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
14048 // CHECK20-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
14049 // CHECK20:       omp_offload.failed:
14050 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR3]]
14051 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT]]
14052 // CHECK20:       omp_offload.cont:
14053 // CHECK20-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
14054 // CHECK20-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
14055 // CHECK20-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
14056 // CHECK20-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
14057 // CHECK20-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
14058 // CHECK20-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
14059 // CHECK20-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
14060 // CHECK20-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
14061 // CHECK20-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
14062 // CHECK20-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
14063 // CHECK20:       omp_if.then:
14064 // CHECK20-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
14065 // CHECK20-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
14066 // CHECK20-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
14067 // CHECK20-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
14068 // CHECK20-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
14069 // CHECK20-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
14070 // CHECK20-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
14071 // CHECK20-NEXT:    store i8* null, i8** [[TMP77]], align 4
14072 // CHECK20-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
14073 // CHECK20-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
14074 // CHECK20-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
14075 // CHECK20-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
14076 // CHECK20-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
14077 // CHECK20-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
14078 // CHECK20-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
14079 // CHECK20-NEXT:    store i8* null, i8** [[TMP82]], align 4
14080 // CHECK20-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
14081 // CHECK20-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
14082 // CHECK20-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14083 // CHECK20-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
14084 // CHECK20-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
14085 // CHECK20:       omp_offload.failed15:
14086 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
14087 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
14088 // CHECK20:       omp_offload.cont16:
14089 // CHECK20-NEXT:    br label [[OMP_IF_END:%.*]]
14090 // CHECK20:       omp_if.else:
14091 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
14092 // CHECK20-NEXT:    br label [[OMP_IF_END]]
14093 // CHECK20:       omp_if.end:
14094 // CHECK20-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
14095 // CHECK20-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED17]], align 4
14096 // CHECK20-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED17]], align 4
14097 // CHECK20-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
14098 // CHECK20-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP89]], 20
14099 // CHECK20-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
14100 // CHECK20:       omp_if.then19:
14101 // CHECK20-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
14102 // CHECK20-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
14103 // CHECK20-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
14104 // CHECK20-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
14105 // CHECK20-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
14106 // CHECK20-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
14107 // CHECK20-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
14108 // CHECK20-NEXT:    store i32 [[TMP88]], i32* [[TMP96]], align 4
14109 // CHECK20-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
14110 // CHECK20-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
14111 // CHECK20-NEXT:    store i32 [[TMP88]], i32* [[TMP98]], align 4
14112 // CHECK20-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
14113 // CHECK20-NEXT:    store i64 4, i64* [[TMP99]], align 4
14114 // CHECK20-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
14115 // CHECK20-NEXT:    store i8* null, i8** [[TMP100]], align 4
14116 // CHECK20-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
14117 // CHECK20-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
14118 // CHECK20-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
14119 // CHECK20-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
14120 // CHECK20-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
14121 // CHECK20-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
14122 // CHECK20-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
14123 // CHECK20-NEXT:    store i64 40, i64* [[TMP105]], align 4
14124 // CHECK20-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
14125 // CHECK20-NEXT:    store i8* null, i8** [[TMP106]], align 4
14126 // CHECK20-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
14127 // CHECK20-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
14128 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP108]], align 4
14129 // CHECK20-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
14130 // CHECK20-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
14131 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP110]], align 4
14132 // CHECK20-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
14133 // CHECK20-NEXT:    store i64 4, i64* [[TMP111]], align 4
14134 // CHECK20-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
14135 // CHECK20-NEXT:    store i8* null, i8** [[TMP112]], align 4
14136 // CHECK20-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
14137 // CHECK20-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
14138 // CHECK20-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
14139 // CHECK20-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
14140 // CHECK20-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to float**
14141 // CHECK20-NEXT:    store float* [[VLA]], float** [[TMP116]], align 4
14142 // CHECK20-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
14143 // CHECK20-NEXT:    store i64 [[TMP91]], i64* [[TMP117]], align 4
14144 // CHECK20-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
14145 // CHECK20-NEXT:    store i8* null, i8** [[TMP118]], align 4
14146 // CHECK20-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
14147 // CHECK20-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
14148 // CHECK20-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
14149 // CHECK20-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
14150 // CHECK20-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [5 x [10 x double]]**
14151 // CHECK20-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP122]], align 4
14152 // CHECK20-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
14153 // CHECK20-NEXT:    store i64 400, i64* [[TMP123]], align 4
14154 // CHECK20-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 4
14155 // CHECK20-NEXT:    store i8* null, i8** [[TMP124]], align 4
14156 // CHECK20-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
14157 // CHECK20-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i32*
14158 // CHECK20-NEXT:    store i32 5, i32* [[TMP126]], align 4
14159 // CHECK20-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
14160 // CHECK20-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
14161 // CHECK20-NEXT:    store i32 5, i32* [[TMP128]], align 4
14162 // CHECK20-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 5
14163 // CHECK20-NEXT:    store i64 4, i64* [[TMP129]], align 4
14164 // CHECK20-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 5
14165 // CHECK20-NEXT:    store i8* null, i8** [[TMP130]], align 4
14166 // CHECK20-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
14167 // CHECK20-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
14168 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[TMP132]], align 4
14169 // CHECK20-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
14170 // CHECK20-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
14171 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[TMP134]], align 4
14172 // CHECK20-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 6
14173 // CHECK20-NEXT:    store i64 4, i64* [[TMP135]], align 4
14174 // CHECK20-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 6
14175 // CHECK20-NEXT:    store i8* null, i8** [[TMP136]], align 4
14176 // CHECK20-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
14177 // CHECK20-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
14178 // CHECK20-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 4
14179 // CHECK20-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
14180 // CHECK20-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to double**
14181 // CHECK20-NEXT:    store double* [[VLA1]], double** [[TMP140]], align 4
14182 // CHECK20-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
14183 // CHECK20-NEXT:    store i64 [[TMP94]], i64* [[TMP141]], align 4
14184 // CHECK20-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 7
14185 // CHECK20-NEXT:    store i8* null, i8** [[TMP142]], align 4
14186 // CHECK20-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
14187 // CHECK20-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
14188 // CHECK20-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 4
14189 // CHECK20-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
14190 // CHECK20-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %struct.TT**
14191 // CHECK20-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP146]], align 4
14192 // CHECK20-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 8
14193 // CHECK20-NEXT:    store i64 12, i64* [[TMP147]], align 4
14194 // CHECK20-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 8
14195 // CHECK20-NEXT:    store i8* null, i8** [[TMP148]], align 4
14196 // CHECK20-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
14197 // CHECK20-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
14198 // CHECK20-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
14199 // CHECK20-NEXT:    [[TMP152:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP149]], i8** [[TMP150]], i64* [[TMP151]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14200 // CHECK20-NEXT:    [[TMP153:%.*]] = icmp ne i32 [[TMP152]], 0
14201 // CHECK20-NEXT:    br i1 [[TMP153]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
14202 // CHECK20:       omp_offload.failed23:
14203 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
14204 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
14205 // CHECK20:       omp_offload.cont24:
14206 // CHECK20-NEXT:    br label [[OMP_IF_END26:%.*]]
14207 // CHECK20:       omp_if.else25:
14208 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
14209 // CHECK20-NEXT:    br label [[OMP_IF_END26]]
14210 // CHECK20:       omp_if.end26:
14211 // CHECK20-NEXT:    store i32 0, i32* [[NN]], align 4
14212 // CHECK20-NEXT:    [[TMP154:%.*]] = load i32, i32* [[NN]], align 4
14213 // CHECK20-NEXT:    store i32 [[TMP154]], i32* [[NN_CASTED]], align 4
14214 // CHECK20-NEXT:    [[TMP155:%.*]] = load i32, i32* [[NN_CASTED]], align 4
14215 // CHECK20-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
14216 // CHECK20-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
14217 // CHECK20-NEXT:    store i32 [[TMP155]], i32* [[TMP157]], align 4
14218 // CHECK20-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
14219 // CHECK20-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32*
14220 // CHECK20-NEXT:    store i32 [[TMP155]], i32* [[TMP159]], align 4
14221 // CHECK20-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
14222 // CHECK20-NEXT:    store i8* null, i8** [[TMP160]], align 4
14223 // CHECK20-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
14224 // CHECK20-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
14225 // CHECK20-NEXT:    [[TMP163:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP161]], i8** [[TMP162]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.13, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14226 // CHECK20-NEXT:    [[TMP164:%.*]] = icmp ne i32 [[TMP163]], 0
14227 // CHECK20-NEXT:    br i1 [[TMP164]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
14228 // CHECK20:       omp_offload.failed30:
14229 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i32 [[TMP155]]) #[[ATTR3]]
14230 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
14231 // CHECK20:       omp_offload.cont31:
14232 // CHECK20-NEXT:    [[TMP165:%.*]] = load i32, i32* [[NN]], align 4
14233 // CHECK20-NEXT:    store i32 [[TMP165]], i32* [[NN_CASTED32]], align 4
14234 // CHECK20-NEXT:    [[TMP166:%.*]] = load i32, i32* [[NN_CASTED32]], align 4
14235 // CHECK20-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
14236 // CHECK20-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i32*
14237 // CHECK20-NEXT:    store i32 [[TMP166]], i32* [[TMP168]], align 4
14238 // CHECK20-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
14239 // CHECK20-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i32*
14240 // CHECK20-NEXT:    store i32 [[TMP166]], i32* [[TMP170]], align 4
14241 // CHECK20-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
14242 // CHECK20-NEXT:    store i8* null, i8** [[TMP171]], align 4
14243 // CHECK20-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
14244 // CHECK20-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
14245 // CHECK20-NEXT:    [[TMP174:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP172]], i8** [[TMP173]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14246 // CHECK20-NEXT:    [[TMP175:%.*]] = icmp ne i32 [[TMP174]], 0
14247 // CHECK20-NEXT:    br i1 [[TMP175]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
14248 // CHECK20:       omp_offload.failed36:
14249 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i32 [[TMP166]]) #[[ATTR3]]
14250 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
14251 // CHECK20:       omp_offload.cont37:
14252 // CHECK20-NEXT:    [[TMP176:%.*]] = load i32, i32* [[A]], align 4
14253 // CHECK20-NEXT:    [[TMP177:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
14254 // CHECK20-NEXT:    call void @llvm.stackrestore(i8* [[TMP177]])
14255 // CHECK20-NEXT:    ret i32 [[TMP176]]
14256 //
14257 //
14258 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
14259 // CHECK20-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR2:[0-9]+]] {
14260 // CHECK20-NEXT:  entry:
14261 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
14262 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
14263 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
14264 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
14265 // CHECK20-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
14266 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
14267 // CHECK20-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
14268 // CHECK20-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
14269 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
14270 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
14271 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
14272 // CHECK20-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
14273 // CHECK20-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
14274 // CHECK20-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
14275 // CHECK20-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
14276 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
14277 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
14278 // CHECK20-NEXT:    ret void
14279 //
14280 //
14281 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined.
14282 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
14283 // CHECK20-NEXT:  entry:
14284 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14285 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14286 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
14287 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14288 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14289 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
14290 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
14291 // CHECK20-NEXT:    ret void
14292 //
14293 //
14294 // CHECK20-LABEL: define {{[^@]+}}@.omp_task_privates_map.
14295 // CHECK20-SAME: (%struct..kmp_privates.t* noalias [[TMP0:%.*]], i16** noalias [[TMP1:%.*]], [3 x i8*]** noalias [[TMP2:%.*]], [3 x i8*]** noalias [[TMP3:%.*]], [3 x i64]** noalias [[TMP4:%.*]]) #[[ATTR4:[0-9]+]] {
14296 // CHECK20-NEXT:  entry:
14297 // CHECK20-NEXT:    [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 4
14298 // CHECK20-NEXT:    [[DOTADDR1:%.*]] = alloca i16**, align 4
14299 // CHECK20-NEXT:    [[DOTADDR2:%.*]] = alloca [3 x i8*]**, align 4
14300 // CHECK20-NEXT:    [[DOTADDR3:%.*]] = alloca [3 x i8*]**, align 4
14301 // CHECK20-NEXT:    [[DOTADDR4:%.*]] = alloca [3 x i64]**, align 4
14302 // CHECK20-NEXT:    store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 4
14303 // CHECK20-NEXT:    store i16** [[TMP1]], i16*** [[DOTADDR1]], align 4
14304 // CHECK20-NEXT:    store [3 x i8*]** [[TMP2]], [3 x i8*]*** [[DOTADDR2]], align 4
14305 // CHECK20-NEXT:    store [3 x i8*]** [[TMP3]], [3 x i8*]*** [[DOTADDR3]], align 4
14306 // CHECK20-NEXT:    store [3 x i64]** [[TMP4]], [3 x i64]*** [[DOTADDR4]], align 4
14307 // CHECK20-NEXT:    [[TMP5:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 4
14308 // CHECK20-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 0
14309 // CHECK20-NEXT:    [[TMP7:%.*]] = load [3 x i64]**, [3 x i64]*** [[DOTADDR4]], align 4
14310 // CHECK20-NEXT:    store [3 x i64]* [[TMP6]], [3 x i64]** [[TMP7]], align 4
14311 // CHECK20-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 1
14312 // CHECK20-NEXT:    [[TMP9:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR2]], align 4
14313 // CHECK20-NEXT:    store [3 x i8*]* [[TMP8]], [3 x i8*]** [[TMP9]], align 4
14314 // CHECK20-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 2
14315 // CHECK20-NEXT:    [[TMP11:%.*]] = load [3 x i8*]**, [3 x i8*]*** [[DOTADDR3]], align 4
14316 // CHECK20-NEXT:    store [3 x i8*]* [[TMP10]], [3 x i8*]** [[TMP11]], align 4
14317 // CHECK20-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP5]], i32 0, i32 3
14318 // CHECK20-NEXT:    [[TMP13:%.*]] = load i16**, i16*** [[DOTADDR1]], align 4
14319 // CHECK20-NEXT:    store i16* [[TMP12]], i16** [[TMP13]], align 4
14320 // CHECK20-NEXT:    ret void
14321 //
14322 //
14323 // CHECK20-LABEL: define {{[^@]+}}@.omp_task_entry.
14324 // CHECK20-SAME: (i32 [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias [[TMP1:%.*]]) #[[ATTR5:[0-9]+]] {
14325 // CHECK20-NEXT:  entry:
14326 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
14327 // CHECK20-NEXT:    [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 4
14328 // CHECK20-NEXT:    [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 4
14329 // CHECK20-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
14330 // CHECK20-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
14331 // CHECK20-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
14332 // CHECK20-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i16*, align 4
14333 // CHECK20-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR1_I:%.*]] = alloca [3 x i8*]*, align 4
14334 // CHECK20-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR2_I:%.*]] = alloca [3 x i8*]*, align 4
14335 // CHECK20-NEXT:    [[DOTFIRSTPRIV_PTR_ADDR3_I:%.*]] = alloca [3 x i64]*, align 4
14336 // CHECK20-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
14337 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
14338 // CHECK20-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
14339 // CHECK20-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
14340 // CHECK20-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
14341 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
14342 // CHECK20-NEXT:    store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
14343 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
14344 // CHECK20-NEXT:    [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 4
14345 // CHECK20-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
14346 // CHECK20-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
14347 // CHECK20-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
14348 // CHECK20-NEXT:    [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 4
14349 // CHECK20-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
14350 // CHECK20-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 1
14351 // CHECK20-NEXT:    [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
14352 // CHECK20-NEXT:    [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
14353 // CHECK20-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META15:![0-9]+]])
14354 // CHECK20-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
14355 // CHECK20-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
14356 // CHECK20-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META22:![0-9]+]])
14357 // CHECK20-NEXT:    store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !24
14358 // CHECK20-NEXT:    store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 4, !noalias !24
14359 // CHECK20-NEXT:    store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
14360 // CHECK20-NEXT:    store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
14361 // CHECK20-NEXT:    store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !24
14362 // CHECK20-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
14363 // CHECK20-NEXT:    [[TMP12:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !24
14364 // CHECK20-NEXT:    [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 4, !noalias !24
14365 // CHECK20-NEXT:    [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 4, !noalias !24
14366 // CHECK20-NEXT:    [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i16**, [3 x i8*]**, [3 x i8*]**, [3 x i64]**)*
14367 // CHECK20-NEXT:    call void [[TMP15]](i8* [[TMP14]], i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]]) #[[ATTR3]]
14368 // CHECK20-NEXT:    [[TMP16:%.*]] = load i16*, i16** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 4, !noalias !24
14369 // CHECK20-NEXT:    [[TMP17:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR1_I]], align 4, !noalias !24
14370 // CHECK20-NEXT:    [[TMP18:%.*]] = load [3 x i8*]*, [3 x i8*]** [[DOTFIRSTPRIV_PTR_ADDR2_I]], align 4, !noalias !24
14371 // CHECK20-NEXT:    [[TMP19:%.*]] = load [3 x i64]*, [3 x i64]** [[DOTFIRSTPRIV_PTR_ADDR3_I]], align 4, !noalias !24
14372 // CHECK20-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP17]], i32 0, i32 0
14373 // CHECK20-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[TMP18]], i32 0, i32 0
14374 // CHECK20-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i32 0, i32 0
14375 // CHECK20-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
14376 // CHECK20-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
14377 // CHECK20-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
14378 // CHECK20-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
14379 // CHECK20-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]]) #[[ATTR3]]
14380 // CHECK20-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
14381 // CHECK20-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
14382 // CHECK20:       omp_offload.failed.i:
14383 // CHECK20-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
14384 // CHECK20-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
14385 // CHECK20-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !24
14386 // CHECK20-NEXT:    [[TMP30:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !24
14387 // CHECK20-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
14388 // CHECK20-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
14389 // CHECK20-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !24
14390 // CHECK20-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
14391 // CHECK20-NEXT:    store i32 [[TMP33]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
14392 // CHECK20-NEXT:    [[TMP34:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !24
14393 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i32 [[TMP30]], i32 [[TMP32]], i32 [[TMP34]]) #[[ATTR3]]
14394 // CHECK20-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
14395 // CHECK20:       .omp_outlined..1.exit:
14396 // CHECK20-NEXT:    ret i32 0
14397 //
14398 //
14399 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l105
14400 // CHECK20-SAME: (i32 [[A:%.*]]) #[[ATTR2]] {
14401 // CHECK20-NEXT:  entry:
14402 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
14403 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
14404 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
14405 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
14406 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
14407 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
14408 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]])
14409 // CHECK20-NEXT:    ret void
14410 //
14411 //
14412 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..2
14413 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]]) #[[ATTR2]] {
14414 // CHECK20-NEXT:  entry:
14415 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14416 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14417 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
14418 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14419 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14420 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
14421 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
14422 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
14423 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
14424 // CHECK20-NEXT:    ret void
14425 //
14426 //
14427 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
14428 // CHECK20-SAME: (i32 [[AA:%.*]]) #[[ATTR2]] {
14429 // CHECK20-NEXT:  entry:
14430 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
14431 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
14432 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
14433 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
14434 // CHECK20-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
14435 // CHECK20-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
14436 // CHECK20-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
14437 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
14438 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP1]])
14439 // CHECK20-NEXT:    ret void
14440 //
14441 //
14442 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..3
14443 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
14444 // CHECK20-NEXT:  entry:
14445 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14446 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14447 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
14448 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14449 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14450 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
14451 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
14452 // CHECK20-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
14453 // CHECK20-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
14454 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
14455 // CHECK20-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
14456 // CHECK20-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
14457 // CHECK20-NEXT:    ret void
14458 //
14459 //
14460 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
14461 // CHECK20-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
14462 // CHECK20-NEXT:  entry:
14463 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
14464 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
14465 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
14466 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
14467 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
14468 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
14469 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
14470 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
14471 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
14472 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
14473 // CHECK20-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
14474 // CHECK20-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
14475 // CHECK20-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
14476 // CHECK20-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
14477 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
14478 // CHECK20-NEXT:    ret void
14479 //
14480 //
14481 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..6
14482 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR2]] {
14483 // CHECK20-NEXT:  entry:
14484 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14485 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14486 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
14487 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
14488 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14489 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14490 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
14491 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
14492 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
14493 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
14494 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
14495 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
14496 // CHECK20-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
14497 // CHECK20-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
14498 // CHECK20-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
14499 // CHECK20-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
14500 // CHECK20-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
14501 // CHECK20-NEXT:    ret void
14502 //
14503 //
14504 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
14505 // CHECK20-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
14506 // CHECK20-NEXT:  entry:
14507 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
14508 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
14509 // CHECK20-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
14510 // CHECK20-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
14511 // CHECK20-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
14512 // CHECK20-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
14513 // CHECK20-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
14514 // CHECK20-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
14515 // CHECK20-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
14516 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
14517 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
14518 // CHECK20-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
14519 // CHECK20-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
14520 // CHECK20-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
14521 // CHECK20-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
14522 // CHECK20-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
14523 // CHECK20-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
14524 // CHECK20-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
14525 // CHECK20-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
14526 // CHECK20-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
14527 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
14528 // CHECK20-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
14529 // CHECK20-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
14530 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
14531 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
14532 // CHECK20-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
14533 // CHECK20-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
14534 // CHECK20-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
14535 // CHECK20-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
14536 // CHECK20-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
14537 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
14538 // CHECK20-NEXT:    ret void
14539 //
14540 //
14541 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..9
14542 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR2]] {
14543 // CHECK20-NEXT:  entry:
14544 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14545 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14546 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
14547 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
14548 // CHECK20-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
14549 // CHECK20-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
14550 // CHECK20-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
14551 // CHECK20-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
14552 // CHECK20-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
14553 // CHECK20-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
14554 // CHECK20-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
14555 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14556 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14557 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
14558 // CHECK20-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
14559 // CHECK20-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
14560 // CHECK20-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
14561 // CHECK20-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
14562 // CHECK20-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
14563 // CHECK20-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
14564 // CHECK20-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
14565 // CHECK20-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
14566 // CHECK20-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
14567 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
14568 // CHECK20-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
14569 // CHECK20-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
14570 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
14571 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
14572 // CHECK20-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
14573 // CHECK20-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
14574 // CHECK20-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
14575 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
14576 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
14577 // CHECK20-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
14578 // CHECK20-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
14579 // CHECK20-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
14580 // CHECK20-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
14581 // CHECK20-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
14582 // CHECK20-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
14583 // CHECK20-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
14584 // CHECK20-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
14585 // CHECK20-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
14586 // CHECK20-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
14587 // CHECK20-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
14588 // CHECK20-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
14589 // CHECK20-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
14590 // CHECK20-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
14591 // CHECK20-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
14592 // CHECK20-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
14593 // CHECK20-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
14594 // CHECK20-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
14595 // CHECK20-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
14596 // CHECK20-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
14597 // CHECK20-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
14598 // CHECK20-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
14599 // CHECK20-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
14600 // CHECK20-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
14601 // CHECK20-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
14602 // CHECK20-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
14603 // CHECK20-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
14604 // CHECK20-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
14605 // CHECK20-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
14606 // CHECK20-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
14607 // CHECK20-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
14608 // CHECK20-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
14609 // CHECK20-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
14610 // CHECK20-NEXT:    ret void
14611 //
14612 //
14613 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
14614 // CHECK20-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
14615 // CHECK20-NEXT:  entry:
14616 // CHECK20-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
14617 // CHECK20-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
14618 // CHECK20-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
14619 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
14620 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
14621 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
14622 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP1]])
14623 // CHECK20-NEXT:    ret void
14624 //
14625 //
14626 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..11
14627 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
14628 // CHECK20-NEXT:  entry:
14629 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14630 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14631 // CHECK20-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
14632 // CHECK20-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
14633 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14634 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14635 // CHECK20-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
14636 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
14637 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
14638 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
14639 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..12 to void (i32*, i32*, ...)*), i32 [[TMP1]])
14640 // CHECK20-NEXT:    ret void
14641 //
14642 //
14643 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..12
14644 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
14645 // CHECK20-NEXT:  entry:
14646 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14647 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14648 // CHECK20-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
14649 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14650 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14651 // CHECK20-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
14652 // CHECK20-NEXT:    ret void
14653 //
14654 //
14655 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
14656 // CHECK20-SAME: (i32 [[NN:%.*]]) #[[ATTR2]] {
14657 // CHECK20-NEXT:  entry:
14658 // CHECK20-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
14659 // CHECK20-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
14660 // CHECK20-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
14661 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
14662 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
14663 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
14664 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i32 [[TMP1]])
14665 // CHECK20-NEXT:    ret void
14666 //
14667 //
14668 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..15
14669 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR2]] {
14670 // CHECK20-NEXT:  entry:
14671 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14672 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14673 // CHECK20-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
14674 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14675 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14676 // CHECK20-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
14677 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..16 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
14678 // CHECK20-NEXT:    ret void
14679 //
14680 //
14681 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..16
14682 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR2]] {
14683 // CHECK20-NEXT:  entry:
14684 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14685 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14686 // CHECK20-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
14687 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14688 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14689 // CHECK20-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
14690 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
14691 // CHECK20-NEXT:    ret void
14692 //
14693 //
14694 // CHECK20-LABEL: define {{[^@]+}}@_Z6bazzzziPi
14695 // CHECK20-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
14696 // CHECK20-NEXT:  entry:
14697 // CHECK20-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
14698 // CHECK20-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
14699 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
14700 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
14701 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
14702 // CHECK20-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
14703 // CHECK20-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
14704 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
14705 // CHECK20-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
14706 // CHECK20-NEXT:    [[TMP2:%.*]] = bitcast i8** [[TMP1]] to i32*
14707 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[TMP2]], align 4
14708 // CHECK20-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
14709 // CHECK20-NEXT:    [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32*
14710 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[TMP4]], align 4
14711 // CHECK20-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
14712 // CHECK20-NEXT:    store i8* null, i8** [[TMP5]], align 4
14713 // CHECK20-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
14714 // CHECK20-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
14715 // CHECK20-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14716 // CHECK20-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
14717 // CHECK20-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
14718 // CHECK20:       omp_offload.failed:
14719 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i32 [[TMP0]]) #[[ATTR3]]
14720 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT]]
14721 // CHECK20:       omp_offload.cont:
14722 // CHECK20-NEXT:    ret void
14723 //
14724 //
14725 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
14726 // CHECK20-SAME: (i32 [[VLA:%.*]]) #[[ATTR2]] {
14727 // CHECK20-NEXT:  entry:
14728 // CHECK20-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
14729 // CHECK20-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
14730 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
14731 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..19 to void (i32*, i32*, ...)*), i32 [[TMP0]])
14732 // CHECK20-NEXT:    ret void
14733 //
14734 //
14735 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..19
14736 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR2]] {
14737 // CHECK20-NEXT:  entry:
14738 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
14739 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
14740 // CHECK20-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
14741 // CHECK20-NEXT:    [[F:%.*]] = alloca i32*, align 4
14742 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
14743 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
14744 // CHECK20-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
14745 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
14746 // CHECK20-NEXT:    ret void
14747 //
14748 //
14749 // CHECK20-LABEL: define {{[^@]+}}@_Z3bari
14750 // CHECK20-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
14751 // CHECK20-NEXT:  entry:
14752 // CHECK20-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
14753 // CHECK20-NEXT:    [[A:%.*]] = alloca i32, align 4
14754 // CHECK20-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
14755 // CHECK20-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
14756 // CHECK20-NEXT:    store i32 0, i32* [[A]], align 4
14757 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
14758 // CHECK20-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
14759 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
14760 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
14761 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
14762 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
14763 // CHECK20-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
14764 // CHECK20-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
14765 // CHECK20-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
14766 // CHECK20-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
14767 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
14768 // CHECK20-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
14769 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
14770 // CHECK20-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
14771 // CHECK20-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
14772 // CHECK20-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
14773 // CHECK20-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
14774 // CHECK20-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
14775 // CHECK20-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
14776 // CHECK20-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
14777 // CHECK20-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
14778 // CHECK20-NEXT:    ret i32 [[TMP8]]
14779 //
14780 //
14781 // CHECK20-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
14782 // CHECK20-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
14783 // CHECK20-NEXT:  entry:
14784 // CHECK20-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
14785 // CHECK20-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
14786 // CHECK20-NEXT:    [[B:%.*]] = alloca i32, align 4
14787 // CHECK20-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
14788 // CHECK20-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
14789 // CHECK20-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
14790 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [5 x i8*], align 4
14791 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [5 x i8*], align 4
14792 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [5 x i8*], align 4
14793 // CHECK20-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [5 x i64], align 4
14794 // CHECK20-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
14795 // CHECK20-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
14796 // CHECK20-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
14797 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
14798 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
14799 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
14800 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
14801 // CHECK20-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
14802 // CHECK20-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
14803 // CHECK20-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
14804 // CHECK20-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
14805 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
14806 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
14807 // CHECK20-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
14808 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
14809 // CHECK20-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
14810 // CHECK20-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 60
14811 // CHECK20-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
14812 // CHECK20:       omp_if.then:
14813 // CHECK20-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
14814 // CHECK20-NEXT:    [[TMP7:%.*]] = mul nuw i32 2, [[TMP1]]
14815 // CHECK20-NEXT:    [[TMP8:%.*]] = mul nuw i32 [[TMP7]], 2
14816 // CHECK20-NEXT:    [[TMP9:%.*]] = sext i32 [[TMP8]] to i64
14817 // CHECK20-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
14818 // CHECK20-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to %struct.S1**
14819 // CHECK20-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP11]], align 4
14820 // CHECK20-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
14821 // CHECK20-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to double**
14822 // CHECK20-NEXT:    store double* [[A]], double** [[TMP13]], align 4
14823 // CHECK20-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
14824 // CHECK20-NEXT:    store i64 8, i64* [[TMP14]], align 4
14825 // CHECK20-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
14826 // CHECK20-NEXT:    store i8* null, i8** [[TMP15]], align 4
14827 // CHECK20-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
14828 // CHECK20-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
14829 // CHECK20-NEXT:    store i32 [[TMP5]], i32* [[TMP17]], align 4
14830 // CHECK20-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
14831 // CHECK20-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
14832 // CHECK20-NEXT:    store i32 [[TMP5]], i32* [[TMP19]], align 4
14833 // CHECK20-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 1
14834 // CHECK20-NEXT:    store i64 4, i64* [[TMP20]], align 4
14835 // CHECK20-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
14836 // CHECK20-NEXT:    store i8* null, i8** [[TMP21]], align 4
14837 // CHECK20-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
14838 // CHECK20-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
14839 // CHECK20-NEXT:    store i32 2, i32* [[TMP23]], align 4
14840 // CHECK20-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
14841 // CHECK20-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
14842 // CHECK20-NEXT:    store i32 2, i32* [[TMP25]], align 4
14843 // CHECK20-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 2
14844 // CHECK20-NEXT:    store i64 4, i64* [[TMP26]], align 4
14845 // CHECK20-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
14846 // CHECK20-NEXT:    store i8* null, i8** [[TMP27]], align 4
14847 // CHECK20-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
14848 // CHECK20-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
14849 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP29]], align 4
14850 // CHECK20-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
14851 // CHECK20-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
14852 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP31]], align 4
14853 // CHECK20-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
14854 // CHECK20-NEXT:    store i64 4, i64* [[TMP32]], align 4
14855 // CHECK20-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
14856 // CHECK20-NEXT:    store i8* null, i8** [[TMP33]], align 4
14857 // CHECK20-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 4
14858 // CHECK20-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i16**
14859 // CHECK20-NEXT:    store i16* [[VLA]], i16** [[TMP35]], align 4
14860 // CHECK20-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 4
14861 // CHECK20-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i16**
14862 // CHECK20-NEXT:    store i16* [[VLA]], i16** [[TMP37]], align 4
14863 // CHECK20-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 4
14864 // CHECK20-NEXT:    store i64 [[TMP9]], i64* [[TMP38]], align 4
14865 // CHECK20-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 4
14866 // CHECK20-NEXT:    store i8* null, i8** [[TMP39]], align 4
14867 // CHECK20-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
14868 // CHECK20-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
14869 // CHECK20-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
14870 // CHECK20-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP40]], i8** [[TMP41]], i64* [[TMP42]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14871 // CHECK20-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
14872 // CHECK20-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
14873 // CHECK20:       omp_offload.failed:
14874 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
14875 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT]]
14876 // CHECK20:       omp_offload.cont:
14877 // CHECK20-NEXT:    br label [[OMP_IF_END:%.*]]
14878 // CHECK20:       omp_if.else:
14879 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
14880 // CHECK20-NEXT:    br label [[OMP_IF_END]]
14881 // CHECK20:       omp_if.end:
14882 // CHECK20-NEXT:    [[TMP45:%.*]] = mul nsw i32 1, [[TMP1]]
14883 // CHECK20-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP45]]
14884 // CHECK20-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
14885 // CHECK20-NEXT:    [[TMP46:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
14886 // CHECK20-NEXT:    [[CONV:%.*]] = sext i16 [[TMP46]] to i32
14887 // CHECK20-NEXT:    [[TMP47:%.*]] = load i32, i32* [[B]], align 4
14888 // CHECK20-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP47]]
14889 // CHECK20-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
14890 // CHECK20-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
14891 // CHECK20-NEXT:    ret i32 [[ADD3]]
14892 //
14893 //
14894 // CHECK20-LABEL: define {{[^@]+}}@_ZL7fstatici
14895 // CHECK20-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
14896 // CHECK20-NEXT:  entry:
14897 // CHECK20-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
14898 // CHECK20-NEXT:    [[A:%.*]] = alloca i32, align 4
14899 // CHECK20-NEXT:    [[AA:%.*]] = alloca i16, align 2
14900 // CHECK20-NEXT:    [[AAA:%.*]] = alloca i8, align 1
14901 // CHECK20-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
14902 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
14903 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
14904 // CHECK20-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
14905 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [4 x i8*], align 4
14906 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [4 x i8*], align 4
14907 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [4 x i8*], align 4
14908 // CHECK20-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
14909 // CHECK20-NEXT:    store i32 0, i32* [[A]], align 4
14910 // CHECK20-NEXT:    store i16 0, i16* [[AA]], align 2
14911 // CHECK20-NEXT:    store i8 0, i8* [[AAA]], align 1
14912 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
14913 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
14914 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
14915 // CHECK20-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
14916 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
14917 // CHECK20-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
14918 // CHECK20-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
14919 // CHECK20-NEXT:    [[TMP4:%.*]] = load i8, i8* [[AAA]], align 1
14920 // CHECK20-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
14921 // CHECK20-NEXT:    store i8 [[TMP4]], i8* [[CONV1]], align 1
14922 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
14923 // CHECK20-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
14924 // CHECK20-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 50
14925 // CHECK20-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
14926 // CHECK20:       omp_if.then:
14927 // CHECK20-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
14928 // CHECK20-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
14929 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
14930 // CHECK20-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
14931 // CHECK20-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
14932 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP10]], align 4
14933 // CHECK20-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
14934 // CHECK20-NEXT:    store i8* null, i8** [[TMP11]], align 4
14935 // CHECK20-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
14936 // CHECK20-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
14937 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
14938 // CHECK20-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
14939 // CHECK20-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
14940 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[TMP15]], align 4
14941 // CHECK20-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
14942 // CHECK20-NEXT:    store i8* null, i8** [[TMP16]], align 4
14943 // CHECK20-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
14944 // CHECK20-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
14945 // CHECK20-NEXT:    store i32 [[TMP5]], i32* [[TMP18]], align 4
14946 // CHECK20-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
14947 // CHECK20-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
14948 // CHECK20-NEXT:    store i32 [[TMP5]], i32* [[TMP20]], align 4
14949 // CHECK20-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
14950 // CHECK20-NEXT:    store i8* null, i8** [[TMP21]], align 4
14951 // CHECK20-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 3
14952 // CHECK20-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to [10 x i32]**
14953 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP23]], align 4
14954 // CHECK20-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 3
14955 // CHECK20-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [10 x i32]**
14956 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP25]], align 4
14957 // CHECK20-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 3
14958 // CHECK20-NEXT:    store i8* null, i8** [[TMP26]], align 4
14959 // CHECK20-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
14960 // CHECK20-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
14961 // CHECK20-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.25, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.26, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
14962 // CHECK20-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
14963 // CHECK20-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
14964 // CHECK20:       omp_offload.failed:
14965 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
14966 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT]]
14967 // CHECK20:       omp_offload.cont:
14968 // CHECK20-NEXT:    br label [[OMP_IF_END:%.*]]
14969 // CHECK20:       omp_if.else:
14970 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
14971 // CHECK20-NEXT:    br label [[OMP_IF_END]]
14972 // CHECK20:       omp_if.end:
14973 // CHECK20-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
14974 // CHECK20-NEXT:    ret i32 [[TMP31]]
14975 //
14976 //
14977 // CHECK20-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
14978 // CHECK20-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
14979 // CHECK20-NEXT:  entry:
14980 // CHECK20-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
14981 // CHECK20-NEXT:    [[A:%.*]] = alloca i32, align 4
14982 // CHECK20-NEXT:    [[AA:%.*]] = alloca i16, align 2
14983 // CHECK20-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
14984 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
14985 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
14986 // CHECK20-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [3 x i8*], align 4
14987 // CHECK20-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [3 x i8*], align 4
14988 // CHECK20-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
14989 // CHECK20-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
14990 // CHECK20-NEXT:    store i32 0, i32* [[A]], align 4
14991 // CHECK20-NEXT:    store i16 0, i16* [[AA]], align 2
14992 // CHECK20-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
14993 // CHECK20-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
14994 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
14995 // CHECK20-NEXT:    [[TMP2:%.*]] = load i16, i16* [[AA]], align 2
14996 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
14997 // CHECK20-NEXT:    store i16 [[TMP2]], i16* [[CONV]], align 2
14998 // CHECK20-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
14999 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
15000 // CHECK20-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 40
15001 // CHECK20-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
15002 // CHECK20:       omp_if.then:
15003 // CHECK20-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
15004 // CHECK20-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
15005 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP6]], align 4
15006 // CHECK20-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
15007 // CHECK20-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
15008 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[TMP8]], align 4
15009 // CHECK20-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
15010 // CHECK20-NEXT:    store i8* null, i8** [[TMP9]], align 4
15011 // CHECK20-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
15012 // CHECK20-NEXT:    [[TMP11:%.*]] = bitcast i8** [[TMP10]] to i32*
15013 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[TMP11]], align 4
15014 // CHECK20-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
15015 // CHECK20-NEXT:    [[TMP13:%.*]] = bitcast i8** [[TMP12]] to i32*
15016 // CHECK20-NEXT:    store i32 [[TMP3]], i32* [[TMP13]], align 4
15017 // CHECK20-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
15018 // CHECK20-NEXT:    store i8* null, i8** [[TMP14]], align 4
15019 // CHECK20-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
15020 // CHECK20-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to [10 x i32]**
15021 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP16]], align 4
15022 // CHECK20-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
15023 // CHECK20-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
15024 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[TMP18]], align 4
15025 // CHECK20-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
15026 // CHECK20-NEXT:    store i8* null, i8** [[TMP19]], align 4
15027 // CHECK20-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
15028 // CHECK20-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
15029 // CHECK20-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
15030 // CHECK20-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
15031 // CHECK20-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
15032 // CHECK20:       omp_offload.failed:
15033 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
15034 // CHECK20-NEXT:    br label [[OMP_OFFLOAD_CONT]]
15035 // CHECK20:       omp_offload.cont:
15036 // CHECK20-NEXT:    br label [[OMP_IF_END:%.*]]
15037 // CHECK20:       omp_if.else:
15038 // CHECK20-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
15039 // CHECK20-NEXT:    br label [[OMP_IF_END]]
15040 // CHECK20:       omp_if.end:
15041 // CHECK20-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
15042 // CHECK20-NEXT:    ret i32 [[TMP24]]
15043 //
15044 //
15045 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
15046 // CHECK20-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
15047 // CHECK20-NEXT:  entry:
15048 // CHECK20-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
15049 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
15050 // CHECK20-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
15051 // CHECK20-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
15052 // CHECK20-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
15053 // CHECK20-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
15054 // CHECK20-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
15055 // CHECK20-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
15056 // CHECK20-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
15057 // CHECK20-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
15058 // CHECK20-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
15059 // CHECK20-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
15060 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
15061 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
15062 // CHECK20-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
15063 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
15064 // CHECK20-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
15065 // CHECK20-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
15066 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..22 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
15067 // CHECK20-NEXT:    ret void
15068 //
15069 //
15070 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..22
15071 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR2]] {
15072 // CHECK20-NEXT:  entry:
15073 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
15074 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
15075 // CHECK20-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
15076 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
15077 // CHECK20-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
15078 // CHECK20-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
15079 // CHECK20-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
15080 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
15081 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
15082 // CHECK20-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
15083 // CHECK20-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
15084 // CHECK20-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
15085 // CHECK20-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
15086 // CHECK20-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
15087 // CHECK20-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
15088 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
15089 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
15090 // CHECK20-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
15091 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
15092 // CHECK20-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
15093 // CHECK20-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
15094 // CHECK20-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
15095 // CHECK20-NEXT:    store double [[ADD]], double* [[A]], align 4
15096 // CHECK20-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
15097 // CHECK20-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
15098 // CHECK20-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
15099 // CHECK20-NEXT:    store double [[INC]], double* [[A3]], align 4
15100 // CHECK20-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
15101 // CHECK20-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
15102 // CHECK20-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
15103 // CHECK20-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
15104 // CHECK20-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
15105 // CHECK20-NEXT:    ret void
15106 //
15107 //
15108 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
15109 // CHECK20-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
15110 // CHECK20-NEXT:  entry:
15111 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
15112 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
15113 // CHECK20-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
15114 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
15115 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
15116 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
15117 // CHECK20-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
15118 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
15119 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
15120 // CHECK20-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
15121 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
15122 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
15123 // CHECK20-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
15124 // CHECK20-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
15125 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
15126 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
15127 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
15128 // CHECK20-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
15129 // CHECK20-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
15130 // CHECK20-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
15131 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
15132 // CHECK20-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
15133 // CHECK20-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
15134 // CHECK20-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
15135 // CHECK20-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
15136 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
15137 // CHECK20-NEXT:    ret void
15138 //
15139 //
15140 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..24
15141 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
15142 // CHECK20-NEXT:  entry:
15143 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
15144 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
15145 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
15146 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
15147 // CHECK20-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
15148 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
15149 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
15150 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
15151 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
15152 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
15153 // CHECK20-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
15154 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
15155 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
15156 // CHECK20-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
15157 // CHECK20-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
15158 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
15159 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
15160 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
15161 // CHECK20-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
15162 // CHECK20-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
15163 // CHECK20-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
15164 // CHECK20-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
15165 // CHECK20-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
15166 // CHECK20-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
15167 // CHECK20-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
15168 // CHECK20-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
15169 // CHECK20-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
15170 // CHECK20-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
15171 // CHECK20-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
15172 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15173 // CHECK20-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
15174 // CHECK20-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
15175 // CHECK20-NEXT:    ret void
15176 //
15177 //
15178 // CHECK20-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
15179 // CHECK20-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
15180 // CHECK20-NEXT:  entry:
15181 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
15182 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
15183 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
15184 // CHECK20-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
15185 // CHECK20-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
15186 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
15187 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
15188 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
15189 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
15190 // CHECK20-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
15191 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
15192 // CHECK20-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
15193 // CHECK20-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
15194 // CHECK20-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
15195 // CHECK20-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
15196 // CHECK20-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
15197 // CHECK20-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
15198 // CHECK20-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..27 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
15199 // CHECK20-NEXT:    ret void
15200 //
15201 //
15202 // CHECK20-LABEL: define {{[^@]+}}@.omp_outlined..27
15203 // CHECK20-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR2]] {
15204 // CHECK20-NEXT:  entry:
15205 // CHECK20-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
15206 // CHECK20-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
15207 // CHECK20-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
15208 // CHECK20-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
15209 // CHECK20-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
15210 // CHECK20-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
15211 // CHECK20-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
15212 // CHECK20-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
15213 // CHECK20-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
15214 // CHECK20-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
15215 // CHECK20-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
15216 // CHECK20-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
15217 // CHECK20-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
15218 // CHECK20-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
15219 // CHECK20-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
15220 // CHECK20-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
15221 // CHECK20-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
15222 // CHECK20-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
15223 // CHECK20-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
15224 // CHECK20-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
15225 // CHECK20-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
15226 // CHECK20-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15227 // CHECK20-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
15228 // CHECK20-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
15229 // CHECK20-NEXT:    ret void
15230 //
15231 //
15232 // CHECK20-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
15233 // CHECK20-SAME: () #[[ATTR4]] {
15234 // CHECK20-NEXT:  entry:
15235 // CHECK20-NEXT:    call void @__tgt_register_requires(i64 1)
15236 // CHECK20-NEXT:    ret void
15237 //
15238 //
15239 // CHECK21-LABEL: define {{[^@]+}}@_Z3fooi
15240 // CHECK21-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
15241 // CHECK21-NEXT:  entry:
15242 // CHECK21-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15243 // CHECK21-NEXT:    [[A:%.*]] = alloca i32, align 4
15244 // CHECK21-NEXT:    [[AA:%.*]] = alloca i16, align 2
15245 // CHECK21-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
15246 // CHECK21-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
15247 // CHECK21-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
15248 // CHECK21-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
15249 // CHECK21-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
15250 // CHECK21-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
15251 // CHECK21-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
15252 // CHECK21-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
15253 // CHECK21-NEXT:    [[NN:%.*]] = alloca i32, align 4
15254 // CHECK21-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15255 // CHECK21-NEXT:    store i32 0, i32* [[A]], align 4
15256 // CHECK21-NEXT:    store i16 0, i16* [[AA]], align 2
15257 // CHECK21-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15258 // CHECK21-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
15259 // CHECK21-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
15260 // CHECK21-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
15261 // CHECK21-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
15262 // CHECK21-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
15263 // CHECK21-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
15264 // CHECK21-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
15265 // CHECK21-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
15266 // CHECK21-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
15267 // CHECK21-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
15268 // CHECK21-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
15269 // CHECK21-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
15270 // CHECK21-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
15271 // CHECK21-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
15272 // CHECK21-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
15273 // CHECK21-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
15274 // CHECK21-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15275 // CHECK21-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
15276 // CHECK21-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
15277 // CHECK21-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
15278 // CHECK21-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
15279 // CHECK21-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
15280 // CHECK21-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
15281 // CHECK21-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
15282 // CHECK21-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
15283 // CHECK21-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
15284 // CHECK21-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
15285 // CHECK21-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
15286 // CHECK21-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
15287 // CHECK21-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
15288 // CHECK21-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
15289 // CHECK21-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
15290 // CHECK21-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
15291 // CHECK21-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
15292 // CHECK21-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
15293 // CHECK21-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
15294 // CHECK21-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
15295 // CHECK21-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
15296 // CHECK21-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
15297 // CHECK21-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
15298 // CHECK21-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
15299 // CHECK21-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
15300 // CHECK21-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
15301 // CHECK21-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
15302 // CHECK21-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
15303 // CHECK21-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
15304 // CHECK21-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
15305 // CHECK21-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
15306 // CHECK21-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
15307 // CHECK21-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
15308 // CHECK21-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
15309 // CHECK21-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
15310 // CHECK21-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
15311 // CHECK21-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
15312 // CHECK21-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
15313 // CHECK21-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
15314 // CHECK21-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
15315 // CHECK21-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
15316 // CHECK21-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
15317 // CHECK21-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
15318 // CHECK21-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
15319 // CHECK21-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
15320 // CHECK21-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
15321 // CHECK21-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
15322 // CHECK21-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
15323 // CHECK21-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
15324 // CHECK21-NEXT:    store i32 0, i32* [[NN]], align 4
15325 // CHECK21-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
15326 // CHECK21-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
15327 // CHECK21-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
15328 // CHECK21-NEXT:    ret i32 [[TMP20]]
15329 //
15330 //
15331 // CHECK21-LABEL: define {{[^@]+}}@_Z6bazzzziPi
15332 // CHECK21-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
15333 // CHECK21-NEXT:  entry:
15334 // CHECK21-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15335 // CHECK21-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
15336 // CHECK21-NEXT:    [[F1:%.*]] = alloca i32*, align 8
15337 // CHECK21-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15338 // CHECK21-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
15339 // CHECK21-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15340 // CHECK21-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
15341 // CHECK21-NEXT:    ret void
15342 //
15343 //
15344 // CHECK21-LABEL: define {{[^@]+}}@_Z3bari
15345 // CHECK21-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
15346 // CHECK21-NEXT:  entry:
15347 // CHECK21-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15348 // CHECK21-NEXT:    [[A:%.*]] = alloca i32, align 4
15349 // CHECK21-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
15350 // CHECK21-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15351 // CHECK21-NEXT:    store i32 0, i32* [[A]], align 4
15352 // CHECK21-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15353 // CHECK21-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
15354 // CHECK21-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
15355 // CHECK21-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
15356 // CHECK21-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15357 // CHECK21-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
15358 // CHECK21-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
15359 // CHECK21-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
15360 // CHECK21-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
15361 // CHECK21-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
15362 // CHECK21-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
15363 // CHECK21-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
15364 // CHECK21-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
15365 // CHECK21-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
15366 // CHECK21-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
15367 // CHECK21-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
15368 // CHECK21-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
15369 // CHECK21-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
15370 // CHECK21-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
15371 // CHECK21-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
15372 // CHECK21-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
15373 // CHECK21-NEXT:    ret i32 [[TMP8]]
15374 //
15375 //
15376 // CHECK21-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
15377 // CHECK21-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
15378 // CHECK21-NEXT:  entry:
15379 // CHECK21-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
15380 // CHECK21-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15381 // CHECK21-NEXT:    [[B:%.*]] = alloca i32, align 4
15382 // CHECK21-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
15383 // CHECK21-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
15384 // CHECK21-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
15385 // CHECK21-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15386 // CHECK21-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
15387 // CHECK21-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15388 // CHECK21-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15389 // CHECK21-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
15390 // CHECK21-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
15391 // CHECK21-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
15392 // CHECK21-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
15393 // CHECK21-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
15394 // CHECK21-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
15395 // CHECK21-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
15396 // CHECK21-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
15397 // CHECK21-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
15398 // CHECK21-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
15399 // CHECK21-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
15400 // CHECK21-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
15401 // CHECK21-NEXT:    store double [[ADD2]], double* [[A]], align 8
15402 // CHECK21-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
15403 // CHECK21-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
15404 // CHECK21-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
15405 // CHECK21-NEXT:    store double [[INC]], double* [[A3]], align 8
15406 // CHECK21-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
15407 // CHECK21-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
15408 // CHECK21-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
15409 // CHECK21-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
15410 // CHECK21-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
15411 // CHECK21-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
15412 // CHECK21-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
15413 // CHECK21-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
15414 // CHECK21-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
15415 // CHECK21-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
15416 // CHECK21-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
15417 // CHECK21-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
15418 // CHECK21-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
15419 // CHECK21-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
15420 // CHECK21-NEXT:    ret i32 [[ADD9]]
15421 //
15422 //
15423 // CHECK21-LABEL: define {{[^@]+}}@_ZL7fstatici
15424 // CHECK21-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
15425 // CHECK21-NEXT:  entry:
15426 // CHECK21-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15427 // CHECK21-NEXT:    [[A:%.*]] = alloca i32, align 4
15428 // CHECK21-NEXT:    [[AA:%.*]] = alloca i16, align 2
15429 // CHECK21-NEXT:    [[AAA:%.*]] = alloca i8, align 1
15430 // CHECK21-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
15431 // CHECK21-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15432 // CHECK21-NEXT:    store i32 0, i32* [[A]], align 4
15433 // CHECK21-NEXT:    store i16 0, i16* [[AA]], align 2
15434 // CHECK21-NEXT:    store i8 0, i8* [[AAA]], align 1
15435 // CHECK21-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
15436 // CHECK21-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15437 // CHECK21-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15438 // CHECK21-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
15439 // CHECK21-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
15440 // CHECK21-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
15441 // CHECK21-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
15442 // CHECK21-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
15443 // CHECK21-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
15444 // CHECK21-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
15445 // CHECK21-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
15446 // CHECK21-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
15447 // CHECK21-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
15448 // CHECK21-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
15449 // CHECK21-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15450 // CHECK21-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
15451 // CHECK21-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
15452 // CHECK21-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
15453 // CHECK21-NEXT:    ret i32 [[TMP4]]
15454 //
15455 //
15456 // CHECK21-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
15457 // CHECK21-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
15458 // CHECK21-NEXT:  entry:
15459 // CHECK21-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15460 // CHECK21-NEXT:    [[A:%.*]] = alloca i32, align 4
15461 // CHECK21-NEXT:    [[AA:%.*]] = alloca i16, align 2
15462 // CHECK21-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
15463 // CHECK21-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15464 // CHECK21-NEXT:    store i32 0, i32* [[A]], align 4
15465 // CHECK21-NEXT:    store i16 0, i16* [[AA]], align 2
15466 // CHECK21-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
15467 // CHECK21-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15468 // CHECK21-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15469 // CHECK21-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
15470 // CHECK21-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
15471 // CHECK21-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
15472 // CHECK21-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
15473 // CHECK21-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
15474 // CHECK21-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
15475 // CHECK21-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15476 // CHECK21-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
15477 // CHECK21-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
15478 // CHECK21-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
15479 // CHECK21-NEXT:    ret i32 [[TMP3]]
15480 //
15481 //
15482 // CHECK22-LABEL: define {{[^@]+}}@_Z3fooi
15483 // CHECK22-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
15484 // CHECK22-NEXT:  entry:
15485 // CHECK22-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15486 // CHECK22-NEXT:    [[A:%.*]] = alloca i32, align 4
15487 // CHECK22-NEXT:    [[AA:%.*]] = alloca i16, align 2
15488 // CHECK22-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
15489 // CHECK22-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
15490 // CHECK22-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
15491 // CHECK22-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
15492 // CHECK22-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
15493 // CHECK22-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
15494 // CHECK22-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
15495 // CHECK22-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
15496 // CHECK22-NEXT:    [[NN:%.*]] = alloca i32, align 4
15497 // CHECK22-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15498 // CHECK22-NEXT:    store i32 0, i32* [[A]], align 4
15499 // CHECK22-NEXT:    store i16 0, i16* [[AA]], align 2
15500 // CHECK22-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15501 // CHECK22-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
15502 // CHECK22-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
15503 // CHECK22-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
15504 // CHECK22-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
15505 // CHECK22-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
15506 // CHECK22-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
15507 // CHECK22-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
15508 // CHECK22-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
15509 // CHECK22-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
15510 // CHECK22-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
15511 // CHECK22-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
15512 // CHECK22-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
15513 // CHECK22-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
15514 // CHECK22-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
15515 // CHECK22-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
15516 // CHECK22-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
15517 // CHECK22-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15518 // CHECK22-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
15519 // CHECK22-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
15520 // CHECK22-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
15521 // CHECK22-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
15522 // CHECK22-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
15523 // CHECK22-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
15524 // CHECK22-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
15525 // CHECK22-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
15526 // CHECK22-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
15527 // CHECK22-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
15528 // CHECK22-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
15529 // CHECK22-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
15530 // CHECK22-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
15531 // CHECK22-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
15532 // CHECK22-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
15533 // CHECK22-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
15534 // CHECK22-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
15535 // CHECK22-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
15536 // CHECK22-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
15537 // CHECK22-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
15538 // CHECK22-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
15539 // CHECK22-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
15540 // CHECK22-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
15541 // CHECK22-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
15542 // CHECK22-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
15543 // CHECK22-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
15544 // CHECK22-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
15545 // CHECK22-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
15546 // CHECK22-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
15547 // CHECK22-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
15548 // CHECK22-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
15549 // CHECK22-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
15550 // CHECK22-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
15551 // CHECK22-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
15552 // CHECK22-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
15553 // CHECK22-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
15554 // CHECK22-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
15555 // CHECK22-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
15556 // CHECK22-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
15557 // CHECK22-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
15558 // CHECK22-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
15559 // CHECK22-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
15560 // CHECK22-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
15561 // CHECK22-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
15562 // CHECK22-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
15563 // CHECK22-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
15564 // CHECK22-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
15565 // CHECK22-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
15566 // CHECK22-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
15567 // CHECK22-NEXT:    store i32 0, i32* [[NN]], align 4
15568 // CHECK22-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
15569 // CHECK22-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
15570 // CHECK22-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
15571 // CHECK22-NEXT:    ret i32 [[TMP20]]
15572 //
15573 //
15574 // CHECK22-LABEL: define {{[^@]+}}@_Z6bazzzziPi
15575 // CHECK22-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
15576 // CHECK22-NEXT:  entry:
15577 // CHECK22-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15578 // CHECK22-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
15579 // CHECK22-NEXT:    [[F1:%.*]] = alloca i32*, align 8
15580 // CHECK22-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15581 // CHECK22-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
15582 // CHECK22-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15583 // CHECK22-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
15584 // CHECK22-NEXT:    ret void
15585 //
15586 //
15587 // CHECK22-LABEL: define {{[^@]+}}@_Z3bari
15588 // CHECK22-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
15589 // CHECK22-NEXT:  entry:
15590 // CHECK22-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15591 // CHECK22-NEXT:    [[A:%.*]] = alloca i32, align 4
15592 // CHECK22-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
15593 // CHECK22-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15594 // CHECK22-NEXT:    store i32 0, i32* [[A]], align 4
15595 // CHECK22-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15596 // CHECK22-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
15597 // CHECK22-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
15598 // CHECK22-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
15599 // CHECK22-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15600 // CHECK22-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
15601 // CHECK22-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
15602 // CHECK22-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
15603 // CHECK22-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
15604 // CHECK22-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
15605 // CHECK22-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
15606 // CHECK22-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
15607 // CHECK22-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
15608 // CHECK22-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
15609 // CHECK22-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
15610 // CHECK22-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
15611 // CHECK22-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
15612 // CHECK22-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
15613 // CHECK22-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
15614 // CHECK22-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
15615 // CHECK22-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
15616 // CHECK22-NEXT:    ret i32 [[TMP8]]
15617 //
15618 //
15619 // CHECK22-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
15620 // CHECK22-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
15621 // CHECK22-NEXT:  entry:
15622 // CHECK22-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
15623 // CHECK22-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15624 // CHECK22-NEXT:    [[B:%.*]] = alloca i32, align 4
15625 // CHECK22-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
15626 // CHECK22-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
15627 // CHECK22-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
15628 // CHECK22-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15629 // CHECK22-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
15630 // CHECK22-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15631 // CHECK22-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15632 // CHECK22-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
15633 // CHECK22-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
15634 // CHECK22-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
15635 // CHECK22-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
15636 // CHECK22-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
15637 // CHECK22-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
15638 // CHECK22-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
15639 // CHECK22-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
15640 // CHECK22-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
15641 // CHECK22-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
15642 // CHECK22-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
15643 // CHECK22-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
15644 // CHECK22-NEXT:    store double [[ADD2]], double* [[A]], align 8
15645 // CHECK22-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
15646 // CHECK22-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
15647 // CHECK22-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
15648 // CHECK22-NEXT:    store double [[INC]], double* [[A3]], align 8
15649 // CHECK22-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
15650 // CHECK22-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
15651 // CHECK22-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
15652 // CHECK22-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
15653 // CHECK22-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
15654 // CHECK22-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
15655 // CHECK22-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
15656 // CHECK22-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
15657 // CHECK22-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
15658 // CHECK22-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
15659 // CHECK22-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
15660 // CHECK22-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
15661 // CHECK22-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
15662 // CHECK22-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
15663 // CHECK22-NEXT:    ret i32 [[ADD9]]
15664 //
15665 //
15666 // CHECK22-LABEL: define {{[^@]+}}@_ZL7fstatici
15667 // CHECK22-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
15668 // CHECK22-NEXT:  entry:
15669 // CHECK22-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15670 // CHECK22-NEXT:    [[A:%.*]] = alloca i32, align 4
15671 // CHECK22-NEXT:    [[AA:%.*]] = alloca i16, align 2
15672 // CHECK22-NEXT:    [[AAA:%.*]] = alloca i8, align 1
15673 // CHECK22-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
15674 // CHECK22-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15675 // CHECK22-NEXT:    store i32 0, i32* [[A]], align 4
15676 // CHECK22-NEXT:    store i16 0, i16* [[AA]], align 2
15677 // CHECK22-NEXT:    store i8 0, i8* [[AAA]], align 1
15678 // CHECK22-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
15679 // CHECK22-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15680 // CHECK22-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15681 // CHECK22-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
15682 // CHECK22-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
15683 // CHECK22-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
15684 // CHECK22-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
15685 // CHECK22-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
15686 // CHECK22-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
15687 // CHECK22-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
15688 // CHECK22-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
15689 // CHECK22-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
15690 // CHECK22-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
15691 // CHECK22-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
15692 // CHECK22-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15693 // CHECK22-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
15694 // CHECK22-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
15695 // CHECK22-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
15696 // CHECK22-NEXT:    ret i32 [[TMP4]]
15697 //
15698 //
15699 // CHECK22-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
15700 // CHECK22-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
15701 // CHECK22-NEXT:  entry:
15702 // CHECK22-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15703 // CHECK22-NEXT:    [[A:%.*]] = alloca i32, align 4
15704 // CHECK22-NEXT:    [[AA:%.*]] = alloca i16, align 2
15705 // CHECK22-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
15706 // CHECK22-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15707 // CHECK22-NEXT:    store i32 0, i32* [[A]], align 4
15708 // CHECK22-NEXT:    store i16 0, i16* [[AA]], align 2
15709 // CHECK22-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
15710 // CHECK22-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15711 // CHECK22-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15712 // CHECK22-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
15713 // CHECK22-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
15714 // CHECK22-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
15715 // CHECK22-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
15716 // CHECK22-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
15717 // CHECK22-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
15718 // CHECK22-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15719 // CHECK22-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
15720 // CHECK22-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
15721 // CHECK22-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
15722 // CHECK22-NEXT:    ret i32 [[TMP3]]
15723 //
15724 //
15725 // CHECK23-LABEL: define {{[^@]+}}@_Z3fooi
15726 // CHECK23-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
15727 // CHECK23-NEXT:  entry:
15728 // CHECK23-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15729 // CHECK23-NEXT:    [[A:%.*]] = alloca i32, align 4
15730 // CHECK23-NEXT:    [[AA:%.*]] = alloca i16, align 2
15731 // CHECK23-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
15732 // CHECK23-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
15733 // CHECK23-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
15734 // CHECK23-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
15735 // CHECK23-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
15736 // CHECK23-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
15737 // CHECK23-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
15738 // CHECK23-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
15739 // CHECK23-NEXT:    [[NN:%.*]] = alloca i32, align 4
15740 // CHECK23-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15741 // CHECK23-NEXT:    store i32 0, i32* [[A]], align 4
15742 // CHECK23-NEXT:    store i16 0, i16* [[AA]], align 2
15743 // CHECK23-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15744 // CHECK23-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
15745 // CHECK23-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
15746 // CHECK23-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
15747 // CHECK23-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
15748 // CHECK23-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
15749 // CHECK23-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
15750 // CHECK23-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
15751 // CHECK23-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
15752 // CHECK23-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
15753 // CHECK23-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
15754 // CHECK23-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
15755 // CHECK23-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
15756 // CHECK23-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
15757 // CHECK23-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
15758 // CHECK23-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15759 // CHECK23-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
15760 // CHECK23-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
15761 // CHECK23-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
15762 // CHECK23-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
15763 // CHECK23-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
15764 // CHECK23-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
15765 // CHECK23-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
15766 // CHECK23-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
15767 // CHECK23-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
15768 // CHECK23-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
15769 // CHECK23-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
15770 // CHECK23-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
15771 // CHECK23-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
15772 // CHECK23-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
15773 // CHECK23-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
15774 // CHECK23-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
15775 // CHECK23-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
15776 // CHECK23-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
15777 // CHECK23-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
15778 // CHECK23-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
15779 // CHECK23-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
15780 // CHECK23-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
15781 // CHECK23-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
15782 // CHECK23-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
15783 // CHECK23-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
15784 // CHECK23-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
15785 // CHECK23-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
15786 // CHECK23-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
15787 // CHECK23-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
15788 // CHECK23-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
15789 // CHECK23-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
15790 // CHECK23-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
15791 // CHECK23-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
15792 // CHECK23-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
15793 // CHECK23-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
15794 // CHECK23-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
15795 // CHECK23-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
15796 // CHECK23-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
15797 // CHECK23-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
15798 // CHECK23-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
15799 // CHECK23-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
15800 // CHECK23-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
15801 // CHECK23-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
15802 // CHECK23-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
15803 // CHECK23-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
15804 // CHECK23-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
15805 // CHECK23-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
15806 // CHECK23-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
15807 // CHECK23-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
15808 // CHECK23-NEXT:    store i32 0, i32* [[NN]], align 4
15809 // CHECK23-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
15810 // CHECK23-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
15811 // CHECK23-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
15812 // CHECK23-NEXT:    ret i32 [[TMP18]]
15813 //
15814 //
15815 // CHECK23-LABEL: define {{[^@]+}}@_Z6bazzzziPi
15816 // CHECK23-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
15817 // CHECK23-NEXT:  entry:
15818 // CHECK23-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15819 // CHECK23-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
15820 // CHECK23-NEXT:    [[F1:%.*]] = alloca i32*, align 4
15821 // CHECK23-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15822 // CHECK23-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
15823 // CHECK23-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15824 // CHECK23-NEXT:    ret void
15825 //
15826 //
15827 // CHECK23-LABEL: define {{[^@]+}}@_Z3bari
15828 // CHECK23-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
15829 // CHECK23-NEXT:  entry:
15830 // CHECK23-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15831 // CHECK23-NEXT:    [[A:%.*]] = alloca i32, align 4
15832 // CHECK23-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
15833 // CHECK23-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15834 // CHECK23-NEXT:    store i32 0, i32* [[A]], align 4
15835 // CHECK23-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15836 // CHECK23-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
15837 // CHECK23-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
15838 // CHECK23-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
15839 // CHECK23-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15840 // CHECK23-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
15841 // CHECK23-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
15842 // CHECK23-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
15843 // CHECK23-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
15844 // CHECK23-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
15845 // CHECK23-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
15846 // CHECK23-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
15847 // CHECK23-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
15848 // CHECK23-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
15849 // CHECK23-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
15850 // CHECK23-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
15851 // CHECK23-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
15852 // CHECK23-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
15853 // CHECK23-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
15854 // CHECK23-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
15855 // CHECK23-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
15856 // CHECK23-NEXT:    ret i32 [[TMP8]]
15857 //
15858 //
15859 // CHECK23-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
15860 // CHECK23-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
15861 // CHECK23-NEXT:  entry:
15862 // CHECK23-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
15863 // CHECK23-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15864 // CHECK23-NEXT:    [[B:%.*]] = alloca i32, align 4
15865 // CHECK23-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
15866 // CHECK23-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
15867 // CHECK23-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
15868 // CHECK23-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15869 // CHECK23-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
15870 // CHECK23-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15871 // CHECK23-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15872 // CHECK23-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
15873 // CHECK23-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
15874 // CHECK23-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
15875 // CHECK23-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
15876 // CHECK23-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
15877 // CHECK23-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
15878 // CHECK23-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
15879 // CHECK23-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
15880 // CHECK23-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
15881 // CHECK23-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
15882 // CHECK23-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
15883 // CHECK23-NEXT:    store double [[ADD2]], double* [[A]], align 4
15884 // CHECK23-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
15885 // CHECK23-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
15886 // CHECK23-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
15887 // CHECK23-NEXT:    store double [[INC]], double* [[A3]], align 4
15888 // CHECK23-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
15889 // CHECK23-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
15890 // CHECK23-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
15891 // CHECK23-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
15892 // CHECK23-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
15893 // CHECK23-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
15894 // CHECK23-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
15895 // CHECK23-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
15896 // CHECK23-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
15897 // CHECK23-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
15898 // CHECK23-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
15899 // CHECK23-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
15900 // CHECK23-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
15901 // CHECK23-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
15902 // CHECK23-NEXT:    ret i32 [[ADD9]]
15903 //
15904 //
15905 // CHECK23-LABEL: define {{[^@]+}}@_ZL7fstatici
15906 // CHECK23-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
15907 // CHECK23-NEXT:  entry:
15908 // CHECK23-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15909 // CHECK23-NEXT:    [[A:%.*]] = alloca i32, align 4
15910 // CHECK23-NEXT:    [[AA:%.*]] = alloca i16, align 2
15911 // CHECK23-NEXT:    [[AAA:%.*]] = alloca i8, align 1
15912 // CHECK23-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
15913 // CHECK23-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15914 // CHECK23-NEXT:    store i32 0, i32* [[A]], align 4
15915 // CHECK23-NEXT:    store i16 0, i16* [[AA]], align 2
15916 // CHECK23-NEXT:    store i8 0, i8* [[AAA]], align 1
15917 // CHECK23-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
15918 // CHECK23-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15919 // CHECK23-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15920 // CHECK23-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
15921 // CHECK23-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
15922 // CHECK23-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
15923 // CHECK23-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
15924 // CHECK23-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
15925 // CHECK23-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
15926 // CHECK23-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
15927 // CHECK23-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
15928 // CHECK23-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
15929 // CHECK23-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
15930 // CHECK23-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
15931 // CHECK23-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15932 // CHECK23-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
15933 // CHECK23-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
15934 // CHECK23-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
15935 // CHECK23-NEXT:    ret i32 [[TMP4]]
15936 //
15937 //
15938 // CHECK23-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
15939 // CHECK23-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
15940 // CHECK23-NEXT:  entry:
15941 // CHECK23-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15942 // CHECK23-NEXT:    [[A:%.*]] = alloca i32, align 4
15943 // CHECK23-NEXT:    [[AA:%.*]] = alloca i16, align 2
15944 // CHECK23-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
15945 // CHECK23-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15946 // CHECK23-NEXT:    store i32 0, i32* [[A]], align 4
15947 // CHECK23-NEXT:    store i16 0, i16* [[AA]], align 2
15948 // CHECK23-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
15949 // CHECK23-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
15950 // CHECK23-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15951 // CHECK23-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
15952 // CHECK23-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
15953 // CHECK23-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
15954 // CHECK23-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
15955 // CHECK23-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
15956 // CHECK23-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
15957 // CHECK23-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
15958 // CHECK23-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
15959 // CHECK23-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
15960 // CHECK23-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
15961 // CHECK23-NEXT:    ret i32 [[TMP3]]
15962 //
15963 //
15964 // CHECK24-LABEL: define {{[^@]+}}@_Z3fooi
15965 // CHECK24-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
15966 // CHECK24-NEXT:  entry:
15967 // CHECK24-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
15968 // CHECK24-NEXT:    [[A:%.*]] = alloca i32, align 4
15969 // CHECK24-NEXT:    [[AA:%.*]] = alloca i16, align 2
15970 // CHECK24-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
15971 // CHECK24-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
15972 // CHECK24-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
15973 // CHECK24-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
15974 // CHECK24-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
15975 // CHECK24-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
15976 // CHECK24-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
15977 // CHECK24-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
15978 // CHECK24-NEXT:    [[NN:%.*]] = alloca i32, align 4
15979 // CHECK24-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
15980 // CHECK24-NEXT:    store i32 0, i32* [[A]], align 4
15981 // CHECK24-NEXT:    store i16 0, i16* [[AA]], align 2
15982 // CHECK24-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
15983 // CHECK24-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
15984 // CHECK24-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
15985 // CHECK24-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
15986 // CHECK24-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
15987 // CHECK24-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
15988 // CHECK24-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
15989 // CHECK24-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
15990 // CHECK24-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
15991 // CHECK24-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
15992 // CHECK24-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
15993 // CHECK24-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
15994 // CHECK24-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
15995 // CHECK24-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
15996 // CHECK24-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
15997 // CHECK24-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
15998 // CHECK24-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
15999 // CHECK24-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
16000 // CHECK24-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
16001 // CHECK24-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
16002 // CHECK24-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
16003 // CHECK24-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
16004 // CHECK24-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
16005 // CHECK24-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
16006 // CHECK24-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
16007 // CHECK24-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
16008 // CHECK24-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
16009 // CHECK24-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
16010 // CHECK24-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
16011 // CHECK24-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
16012 // CHECK24-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
16013 // CHECK24-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
16014 // CHECK24-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
16015 // CHECK24-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
16016 // CHECK24-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
16017 // CHECK24-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
16018 // CHECK24-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
16019 // CHECK24-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
16020 // CHECK24-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
16021 // CHECK24-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
16022 // CHECK24-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
16023 // CHECK24-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
16024 // CHECK24-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
16025 // CHECK24-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
16026 // CHECK24-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
16027 // CHECK24-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
16028 // CHECK24-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
16029 // CHECK24-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
16030 // CHECK24-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
16031 // CHECK24-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
16032 // CHECK24-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
16033 // CHECK24-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
16034 // CHECK24-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
16035 // CHECK24-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
16036 // CHECK24-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
16037 // CHECK24-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
16038 // CHECK24-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
16039 // CHECK24-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
16040 // CHECK24-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
16041 // CHECK24-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
16042 // CHECK24-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
16043 // CHECK24-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
16044 // CHECK24-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
16045 // CHECK24-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
16046 // CHECK24-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
16047 // CHECK24-NEXT:    store i32 0, i32* [[NN]], align 4
16048 // CHECK24-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
16049 // CHECK24-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
16050 // CHECK24-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
16051 // CHECK24-NEXT:    ret i32 [[TMP18]]
16052 //
16053 //
16054 // CHECK24-LABEL: define {{[^@]+}}@_Z6bazzzziPi
16055 // CHECK24-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
16056 // CHECK24-NEXT:  entry:
16057 // CHECK24-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
16058 // CHECK24-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
16059 // CHECK24-NEXT:    [[F1:%.*]] = alloca i32*, align 4
16060 // CHECK24-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
16061 // CHECK24-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
16062 // CHECK24-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
16063 // CHECK24-NEXT:    ret void
16064 //
16065 //
16066 // CHECK24-LABEL: define {{[^@]+}}@_Z3bari
16067 // CHECK24-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
16068 // CHECK24-NEXT:  entry:
16069 // CHECK24-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
16070 // CHECK24-NEXT:    [[A:%.*]] = alloca i32, align 4
16071 // CHECK24-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
16072 // CHECK24-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
16073 // CHECK24-NEXT:    store i32 0, i32* [[A]], align 4
16074 // CHECK24-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
16075 // CHECK24-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
16076 // CHECK24-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
16077 // CHECK24-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
16078 // CHECK24-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
16079 // CHECK24-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
16080 // CHECK24-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
16081 // CHECK24-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
16082 // CHECK24-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
16083 // CHECK24-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
16084 // CHECK24-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
16085 // CHECK24-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
16086 // CHECK24-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
16087 // CHECK24-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
16088 // CHECK24-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
16089 // CHECK24-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
16090 // CHECK24-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
16091 // CHECK24-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
16092 // CHECK24-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
16093 // CHECK24-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
16094 // CHECK24-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
16095 // CHECK24-NEXT:    ret i32 [[TMP8]]
16096 //
16097 //
16098 // CHECK24-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
16099 // CHECK24-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
16100 // CHECK24-NEXT:  entry:
16101 // CHECK24-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
16102 // CHECK24-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
16103 // CHECK24-NEXT:    [[B:%.*]] = alloca i32, align 4
16104 // CHECK24-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
16105 // CHECK24-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
16106 // CHECK24-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
16107 // CHECK24-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
16108 // CHECK24-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
16109 // CHECK24-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
16110 // CHECK24-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
16111 // CHECK24-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
16112 // CHECK24-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
16113 // CHECK24-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
16114 // CHECK24-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
16115 // CHECK24-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
16116 // CHECK24-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
16117 // CHECK24-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
16118 // CHECK24-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
16119 // CHECK24-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
16120 // CHECK24-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
16121 // CHECK24-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
16122 // CHECK24-NEXT:    store double [[ADD2]], double* [[A]], align 4
16123 // CHECK24-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
16124 // CHECK24-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
16125 // CHECK24-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
16126 // CHECK24-NEXT:    store double [[INC]], double* [[A3]], align 4
16127 // CHECK24-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
16128 // CHECK24-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
16129 // CHECK24-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
16130 // CHECK24-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
16131 // CHECK24-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
16132 // CHECK24-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
16133 // CHECK24-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
16134 // CHECK24-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
16135 // CHECK24-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
16136 // CHECK24-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
16137 // CHECK24-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
16138 // CHECK24-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
16139 // CHECK24-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
16140 // CHECK24-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
16141 // CHECK24-NEXT:    ret i32 [[ADD9]]
16142 //
16143 //
16144 // CHECK24-LABEL: define {{[^@]+}}@_ZL7fstatici
16145 // CHECK24-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
16146 // CHECK24-NEXT:  entry:
16147 // CHECK24-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
16148 // CHECK24-NEXT:    [[A:%.*]] = alloca i32, align 4
16149 // CHECK24-NEXT:    [[AA:%.*]] = alloca i16, align 2
16150 // CHECK24-NEXT:    [[AAA:%.*]] = alloca i8, align 1
16151 // CHECK24-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
16152 // CHECK24-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
16153 // CHECK24-NEXT:    store i32 0, i32* [[A]], align 4
16154 // CHECK24-NEXT:    store i16 0, i16* [[AA]], align 2
16155 // CHECK24-NEXT:    store i8 0, i8* [[AAA]], align 1
16156 // CHECK24-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
16157 // CHECK24-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
16158 // CHECK24-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
16159 // CHECK24-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
16160 // CHECK24-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
16161 // CHECK24-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
16162 // CHECK24-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
16163 // CHECK24-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
16164 // CHECK24-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
16165 // CHECK24-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
16166 // CHECK24-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
16167 // CHECK24-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
16168 // CHECK24-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
16169 // CHECK24-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
16170 // CHECK24-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
16171 // CHECK24-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
16172 // CHECK24-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
16173 // CHECK24-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
16174 // CHECK24-NEXT:    ret i32 [[TMP4]]
16175 //
16176 //
16177 // CHECK24-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
16178 // CHECK24-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
16179 // CHECK24-NEXT:  entry:
16180 // CHECK24-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
16181 // CHECK24-NEXT:    [[A:%.*]] = alloca i32, align 4
16182 // CHECK24-NEXT:    [[AA:%.*]] = alloca i16, align 2
16183 // CHECK24-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
16184 // CHECK24-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
16185 // CHECK24-NEXT:    store i32 0, i32* [[A]], align 4
16186 // CHECK24-NEXT:    store i16 0, i16* [[AA]], align 2
16187 // CHECK24-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
16188 // CHECK24-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
16189 // CHECK24-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
16190 // CHECK24-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
16191 // CHECK24-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
16192 // CHECK24-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
16193 // CHECK24-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
16194 // CHECK24-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
16195 // CHECK24-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
16196 // CHECK24-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
16197 // CHECK24-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
16198 // CHECK24-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
16199 // CHECK24-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
16200 // CHECK24-NEXT:    ret i32 [[TMP3]]
16201 //
16202 //
16203 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
16204 // CHECK25-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
16205 // CHECK25-NEXT:  entry:
16206 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16207 // CHECK25-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
16208 // CHECK25-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
16209 // CHECK25-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16210 // CHECK25-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
16211 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16212 // CHECK25-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
16213 // CHECK25-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
16214 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16215 // CHECK25-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
16216 // CHECK25-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
16217 // CHECK25-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
16218 // CHECK25-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
16219 // CHECK25-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
16220 // CHECK25-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
16221 // CHECK25-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16222 // CHECK25-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
16223 // CHECK25-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16224 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
16225 // CHECK25-NEXT:    ret void
16226 //
16227 //
16228 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined.
16229 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16230 // CHECK25-NEXT:  entry:
16231 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16232 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16233 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16234 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16235 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16236 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16237 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16238 // CHECK25-NEXT:    ret void
16239 //
16240 //
16241 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
16242 // CHECK25-SAME: (i64 [[AA:%.*]]) #[[ATTR0]] {
16243 // CHECK25-NEXT:  entry:
16244 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16245 // CHECK25-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16246 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16247 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16248 // CHECK25-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
16249 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16250 // CHECK25-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
16251 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16252 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP1]])
16253 // CHECK25-NEXT:    ret void
16254 //
16255 //
16256 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..1
16257 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16258 // CHECK25-NEXT:  entry:
16259 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16260 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16261 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16262 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16263 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16264 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16265 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16266 // CHECK25-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
16267 // CHECK25-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
16268 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
16269 // CHECK25-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
16270 // CHECK25-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
16271 // CHECK25-NEXT:    ret void
16272 //
16273 //
16274 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
16275 // CHECK25-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16276 // CHECK25-NEXT:  entry:
16277 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16278 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16279 // CHECK25-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
16280 // CHECK25-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16281 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16282 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16283 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16284 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16285 // CHECK25-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16286 // CHECK25-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
16287 // CHECK25-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
16288 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
16289 // CHECK25-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
16290 // CHECK25-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16291 // CHECK25-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
16292 // CHECK25-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16293 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
16294 // CHECK25-NEXT:    ret void
16295 //
16296 //
16297 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..2
16298 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16299 // CHECK25-NEXT:  entry:
16300 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16301 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16302 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16303 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16304 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16305 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16306 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16307 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16308 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16309 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16310 // CHECK25-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16311 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
16312 // CHECK25-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
16313 // CHECK25-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
16314 // CHECK25-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
16315 // CHECK25-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
16316 // CHECK25-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
16317 // CHECK25-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
16318 // CHECK25-NEXT:    ret void
16319 //
16320 //
16321 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
16322 // CHECK25-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
16323 // CHECK25-NEXT:  entry:
16324 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16325 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
16326 // CHECK25-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16327 // CHECK25-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
16328 // CHECK25-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
16329 // CHECK25-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
16330 // CHECK25-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
16331 // CHECK25-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
16332 // CHECK25-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
16333 // CHECK25-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
16334 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16335 // CHECK25-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
16336 // CHECK25-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16337 // CHECK25-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
16338 // CHECK25-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
16339 // CHECK25-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
16340 // CHECK25-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
16341 // CHECK25-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
16342 // CHECK25-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
16343 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16344 // CHECK25-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
16345 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16346 // CHECK25-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
16347 // CHECK25-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
16348 // CHECK25-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
16349 // CHECK25-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
16350 // CHECK25-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
16351 // CHECK25-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
16352 // CHECK25-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
16353 // CHECK25-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
16354 // CHECK25-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
16355 // CHECK25-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
16356 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
16357 // CHECK25-NEXT:    ret void
16358 //
16359 //
16360 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..3
16361 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
16362 // CHECK25-NEXT:  entry:
16363 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16364 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16365 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16366 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
16367 // CHECK25-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16368 // CHECK25-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
16369 // CHECK25-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
16370 // CHECK25-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
16371 // CHECK25-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
16372 // CHECK25-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
16373 // CHECK25-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
16374 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16375 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16376 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16377 // CHECK25-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
16378 // CHECK25-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16379 // CHECK25-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
16380 // CHECK25-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
16381 // CHECK25-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
16382 // CHECK25-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
16383 // CHECK25-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
16384 // CHECK25-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
16385 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16386 // CHECK25-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
16387 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16388 // CHECK25-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
16389 // CHECK25-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
16390 // CHECK25-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
16391 // CHECK25-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
16392 // CHECK25-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
16393 // CHECK25-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
16394 // CHECK25-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
16395 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
16396 // CHECK25-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
16397 // CHECK25-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
16398 // CHECK25-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
16399 // CHECK25-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
16400 // CHECK25-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
16401 // CHECK25-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
16402 // CHECK25-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
16403 // CHECK25-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
16404 // CHECK25-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
16405 // CHECK25-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
16406 // CHECK25-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
16407 // CHECK25-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
16408 // CHECK25-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
16409 // CHECK25-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
16410 // CHECK25-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
16411 // CHECK25-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
16412 // CHECK25-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
16413 // CHECK25-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
16414 // CHECK25-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
16415 // CHECK25-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
16416 // CHECK25-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
16417 // CHECK25-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
16418 // CHECK25-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
16419 // CHECK25-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
16420 // CHECK25-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
16421 // CHECK25-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
16422 // CHECK25-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
16423 // CHECK25-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
16424 // CHECK25-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
16425 // CHECK25-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
16426 // CHECK25-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
16427 // CHECK25-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
16428 // CHECK25-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
16429 // CHECK25-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
16430 // CHECK25-NEXT:    ret void
16431 //
16432 //
16433 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
16434 // CHECK25-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
16435 // CHECK25-NEXT:  entry:
16436 // CHECK25-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16437 // CHECK25-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
16438 // CHECK25-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16439 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16440 // CHECK25-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16441 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
16442 // CHECK25-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
16443 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
16444 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i64 [[TMP1]])
16445 // CHECK25-NEXT:    ret void
16446 //
16447 //
16448 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..4
16449 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
16450 // CHECK25-NEXT:  entry:
16451 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16452 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16453 // CHECK25-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16454 // CHECK25-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
16455 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16456 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16457 // CHECK25-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16458 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16459 // CHECK25-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16460 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
16461 // CHECK25-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
16462 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
16463 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP1]])
16464 // CHECK25-NEXT:    ret void
16465 //
16466 //
16467 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..5
16468 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
16469 // CHECK25-NEXT:  entry:
16470 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16471 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16472 // CHECK25-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16473 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16474 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16475 // CHECK25-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16476 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16477 // CHECK25-NEXT:    ret void
16478 //
16479 //
16480 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
16481 // CHECK25-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
16482 // CHECK25-NEXT:  entry:
16483 // CHECK25-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16484 // CHECK25-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
16485 // CHECK25-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16486 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16487 // CHECK25-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16488 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
16489 // CHECK25-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
16490 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
16491 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]])
16492 // CHECK25-NEXT:    ret void
16493 //
16494 //
16495 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..6
16496 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
16497 // CHECK25-NEXT:  entry:
16498 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16499 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16500 // CHECK25-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16501 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16502 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16503 // CHECK25-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16504 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16505 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[CONV]])
16506 // CHECK25-NEXT:    ret void
16507 //
16508 //
16509 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..7
16510 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
16511 // CHECK25-NEXT:  entry:
16512 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16513 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16514 // CHECK25-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
16515 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16516 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16517 // CHECK25-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
16518 // CHECK25-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
16519 // CHECK25-NEXT:    ret void
16520 //
16521 //
16522 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
16523 // CHECK25-SAME: (i64 [[VLA:%.*]]) #[[ATTR0]] {
16524 // CHECK25-NEXT:  entry:
16525 // CHECK25-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16526 // CHECK25-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16527 // CHECK25-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16528 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i64 [[TMP0]])
16529 // CHECK25-NEXT:    ret void
16530 //
16531 //
16532 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..8
16533 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR0]] {
16534 // CHECK25-NEXT:  entry:
16535 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16536 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16537 // CHECK25-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16538 // CHECK25-NEXT:    [[F:%.*]] = alloca i32*, align 8
16539 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16540 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16541 // CHECK25-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16542 // CHECK25-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16543 // CHECK25-NEXT:    ret void
16544 //
16545 //
16546 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
16547 // CHECK25-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
16548 // CHECK25-NEXT:  entry:
16549 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16550 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16551 // CHECK25-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
16552 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
16553 // CHECK25-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
16554 // CHECK25-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16555 // CHECK25-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
16556 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16557 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16558 // CHECK25-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
16559 // CHECK25-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
16560 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16561 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16562 // CHECK25-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
16563 // CHECK25-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
16564 // CHECK25-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
16565 // CHECK25-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
16566 // CHECK25-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
16567 // CHECK25-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
16568 // CHECK25-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
16569 // CHECK25-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16570 // CHECK25-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
16571 // CHECK25-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16572 // CHECK25-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
16573 // CHECK25-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
16574 // CHECK25-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
16575 // CHECK25-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
16576 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
16577 // CHECK25-NEXT:    ret void
16578 //
16579 //
16580 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..9
16581 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
16582 // CHECK25-NEXT:  entry:
16583 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16584 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16585 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16586 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16587 // CHECK25-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
16588 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
16589 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16590 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16591 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16592 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16593 // CHECK25-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
16594 // CHECK25-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
16595 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16596 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16597 // CHECK25-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
16598 // CHECK25-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
16599 // CHECK25-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
16600 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
16601 // CHECK25-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
16602 // CHECK25-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
16603 // CHECK25-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
16604 // CHECK25-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
16605 // CHECK25-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
16606 // CHECK25-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
16607 // CHECK25-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
16608 // CHECK25-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
16609 // CHECK25-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
16610 // CHECK25-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
16611 // CHECK25-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
16612 // CHECK25-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
16613 // CHECK25-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
16614 // CHECK25-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
16615 // CHECK25-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
16616 // CHECK25-NEXT:    ret void
16617 //
16618 //
16619 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
16620 // CHECK25-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
16621 // CHECK25-NEXT:  entry:
16622 // CHECK25-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
16623 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
16624 // CHECK25-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16625 // CHECK25-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
16626 // CHECK25-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
16627 // CHECK25-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
16628 // CHECK25-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
16629 // CHECK25-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
16630 // CHECK25-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16631 // CHECK25-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
16632 // CHECK25-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
16633 // CHECK25-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
16634 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
16635 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16636 // CHECK25-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
16637 // CHECK25-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
16638 // CHECK25-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
16639 // CHECK25-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
16640 // CHECK25-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
16641 // CHECK25-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
16642 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
16643 // CHECK25-NEXT:    ret void
16644 //
16645 //
16646 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..10
16647 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
16648 // CHECK25-NEXT:  entry:
16649 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16650 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16651 // CHECK25-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
16652 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
16653 // CHECK25-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16654 // CHECK25-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
16655 // CHECK25-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
16656 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16657 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16658 // CHECK25-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
16659 // CHECK25-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
16660 // CHECK25-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16661 // CHECK25-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
16662 // CHECK25-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
16663 // CHECK25-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
16664 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
16665 // CHECK25-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16666 // CHECK25-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
16667 // CHECK25-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
16668 // CHECK25-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
16669 // CHECK25-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
16670 // CHECK25-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
16671 // CHECK25-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
16672 // CHECK25-NEXT:    store double [[ADD]], double* [[A]], align 8
16673 // CHECK25-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
16674 // CHECK25-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
16675 // CHECK25-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
16676 // CHECK25-NEXT:    store double [[INC]], double* [[A4]], align 8
16677 // CHECK25-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
16678 // CHECK25-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
16679 // CHECK25-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
16680 // CHECK25-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
16681 // CHECK25-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
16682 // CHECK25-NEXT:    ret void
16683 //
16684 //
16685 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
16686 // CHECK25-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
16687 // CHECK25-NEXT:  entry:
16688 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16689 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16690 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
16691 // CHECK25-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
16692 // CHECK25-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16693 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16694 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16695 // CHECK25-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
16696 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16697 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16698 // CHECK25-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
16699 // CHECK25-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
16700 // CHECK25-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
16701 // CHECK25-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
16702 // CHECK25-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
16703 // CHECK25-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
16704 // CHECK25-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16705 // CHECK25-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
16706 // CHECK25-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16707 // CHECK25-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
16708 // CHECK25-NEXT:    ret void
16709 //
16710 //
16711 // CHECK25-LABEL: define {{[^@]+}}@.omp_outlined..11
16712 // CHECK25-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
16713 // CHECK25-NEXT:  entry:
16714 // CHECK25-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16715 // CHECK25-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16716 // CHECK25-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16717 // CHECK25-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16718 // CHECK25-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
16719 // CHECK25-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16720 // CHECK25-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16721 // CHECK25-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16722 // CHECK25-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16723 // CHECK25-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
16724 // CHECK25-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16725 // CHECK25-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16726 // CHECK25-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
16727 // CHECK25-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
16728 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
16729 // CHECK25-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
16730 // CHECK25-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
16731 // CHECK25-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
16732 // CHECK25-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
16733 // CHECK25-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
16734 // CHECK25-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
16735 // CHECK25-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
16736 // CHECK25-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
16737 // CHECK25-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
16738 // CHECK25-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
16739 // CHECK25-NEXT:    ret void
16740 //
16741 //
16742 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
16743 // CHECK26-SAME: (i64 [[AA:%.*]], i64 [[DOTCAPTURE_EXPR_:%.*]], i64 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
16744 // CHECK26-NEXT:  entry:
16745 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16746 // CHECK26-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i64, align 8
16747 // CHECK26-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i64, align 8
16748 // CHECK26-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16749 // CHECK26-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
16750 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16751 // CHECK26-NEXT:    store i64 [[DOTCAPTURE_EXPR_]], i64* [[DOTCAPTURE_EXPR__ADDR]], align 8
16752 // CHECK26-NEXT:    store i64 [[DOTCAPTURE_EXPR_1]], i64* [[DOTCAPTURE_EXPR__ADDR2]], align 8
16753 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16754 // CHECK26-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR]] to i32*
16755 // CHECK26-NEXT:    [[CONV4:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__ADDR2]] to i32*
16756 // CHECK26-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV3]], align 8
16757 // CHECK26-NEXT:    [[TMP2:%.*]] = load i32, i32* [[CONV4]], align 8
16758 // CHECK26-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
16759 // CHECK26-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 8
16760 // CHECK26-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16761 // CHECK26-NEXT:    store i16 [[TMP3]], i16* [[CONV5]], align 2
16762 // CHECK26-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16763 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined. to void (i32*, i32*, ...)*), i64 [[TMP4]])
16764 // CHECK26-NEXT:    ret void
16765 //
16766 //
16767 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined.
16768 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16769 // CHECK26-NEXT:  entry:
16770 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16771 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16772 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16773 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16774 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16775 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16776 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16777 // CHECK26-NEXT:    ret void
16778 //
16779 //
16780 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
16781 // CHECK26-SAME: (i64 [[AA:%.*]]) #[[ATTR0]] {
16782 // CHECK26-NEXT:  entry:
16783 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16784 // CHECK26-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16785 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16786 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16787 // CHECK26-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
16788 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16789 // CHECK26-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
16790 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16791 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP1]])
16792 // CHECK26-NEXT:    ret void
16793 //
16794 //
16795 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..1
16796 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16797 // CHECK26-NEXT:  entry:
16798 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16799 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16800 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16801 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16802 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16803 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16804 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16805 // CHECK26-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 8
16806 // CHECK26-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
16807 // CHECK26-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
16808 // CHECK26-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
16809 // CHECK26-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 8
16810 // CHECK26-NEXT:    ret void
16811 //
16812 //
16813 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
16814 // CHECK26-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16815 // CHECK26-NEXT:  entry:
16816 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16817 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16818 // CHECK26-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
16819 // CHECK26-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
16820 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16821 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16822 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16823 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16824 // CHECK26-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16825 // CHECK26-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
16826 // CHECK26-NEXT:    store i32 [[TMP0]], i32* [[CONV2]], align 4
16827 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[A_CASTED]], align 8
16828 // CHECK26-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
16829 // CHECK26-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
16830 // CHECK26-NEXT:    store i16 [[TMP2]], i16* [[CONV3]], align 2
16831 // CHECK26-NEXT:    [[TMP3:%.*]] = load i64, i64* [[AA_CASTED]], align 8
16832 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i64 [[TMP1]], i64 [[TMP3]])
16833 // CHECK26-NEXT:    ret void
16834 //
16835 //
16836 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..2
16837 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]]) #[[ATTR0]] {
16838 // CHECK26-NEXT:  entry:
16839 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16840 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16841 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16842 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
16843 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16844 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16845 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16846 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
16847 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16848 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
16849 // CHECK26-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16850 // CHECK26-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
16851 // CHECK26-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
16852 // CHECK26-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV1]], align 8
16853 // CHECK26-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP1]] to i32
16854 // CHECK26-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
16855 // CHECK26-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
16856 // CHECK26-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
16857 // CHECK26-NEXT:    ret void
16858 //
16859 //
16860 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
16861 // CHECK26-SAME: (i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
16862 // CHECK26-NEXT:  entry:
16863 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16864 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
16865 // CHECK26-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16866 // CHECK26-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
16867 // CHECK26-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
16868 // CHECK26-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
16869 // CHECK26-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
16870 // CHECK26-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
16871 // CHECK26-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
16872 // CHECK26-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
16873 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16874 // CHECK26-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
16875 // CHECK26-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16876 // CHECK26-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
16877 // CHECK26-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
16878 // CHECK26-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
16879 // CHECK26-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
16880 // CHECK26-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
16881 // CHECK26-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
16882 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16883 // CHECK26-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
16884 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16885 // CHECK26-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
16886 // CHECK26-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
16887 // CHECK26-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
16888 // CHECK26-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
16889 // CHECK26-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
16890 // CHECK26-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
16891 // CHECK26-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
16892 // CHECK26-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED]] to i32*
16893 // CHECK26-NEXT:    store i32 [[TMP8]], i32* [[CONV5]], align 4
16894 // CHECK26-NEXT:    [[TMP9:%.*]] = load i64, i64* [[A_CASTED]], align 8
16895 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, [10 x float]*, i64, float*, [5 x [10 x double]]*, i64, i64, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i64 [[TMP9]], [10 x float]* [[TMP0]], i64 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i64 [[TMP4]], i64 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
16896 // CHECK26-NEXT:    ret void
16897 //
16898 //
16899 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..3
16900 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i64 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 8 dereferenceable(400) [[C:%.*]], i64 [[VLA1:%.*]], i64 [[VLA3:%.*]], double* nonnull align 8 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 8 dereferenceable(16) [[D:%.*]]) #[[ATTR0]] {
16901 // CHECK26-NEXT:  entry:
16902 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16903 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16904 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
16905 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 8
16906 // CHECK26-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
16907 // CHECK26-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 8
16908 // CHECK26-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 8
16909 // CHECK26-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
16910 // CHECK26-NEXT:    [[VLA_ADDR4:%.*]] = alloca i64, align 8
16911 // CHECK26-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 8
16912 // CHECK26-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 8
16913 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16914 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16915 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
16916 // CHECK26-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 8
16917 // CHECK26-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
16918 // CHECK26-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 8
16919 // CHECK26-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 8
16920 // CHECK26-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
16921 // CHECK26-NEXT:    store i64 [[VLA3]], i64* [[VLA_ADDR4]], align 8
16922 // CHECK26-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 8
16923 // CHECK26-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 8
16924 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
16925 // CHECK26-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 8
16926 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
16927 // CHECK26-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 8
16928 // CHECK26-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 8
16929 // CHECK26-NEXT:    [[TMP4:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
16930 // CHECK26-NEXT:    [[TMP5:%.*]] = load i64, i64* [[VLA_ADDR4]], align 8
16931 // CHECK26-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 8
16932 // CHECK26-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 8
16933 // CHECK26-NEXT:    [[TMP8:%.*]] = load i32, i32* [[CONV]], align 8
16934 // CHECK26-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
16935 // CHECK26-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
16936 // CHECK26-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i64 0, i64 2
16937 // CHECK26-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
16938 // CHECK26-NEXT:    [[CONV5:%.*]] = fpext float [[TMP9]] to double
16939 // CHECK26-NEXT:    [[ADD6:%.*]] = fadd double [[CONV5]], 1.000000e+00
16940 // CHECK26-NEXT:    [[CONV7:%.*]] = fptrunc double [[ADD6]] to float
16941 // CHECK26-NEXT:    store float [[CONV7]], float* [[ARRAYIDX]], align 4
16942 // CHECK26-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds float, float* [[TMP2]], i64 3
16943 // CHECK26-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX8]], align 4
16944 // CHECK26-NEXT:    [[CONV9:%.*]] = fpext float [[TMP10]] to double
16945 // CHECK26-NEXT:    [[ADD10:%.*]] = fadd double [[CONV9]], 1.000000e+00
16946 // CHECK26-NEXT:    [[CONV11:%.*]] = fptrunc double [[ADD10]] to float
16947 // CHECK26-NEXT:    store float [[CONV11]], float* [[ARRAYIDX8]], align 4
16948 // CHECK26-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i64 0, i64 1
16949 // CHECK26-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX12]], i64 0, i64 2
16950 // CHECK26-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX13]], align 8
16951 // CHECK26-NEXT:    [[ADD14:%.*]] = fadd double [[TMP11]], 1.000000e+00
16952 // CHECK26-NEXT:    store double [[ADD14]], double* [[ARRAYIDX13]], align 8
16953 // CHECK26-NEXT:    [[TMP12:%.*]] = mul nsw i64 1, [[TMP5]]
16954 // CHECK26-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[TMP6]], i64 [[TMP12]]
16955 // CHECK26-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX15]], i64 3
16956 // CHECK26-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX16]], align 8
16957 // CHECK26-NEXT:    [[ADD17:%.*]] = fadd double [[TMP13]], 1.000000e+00
16958 // CHECK26-NEXT:    store double [[ADD17]], double* [[ARRAYIDX16]], align 8
16959 // CHECK26-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
16960 // CHECK26-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 8
16961 // CHECK26-NEXT:    [[ADD18:%.*]] = add nsw i64 [[TMP14]], 1
16962 // CHECK26-NEXT:    store i64 [[ADD18]], i64* [[X]], align 8
16963 // CHECK26-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
16964 // CHECK26-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 8
16965 // CHECK26-NEXT:    [[CONV19:%.*]] = sext i8 [[TMP15]] to i32
16966 // CHECK26-NEXT:    [[ADD20:%.*]] = add nsw i32 [[CONV19]], 1
16967 // CHECK26-NEXT:    [[CONV21:%.*]] = trunc i32 [[ADD20]] to i8
16968 // CHECK26-NEXT:    store i8 [[CONV21]], i8* [[Y]], align 8
16969 // CHECK26-NEXT:    ret void
16970 //
16971 //
16972 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
16973 // CHECK26-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
16974 // CHECK26-NEXT:  entry:
16975 // CHECK26-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16976 // CHECK26-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
16977 // CHECK26-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16978 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16979 // CHECK26-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16980 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
16981 // CHECK26-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
16982 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
16983 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i64 [[TMP1]])
16984 // CHECK26-NEXT:    ret void
16985 //
16986 //
16987 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..4
16988 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
16989 // CHECK26-NEXT:  entry:
16990 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
16991 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
16992 // CHECK26-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
16993 // CHECK26-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
16994 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
16995 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
16996 // CHECK26-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
16997 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
16998 // CHECK26-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
16999 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
17000 // CHECK26-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
17001 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
17002 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP1]])
17003 // CHECK26-NEXT:    ret void
17004 //
17005 //
17006 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..5
17007 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
17008 // CHECK26-NEXT:  entry:
17009 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17010 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17011 // CHECK26-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
17012 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17013 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17014 // CHECK26-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
17015 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
17016 // CHECK26-NEXT:    ret void
17017 //
17018 //
17019 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
17020 // CHECK26-SAME: (i64 [[NN:%.*]]) #[[ATTR0]] {
17021 // CHECK26-NEXT:  entry:
17022 // CHECK26-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
17023 // CHECK26-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
17024 // CHECK26-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
17025 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
17026 // CHECK26-NEXT:    [[TMP0:%.*]] = load i32, i32* [[CONV]], align 8
17027 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
17028 // CHECK26-NEXT:    store i32 [[TMP0]], i32* [[CONV1]], align 4
17029 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[NN_CASTED]], align 8
17030 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i64 [[TMP1]])
17031 // CHECK26-NEXT:    ret void
17032 //
17033 //
17034 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..6
17035 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[NN:%.*]]) #[[ATTR0]] {
17036 // CHECK26-NEXT:  entry:
17037 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17038 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17039 // CHECK26-NEXT:    [[NN_ADDR:%.*]] = alloca i64, align 8
17040 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17041 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17042 // CHECK26-NEXT:    store i64 [[NN]], i64* [[NN_ADDR]], align 8
17043 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[NN_ADDR]] to i32*
17044 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[CONV]])
17045 // CHECK26-NEXT:    ret void
17046 //
17047 //
17048 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..7
17049 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
17050 // CHECK26-NEXT:  entry:
17051 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17052 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17053 // CHECK26-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 8
17054 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17055 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17056 // CHECK26-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 8
17057 // CHECK26-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 8
17058 // CHECK26-NEXT:    ret void
17059 //
17060 //
17061 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
17062 // CHECK26-SAME: (i64 [[VLA:%.*]]) #[[ATTR0]] {
17063 // CHECK26-NEXT:  entry:
17064 // CHECK26-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
17065 // CHECK26-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
17066 // CHECK26-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
17067 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i64 [[TMP0]])
17068 // CHECK26-NEXT:    ret void
17069 //
17070 //
17071 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..8
17072 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[VLA:%.*]]) #[[ATTR0]] {
17073 // CHECK26-NEXT:  entry:
17074 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17075 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17076 // CHECK26-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
17077 // CHECK26-NEXT:    [[F:%.*]] = alloca i32*, align 8
17078 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17079 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17080 // CHECK26-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
17081 // CHECK26-NEXT:    [[TMP0:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
17082 // CHECK26-NEXT:    ret void
17083 //
17084 //
17085 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
17086 // CHECK26-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17087 // CHECK26-NEXT:  entry:
17088 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
17089 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
17090 // CHECK26-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
17091 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
17092 // CHECK26-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
17093 // CHECK26-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
17094 // CHECK26-NEXT:    [[AAA_CASTED:%.*]] = alloca i64, align 8
17095 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
17096 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
17097 // CHECK26-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
17098 // CHECK26-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
17099 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
17100 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
17101 // CHECK26-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
17102 // CHECK26-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
17103 // CHECK26-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
17104 // CHECK26-NEXT:    [[CONV3:%.*]] = bitcast i64* [[A_CASTED]] to i32*
17105 // CHECK26-NEXT:    store i32 [[TMP1]], i32* [[CONV3]], align 4
17106 // CHECK26-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
17107 // CHECK26-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
17108 // CHECK26-NEXT:    [[CONV4:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
17109 // CHECK26-NEXT:    store i16 [[TMP3]], i16* [[CONV4]], align 2
17110 // CHECK26-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
17111 // CHECK26-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV2]], align 8
17112 // CHECK26-NEXT:    [[CONV5:%.*]] = bitcast i64* [[AAA_CASTED]] to i8*
17113 // CHECK26-NEXT:    store i8 [[TMP5]], i8* [[CONV5]], align 1
17114 // CHECK26-NEXT:    [[TMP6:%.*]] = load i64, i64* [[AAA_CASTED]], align 8
17115 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, i64, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [10 x i32]* [[TMP0]])
17116 // CHECK26-NEXT:    ret void
17117 //
17118 //
17119 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..9
17120 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], i64 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17121 // CHECK26-NEXT:  entry:
17122 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17123 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17124 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
17125 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
17126 // CHECK26-NEXT:    [[AAA_ADDR:%.*]] = alloca i64, align 8
17127 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
17128 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17129 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17130 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
17131 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
17132 // CHECK26-NEXT:    store i64 [[AAA]], i64* [[AAA_ADDR]], align 8
17133 // CHECK26-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
17134 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
17135 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
17136 // CHECK26-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AAA_ADDR]] to i8*
17137 // CHECK26-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
17138 // CHECK26-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
17139 // CHECK26-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
17140 // CHECK26-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
17141 // CHECK26-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
17142 // CHECK26-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP2]] to i32
17143 // CHECK26-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
17144 // CHECK26-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i16
17145 // CHECK26-NEXT:    store i16 [[CONV5]], i16* [[CONV1]], align 8
17146 // CHECK26-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV2]], align 8
17147 // CHECK26-NEXT:    [[CONV6:%.*]] = sext i8 [[TMP3]] to i32
17148 // CHECK26-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
17149 // CHECK26-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i8
17150 // CHECK26-NEXT:    store i8 [[CONV8]], i8* [[CONV2]], align 8
17151 // CHECK26-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
17152 // CHECK26-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
17153 // CHECK26-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP4]], 1
17154 // CHECK26-NEXT:    store i32 [[ADD9]], i32* [[ARRAYIDX]], align 4
17155 // CHECK26-NEXT:    ret void
17156 //
17157 //
17158 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
17159 // CHECK26-SAME: (%struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
17160 // CHECK26-NEXT:  entry:
17161 // CHECK26-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
17162 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
17163 // CHECK26-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
17164 // CHECK26-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
17165 // CHECK26-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
17166 // CHECK26-NEXT:    [[B_CASTED:%.*]] = alloca i64, align 8
17167 // CHECK26-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
17168 // CHECK26-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
17169 // CHECK26-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
17170 // CHECK26-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
17171 // CHECK26-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
17172 // CHECK26-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
17173 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
17174 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
17175 // CHECK26-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
17176 // CHECK26-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
17177 // CHECK26-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
17178 // CHECK26-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i32*
17179 // CHECK26-NEXT:    store i32 [[TMP4]], i32* [[CONV3]], align 4
17180 // CHECK26-NEXT:    [[TMP5:%.*]] = load i64, i64* [[B_CASTED]], align 8
17181 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i64, i64, i64, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i64 [[TMP5]], i64 [[TMP1]], i64 [[TMP2]], i16* [[TMP3]])
17182 // CHECK26-NEXT:    ret void
17183 //
17184 //
17185 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..10
17186 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i64 [[B:%.*]], i64 [[VLA:%.*]], i64 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
17187 // CHECK26-NEXT:  entry:
17188 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17189 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17190 // CHECK26-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
17191 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca i64, align 8
17192 // CHECK26-NEXT:    [[VLA_ADDR:%.*]] = alloca i64, align 8
17193 // CHECK26-NEXT:    [[VLA_ADDR2:%.*]] = alloca i64, align 8
17194 // CHECK26-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 8
17195 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17196 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17197 // CHECK26-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
17198 // CHECK26-NEXT:    store i64 [[B]], i64* [[B_ADDR]], align 8
17199 // CHECK26-NEXT:    store i64 [[VLA]], i64* [[VLA_ADDR]], align 8
17200 // CHECK26-NEXT:    store i64 [[VLA1]], i64* [[VLA_ADDR2]], align 8
17201 // CHECK26-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 8
17202 // CHECK26-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
17203 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[B_ADDR]] to i32*
17204 // CHECK26-NEXT:    [[TMP1:%.*]] = load i64, i64* [[VLA_ADDR]], align 8
17205 // CHECK26-NEXT:    [[TMP2:%.*]] = load i64, i64* [[VLA_ADDR2]], align 8
17206 // CHECK26-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 8
17207 // CHECK26-NEXT:    [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8
17208 // CHECK26-NEXT:    [[CONV3:%.*]] = sitofp i32 [[TMP4]] to double
17209 // CHECK26-NEXT:    [[ADD:%.*]] = fadd double [[CONV3]], 1.500000e+00
17210 // CHECK26-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
17211 // CHECK26-NEXT:    store double [[ADD]], double* [[A]], align 8
17212 // CHECK26-NEXT:    [[A4:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
17213 // CHECK26-NEXT:    [[TMP5:%.*]] = load double, double* [[A4]], align 8
17214 // CHECK26-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
17215 // CHECK26-NEXT:    store double [[INC]], double* [[A4]], align 8
17216 // CHECK26-NEXT:    [[CONV5:%.*]] = fptosi double [[INC]] to i16
17217 // CHECK26-NEXT:    [[TMP6:%.*]] = mul nsw i64 1, [[TMP2]]
17218 // CHECK26-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i64 [[TMP6]]
17219 // CHECK26-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
17220 // CHECK26-NEXT:    store i16 [[CONV5]], i16* [[ARRAYIDX6]], align 2
17221 // CHECK26-NEXT:    ret void
17222 //
17223 //
17224 // CHECK26-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
17225 // CHECK26-SAME: (i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17226 // CHECK26-NEXT:  entry:
17227 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
17228 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
17229 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
17230 // CHECK26-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
17231 // CHECK26-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8
17232 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
17233 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
17234 // CHECK26-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
17235 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
17236 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
17237 // CHECK26-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
17238 // CHECK26-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
17239 // CHECK26-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED]] to i32*
17240 // CHECK26-NEXT:    store i32 [[TMP1]], i32* [[CONV2]], align 4
17241 // CHECK26-NEXT:    [[TMP2:%.*]] = load i64, i64* [[A_CASTED]], align 8
17242 // CHECK26-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV1]], align 8
17243 // CHECK26-NEXT:    [[CONV3:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
17244 // CHECK26-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
17245 // CHECK26-NEXT:    [[TMP4:%.*]] = load i64, i64* [[AA_CASTED]], align 8
17246 // CHECK26-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP2]], i64 [[TMP4]], [10 x i32]* [[TMP0]])
17247 // CHECK26-NEXT:    ret void
17248 //
17249 //
17250 // CHECK26-LABEL: define {{[^@]+}}@.omp_outlined..11
17251 // CHECK26-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[A:%.*]], i64 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17252 // CHECK26-NEXT:  entry:
17253 // CHECK26-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
17254 // CHECK26-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
17255 // CHECK26-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8
17256 // CHECK26-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8
17257 // CHECK26-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 8
17258 // CHECK26-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
17259 // CHECK26-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
17260 // CHECK26-NEXT:    store i64 [[A]], i64* [[A_ADDR]], align 8
17261 // CHECK26-NEXT:    store i64 [[AA]], i64* [[AA_ADDR]], align 8
17262 // CHECK26-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 8
17263 // CHECK26-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*
17264 // CHECK26-NEXT:    [[CONV1:%.*]] = bitcast i64* [[AA_ADDR]] to i16*
17265 // CHECK26-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 8
17266 // CHECK26-NEXT:    [[TMP1:%.*]] = load i32, i32* [[CONV]], align 8
17267 // CHECK26-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
17268 // CHECK26-NEXT:    store i32 [[ADD]], i32* [[CONV]], align 8
17269 // CHECK26-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV1]], align 8
17270 // CHECK26-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
17271 // CHECK26-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
17272 // CHECK26-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
17273 // CHECK26-NEXT:    store i16 [[CONV4]], i16* [[CONV1]], align 8
17274 // CHECK26-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i64 0, i64 2
17275 // CHECK26-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
17276 // CHECK26-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP3]], 1
17277 // CHECK26-NEXT:    store i32 [[ADD5]], i32* [[ARRAYIDX]], align 4
17278 // CHECK26-NEXT:    ret void
17279 //
17280 //
17281 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
17282 // CHECK27-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
17283 // CHECK27-NEXT:  entry:
17284 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17285 // CHECK27-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
17286 // CHECK27-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
17287 // CHECK27-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17288 // CHECK27-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
17289 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17290 // CHECK27-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
17291 // CHECK27-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
17292 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17293 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
17294 // CHECK27-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
17295 // CHECK27-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
17296 // CHECK27-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
17297 // CHECK27-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17298 // CHECK27-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
17299 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17300 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
17301 // CHECK27-NEXT:    ret void
17302 //
17303 //
17304 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined.
17305 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17306 // CHECK27-NEXT:  entry:
17307 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17308 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17309 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17310 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17311 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17312 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17313 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17314 // CHECK27-NEXT:    ret void
17315 //
17316 //
17317 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
17318 // CHECK27-SAME: (i32 [[AA:%.*]]) #[[ATTR0]] {
17319 // CHECK27-NEXT:  entry:
17320 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17321 // CHECK27-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17322 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17323 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17324 // CHECK27-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
17325 // CHECK27-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17326 // CHECK27-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
17327 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17328 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i32 [[TMP1]])
17329 // CHECK27-NEXT:    ret void
17330 //
17331 //
17332 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..1
17333 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17334 // CHECK27-NEXT:  entry:
17335 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17336 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17337 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17338 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17339 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17340 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17341 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17342 // CHECK27-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
17343 // CHECK27-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
17344 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
17345 // CHECK27-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
17346 // CHECK27-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
17347 // CHECK27-NEXT:    ret void
17348 //
17349 //
17350 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
17351 // CHECK27-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17352 // CHECK27-NEXT:  entry:
17353 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17354 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17355 // CHECK27-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
17356 // CHECK27-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17357 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17358 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17359 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17360 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
17361 // CHECK27-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
17362 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
17363 // CHECK27-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
17364 // CHECK27-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17365 // CHECK27-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
17366 // CHECK27-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17367 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
17368 // CHECK27-NEXT:    ret void
17369 //
17370 //
17371 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..2
17372 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17373 // CHECK27-NEXT:  entry:
17374 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17375 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17376 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17377 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17378 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17379 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17380 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17381 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17382 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17383 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
17384 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
17385 // CHECK27-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
17386 // CHECK27-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
17387 // CHECK27-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
17388 // CHECK27-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
17389 // CHECK27-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
17390 // CHECK27-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
17391 // CHECK27-NEXT:    ret void
17392 //
17393 //
17394 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
17395 // CHECK27-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
17396 // CHECK27-NEXT:  entry:
17397 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17398 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
17399 // CHECK27-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17400 // CHECK27-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
17401 // CHECK27-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
17402 // CHECK27-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
17403 // CHECK27-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
17404 // CHECK27-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
17405 // CHECK27-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
17406 // CHECK27-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
17407 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17408 // CHECK27-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
17409 // CHECK27-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17410 // CHECK27-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
17411 // CHECK27-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
17412 // CHECK27-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
17413 // CHECK27-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
17414 // CHECK27-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
17415 // CHECK27-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
17416 // CHECK27-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
17417 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17418 // CHECK27-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
17419 // CHECK27-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
17420 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
17421 // CHECK27-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
17422 // CHECK27-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
17423 // CHECK27-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
17424 // CHECK27-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
17425 // CHECK27-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
17426 // CHECK27-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
17427 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
17428 // CHECK27-NEXT:    ret void
17429 //
17430 //
17431 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..3
17432 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
17433 // CHECK27-NEXT:  entry:
17434 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17435 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17436 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17437 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
17438 // CHECK27-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17439 // CHECK27-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
17440 // CHECK27-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
17441 // CHECK27-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
17442 // CHECK27-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
17443 // CHECK27-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
17444 // CHECK27-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
17445 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17446 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17447 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17448 // CHECK27-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
17449 // CHECK27-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17450 // CHECK27-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
17451 // CHECK27-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
17452 // CHECK27-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
17453 // CHECK27-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
17454 // CHECK27-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
17455 // CHECK27-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
17456 // CHECK27-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
17457 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17458 // CHECK27-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
17459 // CHECK27-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
17460 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
17461 // CHECK27-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
17462 // CHECK27-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
17463 // CHECK27-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
17464 // CHECK27-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
17465 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
17466 // CHECK27-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
17467 // CHECK27-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
17468 // CHECK27-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
17469 // CHECK27-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
17470 // CHECK27-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
17471 // CHECK27-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
17472 // CHECK27-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
17473 // CHECK27-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
17474 // CHECK27-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
17475 // CHECK27-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
17476 // CHECK27-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
17477 // CHECK27-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
17478 // CHECK27-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
17479 // CHECK27-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
17480 // CHECK27-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
17481 // CHECK27-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
17482 // CHECK27-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
17483 // CHECK27-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
17484 // CHECK27-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
17485 // CHECK27-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
17486 // CHECK27-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
17487 // CHECK27-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
17488 // CHECK27-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
17489 // CHECK27-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
17490 // CHECK27-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
17491 // CHECK27-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
17492 // CHECK27-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
17493 // CHECK27-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
17494 // CHECK27-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
17495 // CHECK27-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
17496 // CHECK27-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
17497 // CHECK27-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
17498 // CHECK27-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
17499 // CHECK27-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
17500 // CHECK27-NEXT:    ret void
17501 //
17502 //
17503 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
17504 // CHECK27-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
17505 // CHECK27-NEXT:  entry:
17506 // CHECK27-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
17507 // CHECK27-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
17508 // CHECK27-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
17509 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
17510 // CHECK27-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
17511 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
17512 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i32 [[TMP1]])
17513 // CHECK27-NEXT:    ret void
17514 //
17515 //
17516 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..4
17517 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
17518 // CHECK27-NEXT:  entry:
17519 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17520 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17521 // CHECK27-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
17522 // CHECK27-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
17523 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17524 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17525 // CHECK27-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
17526 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
17527 // CHECK27-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
17528 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
17529 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i32 [[TMP1]])
17530 // CHECK27-NEXT:    ret void
17531 //
17532 //
17533 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..5
17534 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
17535 // CHECK27-NEXT:  entry:
17536 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17537 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17538 // CHECK27-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
17539 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17540 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17541 // CHECK27-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
17542 // CHECK27-NEXT:    ret void
17543 //
17544 //
17545 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
17546 // CHECK27-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
17547 // CHECK27-NEXT:  entry:
17548 // CHECK27-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
17549 // CHECK27-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
17550 // CHECK27-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
17551 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
17552 // CHECK27-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
17553 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
17554 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]])
17555 // CHECK27-NEXT:    ret void
17556 //
17557 //
17558 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..6
17559 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
17560 // CHECK27-NEXT:  entry:
17561 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17562 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17563 // CHECK27-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
17564 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17565 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17566 // CHECK27-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
17567 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
17568 // CHECK27-NEXT:    ret void
17569 //
17570 //
17571 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..7
17572 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
17573 // CHECK27-NEXT:  entry:
17574 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17575 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17576 // CHECK27-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
17577 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17578 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17579 // CHECK27-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
17580 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
17581 // CHECK27-NEXT:    ret void
17582 //
17583 //
17584 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
17585 // CHECK27-SAME: (i32 [[VLA:%.*]]) #[[ATTR0]] {
17586 // CHECK27-NEXT:  entry:
17587 // CHECK27-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17588 // CHECK27-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17589 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17590 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32 [[TMP0]])
17591 // CHECK27-NEXT:    ret void
17592 //
17593 //
17594 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..8
17595 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR0]] {
17596 // CHECK27-NEXT:  entry:
17597 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17598 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17599 // CHECK27-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17600 // CHECK27-NEXT:    [[F:%.*]] = alloca i32*, align 4
17601 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17602 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17603 // CHECK27-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17604 // CHECK27-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17605 // CHECK27-NEXT:    ret void
17606 //
17607 //
17608 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
17609 // CHECK27-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17610 // CHECK27-NEXT:  entry:
17611 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17612 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17613 // CHECK27-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
17614 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
17615 // CHECK27-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
17616 // CHECK27-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17617 // CHECK27-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
17618 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17619 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17620 // CHECK27-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
17621 // CHECK27-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
17622 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17623 // CHECK27-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
17624 // CHECK27-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
17625 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
17626 // CHECK27-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
17627 // CHECK27-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
17628 // CHECK27-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
17629 // CHECK27-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17630 // CHECK27-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
17631 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17632 // CHECK27-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
17633 // CHECK27-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
17634 // CHECK27-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
17635 // CHECK27-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
17636 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
17637 // CHECK27-NEXT:    ret void
17638 //
17639 //
17640 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..9
17641 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17642 // CHECK27-NEXT:  entry:
17643 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17644 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17645 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17646 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17647 // CHECK27-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
17648 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
17649 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17650 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17651 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17652 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17653 // CHECK27-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
17654 // CHECK27-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
17655 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17656 // CHECK27-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
17657 // CHECK27-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
17658 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
17659 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
17660 // CHECK27-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
17661 // CHECK27-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
17662 // CHECK27-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
17663 // CHECK27-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
17664 // CHECK27-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
17665 // CHECK27-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
17666 // CHECK27-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
17667 // CHECK27-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
17668 // CHECK27-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
17669 // CHECK27-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
17670 // CHECK27-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
17671 // CHECK27-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
17672 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
17673 // CHECK27-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
17674 // CHECK27-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
17675 // CHECK27-NEXT:    ret void
17676 //
17677 //
17678 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
17679 // CHECK27-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
17680 // CHECK27-NEXT:  entry:
17681 // CHECK27-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
17682 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
17683 // CHECK27-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17684 // CHECK27-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
17685 // CHECK27-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
17686 // CHECK27-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
17687 // CHECK27-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
17688 // CHECK27-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
17689 // CHECK27-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17690 // CHECK27-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
17691 // CHECK27-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
17692 // CHECK27-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
17693 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17694 // CHECK27-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
17695 // CHECK27-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
17696 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
17697 // CHECK27-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
17698 // CHECK27-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
17699 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
17700 // CHECK27-NEXT:    ret void
17701 //
17702 //
17703 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..10
17704 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
17705 // CHECK27-NEXT:  entry:
17706 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17707 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17708 // CHECK27-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
17709 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
17710 // CHECK27-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17711 // CHECK27-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
17712 // CHECK27-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
17713 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17714 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17715 // CHECK27-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
17716 // CHECK27-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
17717 // CHECK27-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17718 // CHECK27-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
17719 // CHECK27-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
17720 // CHECK27-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
17721 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17722 // CHECK27-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
17723 // CHECK27-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
17724 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
17725 // CHECK27-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
17726 // CHECK27-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
17727 // CHECK27-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
17728 // CHECK27-NEXT:    store double [[ADD]], double* [[A]], align 4
17729 // CHECK27-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
17730 // CHECK27-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
17731 // CHECK27-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
17732 // CHECK27-NEXT:    store double [[INC]], double* [[A3]], align 4
17733 // CHECK27-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
17734 // CHECK27-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
17735 // CHECK27-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
17736 // CHECK27-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
17737 // CHECK27-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
17738 // CHECK27-NEXT:    ret void
17739 //
17740 //
17741 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
17742 // CHECK27-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17743 // CHECK27-NEXT:  entry:
17744 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17745 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17746 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
17747 // CHECK27-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
17748 // CHECK27-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17749 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17750 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17751 // CHECK27-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
17752 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17753 // CHECK27-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
17754 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
17755 // CHECK27-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
17756 // CHECK27-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
17757 // CHECK27-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
17758 // CHECK27-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17759 // CHECK27-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
17760 // CHECK27-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17761 // CHECK27-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
17762 // CHECK27-NEXT:    ret void
17763 //
17764 //
17765 // CHECK27-LABEL: define {{[^@]+}}@.omp_outlined..11
17766 // CHECK27-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
17767 // CHECK27-NEXT:  entry:
17768 // CHECK27-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17769 // CHECK27-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17770 // CHECK27-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17771 // CHECK27-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17772 // CHECK27-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
17773 // CHECK27-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17774 // CHECK27-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17775 // CHECK27-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17776 // CHECK27-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17777 // CHECK27-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
17778 // CHECK27-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17779 // CHECK27-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
17780 // CHECK27-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
17781 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
17782 // CHECK27-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
17783 // CHECK27-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
17784 // CHECK27-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
17785 // CHECK27-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
17786 // CHECK27-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
17787 // CHECK27-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
17788 // CHECK27-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
17789 // CHECK27-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
17790 // CHECK27-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
17791 // CHECK27-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
17792 // CHECK27-NEXT:    ret void
17793 //
17794 //
17795 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
17796 // CHECK28-SAME: (i32 [[AA:%.*]], i32 [[DOTCAPTURE_EXPR_:%.*]], i32 [[DOTCAPTURE_EXPR_1:%.*]]) #[[ATTR0:[0-9]+]] {
17797 // CHECK28-NEXT:  entry:
17798 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17799 // CHECK28-NEXT:    [[DOTCAPTURE_EXPR__ADDR:%.*]] = alloca i32, align 4
17800 // CHECK28-NEXT:    [[DOTCAPTURE_EXPR__ADDR2:%.*]] = alloca i32, align 4
17801 // CHECK28-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17802 // CHECK28-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
17803 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17804 // CHECK28-NEXT:    store i32 [[DOTCAPTURE_EXPR_]], i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
17805 // CHECK28-NEXT:    store i32 [[DOTCAPTURE_EXPR_1]], i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
17806 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17807 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR]], align 4
17808 // CHECK28-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__ADDR2]], align 4
17809 // CHECK28-NEXT:    call void @__kmpc_push_num_teams(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP1]], i32 [[TMP2]])
17810 // CHECK28-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
17811 // CHECK28-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17812 // CHECK28-NEXT:    store i16 [[TMP3]], i16* [[CONV3]], align 2
17813 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17814 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined. to void (i32*, i32*, ...)*), i32 [[TMP4]])
17815 // CHECK28-NEXT:    ret void
17816 //
17817 //
17818 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined.
17819 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17820 // CHECK28-NEXT:  entry:
17821 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17822 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17823 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17824 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17825 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17826 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17827 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17828 // CHECK28-NEXT:    ret void
17829 //
17830 //
17831 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111
17832 // CHECK28-SAME: (i32 [[AA:%.*]]) #[[ATTR0]] {
17833 // CHECK28-NEXT:  entry:
17834 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17835 // CHECK28-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17836 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17837 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17838 // CHECK28-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
17839 // CHECK28-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17840 // CHECK28-NEXT:    store i16 [[TMP0]], i16* [[CONV1]], align 2
17841 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17842 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i32 [[TMP1]])
17843 // CHECK28-NEXT:    ret void
17844 //
17845 //
17846 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..1
17847 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17848 // CHECK28-NEXT:  entry:
17849 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17850 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17851 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17852 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17853 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17854 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17855 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17856 // CHECK28-NEXT:    [[TMP0:%.*]] = load i16, i16* [[CONV]], align 4
17857 // CHECK28-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP0]] to i32
17858 // CHECK28-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV1]], 1
17859 // CHECK28-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD]] to i16
17860 // CHECK28-NEXT:    store i16 [[CONV2]], i16* [[CONV]], align 4
17861 // CHECK28-NEXT:    ret void
17862 //
17863 //
17864 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118
17865 // CHECK28-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17866 // CHECK28-NEXT:  entry:
17867 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17868 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17869 // CHECK28-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
17870 // CHECK28-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
17871 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17872 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17873 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17874 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
17875 // CHECK28-NEXT:    store i32 [[TMP0]], i32* [[A_CASTED]], align 4
17876 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_CASTED]], align 4
17877 // CHECK28-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
17878 // CHECK28-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
17879 // CHECK28-NEXT:    store i16 [[TMP2]], i16* [[CONV1]], align 2
17880 // CHECK28-NEXT:    [[TMP3:%.*]] = load i32, i32* [[AA_CASTED]], align 4
17881 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32)* @.omp_outlined..2 to void (i32*, i32*, ...)*), i32 [[TMP1]], i32 [[TMP3]])
17882 // CHECK28-NEXT:    ret void
17883 //
17884 //
17885 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..2
17886 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]]) #[[ATTR0]] {
17887 // CHECK28-NEXT:  entry:
17888 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17889 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17890 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17891 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
17892 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17893 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17894 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17895 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
17896 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
17897 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
17898 // CHECK28-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
17899 // CHECK28-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
17900 // CHECK28-NEXT:    [[TMP1:%.*]] = load i16, i16* [[CONV]], align 4
17901 // CHECK28-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP1]] to i32
17902 // CHECK28-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
17903 // CHECK28-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
17904 // CHECK28-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
17905 // CHECK28-NEXT:    ret void
17906 //
17907 //
17908 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142
17909 // CHECK28-SAME: (i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
17910 // CHECK28-NEXT:  entry:
17911 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17912 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
17913 // CHECK28-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17914 // CHECK28-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
17915 // CHECK28-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
17916 // CHECK28-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
17917 // CHECK28-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
17918 // CHECK28-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
17919 // CHECK28-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
17920 // CHECK28-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
17921 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17922 // CHECK28-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
17923 // CHECK28-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17924 // CHECK28-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
17925 // CHECK28-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
17926 // CHECK28-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
17927 // CHECK28-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
17928 // CHECK28-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
17929 // CHECK28-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
17930 // CHECK28-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
17931 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17932 // CHECK28-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
17933 // CHECK28-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
17934 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
17935 // CHECK28-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
17936 // CHECK28-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
17937 // CHECK28-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
17938 // CHECK28-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
17939 // CHECK28-NEXT:    store i32 [[TMP8]], i32* [[A_CASTED]], align 4
17940 // CHECK28-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A_CASTED]], align 4
17941 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 9, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, [10 x float]*, i32, float*, [5 x [10 x double]]*, i32, i32, double*, %struct.TT*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32 [[TMP9]], [10 x float]* [[TMP0]], i32 [[TMP1]], float* [[TMP2]], [5 x [10 x double]]* [[TMP3]], i32 [[TMP4]], i32 [[TMP5]], double* [[TMP6]], %struct.TT* [[TMP7]])
17942 // CHECK28-NEXT:    ret void
17943 //
17944 //
17945 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..3
17946 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], [10 x float]* nonnull align 4 dereferenceable(40) [[B:%.*]], i32 [[VLA:%.*]], float* nonnull align 4 dereferenceable(4) [[BN:%.*]], [5 x [10 x double]]* nonnull align 4 dereferenceable(400) [[C:%.*]], i32 [[VLA1:%.*]], i32 [[VLA3:%.*]], double* nonnull align 4 dereferenceable(8) [[CN:%.*]], %struct.TT* nonnull align 4 dereferenceable(12) [[D:%.*]]) #[[ATTR0]] {
17947 // CHECK28-NEXT:  entry:
17948 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
17949 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
17950 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
17951 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca [10 x float]*, align 4
17952 // CHECK28-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
17953 // CHECK28-NEXT:    [[BN_ADDR:%.*]] = alloca float*, align 4
17954 // CHECK28-NEXT:    [[C_ADDR:%.*]] = alloca [5 x [10 x double]]*, align 4
17955 // CHECK28-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
17956 // CHECK28-NEXT:    [[VLA_ADDR4:%.*]] = alloca i32, align 4
17957 // CHECK28-NEXT:    [[CN_ADDR:%.*]] = alloca double*, align 4
17958 // CHECK28-NEXT:    [[D_ADDR:%.*]] = alloca %struct.TT*, align 4
17959 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
17960 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
17961 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
17962 // CHECK28-NEXT:    store [10 x float]* [[B]], [10 x float]** [[B_ADDR]], align 4
17963 // CHECK28-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
17964 // CHECK28-NEXT:    store float* [[BN]], float** [[BN_ADDR]], align 4
17965 // CHECK28-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[C_ADDR]], align 4
17966 // CHECK28-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
17967 // CHECK28-NEXT:    store i32 [[VLA3]], i32* [[VLA_ADDR4]], align 4
17968 // CHECK28-NEXT:    store double* [[CN]], double** [[CN_ADDR]], align 4
17969 // CHECK28-NEXT:    store %struct.TT* [[D]], %struct.TT** [[D_ADDR]], align 4
17970 // CHECK28-NEXT:    [[TMP0:%.*]] = load [10 x float]*, [10 x float]** [[B_ADDR]], align 4
17971 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
17972 // CHECK28-NEXT:    [[TMP2:%.*]] = load float*, float** [[BN_ADDR]], align 4
17973 // CHECK28-NEXT:    [[TMP3:%.*]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], align 4
17974 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
17975 // CHECK28-NEXT:    [[TMP5:%.*]] = load i32, i32* [[VLA_ADDR4]], align 4
17976 // CHECK28-NEXT:    [[TMP6:%.*]] = load double*, double** [[CN_ADDR]], align 4
17977 // CHECK28-NEXT:    [[TMP7:%.*]] = load %struct.TT*, %struct.TT** [[D_ADDR]], align 4
17978 // CHECK28-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_ADDR]], align 4
17979 // CHECK28-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
17980 // CHECK28-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
17981 // CHECK28-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[TMP0]], i32 0, i32 2
17982 // CHECK28-NEXT:    [[TMP9:%.*]] = load float, float* [[ARRAYIDX]], align 4
17983 // CHECK28-NEXT:    [[CONV:%.*]] = fpext float [[TMP9]] to double
17984 // CHECK28-NEXT:    [[ADD5:%.*]] = fadd double [[CONV]], 1.000000e+00
17985 // CHECK28-NEXT:    [[CONV6:%.*]] = fptrunc double [[ADD5]] to float
17986 // CHECK28-NEXT:    store float [[CONV6]], float* [[ARRAYIDX]], align 4
17987 // CHECK28-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds float, float* [[TMP2]], i32 3
17988 // CHECK28-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX7]], align 4
17989 // CHECK28-NEXT:    [[CONV8:%.*]] = fpext float [[TMP10]] to double
17990 // CHECK28-NEXT:    [[ADD9:%.*]] = fadd double [[CONV8]], 1.000000e+00
17991 // CHECK28-NEXT:    [[CONV10:%.*]] = fptrunc double [[ADD9]] to float
17992 // CHECK28-NEXT:    store float [[CONV10]], float* [[ARRAYIDX7]], align 4
17993 // CHECK28-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[TMP3]], i32 0, i32 1
17994 // CHECK28-NEXT:    [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX11]], i32 0, i32 2
17995 // CHECK28-NEXT:    [[TMP11:%.*]] = load double, double* [[ARRAYIDX12]], align 8
17996 // CHECK28-NEXT:    [[ADD13:%.*]] = fadd double [[TMP11]], 1.000000e+00
17997 // CHECK28-NEXT:    store double [[ADD13]], double* [[ARRAYIDX12]], align 8
17998 // CHECK28-NEXT:    [[TMP12:%.*]] = mul nsw i32 1, [[TMP5]]
17999 // CHECK28-NEXT:    [[ARRAYIDX14:%.*]] = getelementptr inbounds double, double* [[TMP6]], i32 [[TMP12]]
18000 // CHECK28-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX14]], i32 3
18001 // CHECK28-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX15]], align 8
18002 // CHECK28-NEXT:    [[ADD16:%.*]] = fadd double [[TMP13]], 1.000000e+00
18003 // CHECK28-NEXT:    store double [[ADD16]], double* [[ARRAYIDX15]], align 8
18004 // CHECK28-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT:%.*]], %struct.TT* [[TMP7]], i32 0, i32 0
18005 // CHECK28-NEXT:    [[TMP14:%.*]] = load i64, i64* [[X]], align 4
18006 // CHECK28-NEXT:    [[ADD17:%.*]] = add nsw i64 [[TMP14]], 1
18007 // CHECK28-NEXT:    store i64 [[ADD17]], i64* [[X]], align 4
18008 // CHECK28-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[TMP7]], i32 0, i32 1
18009 // CHECK28-NEXT:    [[TMP15:%.*]] = load i8, i8* [[Y]], align 4
18010 // CHECK28-NEXT:    [[CONV18:%.*]] = sext i8 [[TMP15]] to i32
18011 // CHECK28-NEXT:    [[ADD19:%.*]] = add nsw i32 [[CONV18]], 1
18012 // CHECK28-NEXT:    [[CONV20:%.*]] = trunc i32 [[ADD19]] to i8
18013 // CHECK28-NEXT:    store i8 [[CONV20]], i8* [[Y]], align 4
18014 // CHECK28-NEXT:    ret void
18015 //
18016 //
18017 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154
18018 // CHECK28-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
18019 // CHECK28-NEXT:  entry:
18020 // CHECK28-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
18021 // CHECK28-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
18022 // CHECK28-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
18023 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
18024 // CHECK28-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
18025 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
18026 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i32 [[TMP1]])
18027 // CHECK28-NEXT:    ret void
18028 //
18029 //
18030 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..4
18031 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
18032 // CHECK28-NEXT:  entry:
18033 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18034 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18035 // CHECK28-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
18036 // CHECK28-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
18037 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18038 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18039 // CHECK28-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
18040 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
18041 // CHECK28-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
18042 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
18043 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i32 [[TMP1]])
18044 // CHECK28-NEXT:    ret void
18045 //
18046 //
18047 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..5
18048 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
18049 // CHECK28-NEXT:  entry:
18050 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18051 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18052 // CHECK28-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
18053 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18054 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18055 // CHECK28-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
18056 // CHECK28-NEXT:    ret void
18057 //
18058 //
18059 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157
18060 // CHECK28-SAME: (i32 [[NN:%.*]]) #[[ATTR0]] {
18061 // CHECK28-NEXT:  entry:
18062 // CHECK28-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
18063 // CHECK28-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
18064 // CHECK28-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
18065 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[NN_ADDR]], align 4
18066 // CHECK28-NEXT:    store i32 [[TMP0]], i32* [[NN_CASTED]], align 4
18067 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[NN_CASTED]], align 4
18068 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32 [[TMP1]])
18069 // CHECK28-NEXT:    ret void
18070 //
18071 //
18072 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..6
18073 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[NN:%.*]]) #[[ATTR0]] {
18074 // CHECK28-NEXT:  entry:
18075 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18076 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18077 // CHECK28-NEXT:    [[NN_ADDR:%.*]] = alloca i32, align 4
18078 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18079 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18080 // CHECK28-NEXT:    store i32 [[NN]], i32* [[NN_ADDR]], align 4
18081 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..7 to void (i32*, i32*, ...)*), i32* [[NN_ADDR]])
18082 // CHECK28-NEXT:    ret void
18083 //
18084 //
18085 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..7
18086 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[NN:%.*]]) #[[ATTR0]] {
18087 // CHECK28-NEXT:  entry:
18088 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18089 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18090 // CHECK28-NEXT:    [[NN_ADDR:%.*]] = alloca i32*, align 4
18091 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18092 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18093 // CHECK28-NEXT:    store i32* [[NN]], i32** [[NN_ADDR]], align 4
18094 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[NN_ADDR]], align 4
18095 // CHECK28-NEXT:    ret void
18096 //
18097 //
18098 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182
18099 // CHECK28-SAME: (i32 [[VLA:%.*]]) #[[ATTR0]] {
18100 // CHECK28-NEXT:  entry:
18101 // CHECK28-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
18102 // CHECK28-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
18103 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
18104 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32 [[TMP0]])
18105 // CHECK28-NEXT:    ret void
18106 //
18107 //
18108 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..8
18109 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[VLA:%.*]]) #[[ATTR0]] {
18110 // CHECK28-NEXT:  entry:
18111 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18112 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18113 // CHECK28-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
18114 // CHECK28-NEXT:    [[F:%.*]] = alloca i32*, align 4
18115 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18116 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18117 // CHECK28-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
18118 // CHECK28-NEXT:    [[TMP0:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
18119 // CHECK28-NEXT:    ret void
18120 //
18121 //
18122 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209
18123 // CHECK28-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
18124 // CHECK28-NEXT:  entry:
18125 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
18126 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
18127 // CHECK28-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
18128 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
18129 // CHECK28-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
18130 // CHECK28-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
18131 // CHECK28-NEXT:    [[AAA_CASTED:%.*]] = alloca i32, align 4
18132 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
18133 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
18134 // CHECK28-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
18135 // CHECK28-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
18136 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
18137 // CHECK28-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
18138 // CHECK28-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
18139 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
18140 // CHECK28-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
18141 // CHECK28-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
18142 // CHECK28-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
18143 // CHECK28-NEXT:    [[CONV2:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
18144 // CHECK28-NEXT:    store i16 [[TMP3]], i16* [[CONV2]], align 2
18145 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
18146 // CHECK28-NEXT:    [[TMP5:%.*]] = load i8, i8* [[CONV1]], align 4
18147 // CHECK28-NEXT:    [[CONV3:%.*]] = bitcast i32* [[AAA_CASTED]] to i8*
18148 // CHECK28-NEXT:    store i8 [[TMP5]], i8* [[CONV3]], align 1
18149 // CHECK28-NEXT:    [[TMP6:%.*]] = load i32, i32* [[AAA_CASTED]], align 4
18150 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 4, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, i32, [10 x i32]*)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [10 x i32]* [[TMP0]])
18151 // CHECK28-NEXT:    ret void
18152 //
18153 //
18154 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..9
18155 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], i32 [[AAA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
18156 // CHECK28-NEXT:  entry:
18157 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18158 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18159 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
18160 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
18161 // CHECK28-NEXT:    [[AAA_ADDR:%.*]] = alloca i32, align 4
18162 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
18163 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18164 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18165 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
18166 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
18167 // CHECK28-NEXT:    store i32 [[AAA]], i32* [[AAA_ADDR]], align 4
18168 // CHECK28-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
18169 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
18170 // CHECK28-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AAA_ADDR]] to i8*
18171 // CHECK28-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
18172 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
18173 // CHECK28-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
18174 // CHECK28-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
18175 // CHECK28-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
18176 // CHECK28-NEXT:    [[CONV2:%.*]] = sext i16 [[TMP2]] to i32
18177 // CHECK28-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV2]], 1
18178 // CHECK28-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
18179 // CHECK28-NEXT:    store i16 [[CONV4]], i16* [[CONV]], align 4
18180 // CHECK28-NEXT:    [[TMP3:%.*]] = load i8, i8* [[CONV1]], align 4
18181 // CHECK28-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP3]] to i32
18182 // CHECK28-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], 1
18183 // CHECK28-NEXT:    [[CONV7:%.*]] = trunc i32 [[ADD6]] to i8
18184 // CHECK28-NEXT:    store i8 [[CONV7]], i8* [[CONV1]], align 4
18185 // CHECK28-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
18186 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
18187 // CHECK28-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP4]], 1
18188 // CHECK28-NEXT:    store i32 [[ADD8]], i32* [[ARRAYIDX]], align 4
18189 // CHECK28-NEXT:    ret void
18190 //
18191 //
18192 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
18193 // CHECK28-SAME: (%struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
18194 // CHECK28-NEXT:  entry:
18195 // CHECK28-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
18196 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
18197 // CHECK28-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
18198 // CHECK28-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
18199 // CHECK28-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
18200 // CHECK28-NEXT:    [[B_CASTED:%.*]] = alloca i32, align 4
18201 // CHECK28-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
18202 // CHECK28-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
18203 // CHECK28-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
18204 // CHECK28-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
18205 // CHECK28-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
18206 // CHECK28-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
18207 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
18208 // CHECK28-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
18209 // CHECK28-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
18210 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
18211 // CHECK28-NEXT:    store i32 [[TMP4]], i32* [[B_CASTED]], align 4
18212 // CHECK28-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B_CASTED]], align 4
18213 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 5, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.S1*, i32, i32, i32, i16*)* @.omp_outlined..10 to void (i32*, i32*, ...)*), %struct.S1* [[TMP0]], i32 [[TMP5]], i32 [[TMP1]], i32 [[TMP2]], i16* [[TMP3]])
18214 // CHECK28-NEXT:    ret void
18215 //
18216 //
18217 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..10
18218 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.S1* [[THIS:%.*]], i32 [[B:%.*]], i32 [[VLA:%.*]], i32 [[VLA1:%.*]], i16* nonnull align 2 dereferenceable(2) [[C:%.*]]) #[[ATTR0]] {
18219 // CHECK28-NEXT:  entry:
18220 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18221 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18222 // CHECK28-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
18223 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca i32, align 4
18224 // CHECK28-NEXT:    [[VLA_ADDR:%.*]] = alloca i32, align 4
18225 // CHECK28-NEXT:    [[VLA_ADDR2:%.*]] = alloca i32, align 4
18226 // CHECK28-NEXT:    [[C_ADDR:%.*]] = alloca i16*, align 4
18227 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18228 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18229 // CHECK28-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
18230 // CHECK28-NEXT:    store i32 [[B]], i32* [[B_ADDR]], align 4
18231 // CHECK28-NEXT:    store i32 [[VLA]], i32* [[VLA_ADDR]], align 4
18232 // CHECK28-NEXT:    store i32 [[VLA1]], i32* [[VLA_ADDR2]], align 4
18233 // CHECK28-NEXT:    store i16* [[C]], i16** [[C_ADDR]], align 4
18234 // CHECK28-NEXT:    [[TMP0:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
18235 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[VLA_ADDR]], align 4
18236 // CHECK28-NEXT:    [[TMP2:%.*]] = load i32, i32* [[VLA_ADDR2]], align 4
18237 // CHECK28-NEXT:    [[TMP3:%.*]] = load i16*, i16** [[C_ADDR]], align 4
18238 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B_ADDR]], align 4
18239 // CHECK28-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
18240 // CHECK28-NEXT:    [[ADD:%.*]] = fadd double [[CONV]], 1.500000e+00
18241 // CHECK28-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP0]], i32 0, i32 0
18242 // CHECK28-NEXT:    store double [[ADD]], double* [[A]], align 4
18243 // CHECK28-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[TMP0]], i32 0, i32 0
18244 // CHECK28-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
18245 // CHECK28-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
18246 // CHECK28-NEXT:    store double [[INC]], double* [[A3]], align 4
18247 // CHECK28-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
18248 // CHECK28-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP2]]
18249 // CHECK28-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[TMP3]], i32 [[TMP6]]
18250 // CHECK28-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
18251 // CHECK28-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
18252 // CHECK28-NEXT:    ret void
18253 //
18254 //
18255 // CHECK28-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192
18256 // CHECK28-SAME: (i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
18257 // CHECK28-NEXT:  entry:
18258 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
18259 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
18260 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
18261 // CHECK28-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
18262 // CHECK28-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4
18263 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
18264 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
18265 // CHECK28-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
18266 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
18267 // CHECK28-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
18268 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
18269 // CHECK28-NEXT:    store i32 [[TMP1]], i32* [[A_CASTED]], align 4
18270 // CHECK28-NEXT:    [[TMP2:%.*]] = load i32, i32* [[A_CASTED]], align 4
18271 // CHECK28-NEXT:    [[TMP3:%.*]] = load i16, i16* [[CONV]], align 4
18272 // CHECK28-NEXT:    [[CONV1:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
18273 // CHECK28-NEXT:    store i16 [[TMP3]], i16* [[CONV1]], align 2
18274 // CHECK28-NEXT:    [[TMP4:%.*]] = load i32, i32* [[AA_CASTED]], align 4
18275 // CHECK28-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB1]], i32 3, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32, i32, [10 x i32]*)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i32 [[TMP2]], i32 [[TMP4]], [10 x i32]* [[TMP0]])
18276 // CHECK28-NEXT:    ret void
18277 //
18278 //
18279 // CHECK28-LABEL: define {{[^@]+}}@.omp_outlined..11
18280 // CHECK28-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32 [[A:%.*]], i32 [[AA:%.*]], [10 x i32]* nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {
18281 // CHECK28-NEXT:  entry:
18282 // CHECK28-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
18283 // CHECK28-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
18284 // CHECK28-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4
18285 // CHECK28-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4
18286 // CHECK28-NEXT:    [[B_ADDR:%.*]] = alloca [10 x i32]*, align 4
18287 // CHECK28-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
18288 // CHECK28-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
18289 // CHECK28-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
18290 // CHECK28-NEXT:    store i32 [[AA]], i32* [[AA_ADDR]], align 4
18291 // CHECK28-NEXT:    store [10 x i32]* [[B]], [10 x i32]** [[B_ADDR]], align 4
18292 // CHECK28-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_ADDR]] to i16*
18293 // CHECK28-NEXT:    [[TMP0:%.*]] = load [10 x i32]*, [10 x i32]** [[B_ADDR]], align 4
18294 // CHECK28-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A_ADDR]], align 4
18295 // CHECK28-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1
18296 // CHECK28-NEXT:    store i32 [[ADD]], i32* [[A_ADDR]], align 4
18297 // CHECK28-NEXT:    [[TMP2:%.*]] = load i16, i16* [[CONV]], align 4
18298 // CHECK28-NEXT:    [[CONV1:%.*]] = sext i16 [[TMP2]] to i32
18299 // CHECK28-NEXT:    [[ADD2:%.*]] = add nsw i32 [[CONV1]], 1
18300 // CHECK28-NEXT:    [[CONV3:%.*]] = trunc i32 [[ADD2]] to i16
18301 // CHECK28-NEXT:    store i16 [[CONV3]], i16* [[CONV]], align 4
18302 // CHECK28-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[TMP0]], i32 0, i32 2
18303 // CHECK28-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
18304 // CHECK28-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], 1
18305 // CHECK28-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX]], align 4
18306 // CHECK28-NEXT:    ret void
18307 //
18308 //
18309 // CHECK29-LABEL: define {{[^@]+}}@_Z3fooi
18310 // CHECK29-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
18311 // CHECK29-NEXT:  entry:
18312 // CHECK29-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18313 // CHECK29-NEXT:    [[A:%.*]] = alloca i32, align 4
18314 // CHECK29-NEXT:    [[AA:%.*]] = alloca i16, align 2
18315 // CHECK29-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
18316 // CHECK29-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
18317 // CHECK29-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
18318 // CHECK29-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
18319 // CHECK29-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
18320 // CHECK29-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
18321 // CHECK29-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
18322 // CHECK29-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
18323 // CHECK29-NEXT:    [[NN:%.*]] = alloca i32, align 4
18324 // CHECK29-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18325 // CHECK29-NEXT:    store i32 0, i32* [[A]], align 4
18326 // CHECK29-NEXT:    store i16 0, i16* [[AA]], align 2
18327 // CHECK29-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18328 // CHECK29-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
18329 // CHECK29-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
18330 // CHECK29-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
18331 // CHECK29-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
18332 // CHECK29-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
18333 // CHECK29-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
18334 // CHECK29-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
18335 // CHECK29-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
18336 // CHECK29-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
18337 // CHECK29-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
18338 // CHECK29-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
18339 // CHECK29-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
18340 // CHECK29-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
18341 // CHECK29-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
18342 // CHECK29-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
18343 // CHECK29-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
18344 // CHECK29-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18345 // CHECK29-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
18346 // CHECK29-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
18347 // CHECK29-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
18348 // CHECK29-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
18349 // CHECK29-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
18350 // CHECK29-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
18351 // CHECK29-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
18352 // CHECK29-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
18353 // CHECK29-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
18354 // CHECK29-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
18355 // CHECK29-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
18356 // CHECK29-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
18357 // CHECK29-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
18358 // CHECK29-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
18359 // CHECK29-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
18360 // CHECK29-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
18361 // CHECK29-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
18362 // CHECK29-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
18363 // CHECK29-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
18364 // CHECK29-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
18365 // CHECK29-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
18366 // CHECK29-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
18367 // CHECK29-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
18368 // CHECK29-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
18369 // CHECK29-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
18370 // CHECK29-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
18371 // CHECK29-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
18372 // CHECK29-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
18373 // CHECK29-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
18374 // CHECK29-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
18375 // CHECK29-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
18376 // CHECK29-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
18377 // CHECK29-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
18378 // CHECK29-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
18379 // CHECK29-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
18380 // CHECK29-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
18381 // CHECK29-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
18382 // CHECK29-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
18383 // CHECK29-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
18384 // CHECK29-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
18385 // CHECK29-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
18386 // CHECK29-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
18387 // CHECK29-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
18388 // CHECK29-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
18389 // CHECK29-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
18390 // CHECK29-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
18391 // CHECK29-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
18392 // CHECK29-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
18393 // CHECK29-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
18394 // CHECK29-NEXT:    store i32 0, i32* [[NN]], align 4
18395 // CHECK29-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
18396 // CHECK29-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
18397 // CHECK29-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
18398 // CHECK29-NEXT:    ret i32 [[TMP20]]
18399 //
18400 //
18401 // CHECK29-LABEL: define {{[^@]+}}@_Z6bazzzziPi
18402 // CHECK29-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
18403 // CHECK29-NEXT:  entry:
18404 // CHECK29-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18405 // CHECK29-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
18406 // CHECK29-NEXT:    [[F1:%.*]] = alloca i32*, align 8
18407 // CHECK29-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18408 // CHECK29-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
18409 // CHECK29-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18410 // CHECK29-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
18411 // CHECK29-NEXT:    ret void
18412 //
18413 //
18414 // CHECK29-LABEL: define {{[^@]+}}@_Z3bari
18415 // CHECK29-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
18416 // CHECK29-NEXT:  entry:
18417 // CHECK29-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18418 // CHECK29-NEXT:    [[A:%.*]] = alloca i32, align 4
18419 // CHECK29-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
18420 // CHECK29-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18421 // CHECK29-NEXT:    store i32 0, i32* [[A]], align 4
18422 // CHECK29-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18423 // CHECK29-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
18424 // CHECK29-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
18425 // CHECK29-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
18426 // CHECK29-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18427 // CHECK29-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
18428 // CHECK29-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
18429 // CHECK29-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
18430 // CHECK29-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
18431 // CHECK29-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
18432 // CHECK29-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
18433 // CHECK29-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
18434 // CHECK29-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
18435 // CHECK29-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
18436 // CHECK29-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
18437 // CHECK29-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
18438 // CHECK29-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
18439 // CHECK29-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
18440 // CHECK29-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
18441 // CHECK29-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
18442 // CHECK29-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
18443 // CHECK29-NEXT:    ret i32 [[TMP8]]
18444 //
18445 //
18446 // CHECK29-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
18447 // CHECK29-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
18448 // CHECK29-NEXT:  entry:
18449 // CHECK29-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
18450 // CHECK29-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18451 // CHECK29-NEXT:    [[B:%.*]] = alloca i32, align 4
18452 // CHECK29-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
18453 // CHECK29-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
18454 // CHECK29-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
18455 // CHECK29-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18456 // CHECK29-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
18457 // CHECK29-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18458 // CHECK29-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18459 // CHECK29-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
18460 // CHECK29-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
18461 // CHECK29-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
18462 // CHECK29-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
18463 // CHECK29-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
18464 // CHECK29-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
18465 // CHECK29-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
18466 // CHECK29-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
18467 // CHECK29-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
18468 // CHECK29-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
18469 // CHECK29-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
18470 // CHECK29-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
18471 // CHECK29-NEXT:    store double [[ADD2]], double* [[A]], align 8
18472 // CHECK29-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
18473 // CHECK29-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
18474 // CHECK29-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
18475 // CHECK29-NEXT:    store double [[INC]], double* [[A3]], align 8
18476 // CHECK29-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
18477 // CHECK29-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
18478 // CHECK29-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
18479 // CHECK29-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
18480 // CHECK29-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
18481 // CHECK29-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
18482 // CHECK29-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
18483 // CHECK29-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
18484 // CHECK29-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
18485 // CHECK29-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
18486 // CHECK29-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
18487 // CHECK29-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
18488 // CHECK29-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
18489 // CHECK29-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
18490 // CHECK29-NEXT:    ret i32 [[ADD9]]
18491 //
18492 //
18493 // CHECK29-LABEL: define {{[^@]+}}@_ZL7fstatici
18494 // CHECK29-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
18495 // CHECK29-NEXT:  entry:
18496 // CHECK29-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18497 // CHECK29-NEXT:    [[A:%.*]] = alloca i32, align 4
18498 // CHECK29-NEXT:    [[AA:%.*]] = alloca i16, align 2
18499 // CHECK29-NEXT:    [[AAA:%.*]] = alloca i8, align 1
18500 // CHECK29-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
18501 // CHECK29-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18502 // CHECK29-NEXT:    store i32 0, i32* [[A]], align 4
18503 // CHECK29-NEXT:    store i16 0, i16* [[AA]], align 2
18504 // CHECK29-NEXT:    store i8 0, i8* [[AAA]], align 1
18505 // CHECK29-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
18506 // CHECK29-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18507 // CHECK29-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18508 // CHECK29-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
18509 // CHECK29-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
18510 // CHECK29-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
18511 // CHECK29-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
18512 // CHECK29-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
18513 // CHECK29-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
18514 // CHECK29-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
18515 // CHECK29-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
18516 // CHECK29-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
18517 // CHECK29-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
18518 // CHECK29-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
18519 // CHECK29-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
18520 // CHECK29-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
18521 // CHECK29-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
18522 // CHECK29-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
18523 // CHECK29-NEXT:    ret i32 [[TMP4]]
18524 //
18525 //
18526 // CHECK29-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
18527 // CHECK29-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
18528 // CHECK29-NEXT:  entry:
18529 // CHECK29-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18530 // CHECK29-NEXT:    [[A:%.*]] = alloca i32, align 4
18531 // CHECK29-NEXT:    [[AA:%.*]] = alloca i16, align 2
18532 // CHECK29-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
18533 // CHECK29-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18534 // CHECK29-NEXT:    store i32 0, i32* [[A]], align 4
18535 // CHECK29-NEXT:    store i16 0, i16* [[AA]], align 2
18536 // CHECK29-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
18537 // CHECK29-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18538 // CHECK29-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18539 // CHECK29-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
18540 // CHECK29-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
18541 // CHECK29-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
18542 // CHECK29-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
18543 // CHECK29-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
18544 // CHECK29-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
18545 // CHECK29-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
18546 // CHECK29-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
18547 // CHECK29-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
18548 // CHECK29-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
18549 // CHECK29-NEXT:    ret i32 [[TMP3]]
18550 //
18551 //
18552 // CHECK30-LABEL: define {{[^@]+}}@_Z3fooi
18553 // CHECK30-SAME: (i32 signext [[N:%.*]]) #[[ATTR0:[0-9]+]] {
18554 // CHECK30-NEXT:  entry:
18555 // CHECK30-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18556 // CHECK30-NEXT:    [[A:%.*]] = alloca i32, align 4
18557 // CHECK30-NEXT:    [[AA:%.*]] = alloca i16, align 2
18558 // CHECK30-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
18559 // CHECK30-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
18560 // CHECK30-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
18561 // CHECK30-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
18562 // CHECK30-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i64, align 8
18563 // CHECK30-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 8
18564 // CHECK30-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
18565 // CHECK30-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
18566 // CHECK30-NEXT:    [[NN:%.*]] = alloca i32, align 4
18567 // CHECK30-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18568 // CHECK30-NEXT:    store i32 0, i32* [[A]], align 4
18569 // CHECK30-NEXT:    store i16 0, i16* [[AA]], align 2
18570 // CHECK30-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18571 // CHECK30-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
18572 // CHECK30-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
18573 // CHECK30-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
18574 // CHECK30-NEXT:    [[VLA:%.*]] = alloca float, i64 [[TMP1]], align 4
18575 // CHECK30-NEXT:    store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
18576 // CHECK30-NEXT:    [[TMP3:%.*]] = load i32, i32* [[N_ADDR]], align 4
18577 // CHECK30-NEXT:    [[TMP4:%.*]] = zext i32 [[TMP3]] to i64
18578 // CHECK30-NEXT:    [[TMP5:%.*]] = mul nuw i64 5, [[TMP4]]
18579 // CHECK30-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP5]], align 8
18580 // CHECK30-NEXT:    store i64 [[TMP4]], i64* [[__VLA_EXPR1]], align 8
18581 // CHECK30-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
18582 // CHECK30-NEXT:    store i32 [[TMP6]], i32* [[DOTCAPTURE_EXPR_]], align 4
18583 // CHECK30-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
18584 // CHECK30-NEXT:    store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_2]], align 4
18585 // CHECK30-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
18586 // CHECK30-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
18587 // CHECK30-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18588 // CHECK30-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
18589 // CHECK30-NEXT:    [[CONV:%.*]] = sext i16 [[TMP9]] to i32
18590 // CHECK30-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
18591 // CHECK30-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
18592 // CHECK30-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
18593 // CHECK30-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
18594 // CHECK30-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP10]], 1
18595 // CHECK30-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
18596 // CHECK30-NEXT:    [[TMP11:%.*]] = load i16, i16* [[AA]], align 2
18597 // CHECK30-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP11]] to i32
18598 // CHECK30-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
18599 // CHECK30-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
18600 // CHECK30-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
18601 // CHECK30-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4
18602 // CHECK30-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP12]], 1
18603 // CHECK30-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
18604 // CHECK30-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i64 0, i64 2
18605 // CHECK30-NEXT:    [[TMP13:%.*]] = load float, float* [[ARRAYIDX]], align 4
18606 // CHECK30-NEXT:    [[CONV10:%.*]] = fpext float [[TMP13]] to double
18607 // CHECK30-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
18608 // CHECK30-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
18609 // CHECK30-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
18610 // CHECK30-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i64 3
18611 // CHECK30-NEXT:    [[TMP14:%.*]] = load float, float* [[ARRAYIDX13]], align 4
18612 // CHECK30-NEXT:    [[CONV14:%.*]] = fpext float [[TMP14]] to double
18613 // CHECK30-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
18614 // CHECK30-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
18615 // CHECK30-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
18616 // CHECK30-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i64 0, i64 1
18617 // CHECK30-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i64 0, i64 2
18618 // CHECK30-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX18]], align 8
18619 // CHECK30-NEXT:    [[ADD19:%.*]] = fadd double [[TMP15]], 1.000000e+00
18620 // CHECK30-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
18621 // CHECK30-NEXT:    [[TMP16:%.*]] = mul nsw i64 1, [[TMP4]]
18622 // CHECK30-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i64 [[TMP16]]
18623 // CHECK30-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i64 3
18624 // CHECK30-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21]], align 8
18625 // CHECK30-NEXT:    [[ADD22:%.*]] = fadd double [[TMP17]], 1.000000e+00
18626 // CHECK30-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
18627 // CHECK30-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
18628 // CHECK30-NEXT:    [[TMP18:%.*]] = load i64, i64* [[X]], align 8
18629 // CHECK30-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP18]], 1
18630 // CHECK30-NEXT:    store i64 [[ADD23]], i64* [[X]], align 8
18631 // CHECK30-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
18632 // CHECK30-NEXT:    [[TMP19:%.*]] = load i8, i8* [[Y]], align 8
18633 // CHECK30-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP19]] to i32
18634 // CHECK30-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
18635 // CHECK30-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
18636 // CHECK30-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 8
18637 // CHECK30-NEXT:    store i32 0, i32* [[NN]], align 4
18638 // CHECK30-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A]], align 4
18639 // CHECK30-NEXT:    [[TMP21:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
18640 // CHECK30-NEXT:    call void @llvm.stackrestore(i8* [[TMP21]])
18641 // CHECK30-NEXT:    ret i32 [[TMP20]]
18642 //
18643 //
18644 // CHECK30-LABEL: define {{[^@]+}}@_Z6bazzzziPi
18645 // CHECK30-SAME: (i32 signext [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
18646 // CHECK30-NEXT:  entry:
18647 // CHECK30-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18648 // CHECK30-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 8
18649 // CHECK30-NEXT:    [[F1:%.*]] = alloca i32*, align 8
18650 // CHECK30-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18651 // CHECK30-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 8
18652 // CHECK30-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18653 // CHECK30-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
18654 // CHECK30-NEXT:    ret void
18655 //
18656 //
18657 // CHECK30-LABEL: define {{[^@]+}}@_Z3bari
18658 // CHECK30-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
18659 // CHECK30-NEXT:  entry:
18660 // CHECK30-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18661 // CHECK30-NEXT:    [[A:%.*]] = alloca i32, align 4
18662 // CHECK30-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 8
18663 // CHECK30-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18664 // CHECK30-NEXT:    store i32 0, i32* [[A]], align 4
18665 // CHECK30-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18666 // CHECK30-NEXT:    [[CALL:%.*]] = call signext i32 @_Z3fooi(i32 signext [[TMP0]])
18667 // CHECK30-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
18668 // CHECK30-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
18669 // CHECK30-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18670 // CHECK30-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
18671 // CHECK30-NEXT:    [[CALL1:%.*]] = call signext i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 8 dereferenceable(8) [[S]], i32 signext [[TMP2]])
18672 // CHECK30-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
18673 // CHECK30-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
18674 // CHECK30-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
18675 // CHECK30-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
18676 // CHECK30-NEXT:    [[CALL3:%.*]] = call signext i32 @_ZL7fstatici(i32 signext [[TMP4]])
18677 // CHECK30-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
18678 // CHECK30-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
18679 // CHECK30-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
18680 // CHECK30-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
18681 // CHECK30-NEXT:    [[CALL5:%.*]] = call signext i32 @_Z9ftemplateIiET_i(i32 signext [[TMP6]])
18682 // CHECK30-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
18683 // CHECK30-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
18684 // CHECK30-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
18685 // CHECK30-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
18686 // CHECK30-NEXT:    ret i32 [[TMP8]]
18687 //
18688 //
18689 // CHECK30-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
18690 // CHECK30-SAME: (%struct.S1* nonnull align 8 dereferenceable(8) [[THIS:%.*]], i32 signext [[N:%.*]]) #[[ATTR0]] comdat align 2 {
18691 // CHECK30-NEXT:  entry:
18692 // CHECK30-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
18693 // CHECK30-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18694 // CHECK30-NEXT:    [[B:%.*]] = alloca i32, align 4
18695 // CHECK30-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 8
18696 // CHECK30-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i64, align 8
18697 // CHECK30-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
18698 // CHECK30-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18699 // CHECK30-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
18700 // CHECK30-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18701 // CHECK30-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18702 // CHECK30-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
18703 // CHECK30-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
18704 // CHECK30-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
18705 // CHECK30-NEXT:    [[TMP3:%.*]] = call i8* @llvm.stacksave()
18706 // CHECK30-NEXT:    store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
18707 // CHECK30-NEXT:    [[TMP4:%.*]] = mul nuw i64 2, [[TMP2]]
18708 // CHECK30-NEXT:    [[VLA:%.*]] = alloca i16, i64 [[TMP4]], align 2
18709 // CHECK30-NEXT:    store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
18710 // CHECK30-NEXT:    [[TMP5:%.*]] = load i32, i32* [[B]], align 4
18711 // CHECK30-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP5]] to double
18712 // CHECK30-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
18713 // CHECK30-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
18714 // CHECK30-NEXT:    store double [[ADD2]], double* [[A]], align 8
18715 // CHECK30-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
18716 // CHECK30-NEXT:    [[TMP6:%.*]] = load double, double* [[A3]], align 8
18717 // CHECK30-NEXT:    [[INC:%.*]] = fadd double [[TMP6]], 1.000000e+00
18718 // CHECK30-NEXT:    store double [[INC]], double* [[A3]], align 8
18719 // CHECK30-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
18720 // CHECK30-NEXT:    [[TMP7:%.*]] = mul nsw i64 1, [[TMP2]]
18721 // CHECK30-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP7]]
18722 // CHECK30-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
18723 // CHECK30-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
18724 // CHECK30-NEXT:    [[TMP8:%.*]] = mul nsw i64 1, [[TMP2]]
18725 // CHECK30-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP8]]
18726 // CHECK30-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i64 1
18727 // CHECK30-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
18728 // CHECK30-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP9]] to i32
18729 // CHECK30-NEXT:    [[TMP10:%.*]] = load i32, i32* [[B]], align 4
18730 // CHECK30-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP10]]
18731 // CHECK30-NEXT:    [[TMP11:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
18732 // CHECK30-NEXT:    call void @llvm.stackrestore(i8* [[TMP11]])
18733 // CHECK30-NEXT:    ret i32 [[ADD9]]
18734 //
18735 //
18736 // CHECK30-LABEL: define {{[^@]+}}@_ZL7fstatici
18737 // CHECK30-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] {
18738 // CHECK30-NEXT:  entry:
18739 // CHECK30-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18740 // CHECK30-NEXT:    [[A:%.*]] = alloca i32, align 4
18741 // CHECK30-NEXT:    [[AA:%.*]] = alloca i16, align 2
18742 // CHECK30-NEXT:    [[AAA:%.*]] = alloca i8, align 1
18743 // CHECK30-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
18744 // CHECK30-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18745 // CHECK30-NEXT:    store i32 0, i32* [[A]], align 4
18746 // CHECK30-NEXT:    store i16 0, i16* [[AA]], align 2
18747 // CHECK30-NEXT:    store i8 0, i8* [[AAA]], align 1
18748 // CHECK30-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
18749 // CHECK30-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18750 // CHECK30-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18751 // CHECK30-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
18752 // CHECK30-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
18753 // CHECK30-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
18754 // CHECK30-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
18755 // CHECK30-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
18756 // CHECK30-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
18757 // CHECK30-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
18758 // CHECK30-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
18759 // CHECK30-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
18760 // CHECK30-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
18761 // CHECK30-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
18762 // CHECK30-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
18763 // CHECK30-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
18764 // CHECK30-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
18765 // CHECK30-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
18766 // CHECK30-NEXT:    ret i32 [[TMP4]]
18767 //
18768 //
18769 // CHECK30-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
18770 // CHECK30-SAME: (i32 signext [[N:%.*]]) #[[ATTR0]] comdat {
18771 // CHECK30-NEXT:  entry:
18772 // CHECK30-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18773 // CHECK30-NEXT:    [[A:%.*]] = alloca i32, align 4
18774 // CHECK30-NEXT:    [[AA:%.*]] = alloca i16, align 2
18775 // CHECK30-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
18776 // CHECK30-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18777 // CHECK30-NEXT:    store i32 0, i32* [[A]], align 4
18778 // CHECK30-NEXT:    store i16 0, i16* [[AA]], align 2
18779 // CHECK30-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
18780 // CHECK30-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18781 // CHECK30-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18782 // CHECK30-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
18783 // CHECK30-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
18784 // CHECK30-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
18785 // CHECK30-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
18786 // CHECK30-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
18787 // CHECK30-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i64 0, i64 2
18788 // CHECK30-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
18789 // CHECK30-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
18790 // CHECK30-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
18791 // CHECK30-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
18792 // CHECK30-NEXT:    ret i32 [[TMP3]]
18793 //
18794 //
18795 // CHECK31-LABEL: define {{[^@]+}}@_Z3fooi
18796 // CHECK31-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
18797 // CHECK31-NEXT:  entry:
18798 // CHECK31-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18799 // CHECK31-NEXT:    [[A:%.*]] = alloca i32, align 4
18800 // CHECK31-NEXT:    [[AA:%.*]] = alloca i16, align 2
18801 // CHECK31-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
18802 // CHECK31-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
18803 // CHECK31-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
18804 // CHECK31-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
18805 // CHECK31-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
18806 // CHECK31-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
18807 // CHECK31-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
18808 // CHECK31-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
18809 // CHECK31-NEXT:    [[NN:%.*]] = alloca i32, align 4
18810 // CHECK31-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18811 // CHECK31-NEXT:    store i32 0, i32* [[A]], align 4
18812 // CHECK31-NEXT:    store i16 0, i16* [[AA]], align 2
18813 // CHECK31-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18814 // CHECK31-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
18815 // CHECK31-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
18816 // CHECK31-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
18817 // CHECK31-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
18818 // CHECK31-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
18819 // CHECK31-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
18820 // CHECK31-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
18821 // CHECK31-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
18822 // CHECK31-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
18823 // CHECK31-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
18824 // CHECK31-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
18825 // CHECK31-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
18826 // CHECK31-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
18827 // CHECK31-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
18828 // CHECK31-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18829 // CHECK31-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
18830 // CHECK31-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
18831 // CHECK31-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
18832 // CHECK31-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
18833 // CHECK31-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
18834 // CHECK31-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
18835 // CHECK31-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
18836 // CHECK31-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
18837 // CHECK31-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
18838 // CHECK31-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
18839 // CHECK31-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
18840 // CHECK31-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
18841 // CHECK31-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
18842 // CHECK31-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
18843 // CHECK31-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
18844 // CHECK31-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
18845 // CHECK31-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
18846 // CHECK31-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
18847 // CHECK31-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
18848 // CHECK31-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
18849 // CHECK31-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
18850 // CHECK31-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
18851 // CHECK31-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
18852 // CHECK31-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
18853 // CHECK31-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
18854 // CHECK31-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
18855 // CHECK31-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
18856 // CHECK31-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
18857 // CHECK31-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
18858 // CHECK31-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
18859 // CHECK31-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
18860 // CHECK31-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
18861 // CHECK31-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
18862 // CHECK31-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
18863 // CHECK31-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
18864 // CHECK31-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
18865 // CHECK31-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
18866 // CHECK31-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
18867 // CHECK31-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
18868 // CHECK31-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
18869 // CHECK31-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
18870 // CHECK31-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
18871 // CHECK31-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
18872 // CHECK31-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
18873 // CHECK31-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
18874 // CHECK31-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
18875 // CHECK31-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
18876 // CHECK31-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
18877 // CHECK31-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
18878 // CHECK31-NEXT:    store i32 0, i32* [[NN]], align 4
18879 // CHECK31-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
18880 // CHECK31-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
18881 // CHECK31-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
18882 // CHECK31-NEXT:    ret i32 [[TMP18]]
18883 //
18884 //
18885 // CHECK31-LABEL: define {{[^@]+}}@_Z6bazzzziPi
18886 // CHECK31-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
18887 // CHECK31-NEXT:  entry:
18888 // CHECK31-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18889 // CHECK31-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
18890 // CHECK31-NEXT:    [[F1:%.*]] = alloca i32*, align 4
18891 // CHECK31-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18892 // CHECK31-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
18893 // CHECK31-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18894 // CHECK31-NEXT:    ret void
18895 //
18896 //
18897 // CHECK31-LABEL: define {{[^@]+}}@_Z3bari
18898 // CHECK31-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
18899 // CHECK31-NEXT:  entry:
18900 // CHECK31-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18901 // CHECK31-NEXT:    [[A:%.*]] = alloca i32, align 4
18902 // CHECK31-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
18903 // CHECK31-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18904 // CHECK31-NEXT:    store i32 0, i32* [[A]], align 4
18905 // CHECK31-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18906 // CHECK31-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
18907 // CHECK31-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
18908 // CHECK31-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
18909 // CHECK31-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18910 // CHECK31-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
18911 // CHECK31-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
18912 // CHECK31-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
18913 // CHECK31-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
18914 // CHECK31-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
18915 // CHECK31-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
18916 // CHECK31-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
18917 // CHECK31-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
18918 // CHECK31-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
18919 // CHECK31-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
18920 // CHECK31-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
18921 // CHECK31-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
18922 // CHECK31-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
18923 // CHECK31-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
18924 // CHECK31-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
18925 // CHECK31-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
18926 // CHECK31-NEXT:    ret i32 [[TMP8]]
18927 //
18928 //
18929 // CHECK31-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
18930 // CHECK31-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
18931 // CHECK31-NEXT:  entry:
18932 // CHECK31-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
18933 // CHECK31-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18934 // CHECK31-NEXT:    [[B:%.*]] = alloca i32, align 4
18935 // CHECK31-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
18936 // CHECK31-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
18937 // CHECK31-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
18938 // CHECK31-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18939 // CHECK31-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
18940 // CHECK31-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
18941 // CHECK31-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18942 // CHECK31-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
18943 // CHECK31-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
18944 // CHECK31-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
18945 // CHECK31-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
18946 // CHECK31-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
18947 // CHECK31-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
18948 // CHECK31-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
18949 // CHECK31-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
18950 // CHECK31-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
18951 // CHECK31-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
18952 // CHECK31-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
18953 // CHECK31-NEXT:    store double [[ADD2]], double* [[A]], align 4
18954 // CHECK31-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
18955 // CHECK31-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
18956 // CHECK31-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
18957 // CHECK31-NEXT:    store double [[INC]], double* [[A3]], align 4
18958 // CHECK31-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
18959 // CHECK31-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
18960 // CHECK31-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
18961 // CHECK31-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
18962 // CHECK31-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
18963 // CHECK31-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
18964 // CHECK31-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
18965 // CHECK31-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
18966 // CHECK31-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
18967 // CHECK31-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
18968 // CHECK31-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
18969 // CHECK31-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
18970 // CHECK31-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
18971 // CHECK31-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
18972 // CHECK31-NEXT:    ret i32 [[ADD9]]
18973 //
18974 //
18975 // CHECK31-LABEL: define {{[^@]+}}@_ZL7fstatici
18976 // CHECK31-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
18977 // CHECK31-NEXT:  entry:
18978 // CHECK31-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
18979 // CHECK31-NEXT:    [[A:%.*]] = alloca i32, align 4
18980 // CHECK31-NEXT:    [[AA:%.*]] = alloca i16, align 2
18981 // CHECK31-NEXT:    [[AAA:%.*]] = alloca i8, align 1
18982 // CHECK31-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
18983 // CHECK31-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
18984 // CHECK31-NEXT:    store i32 0, i32* [[A]], align 4
18985 // CHECK31-NEXT:    store i16 0, i16* [[AA]], align 2
18986 // CHECK31-NEXT:    store i8 0, i8* [[AAA]], align 1
18987 // CHECK31-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
18988 // CHECK31-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
18989 // CHECK31-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
18990 // CHECK31-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
18991 // CHECK31-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
18992 // CHECK31-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
18993 // CHECK31-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
18994 // CHECK31-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
18995 // CHECK31-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
18996 // CHECK31-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
18997 // CHECK31-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
18998 // CHECK31-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
18999 // CHECK31-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
19000 // CHECK31-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
19001 // CHECK31-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
19002 // CHECK31-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
19003 // CHECK31-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
19004 // CHECK31-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
19005 // CHECK31-NEXT:    ret i32 [[TMP4]]
19006 //
19007 //
19008 // CHECK31-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
19009 // CHECK31-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
19010 // CHECK31-NEXT:  entry:
19011 // CHECK31-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19012 // CHECK31-NEXT:    [[A:%.*]] = alloca i32, align 4
19013 // CHECK31-NEXT:    [[AA:%.*]] = alloca i16, align 2
19014 // CHECK31-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
19015 // CHECK31-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19016 // CHECK31-NEXT:    store i32 0, i32* [[A]], align 4
19017 // CHECK31-NEXT:    store i16 0, i16* [[AA]], align 2
19018 // CHECK31-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
19019 // CHECK31-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
19020 // CHECK31-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
19021 // CHECK31-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
19022 // CHECK31-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
19023 // CHECK31-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
19024 // CHECK31-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
19025 // CHECK31-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
19026 // CHECK31-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
19027 // CHECK31-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
19028 // CHECK31-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
19029 // CHECK31-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
19030 // CHECK31-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
19031 // CHECK31-NEXT:    ret i32 [[TMP3]]
19032 //
19033 //
19034 // CHECK32-LABEL: define {{[^@]+}}@_Z3fooi
19035 // CHECK32-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
19036 // CHECK32-NEXT:  entry:
19037 // CHECK32-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19038 // CHECK32-NEXT:    [[A:%.*]] = alloca i32, align 4
19039 // CHECK32-NEXT:    [[AA:%.*]] = alloca i16, align 2
19040 // CHECK32-NEXT:    [[B:%.*]] = alloca [10 x float], align 4
19041 // CHECK32-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
19042 // CHECK32-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
19043 // CHECK32-NEXT:    [[C:%.*]] = alloca [5 x [10 x double]], align 8
19044 // CHECK32-NEXT:    [[__VLA_EXPR1:%.*]] = alloca i32, align 4
19045 // CHECK32-NEXT:    [[D:%.*]] = alloca [[STRUCT_TT:%.*]], align 4
19046 // CHECK32-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
19047 // CHECK32-NEXT:    [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
19048 // CHECK32-NEXT:    [[NN:%.*]] = alloca i32, align 4
19049 // CHECK32-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19050 // CHECK32-NEXT:    store i32 0, i32* [[A]], align 4
19051 // CHECK32-NEXT:    store i16 0, i16* [[AA]], align 2
19052 // CHECK32-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
19053 // CHECK32-NEXT:    [[TMP1:%.*]] = call i8* @llvm.stacksave()
19054 // CHECK32-NEXT:    store i8* [[TMP1]], i8** [[SAVED_STACK]], align 4
19055 // CHECK32-NEXT:    [[VLA:%.*]] = alloca float, i32 [[TMP0]], align 4
19056 // CHECK32-NEXT:    store i32 [[TMP0]], i32* [[__VLA_EXPR0]], align 4
19057 // CHECK32-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
19058 // CHECK32-NEXT:    [[TMP3:%.*]] = mul nuw i32 5, [[TMP2]]
19059 // CHECK32-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP3]], align 8
19060 // CHECK32-NEXT:    store i32 [[TMP2]], i32* [[__VLA_EXPR1]], align 4
19061 // CHECK32-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
19062 // CHECK32-NEXT:    store i32 [[TMP4]], i32* [[DOTCAPTURE_EXPR_]], align 4
19063 // CHECK32-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
19064 // CHECK32-NEXT:    store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_2]], align 4
19065 // CHECK32-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4
19066 // CHECK32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP6]], 1
19067 // CHECK32-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
19068 // CHECK32-NEXT:    [[TMP7:%.*]] = load i16, i16* [[AA]], align 2
19069 // CHECK32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP7]] to i32
19070 // CHECK32-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], 1
19071 // CHECK32-NEXT:    [[CONV4:%.*]] = trunc i32 [[ADD3]] to i16
19072 // CHECK32-NEXT:    store i16 [[CONV4]], i16* [[AA]], align 2
19073 // CHECK32-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
19074 // CHECK32-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP8]], 1
19075 // CHECK32-NEXT:    store i32 [[ADD5]], i32* [[A]], align 4
19076 // CHECK32-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
19077 // CHECK32-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP9]] to i32
19078 // CHECK32-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], 1
19079 // CHECK32-NEXT:    [[CONV8:%.*]] = trunc i32 [[ADD7]] to i16
19080 // CHECK32-NEXT:    store i16 [[CONV8]], i16* [[AA]], align 2
19081 // CHECK32-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4
19082 // CHECK32-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP10]], 1
19083 // CHECK32-NEXT:    store i32 [[ADD9]], i32* [[A]], align 4
19084 // CHECK32-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], [10 x float]* [[B]], i32 0, i32 2
19085 // CHECK32-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX]], align 4
19086 // CHECK32-NEXT:    [[CONV10:%.*]] = fpext float [[TMP11]] to double
19087 // CHECK32-NEXT:    [[ADD11:%.*]] = fadd double [[CONV10]], 1.000000e+00
19088 // CHECK32-NEXT:    [[CONV12:%.*]] = fptrunc double [[ADD11]] to float
19089 // CHECK32-NEXT:    store float [[CONV12]], float* [[ARRAYIDX]], align 4
19090 // CHECK32-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds float, float* [[VLA]], i32 3
19091 // CHECK32-NEXT:    [[TMP12:%.*]] = load float, float* [[ARRAYIDX13]], align 4
19092 // CHECK32-NEXT:    [[CONV14:%.*]] = fpext float [[TMP12]] to double
19093 // CHECK32-NEXT:    [[ADD15:%.*]] = fadd double [[CONV14]], 1.000000e+00
19094 // CHECK32-NEXT:    [[CONV16:%.*]] = fptrunc double [[ADD15]] to float
19095 // CHECK32-NEXT:    store float [[CONV16]], float* [[ARRAYIDX13]], align 4
19096 // CHECK32-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds [5 x [10 x double]], [5 x [10 x double]]* [[C]], i32 0, i32 1
19097 // CHECK32-NEXT:    [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x double], [10 x double]* [[ARRAYIDX17]], i32 0, i32 2
19098 // CHECK32-NEXT:    [[TMP13:%.*]] = load double, double* [[ARRAYIDX18]], align 8
19099 // CHECK32-NEXT:    [[ADD19:%.*]] = fadd double [[TMP13]], 1.000000e+00
19100 // CHECK32-NEXT:    store double [[ADD19]], double* [[ARRAYIDX18]], align 8
19101 // CHECK32-NEXT:    [[TMP14:%.*]] = mul nsw i32 1, [[TMP2]]
19102 // CHECK32-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds double, double* [[VLA1]], i32 [[TMP14]]
19103 // CHECK32-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds double, double* [[ARRAYIDX20]], i32 3
19104 // CHECK32-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX21]], align 8
19105 // CHECK32-NEXT:    [[ADD22:%.*]] = fadd double [[TMP15]], 1.000000e+00
19106 // CHECK32-NEXT:    store double [[ADD22]], double* [[ARRAYIDX21]], align 8
19107 // CHECK32-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 0
19108 // CHECK32-NEXT:    [[TMP16:%.*]] = load i64, i64* [[X]], align 4
19109 // CHECK32-NEXT:    [[ADD23:%.*]] = add nsw i64 [[TMP16]], 1
19110 // CHECK32-NEXT:    store i64 [[ADD23]], i64* [[X]], align 4
19111 // CHECK32-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_TT]], %struct.TT* [[D]], i32 0, i32 1
19112 // CHECK32-NEXT:    [[TMP17:%.*]] = load i8, i8* [[Y]], align 4
19113 // CHECK32-NEXT:    [[CONV24:%.*]] = sext i8 [[TMP17]] to i32
19114 // CHECK32-NEXT:    [[ADD25:%.*]] = add nsw i32 [[CONV24]], 1
19115 // CHECK32-NEXT:    [[CONV26:%.*]] = trunc i32 [[ADD25]] to i8
19116 // CHECK32-NEXT:    store i8 [[CONV26]], i8* [[Y]], align 4
19117 // CHECK32-NEXT:    store i32 0, i32* [[NN]], align 4
19118 // CHECK32-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4
19119 // CHECK32-NEXT:    [[TMP19:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
19120 // CHECK32-NEXT:    call void @llvm.stackrestore(i8* [[TMP19]])
19121 // CHECK32-NEXT:    ret i32 [[TMP18]]
19122 //
19123 //
19124 // CHECK32-LABEL: define {{[^@]+}}@_Z6bazzzziPi
19125 // CHECK32-SAME: (i32 [[N:%.*]], i32* [[F:%.*]]) #[[ATTR0]] {
19126 // CHECK32-NEXT:  entry:
19127 // CHECK32-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19128 // CHECK32-NEXT:    [[F_ADDR:%.*]] = alloca i32*, align 4
19129 // CHECK32-NEXT:    [[F1:%.*]] = alloca i32*, align 4
19130 // CHECK32-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19131 // CHECK32-NEXT:    store i32* [[F]], i32** [[F_ADDR]], align 4
19132 // CHECK32-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
19133 // CHECK32-NEXT:    ret void
19134 //
19135 //
19136 // CHECK32-LABEL: define {{[^@]+}}@_Z3bari
19137 // CHECK32-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
19138 // CHECK32-NEXT:  entry:
19139 // CHECK32-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19140 // CHECK32-NEXT:    [[A:%.*]] = alloca i32, align 4
19141 // CHECK32-NEXT:    [[S:%.*]] = alloca [[STRUCT_S1:%.*]], align 4
19142 // CHECK32-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19143 // CHECK32-NEXT:    store i32 0, i32* [[A]], align 4
19144 // CHECK32-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
19145 // CHECK32-NEXT:    [[CALL:%.*]] = call i32 @_Z3fooi(i32 [[TMP0]])
19146 // CHECK32-NEXT:    [[TMP1:%.*]] = load i32, i32* [[A]], align 4
19147 // CHECK32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], [[CALL]]
19148 // CHECK32-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
19149 // CHECK32-NEXT:    [[TMP2:%.*]] = load i32, i32* [[N_ADDR]], align 4
19150 // CHECK32-NEXT:    [[CALL1:%.*]] = call i32 @_ZN2S12r1Ei(%struct.S1* nonnull align 4 dereferenceable(8) [[S]], i32 [[TMP2]])
19151 // CHECK32-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
19152 // CHECK32-NEXT:    [[ADD2:%.*]] = add nsw i32 [[TMP3]], [[CALL1]]
19153 // CHECK32-NEXT:    store i32 [[ADD2]], i32* [[A]], align 4
19154 // CHECK32-NEXT:    [[TMP4:%.*]] = load i32, i32* [[N_ADDR]], align 4
19155 // CHECK32-NEXT:    [[CALL3:%.*]] = call i32 @_ZL7fstatici(i32 [[TMP4]])
19156 // CHECK32-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4
19157 // CHECK32-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP5]], [[CALL3]]
19158 // CHECK32-NEXT:    store i32 [[ADD4]], i32* [[A]], align 4
19159 // CHECK32-NEXT:    [[TMP6:%.*]] = load i32, i32* [[N_ADDR]], align 4
19160 // CHECK32-NEXT:    [[CALL5:%.*]] = call i32 @_Z9ftemplateIiET_i(i32 [[TMP6]])
19161 // CHECK32-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
19162 // CHECK32-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP7]], [[CALL5]]
19163 // CHECK32-NEXT:    store i32 [[ADD6]], i32* [[A]], align 4
19164 // CHECK32-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4
19165 // CHECK32-NEXT:    ret i32 [[TMP8]]
19166 //
19167 //
19168 // CHECK32-LABEL: define {{[^@]+}}@_ZN2S12r1Ei
19169 // CHECK32-SAME: (%struct.S1* nonnull align 4 dereferenceable(8) [[THIS:%.*]], i32 [[N:%.*]]) #[[ATTR0]] comdat align 2 {
19170 // CHECK32-NEXT:  entry:
19171 // CHECK32-NEXT:    [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 4
19172 // CHECK32-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19173 // CHECK32-NEXT:    [[B:%.*]] = alloca i32, align 4
19174 // CHECK32-NEXT:    [[SAVED_STACK:%.*]] = alloca i8*, align 4
19175 // CHECK32-NEXT:    [[__VLA_EXPR0:%.*]] = alloca i32, align 4
19176 // CHECK32-NEXT:    store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 4
19177 // CHECK32-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19178 // CHECK32-NEXT:    [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 4
19179 // CHECK32-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N_ADDR]], align 4
19180 // CHECK32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
19181 // CHECK32-NEXT:    store i32 [[ADD]], i32* [[B]], align 4
19182 // CHECK32-NEXT:    [[TMP1:%.*]] = load i32, i32* [[N_ADDR]], align 4
19183 // CHECK32-NEXT:    [[TMP2:%.*]] = call i8* @llvm.stacksave()
19184 // CHECK32-NEXT:    store i8* [[TMP2]], i8** [[SAVED_STACK]], align 4
19185 // CHECK32-NEXT:    [[TMP3:%.*]] = mul nuw i32 2, [[TMP1]]
19186 // CHECK32-NEXT:    [[VLA:%.*]] = alloca i16, i32 [[TMP3]], align 2
19187 // CHECK32-NEXT:    store i32 [[TMP1]], i32* [[__VLA_EXPR0]], align 4
19188 // CHECK32-NEXT:    [[TMP4:%.*]] = load i32, i32* [[B]], align 4
19189 // CHECK32-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP4]] to double
19190 // CHECK32-NEXT:    [[ADD2:%.*]] = fadd double [[CONV]], 1.500000e+00
19191 // CHECK32-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[THIS1]], i32 0, i32 0
19192 // CHECK32-NEXT:    store double [[ADD2]], double* [[A]], align 4
19193 // CHECK32-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
19194 // CHECK32-NEXT:    [[TMP5:%.*]] = load double, double* [[A3]], align 4
19195 // CHECK32-NEXT:    [[INC:%.*]] = fadd double [[TMP5]], 1.000000e+00
19196 // CHECK32-NEXT:    store double [[INC]], double* [[A3]], align 4
19197 // CHECK32-NEXT:    [[CONV4:%.*]] = fptosi double [[INC]] to i16
19198 // CHECK32-NEXT:    [[TMP6:%.*]] = mul nsw i32 1, [[TMP1]]
19199 // CHECK32-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP6]]
19200 // CHECK32-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
19201 // CHECK32-NEXT:    store i16 [[CONV4]], i16* [[ARRAYIDX5]], align 2
19202 // CHECK32-NEXT:    [[TMP7:%.*]] = mul nsw i32 1, [[TMP1]]
19203 // CHECK32-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP7]]
19204 // CHECK32-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX6]], i32 1
19205 // CHECK32-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX7]], align 2
19206 // CHECK32-NEXT:    [[CONV8:%.*]] = sext i16 [[TMP8]] to i32
19207 // CHECK32-NEXT:    [[TMP9:%.*]] = load i32, i32* [[B]], align 4
19208 // CHECK32-NEXT:    [[ADD9:%.*]] = add nsw i32 [[CONV8]], [[TMP9]]
19209 // CHECK32-NEXT:    [[TMP10:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
19210 // CHECK32-NEXT:    call void @llvm.stackrestore(i8* [[TMP10]])
19211 // CHECK32-NEXT:    ret i32 [[ADD9]]
19212 //
19213 //
19214 // CHECK32-LABEL: define {{[^@]+}}@_ZL7fstatici
19215 // CHECK32-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
19216 // CHECK32-NEXT:  entry:
19217 // CHECK32-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19218 // CHECK32-NEXT:    [[A:%.*]] = alloca i32, align 4
19219 // CHECK32-NEXT:    [[AA:%.*]] = alloca i16, align 2
19220 // CHECK32-NEXT:    [[AAA:%.*]] = alloca i8, align 1
19221 // CHECK32-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
19222 // CHECK32-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19223 // CHECK32-NEXT:    store i32 0, i32* [[A]], align 4
19224 // CHECK32-NEXT:    store i16 0, i16* [[AA]], align 2
19225 // CHECK32-NEXT:    store i8 0, i8* [[AAA]], align 1
19226 // CHECK32-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
19227 // CHECK32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
19228 // CHECK32-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
19229 // CHECK32-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
19230 // CHECK32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
19231 // CHECK32-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
19232 // CHECK32-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
19233 // CHECK32-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
19234 // CHECK32-NEXT:    [[TMP2:%.*]] = load i8, i8* [[AAA]], align 1
19235 // CHECK32-NEXT:    [[CONV3:%.*]] = sext i8 [[TMP2]] to i32
19236 // CHECK32-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], 1
19237 // CHECK32-NEXT:    [[CONV5:%.*]] = trunc i32 [[ADD4]] to i8
19238 // CHECK32-NEXT:    store i8 [[CONV5]], i8* [[AAA]], align 1
19239 // CHECK32-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
19240 // CHECK32-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
19241 // CHECK32-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP3]], 1
19242 // CHECK32-NEXT:    store i32 [[ADD6]], i32* [[ARRAYIDX]], align 4
19243 // CHECK32-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4
19244 // CHECK32-NEXT:    ret i32 [[TMP4]]
19245 //
19246 //
19247 // CHECK32-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
19248 // CHECK32-SAME: (i32 [[N:%.*]]) #[[ATTR0]] comdat {
19249 // CHECK32-NEXT:  entry:
19250 // CHECK32-NEXT:    [[N_ADDR:%.*]] = alloca i32, align 4
19251 // CHECK32-NEXT:    [[A:%.*]] = alloca i32, align 4
19252 // CHECK32-NEXT:    [[AA:%.*]] = alloca i16, align 2
19253 // CHECK32-NEXT:    [[B:%.*]] = alloca [10 x i32], align 4
19254 // CHECK32-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
19255 // CHECK32-NEXT:    store i32 0, i32* [[A]], align 4
19256 // CHECK32-NEXT:    store i16 0, i16* [[AA]], align 2
19257 // CHECK32-NEXT:    [[TMP0:%.*]] = load i32, i32* [[A]], align 4
19258 // CHECK32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], 1
19259 // CHECK32-NEXT:    store i32 [[ADD]], i32* [[A]], align 4
19260 // CHECK32-NEXT:    [[TMP1:%.*]] = load i16, i16* [[AA]], align 2
19261 // CHECK32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
19262 // CHECK32-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1
19263 // CHECK32-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16
19264 // CHECK32-NEXT:    store i16 [[CONV2]], i16* [[AA]], align 2
19265 // CHECK32-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[B]], i32 0, i32 2
19266 // CHECK32-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
19267 // CHECK32-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP2]], 1
19268 // CHECK32-NEXT:    store i32 [[ADD3]], i32* [[ARRAYIDX]], align 4
19269 // CHECK32-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
19270 // CHECK32-NEXT:    ret i32 [[TMP3]]
19271 //
19272