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 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK1
3 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
4 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK2
5 
6 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK3
7 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
8 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK4
9 
10 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK5
11 // RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
12 // RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK6
13 
14 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK7
15 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
16 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK8
17 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK9
18 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple aarch64-unknown-unknown -emit-pch -o %t %s
19 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK10
20 
21 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK11
22 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple aarch64-unknown-unknown -emit-pch -o %t %s
23 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK12
24 
25 // RUN: %clang_cc1 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK13
26 // RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple aarch64-unknown-unknown -emit-pch -o %t %s
27 // RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK14
28 
29 // RUN: %clang_cc1 -verify -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK15
30 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple aarch64-unknown-unknown -emit-pch -o %t %s
31 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple aarch64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK16
32 
33 // expected-no-diagnostics
34 #ifndef HEADER
35 #define HEADER
36 
37 void fn1();
38 void fn2();
39 void fn3();
40 void fn4();
41 void fn5();
42 void fn6();
43 
44 int Arg;
45 
46 void gtid_test() {
47 #pragma omp target
48 #pragma omp teams
49 #pragma omp distribute parallel for
50   for(int i = 0 ; i < 100; i++) {}
51 
52 #pragma omp target
53 #pragma omp teams
54 #pragma omp distribute parallel for if (parallel: false)
55   for(int i = 0 ; i < 100; i++) {
56     gtid_test();
57   }
58 }
59 
60 
61 template <typename T>
62 int tmain(T Arg) {
63 #pragma omp target
64 #pragma omp teams
65 #pragma omp distribute parallel for if (true)
66   for(int i = 0 ; i < 100; i++) {
67     fn1();
68   }
69 #pragma omp target
70 #pragma omp teams
71 #pragma omp distribute parallel for if (false)
72   for(int i = 0 ; i < 100; i++) {
73     fn2();
74   }
75 #pragma omp target
76 #pragma omp teams
77 #pragma omp distribute parallel for if (parallel: Arg)
78   for(int i = 0 ; i < 100; i++) {
79     fn3();
80   }
81   return 0;
82 }
83 
84 int main() {
85 #pragma omp target
86 #pragma omp teams
87 #pragma omp distribute parallel for if (true)
88   for(int i = 0 ; i < 100; i++) {
89 
90 
91     fn4();
92   }
93 
94 #pragma omp target
95 #pragma omp teams
96 #pragma omp distribute parallel for if (false)
97   for(int i = 0 ; i < 100; i++) {
98 
99 
100     fn5();
101   }
102 
103 #pragma omp target
104 #pragma omp teams
105 #pragma omp distribute parallel for if (Arg)
106   for(int i = 0 ; i < 100; i++) {
107 
108 
109     fn6();
110   }
111 
112   return tmain(Arg);
113 }
114 
115 
116 
117 
118 
119 
120 // call void [[T_OUTLINE_FUN_3:@.+]](
121 
122 #endif
123 // CHECK1-LABEL: define {{[^@]+}}@_Z9gtid_testv
124 // CHECK1-SAME: () #[[ATTR0:[0-9]+]] {
125 // CHECK1-NEXT:  entry:
126 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
127 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
128 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
129 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
130 // CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
131 // CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
132 // CHECK1:       omp_offload.failed:
133 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
134 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
135 // CHECK1:       omp_offload.cont:
136 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
137 // CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
138 // CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
139 // CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
140 // CHECK1:       omp_offload.failed2:
141 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
142 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
143 // CHECK1:       omp_offload.cont3:
144 // CHECK1-NEXT:    ret void
145 //
146 //
147 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
148 // CHECK1-SAME: () #[[ATTR1:[0-9]+]] {
149 // CHECK1-NEXT:  entry:
150 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
151 // CHECK1-NEXT:    ret void
152 //
153 //
154 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined.
155 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
156 // CHECK1-NEXT:  entry:
157 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
158 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
159 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
160 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
161 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
162 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
163 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
164 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
165 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
166 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
167 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
168 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
169 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
170 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
171 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
172 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
173 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
174 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
175 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
176 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
177 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
178 // CHECK1:       cond.true:
179 // CHECK1-NEXT:    br label [[COND_END:%.*]]
180 // CHECK1:       cond.false:
181 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
182 // CHECK1-NEXT:    br label [[COND_END]]
183 // CHECK1:       cond.end:
184 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
185 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
186 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
187 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
188 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
189 // CHECK1:       omp.inner.for.cond:
190 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
191 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
192 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
193 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
194 // CHECK1:       omp.inner.for.body:
195 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
196 // CHECK1-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
197 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
198 // CHECK1-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
199 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
200 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
201 // CHECK1:       omp.inner.for.inc:
202 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
203 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
204 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
205 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
206 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
207 // CHECK1:       omp.inner.for.end:
208 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
209 // CHECK1:       omp.loop.exit:
210 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
211 // CHECK1-NEXT:    ret void
212 //
213 //
214 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..1
215 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
216 // CHECK1-NEXT:  entry:
217 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
218 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
219 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
220 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
221 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
222 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
223 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
224 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
225 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
226 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
227 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
228 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
229 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
230 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
231 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
232 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
233 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
234 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
235 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
236 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
237 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
238 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
239 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
240 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
241 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
242 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
243 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
244 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
245 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
246 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
247 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
248 // CHECK1:       cond.true:
249 // CHECK1-NEXT:    br label [[COND_END:%.*]]
250 // CHECK1:       cond.false:
251 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
252 // CHECK1-NEXT:    br label [[COND_END]]
253 // CHECK1:       cond.end:
254 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
255 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
256 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
257 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
258 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
259 // CHECK1:       omp.inner.for.cond:
260 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
261 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
262 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
263 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
264 // CHECK1:       omp.inner.for.body:
265 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
266 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
267 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
268 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
269 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
270 // CHECK1:       omp.body.continue:
271 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
272 // CHECK1:       omp.inner.for.inc:
273 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
274 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
275 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
276 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
277 // CHECK1:       omp.inner.for.end:
278 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
279 // CHECK1:       omp.loop.exit:
280 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
281 // CHECK1-NEXT:    ret void
282 //
283 //
284 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
285 // CHECK1-SAME: () #[[ATTR1]] {
286 // CHECK1-NEXT:  entry:
287 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
288 // CHECK1-NEXT:    ret void
289 //
290 //
291 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..2
292 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
293 // CHECK1-NEXT:  entry:
294 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
295 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
296 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
297 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
298 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
299 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
300 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
301 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
302 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
303 // CHECK1-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
304 // CHECK1-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
305 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
306 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
307 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
308 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
309 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
310 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
311 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
312 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
313 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
314 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
315 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
316 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
317 // CHECK1:       cond.true:
318 // CHECK1-NEXT:    br label [[COND_END:%.*]]
319 // CHECK1:       cond.false:
320 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
321 // CHECK1-NEXT:    br label [[COND_END]]
322 // CHECK1:       cond.end:
323 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
324 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
325 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
326 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
327 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
328 // CHECK1:       omp.inner.for.cond:
329 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
330 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
331 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
332 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
333 // CHECK1:       omp.inner.for.body:
334 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
335 // CHECK1-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
336 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
337 // CHECK1-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
338 // CHECK1-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
339 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
340 // CHECK1-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
341 // CHECK1-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
342 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
343 // CHECK1:       omp.inner.for.inc:
344 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
345 // CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
346 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
347 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
348 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
349 // CHECK1:       omp.inner.for.end:
350 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
351 // CHECK1:       omp.loop.exit:
352 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
353 // CHECK1-NEXT:    ret void
354 //
355 //
356 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..3
357 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
358 // CHECK1-NEXT:  entry:
359 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
360 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
361 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
362 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
363 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
364 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
365 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
366 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
367 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
368 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
369 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
370 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
371 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
372 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
373 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
374 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
375 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
376 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
377 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
378 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
379 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
380 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
381 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
382 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
383 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
384 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
385 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
386 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
387 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
388 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
389 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
390 // CHECK1:       cond.true:
391 // CHECK1-NEXT:    br label [[COND_END:%.*]]
392 // CHECK1:       cond.false:
393 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
394 // CHECK1-NEXT:    br label [[COND_END]]
395 // CHECK1:       cond.end:
396 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
397 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
398 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
399 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
400 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
401 // CHECK1:       omp.inner.for.cond:
402 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
403 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
404 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
405 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
406 // CHECK1:       omp.inner.for.body:
407 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
408 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
409 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
410 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
411 // CHECK1-NEXT:    call void @_Z9gtid_testv()
412 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
413 // CHECK1:       omp.body.continue:
414 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
415 // CHECK1:       omp.inner.for.inc:
416 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
417 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
418 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
419 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
420 // CHECK1:       omp.inner.for.end:
421 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
422 // CHECK1:       omp.loop.exit:
423 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
424 // CHECK1-NEXT:    ret void
425 //
426 //
427 // CHECK1-LABEL: define {{[^@]+}}@main
428 // CHECK1-SAME: () #[[ATTR3:[0-9]+]] {
429 // CHECK1-NEXT:  entry:
430 // CHECK1-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
431 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
432 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
433 // CHECK1-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
434 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
435 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
436 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
437 // CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
438 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
439 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
440 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
441 // CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
442 // CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
443 // CHECK1:       omp_offload.failed:
444 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
445 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
446 // CHECK1:       omp_offload.cont:
447 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
448 // CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
449 // CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
450 // CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
451 // CHECK1:       omp_offload.failed2:
452 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
453 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
454 // CHECK1:       omp_offload.cont3:
455 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
456 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
457 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
458 // CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
459 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
460 // CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
461 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
462 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
463 // CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
464 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
465 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
466 // CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
467 // CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
468 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
469 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
470 // CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
471 // CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
472 // CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
473 // CHECK1:       omp_offload.failed5:
474 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
475 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
476 // CHECK1:       omp_offload.cont6:
477 // CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
478 // CHECK1-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
479 // CHECK1-NEXT:    ret i32 [[CALL]]
480 //
481 //
482 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
483 // CHECK1-SAME: () #[[ATTR1]] {
484 // CHECK1-NEXT:  entry:
485 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
486 // CHECK1-NEXT:    ret void
487 //
488 //
489 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..4
490 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
491 // CHECK1-NEXT:  entry:
492 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
493 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
494 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
495 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
496 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
497 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
498 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
499 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
500 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
501 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
502 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
503 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
504 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
505 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
506 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
507 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
508 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
509 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
510 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
511 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
512 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
513 // CHECK1:       cond.true:
514 // CHECK1-NEXT:    br label [[COND_END:%.*]]
515 // CHECK1:       cond.false:
516 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
517 // CHECK1-NEXT:    br label [[COND_END]]
518 // CHECK1:       cond.end:
519 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
520 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
521 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
522 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
523 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
524 // CHECK1:       omp.inner.for.cond:
525 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
526 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
527 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
528 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
529 // CHECK1:       omp.inner.for.body:
530 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
531 // CHECK1-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
532 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
533 // CHECK1-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
534 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
535 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
536 // CHECK1:       omp.inner.for.inc:
537 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
538 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
539 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
540 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
541 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
542 // CHECK1:       omp.inner.for.end:
543 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
544 // CHECK1:       omp.loop.exit:
545 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
546 // CHECK1-NEXT:    ret void
547 //
548 //
549 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..5
550 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
551 // CHECK1-NEXT:  entry:
552 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
553 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
554 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
555 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
556 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
557 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
558 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
559 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
560 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
561 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
562 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
563 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
564 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
565 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
566 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
567 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
568 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
569 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
570 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
571 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
572 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
573 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
574 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
575 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
576 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
577 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
578 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
579 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
580 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
581 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
582 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
583 // CHECK1:       cond.true:
584 // CHECK1-NEXT:    br label [[COND_END:%.*]]
585 // CHECK1:       cond.false:
586 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
587 // CHECK1-NEXT:    br label [[COND_END]]
588 // CHECK1:       cond.end:
589 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
590 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
591 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
592 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
593 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
594 // CHECK1:       omp.inner.for.cond:
595 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
596 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
597 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
598 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
599 // CHECK1:       omp.inner.for.body:
600 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
601 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
602 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
603 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
604 // CHECK1-NEXT:    call void @_Z3fn4v()
605 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
606 // CHECK1:       omp.body.continue:
607 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
608 // CHECK1:       omp.inner.for.inc:
609 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
610 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
611 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
612 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
613 // CHECK1:       omp.inner.for.end:
614 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
615 // CHECK1:       omp.loop.exit:
616 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
617 // CHECK1-NEXT:    ret void
618 //
619 //
620 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
621 // CHECK1-SAME: () #[[ATTR1]] {
622 // CHECK1-NEXT:  entry:
623 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
624 // CHECK1-NEXT:    ret void
625 //
626 //
627 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..6
628 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
629 // CHECK1-NEXT:  entry:
630 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
631 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
632 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
633 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
634 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
635 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
636 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
637 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
638 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
639 // CHECK1-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
640 // CHECK1-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
641 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
642 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
643 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
644 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
645 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
646 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
647 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
648 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
649 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
650 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
651 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
652 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
653 // CHECK1:       cond.true:
654 // CHECK1-NEXT:    br label [[COND_END:%.*]]
655 // CHECK1:       cond.false:
656 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
657 // CHECK1-NEXT:    br label [[COND_END]]
658 // CHECK1:       cond.end:
659 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
660 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
661 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
662 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
663 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
664 // CHECK1:       omp.inner.for.cond:
665 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
666 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
667 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
668 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
669 // CHECK1:       omp.inner.for.body:
670 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
671 // CHECK1-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
672 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
673 // CHECK1-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
674 // CHECK1-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
675 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
676 // CHECK1-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
677 // CHECK1-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
678 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
679 // CHECK1:       omp.inner.for.inc:
680 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
681 // CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
682 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
683 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
684 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
685 // CHECK1:       omp.inner.for.end:
686 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
687 // CHECK1:       omp.loop.exit:
688 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
689 // CHECK1-NEXT:    ret void
690 //
691 //
692 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..7
693 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
694 // CHECK1-NEXT:  entry:
695 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
696 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
697 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
698 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
699 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
700 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
701 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
702 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
703 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
704 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
705 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
706 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
707 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
708 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
709 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
710 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
711 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
712 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
713 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
714 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
715 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
716 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
717 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
718 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
719 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
720 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
721 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
722 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
723 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
724 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
725 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
726 // CHECK1:       cond.true:
727 // CHECK1-NEXT:    br label [[COND_END:%.*]]
728 // CHECK1:       cond.false:
729 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
730 // CHECK1-NEXT:    br label [[COND_END]]
731 // CHECK1:       cond.end:
732 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
733 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
734 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
735 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
736 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
737 // CHECK1:       omp.inner.for.cond:
738 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
739 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
740 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
741 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
742 // CHECK1:       omp.inner.for.body:
743 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
744 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
745 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
746 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
747 // CHECK1-NEXT:    call void @_Z3fn5v()
748 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
749 // CHECK1:       omp.body.continue:
750 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
751 // CHECK1:       omp.inner.for.inc:
752 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
753 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
754 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
755 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
756 // CHECK1:       omp.inner.for.end:
757 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
758 // CHECK1:       omp.loop.exit:
759 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
760 // CHECK1-NEXT:    ret void
761 //
762 //
763 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
764 // CHECK1-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
765 // CHECK1-NEXT:  entry:
766 // CHECK1-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
767 // CHECK1-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
768 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
769 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
770 // CHECK1-NEXT:    ret void
771 //
772 //
773 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..8
774 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
775 // CHECK1-NEXT:  entry:
776 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
777 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
778 // CHECK1-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
779 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
780 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
781 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
782 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
783 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
784 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
785 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
786 // CHECK1-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
787 // CHECK1-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
788 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
789 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
790 // CHECK1-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
791 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
792 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
793 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
794 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
795 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
796 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
797 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
798 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
799 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
800 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
801 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
802 // CHECK1:       cond.true:
803 // CHECK1-NEXT:    br label [[COND_END:%.*]]
804 // CHECK1:       cond.false:
805 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
806 // CHECK1-NEXT:    br label [[COND_END]]
807 // CHECK1:       cond.end:
808 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
809 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
810 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
811 // CHECK1-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
812 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
813 // CHECK1:       omp.inner.for.cond:
814 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
815 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
816 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
817 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
818 // CHECK1:       omp.inner.for.body:
819 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
820 // CHECK1-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
821 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
822 // CHECK1-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
823 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
824 // CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
825 // CHECK1-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
826 // CHECK1:       omp_if.then:
827 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
828 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
829 // CHECK1:       omp_if.else:
830 // CHECK1-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
831 // CHECK1-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
832 // CHECK1-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
833 // CHECK1-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
834 // CHECK1-NEXT:    br label [[OMP_IF_END]]
835 // CHECK1:       omp_if.end:
836 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
837 // CHECK1:       omp.inner.for.inc:
838 // CHECK1-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
839 // CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
840 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
841 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
842 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
843 // CHECK1:       omp.inner.for.end:
844 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
845 // CHECK1:       omp.loop.exit:
846 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
847 // CHECK1-NEXT:    ret void
848 //
849 //
850 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..9
851 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
852 // CHECK1-NEXT:  entry:
853 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
854 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
855 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
856 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
857 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
858 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
859 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
860 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
861 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
862 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
863 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
864 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
865 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
866 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
867 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
868 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
869 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
870 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
871 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
872 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
873 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
874 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
875 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
876 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
877 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
878 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
879 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
880 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
881 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
882 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
883 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
884 // CHECK1:       cond.true:
885 // CHECK1-NEXT:    br label [[COND_END:%.*]]
886 // CHECK1:       cond.false:
887 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
888 // CHECK1-NEXT:    br label [[COND_END]]
889 // CHECK1:       cond.end:
890 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
891 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
892 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
893 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
894 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
895 // CHECK1:       omp.inner.for.cond:
896 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
897 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
898 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
899 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
900 // CHECK1:       omp.inner.for.body:
901 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
902 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
903 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
904 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
905 // CHECK1-NEXT:    call void @_Z3fn6v()
906 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
907 // CHECK1:       omp.body.continue:
908 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
909 // CHECK1:       omp.inner.for.inc:
910 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
911 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
912 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
913 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
914 // CHECK1:       omp.inner.for.end:
915 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
916 // CHECK1:       omp.loop.exit:
917 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
918 // CHECK1-NEXT:    ret void
919 //
920 //
921 // CHECK1-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
922 // CHECK1-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
923 // CHECK1-NEXT:  entry:
924 // CHECK1-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
925 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
926 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
927 // CHECK1-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
928 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
929 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
930 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
931 // CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
932 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
933 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
934 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
935 // CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
936 // CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
937 // CHECK1:       omp_offload.failed:
938 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
939 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
940 // CHECK1:       omp_offload.cont:
941 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
942 // CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
943 // CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
944 // CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
945 // CHECK1:       omp_offload.failed2:
946 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
947 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
948 // CHECK1:       omp_offload.cont3:
949 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
950 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
951 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
952 // CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
953 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
954 // CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
955 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
956 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
957 // CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
958 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
959 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
960 // CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
961 // CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
962 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
963 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
964 // CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
965 // CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
966 // CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
967 // CHECK1:       omp_offload.failed5:
968 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
969 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
970 // CHECK1:       omp_offload.cont6:
971 // CHECK1-NEXT:    ret i32 0
972 //
973 //
974 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
975 // CHECK1-SAME: () #[[ATTR1]] {
976 // CHECK1-NEXT:  entry:
977 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
978 // CHECK1-NEXT:    ret void
979 //
980 //
981 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..10
982 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
983 // CHECK1-NEXT:  entry:
984 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
985 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
986 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
987 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
988 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
989 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
990 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
991 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
992 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
993 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
994 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
995 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
996 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
997 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
998 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
999 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1000 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
1001 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1002 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1003 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
1004 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1005 // CHECK1:       cond.true:
1006 // CHECK1-NEXT:    br label [[COND_END:%.*]]
1007 // CHECK1:       cond.false:
1008 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1009 // CHECK1-NEXT:    br label [[COND_END]]
1010 // CHECK1:       cond.end:
1011 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
1012 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1013 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1014 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
1015 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1016 // CHECK1:       omp.inner.for.cond:
1017 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1018 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1019 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
1020 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1021 // CHECK1:       omp.inner.for.body:
1022 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1023 // CHECK1-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
1024 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1025 // CHECK1-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
1026 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
1027 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1028 // CHECK1:       omp.inner.for.inc:
1029 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1030 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1031 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
1032 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1033 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
1034 // CHECK1:       omp.inner.for.end:
1035 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1036 // CHECK1:       omp.loop.exit:
1037 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
1038 // CHECK1-NEXT:    ret void
1039 //
1040 //
1041 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..11
1042 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1043 // CHECK1-NEXT:  entry:
1044 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1045 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1046 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1047 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1048 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1049 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1050 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1051 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
1052 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1053 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1054 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
1055 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1056 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1057 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1058 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1059 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
1060 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
1061 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1062 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
1063 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1064 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
1065 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
1066 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
1067 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1068 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1069 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1070 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
1071 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1072 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1073 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
1074 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1075 // CHECK1:       cond.true:
1076 // CHECK1-NEXT:    br label [[COND_END:%.*]]
1077 // CHECK1:       cond.false:
1078 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1079 // CHECK1-NEXT:    br label [[COND_END]]
1080 // CHECK1:       cond.end:
1081 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
1082 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
1083 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
1084 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
1085 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1086 // CHECK1:       omp.inner.for.cond:
1087 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1088 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1089 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
1090 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1091 // CHECK1:       omp.inner.for.body:
1092 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1093 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
1094 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
1095 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
1096 // CHECK1-NEXT:    call void @_Z3fn1v()
1097 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
1098 // CHECK1:       omp.body.continue:
1099 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1100 // CHECK1:       omp.inner.for.inc:
1101 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1102 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
1103 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
1104 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
1105 // CHECK1:       omp.inner.for.end:
1106 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1107 // CHECK1:       omp.loop.exit:
1108 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
1109 // CHECK1-NEXT:    ret void
1110 //
1111 //
1112 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
1113 // CHECK1-SAME: () #[[ATTR1]] {
1114 // CHECK1-NEXT:  entry:
1115 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
1116 // CHECK1-NEXT:    ret void
1117 //
1118 //
1119 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..12
1120 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
1121 // CHECK1-NEXT:  entry:
1122 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1123 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1124 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1125 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1126 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
1127 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
1128 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1129 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1130 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
1131 // CHECK1-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
1132 // CHECK1-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
1133 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1134 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1135 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
1136 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
1137 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1138 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1139 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1140 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
1141 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1142 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1143 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
1144 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1145 // CHECK1:       cond.true:
1146 // CHECK1-NEXT:    br label [[COND_END:%.*]]
1147 // CHECK1:       cond.false:
1148 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1149 // CHECK1-NEXT:    br label [[COND_END]]
1150 // CHECK1:       cond.end:
1151 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
1152 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1153 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1154 // CHECK1-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
1155 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1156 // CHECK1:       omp.inner.for.cond:
1157 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1158 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1159 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
1160 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1161 // CHECK1:       omp.inner.for.body:
1162 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1163 // CHECK1-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
1164 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1165 // CHECK1-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
1166 // CHECK1-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
1167 // CHECK1-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1168 // CHECK1-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
1169 // CHECK1-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
1170 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1171 // CHECK1:       omp.inner.for.inc:
1172 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1173 // CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1174 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
1175 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1176 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
1177 // CHECK1:       omp.inner.for.end:
1178 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1179 // CHECK1:       omp.loop.exit:
1180 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
1181 // CHECK1-NEXT:    ret void
1182 //
1183 //
1184 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..13
1185 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1186 // CHECK1-NEXT:  entry:
1187 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1188 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1189 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1190 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1191 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1192 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1193 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1194 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
1195 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1196 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1197 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
1198 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1199 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1200 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1201 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1202 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
1203 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
1204 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1205 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
1206 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1207 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
1208 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
1209 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
1210 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1211 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1212 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1213 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
1214 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1215 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1216 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
1217 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1218 // CHECK1:       cond.true:
1219 // CHECK1-NEXT:    br label [[COND_END:%.*]]
1220 // CHECK1:       cond.false:
1221 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1222 // CHECK1-NEXT:    br label [[COND_END]]
1223 // CHECK1:       cond.end:
1224 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
1225 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
1226 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
1227 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
1228 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1229 // CHECK1:       omp.inner.for.cond:
1230 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1231 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1232 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
1233 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1234 // CHECK1:       omp.inner.for.body:
1235 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1236 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
1237 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
1238 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
1239 // CHECK1-NEXT:    call void @_Z3fn2v()
1240 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
1241 // CHECK1:       omp.body.continue:
1242 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1243 // CHECK1:       omp.inner.for.inc:
1244 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1245 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
1246 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
1247 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
1248 // CHECK1:       omp.inner.for.end:
1249 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1250 // CHECK1:       omp.loop.exit:
1251 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
1252 // CHECK1-NEXT:    ret void
1253 //
1254 //
1255 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
1256 // CHECK1-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
1257 // CHECK1-NEXT:  entry:
1258 // CHECK1-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
1259 // CHECK1-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
1260 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
1261 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
1262 // CHECK1-NEXT:    ret void
1263 //
1264 //
1265 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..14
1266 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
1267 // CHECK1-NEXT:  entry:
1268 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1269 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1270 // CHECK1-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
1271 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1272 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1273 // CHECK1-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
1274 // CHECK1-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
1275 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1276 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1277 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
1278 // CHECK1-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
1279 // CHECK1-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
1280 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1281 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1282 // CHECK1-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
1283 // CHECK1-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
1284 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
1285 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
1286 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1287 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1288 // CHECK1-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1289 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
1290 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1291 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1292 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
1293 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1294 // CHECK1:       cond.true:
1295 // CHECK1-NEXT:    br label [[COND_END:%.*]]
1296 // CHECK1:       cond.false:
1297 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1298 // CHECK1-NEXT:    br label [[COND_END]]
1299 // CHECK1:       cond.end:
1300 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
1301 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1302 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1303 // CHECK1-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
1304 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1305 // CHECK1:       omp.inner.for.cond:
1306 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1307 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1308 // CHECK1-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
1309 // CHECK1-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1310 // CHECK1:       omp.inner.for.body:
1311 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1312 // CHECK1-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
1313 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1314 // CHECK1-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
1315 // CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
1316 // CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
1317 // CHECK1-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
1318 // CHECK1:       omp_if.then:
1319 // CHECK1-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
1320 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
1321 // CHECK1:       omp_if.else:
1322 // CHECK1-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
1323 // CHECK1-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1324 // CHECK1-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
1325 // CHECK1-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
1326 // CHECK1-NEXT:    br label [[OMP_IF_END]]
1327 // CHECK1:       omp_if.end:
1328 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1329 // CHECK1:       omp.inner.for.inc:
1330 // CHECK1-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1331 // CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1332 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
1333 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1334 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
1335 // CHECK1:       omp.inner.for.end:
1336 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1337 // CHECK1:       omp.loop.exit:
1338 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
1339 // CHECK1-NEXT:    ret void
1340 //
1341 //
1342 // CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..15
1343 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1344 // CHECK1-NEXT:  entry:
1345 // CHECK1-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1346 // CHECK1-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1347 // CHECK1-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1348 // CHECK1-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1349 // CHECK1-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1350 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1351 // CHECK1-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1352 // CHECK1-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
1353 // CHECK1-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1354 // CHECK1-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1355 // CHECK1-NEXT:    [[I:%.*]] = alloca i32, align 4
1356 // CHECK1-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1357 // CHECK1-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1358 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1359 // CHECK1-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1360 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
1361 // CHECK1-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
1362 // CHECK1-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1363 // CHECK1-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
1364 // CHECK1-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1365 // CHECK1-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
1366 // CHECK1-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
1367 // CHECK1-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
1368 // CHECK1-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1369 // CHECK1-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1370 // CHECK1-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1371 // CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
1372 // CHECK1-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1373 // CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1374 // CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
1375 // CHECK1-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1376 // CHECK1:       cond.true:
1377 // CHECK1-NEXT:    br label [[COND_END:%.*]]
1378 // CHECK1:       cond.false:
1379 // CHECK1-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1380 // CHECK1-NEXT:    br label [[COND_END]]
1381 // CHECK1:       cond.end:
1382 // CHECK1-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
1383 // CHECK1-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
1384 // CHECK1-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
1385 // CHECK1-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
1386 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1387 // CHECK1:       omp.inner.for.cond:
1388 // CHECK1-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1389 // CHECK1-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1390 // CHECK1-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
1391 // CHECK1-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1392 // CHECK1:       omp.inner.for.body:
1393 // CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1394 // CHECK1-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
1395 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
1396 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
1397 // CHECK1-NEXT:    call void @_Z3fn3v()
1398 // CHECK1-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
1399 // CHECK1:       omp.body.continue:
1400 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1401 // CHECK1:       omp.inner.for.inc:
1402 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1403 // CHECK1-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
1404 // CHECK1-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
1405 // CHECK1-NEXT:    br label [[OMP_INNER_FOR_COND]]
1406 // CHECK1:       omp.inner.for.end:
1407 // CHECK1-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1408 // CHECK1:       omp.loop.exit:
1409 // CHECK1-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
1410 // CHECK1-NEXT:    ret void
1411 //
1412 //
1413 // CHECK1-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
1414 // CHECK1-SAME: () #[[ATTR5:[0-9]+]] {
1415 // CHECK1-NEXT:  entry:
1416 // CHECK1-NEXT:    call void @__tgt_register_requires(i64 1)
1417 // CHECK1-NEXT:    ret void
1418 //
1419 //
1420 // CHECK2-LABEL: define {{[^@]+}}@_Z9gtid_testv
1421 // CHECK2-SAME: () #[[ATTR0:[0-9]+]] {
1422 // CHECK2-NEXT:  entry:
1423 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1424 // CHECK2-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
1425 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
1426 // CHECK2-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
1427 // CHECK2-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
1428 // CHECK2-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1429 // CHECK2:       omp_offload.failed:
1430 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
1431 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1432 // CHECK2:       omp_offload.cont:
1433 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
1434 // CHECK2-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
1435 // CHECK2-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
1436 // CHECK2-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
1437 // CHECK2:       omp_offload.failed2:
1438 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
1439 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
1440 // CHECK2:       omp_offload.cont3:
1441 // CHECK2-NEXT:    ret void
1442 //
1443 //
1444 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
1445 // CHECK2-SAME: () #[[ATTR1:[0-9]+]] {
1446 // CHECK2-NEXT:  entry:
1447 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
1448 // CHECK2-NEXT:    ret void
1449 //
1450 //
1451 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined.
1452 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
1453 // CHECK2-NEXT:  entry:
1454 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1455 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1456 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1457 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1458 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
1459 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
1460 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1461 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1462 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1463 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1464 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1465 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
1466 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
1467 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1468 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1469 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1470 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
1471 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1472 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1473 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
1474 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1475 // CHECK2:       cond.true:
1476 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1477 // CHECK2:       cond.false:
1478 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1479 // CHECK2-NEXT:    br label [[COND_END]]
1480 // CHECK2:       cond.end:
1481 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
1482 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1483 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1484 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
1485 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1486 // CHECK2:       omp.inner.for.cond:
1487 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1488 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1489 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
1490 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1491 // CHECK2:       omp.inner.for.body:
1492 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1493 // CHECK2-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
1494 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1495 // CHECK2-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
1496 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
1497 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1498 // CHECK2:       omp.inner.for.inc:
1499 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1500 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1501 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
1502 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1503 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1504 // CHECK2:       omp.inner.for.end:
1505 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1506 // CHECK2:       omp.loop.exit:
1507 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
1508 // CHECK2-NEXT:    ret void
1509 //
1510 //
1511 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..1
1512 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1513 // CHECK2-NEXT:  entry:
1514 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1515 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1516 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1517 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1518 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1519 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1520 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1521 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
1522 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1523 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1524 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1525 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1526 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1527 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1528 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1529 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
1530 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
1531 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1532 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
1533 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1534 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
1535 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
1536 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
1537 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1538 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1539 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1540 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
1541 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1542 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1543 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
1544 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1545 // CHECK2:       cond.true:
1546 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1547 // CHECK2:       cond.false:
1548 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1549 // CHECK2-NEXT:    br label [[COND_END]]
1550 // CHECK2:       cond.end:
1551 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
1552 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
1553 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
1554 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
1555 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1556 // CHECK2:       omp.inner.for.cond:
1557 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1558 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1559 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
1560 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1561 // CHECK2:       omp.inner.for.body:
1562 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1563 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
1564 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
1565 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
1566 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
1567 // CHECK2:       omp.body.continue:
1568 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1569 // CHECK2:       omp.inner.for.inc:
1570 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1571 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
1572 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
1573 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1574 // CHECK2:       omp.inner.for.end:
1575 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1576 // CHECK2:       omp.loop.exit:
1577 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
1578 // CHECK2-NEXT:    ret void
1579 //
1580 //
1581 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
1582 // CHECK2-SAME: () #[[ATTR1]] {
1583 // CHECK2-NEXT:  entry:
1584 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
1585 // CHECK2-NEXT:    ret void
1586 //
1587 //
1588 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..2
1589 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
1590 // CHECK2-NEXT:  entry:
1591 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1592 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1593 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1594 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1595 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
1596 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
1597 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1598 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1599 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1600 // CHECK2-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
1601 // CHECK2-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
1602 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1603 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1604 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
1605 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
1606 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1607 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1608 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1609 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
1610 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1611 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1612 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
1613 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1614 // CHECK2:       cond.true:
1615 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1616 // CHECK2:       cond.false:
1617 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1618 // CHECK2-NEXT:    br label [[COND_END]]
1619 // CHECK2:       cond.end:
1620 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
1621 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1622 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1623 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
1624 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1625 // CHECK2:       omp.inner.for.cond:
1626 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1627 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1628 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
1629 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1630 // CHECK2:       omp.inner.for.body:
1631 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1632 // CHECK2-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
1633 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1634 // CHECK2-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
1635 // CHECK2-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
1636 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1637 // CHECK2-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
1638 // CHECK2-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
1639 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1640 // CHECK2:       omp.inner.for.inc:
1641 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1642 // CHECK2-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1643 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
1644 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1645 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1646 // CHECK2:       omp.inner.for.end:
1647 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1648 // CHECK2:       omp.loop.exit:
1649 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
1650 // CHECK2-NEXT:    ret void
1651 //
1652 //
1653 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..3
1654 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1655 // CHECK2-NEXT:  entry:
1656 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1657 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1658 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1659 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1660 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1661 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1662 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1663 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
1664 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1665 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1666 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1667 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1668 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1669 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1670 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1671 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
1672 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
1673 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1674 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
1675 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1676 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
1677 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
1678 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
1679 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1680 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1681 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1682 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
1683 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1684 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1685 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
1686 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1687 // CHECK2:       cond.true:
1688 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1689 // CHECK2:       cond.false:
1690 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1691 // CHECK2-NEXT:    br label [[COND_END]]
1692 // CHECK2:       cond.end:
1693 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
1694 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
1695 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
1696 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
1697 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1698 // CHECK2:       omp.inner.for.cond:
1699 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1700 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1701 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
1702 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1703 // CHECK2:       omp.inner.for.body:
1704 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1705 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
1706 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
1707 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
1708 // CHECK2-NEXT:    call void @_Z9gtid_testv()
1709 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
1710 // CHECK2:       omp.body.continue:
1711 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1712 // CHECK2:       omp.inner.for.inc:
1713 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1714 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
1715 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
1716 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1717 // CHECK2:       omp.inner.for.end:
1718 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1719 // CHECK2:       omp.loop.exit:
1720 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
1721 // CHECK2-NEXT:    ret void
1722 //
1723 //
1724 // CHECK2-LABEL: define {{[^@]+}}@main
1725 // CHECK2-SAME: () #[[ATTR3:[0-9]+]] {
1726 // CHECK2-NEXT:  entry:
1727 // CHECK2-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
1728 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1729 // CHECK2-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
1730 // CHECK2-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
1731 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
1732 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
1733 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
1734 // CHECK2-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
1735 // CHECK2-NEXT:    store i32 0, i32* [[RETVAL]], align 4
1736 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
1737 // CHECK2-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
1738 // CHECK2-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
1739 // CHECK2-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
1740 // CHECK2:       omp_offload.failed:
1741 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
1742 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
1743 // CHECK2:       omp_offload.cont:
1744 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
1745 // CHECK2-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
1746 // CHECK2-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
1747 // CHECK2-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
1748 // CHECK2:       omp_offload.failed2:
1749 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
1750 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
1751 // CHECK2:       omp_offload.cont3:
1752 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
1753 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
1754 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
1755 // CHECK2-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
1756 // CHECK2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1757 // CHECK2-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
1758 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
1759 // CHECK2-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1760 // CHECK2-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
1761 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
1762 // CHECK2-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
1763 // CHECK2-NEXT:    store i8* null, i8** [[TMP10]], align 8
1764 // CHECK2-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
1765 // CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
1766 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
1767 // CHECK2-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
1768 // CHECK2-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
1769 // CHECK2-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
1770 // CHECK2:       omp_offload.failed5:
1771 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
1772 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
1773 // CHECK2:       omp_offload.cont6:
1774 // CHECK2-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
1775 // CHECK2-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
1776 // CHECK2-NEXT:    ret i32 [[CALL]]
1777 //
1778 //
1779 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
1780 // CHECK2-SAME: () #[[ATTR1]] {
1781 // CHECK2-NEXT:  entry:
1782 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
1783 // CHECK2-NEXT:    ret void
1784 //
1785 //
1786 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..4
1787 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
1788 // CHECK2-NEXT:  entry:
1789 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1790 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1791 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1792 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1793 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
1794 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
1795 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1796 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1797 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1798 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1799 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1800 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
1801 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
1802 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1803 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1804 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1805 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
1806 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1807 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1808 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
1809 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1810 // CHECK2:       cond.true:
1811 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1812 // CHECK2:       cond.false:
1813 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1814 // CHECK2-NEXT:    br label [[COND_END]]
1815 // CHECK2:       cond.end:
1816 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
1817 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1818 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1819 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
1820 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1821 // CHECK2:       omp.inner.for.cond:
1822 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1823 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1824 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
1825 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1826 // CHECK2:       omp.inner.for.body:
1827 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1828 // CHECK2-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
1829 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1830 // CHECK2-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
1831 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
1832 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1833 // CHECK2:       omp.inner.for.inc:
1834 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1835 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1836 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
1837 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1838 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1839 // CHECK2:       omp.inner.for.end:
1840 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1841 // CHECK2:       omp.loop.exit:
1842 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
1843 // CHECK2-NEXT:    ret void
1844 //
1845 //
1846 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..5
1847 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1848 // CHECK2-NEXT:  entry:
1849 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1850 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1851 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1852 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1853 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1854 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1855 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1856 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
1857 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1858 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1859 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1860 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1861 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1862 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1863 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1864 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
1865 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
1866 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
1867 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
1868 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
1869 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
1870 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
1871 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
1872 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1873 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1874 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1875 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
1876 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1877 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1878 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
1879 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1880 // CHECK2:       cond.true:
1881 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1882 // CHECK2:       cond.false:
1883 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1884 // CHECK2-NEXT:    br label [[COND_END]]
1885 // CHECK2:       cond.end:
1886 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
1887 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
1888 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
1889 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
1890 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1891 // CHECK2:       omp.inner.for.cond:
1892 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1893 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
1894 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
1895 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1896 // CHECK2:       omp.inner.for.body:
1897 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1898 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
1899 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
1900 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
1901 // CHECK2-NEXT:    call void @_Z3fn4v()
1902 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
1903 // CHECK2:       omp.body.continue:
1904 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1905 // CHECK2:       omp.inner.for.inc:
1906 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1907 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
1908 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
1909 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1910 // CHECK2:       omp.inner.for.end:
1911 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1912 // CHECK2:       omp.loop.exit:
1913 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
1914 // CHECK2-NEXT:    ret void
1915 //
1916 //
1917 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
1918 // CHECK2-SAME: () #[[ATTR1]] {
1919 // CHECK2-NEXT:  entry:
1920 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
1921 // CHECK2-NEXT:    ret void
1922 //
1923 //
1924 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..6
1925 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
1926 // CHECK2-NEXT:  entry:
1927 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1928 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1929 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1930 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1931 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
1932 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
1933 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
1934 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
1935 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
1936 // CHECK2-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
1937 // CHECK2-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
1938 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
1939 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
1940 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
1941 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
1942 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
1943 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
1944 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1945 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
1946 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
1947 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1948 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
1949 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
1950 // CHECK2:       cond.true:
1951 // CHECK2-NEXT:    br label [[COND_END:%.*]]
1952 // CHECK2:       cond.false:
1953 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1954 // CHECK2-NEXT:    br label [[COND_END]]
1955 // CHECK2:       cond.end:
1956 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
1957 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
1958 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1959 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
1960 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
1961 // CHECK2:       omp.inner.for.cond:
1962 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1963 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1964 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
1965 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
1966 // CHECK2:       omp.inner.for.body:
1967 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
1968 // CHECK2-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
1969 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
1970 // CHECK2-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
1971 // CHECK2-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
1972 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
1973 // CHECK2-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
1974 // CHECK2-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
1975 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
1976 // CHECK2:       omp.inner.for.inc:
1977 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
1978 // CHECK2-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
1979 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
1980 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
1981 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
1982 // CHECK2:       omp.inner.for.end:
1983 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
1984 // CHECK2:       omp.loop.exit:
1985 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
1986 // CHECK2-NEXT:    ret void
1987 //
1988 //
1989 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..7
1990 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
1991 // CHECK2-NEXT:  entry:
1992 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
1993 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
1994 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
1995 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
1996 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
1997 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
1998 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
1999 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
2000 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2001 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2002 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2003 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2004 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2005 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2006 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2007 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
2008 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
2009 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2010 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
2011 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2012 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
2013 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
2014 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
2015 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2016 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2017 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2018 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
2019 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2020 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2021 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
2022 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2023 // CHECK2:       cond.true:
2024 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2025 // CHECK2:       cond.false:
2026 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2027 // CHECK2-NEXT:    br label [[COND_END]]
2028 // CHECK2:       cond.end:
2029 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
2030 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
2031 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
2032 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
2033 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2034 // CHECK2:       omp.inner.for.cond:
2035 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2036 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2037 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
2038 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2039 // CHECK2:       omp.inner.for.body:
2040 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2041 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
2042 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
2043 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
2044 // CHECK2-NEXT:    call void @_Z3fn5v()
2045 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
2046 // CHECK2:       omp.body.continue:
2047 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2048 // CHECK2:       omp.inner.for.inc:
2049 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2050 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
2051 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
2052 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2053 // CHECK2:       omp.inner.for.end:
2054 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2055 // CHECK2:       omp.loop.exit:
2056 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
2057 // CHECK2-NEXT:    ret void
2058 //
2059 //
2060 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
2061 // CHECK2-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
2062 // CHECK2-NEXT:  entry:
2063 // CHECK2-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
2064 // CHECK2-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
2065 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
2066 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
2067 // CHECK2-NEXT:    ret void
2068 //
2069 //
2070 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..8
2071 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
2072 // CHECK2-NEXT:  entry:
2073 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2074 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2075 // CHECK2-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
2076 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2077 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2078 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
2079 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
2080 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2081 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2082 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2083 // CHECK2-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
2084 // CHECK2-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
2085 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2086 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2087 // CHECK2-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
2088 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
2089 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
2090 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
2091 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2092 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2093 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2094 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
2095 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2096 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2097 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
2098 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2099 // CHECK2:       cond.true:
2100 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2101 // CHECK2:       cond.false:
2102 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2103 // CHECK2-NEXT:    br label [[COND_END]]
2104 // CHECK2:       cond.end:
2105 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
2106 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
2107 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2108 // CHECK2-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
2109 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2110 // CHECK2:       omp.inner.for.cond:
2111 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2112 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2113 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
2114 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2115 // CHECK2:       omp.inner.for.body:
2116 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2117 // CHECK2-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
2118 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2119 // CHECK2-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
2120 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
2121 // CHECK2-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
2122 // CHECK2-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
2123 // CHECK2:       omp_if.then:
2124 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
2125 // CHECK2-NEXT:    br label [[OMP_IF_END:%.*]]
2126 // CHECK2:       omp_if.else:
2127 // CHECK2-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
2128 // CHECK2-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2129 // CHECK2-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
2130 // CHECK2-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
2131 // CHECK2-NEXT:    br label [[OMP_IF_END]]
2132 // CHECK2:       omp_if.end:
2133 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2134 // CHECK2:       omp.inner.for.inc:
2135 // CHECK2-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2136 // CHECK2-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
2137 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
2138 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
2139 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2140 // CHECK2:       omp.inner.for.end:
2141 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2142 // CHECK2:       omp.loop.exit:
2143 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
2144 // CHECK2-NEXT:    ret void
2145 //
2146 //
2147 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..9
2148 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
2149 // CHECK2-NEXT:  entry:
2150 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2151 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2152 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
2153 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
2154 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2155 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2156 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
2157 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
2158 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2159 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2160 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2161 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2162 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2163 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2164 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2165 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
2166 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
2167 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2168 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
2169 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2170 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
2171 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
2172 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
2173 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2174 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2175 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2176 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
2177 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2178 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2179 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
2180 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2181 // CHECK2:       cond.true:
2182 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2183 // CHECK2:       cond.false:
2184 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2185 // CHECK2-NEXT:    br label [[COND_END]]
2186 // CHECK2:       cond.end:
2187 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
2188 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
2189 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
2190 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
2191 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2192 // CHECK2:       omp.inner.for.cond:
2193 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2194 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2195 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
2196 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2197 // CHECK2:       omp.inner.for.body:
2198 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2199 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
2200 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
2201 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
2202 // CHECK2-NEXT:    call void @_Z3fn6v()
2203 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
2204 // CHECK2:       omp.body.continue:
2205 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2206 // CHECK2:       omp.inner.for.inc:
2207 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2208 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
2209 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
2210 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2211 // CHECK2:       omp.inner.for.end:
2212 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2213 // CHECK2:       omp.loop.exit:
2214 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
2215 // CHECK2-NEXT:    ret void
2216 //
2217 //
2218 // CHECK2-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
2219 // CHECK2-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
2220 // CHECK2-NEXT:  entry:
2221 // CHECK2-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
2222 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2223 // CHECK2-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
2224 // CHECK2-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
2225 // CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
2226 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
2227 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
2228 // CHECK2-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
2229 // CHECK2-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
2230 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
2231 // CHECK2-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
2232 // CHECK2-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
2233 // CHECK2-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
2234 // CHECK2:       omp_offload.failed:
2235 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
2236 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
2237 // CHECK2:       omp_offload.cont:
2238 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
2239 // CHECK2-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
2240 // CHECK2-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
2241 // CHECK2-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
2242 // CHECK2:       omp_offload.failed2:
2243 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
2244 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
2245 // CHECK2:       omp_offload.cont3:
2246 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
2247 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
2248 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
2249 // CHECK2-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
2250 // CHECK2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2251 // CHECK2-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
2252 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
2253 // CHECK2-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2254 // CHECK2-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
2255 // CHECK2-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
2256 // CHECK2-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
2257 // CHECK2-NEXT:    store i8* null, i8** [[TMP10]], align 8
2258 // CHECK2-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
2259 // CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
2260 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
2261 // CHECK2-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
2262 // CHECK2-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
2263 // CHECK2-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
2264 // CHECK2:       omp_offload.failed5:
2265 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
2266 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
2267 // CHECK2:       omp_offload.cont6:
2268 // CHECK2-NEXT:    ret i32 0
2269 //
2270 //
2271 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
2272 // CHECK2-SAME: () #[[ATTR1]] {
2273 // CHECK2-NEXT:  entry:
2274 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
2275 // CHECK2-NEXT:    ret void
2276 //
2277 //
2278 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..10
2279 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
2280 // CHECK2-NEXT:  entry:
2281 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2282 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2283 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2284 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2285 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
2286 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
2287 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2288 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2289 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2290 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2291 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2292 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
2293 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
2294 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2295 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2296 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2297 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
2298 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2299 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2300 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
2301 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2302 // CHECK2:       cond.true:
2303 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2304 // CHECK2:       cond.false:
2305 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2306 // CHECK2-NEXT:    br label [[COND_END]]
2307 // CHECK2:       cond.end:
2308 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
2309 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
2310 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2311 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
2312 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2313 // CHECK2:       omp.inner.for.cond:
2314 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2315 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2316 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
2317 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2318 // CHECK2:       omp.inner.for.body:
2319 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2320 // CHECK2-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
2321 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2322 // CHECK2-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
2323 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
2324 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2325 // CHECK2:       omp.inner.for.inc:
2326 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2327 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
2328 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
2329 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
2330 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2331 // CHECK2:       omp.inner.for.end:
2332 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2333 // CHECK2:       omp.loop.exit:
2334 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
2335 // CHECK2-NEXT:    ret void
2336 //
2337 //
2338 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..11
2339 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
2340 // CHECK2-NEXT:  entry:
2341 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2342 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2343 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
2344 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
2345 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2346 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2347 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
2348 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
2349 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2350 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2351 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2352 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2353 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2354 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2355 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2356 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
2357 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
2358 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2359 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
2360 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2361 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
2362 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
2363 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
2364 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2365 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2366 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2367 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
2368 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2369 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2370 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
2371 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2372 // CHECK2:       cond.true:
2373 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2374 // CHECK2:       cond.false:
2375 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2376 // CHECK2-NEXT:    br label [[COND_END]]
2377 // CHECK2:       cond.end:
2378 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
2379 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
2380 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
2381 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
2382 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2383 // CHECK2:       omp.inner.for.cond:
2384 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2385 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2386 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
2387 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2388 // CHECK2:       omp.inner.for.body:
2389 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2390 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
2391 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
2392 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
2393 // CHECK2-NEXT:    call void @_Z3fn1v()
2394 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
2395 // CHECK2:       omp.body.continue:
2396 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2397 // CHECK2:       omp.inner.for.inc:
2398 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2399 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
2400 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
2401 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2402 // CHECK2:       omp.inner.for.end:
2403 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2404 // CHECK2:       omp.loop.exit:
2405 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
2406 // CHECK2-NEXT:    ret void
2407 //
2408 //
2409 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
2410 // CHECK2-SAME: () #[[ATTR1]] {
2411 // CHECK2-NEXT:  entry:
2412 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
2413 // CHECK2-NEXT:    ret void
2414 //
2415 //
2416 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..12
2417 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
2418 // CHECK2-NEXT:  entry:
2419 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2420 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2421 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2422 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2423 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
2424 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
2425 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2426 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2427 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2428 // CHECK2-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
2429 // CHECK2-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
2430 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2431 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2432 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
2433 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
2434 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2435 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2436 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2437 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
2438 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2439 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2440 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
2441 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2442 // CHECK2:       cond.true:
2443 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2444 // CHECK2:       cond.false:
2445 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2446 // CHECK2-NEXT:    br label [[COND_END]]
2447 // CHECK2:       cond.end:
2448 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
2449 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
2450 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2451 // CHECK2-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
2452 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2453 // CHECK2:       omp.inner.for.cond:
2454 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2455 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2456 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
2457 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2458 // CHECK2:       omp.inner.for.body:
2459 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2460 // CHECK2-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
2461 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2462 // CHECK2-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
2463 // CHECK2-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
2464 // CHECK2-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2465 // CHECK2-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
2466 // CHECK2-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
2467 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2468 // CHECK2:       omp.inner.for.inc:
2469 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2470 // CHECK2-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
2471 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
2472 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
2473 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2474 // CHECK2:       omp.inner.for.end:
2475 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2476 // CHECK2:       omp.loop.exit:
2477 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
2478 // CHECK2-NEXT:    ret void
2479 //
2480 //
2481 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..13
2482 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
2483 // CHECK2-NEXT:  entry:
2484 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2485 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2486 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
2487 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
2488 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2489 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2490 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
2491 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
2492 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2493 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2494 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2495 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2496 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2497 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2498 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2499 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
2500 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
2501 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2502 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
2503 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2504 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
2505 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
2506 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
2507 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2508 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2509 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2510 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
2511 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2512 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2513 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
2514 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2515 // CHECK2:       cond.true:
2516 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2517 // CHECK2:       cond.false:
2518 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2519 // CHECK2-NEXT:    br label [[COND_END]]
2520 // CHECK2:       cond.end:
2521 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
2522 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
2523 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
2524 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
2525 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2526 // CHECK2:       omp.inner.for.cond:
2527 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2528 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2529 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
2530 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2531 // CHECK2:       omp.inner.for.body:
2532 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2533 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
2534 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
2535 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
2536 // CHECK2-NEXT:    call void @_Z3fn2v()
2537 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
2538 // CHECK2:       omp.body.continue:
2539 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2540 // CHECK2:       omp.inner.for.inc:
2541 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2542 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
2543 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
2544 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2545 // CHECK2:       omp.inner.for.end:
2546 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2547 // CHECK2:       omp.loop.exit:
2548 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
2549 // CHECK2-NEXT:    ret void
2550 //
2551 //
2552 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
2553 // CHECK2-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
2554 // CHECK2-NEXT:  entry:
2555 // CHECK2-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
2556 // CHECK2-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
2557 // CHECK2-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
2558 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
2559 // CHECK2-NEXT:    ret void
2560 //
2561 //
2562 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..14
2563 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
2564 // CHECK2-NEXT:  entry:
2565 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2566 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2567 // CHECK2-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
2568 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2569 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2570 // CHECK2-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
2571 // CHECK2-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
2572 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2573 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2574 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2575 // CHECK2-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
2576 // CHECK2-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
2577 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2578 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2579 // CHECK2-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
2580 // CHECK2-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
2581 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
2582 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
2583 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2584 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2585 // CHECK2-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2586 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
2587 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2588 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2589 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
2590 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2591 // CHECK2:       cond.true:
2592 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2593 // CHECK2:       cond.false:
2594 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2595 // CHECK2-NEXT:    br label [[COND_END]]
2596 // CHECK2:       cond.end:
2597 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
2598 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
2599 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2600 // CHECK2-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
2601 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2602 // CHECK2:       omp.inner.for.cond:
2603 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2604 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2605 // CHECK2-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
2606 // CHECK2-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2607 // CHECK2:       omp.inner.for.body:
2608 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
2609 // CHECK2-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
2610 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
2611 // CHECK2-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
2612 // CHECK2-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
2613 // CHECK2-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
2614 // CHECK2-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
2615 // CHECK2:       omp_if.then:
2616 // CHECK2-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
2617 // CHECK2-NEXT:    br label [[OMP_IF_END:%.*]]
2618 // CHECK2:       omp_if.else:
2619 // CHECK2-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
2620 // CHECK2-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2621 // CHECK2-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
2622 // CHECK2-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
2623 // CHECK2-NEXT:    br label [[OMP_IF_END]]
2624 // CHECK2:       omp_if.end:
2625 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2626 // CHECK2:       omp.inner.for.inc:
2627 // CHECK2-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2628 // CHECK2-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
2629 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
2630 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
2631 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2632 // CHECK2:       omp.inner.for.end:
2633 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2634 // CHECK2:       omp.loop.exit:
2635 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
2636 // CHECK2-NEXT:    ret void
2637 //
2638 //
2639 // CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..15
2640 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
2641 // CHECK2-NEXT:  entry:
2642 // CHECK2-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
2643 // CHECK2-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
2644 // CHECK2-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
2645 // CHECK2-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
2646 // CHECK2-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
2647 // CHECK2-NEXT:    [[TMP:%.*]] = alloca i32, align 4
2648 // CHECK2-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
2649 // CHECK2-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
2650 // CHECK2-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
2651 // CHECK2-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
2652 // CHECK2-NEXT:    [[I:%.*]] = alloca i32, align 4
2653 // CHECK2-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
2654 // CHECK2-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
2655 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2656 // CHECK2-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2657 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
2658 // CHECK2-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
2659 // CHECK2-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
2660 // CHECK2-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
2661 // CHECK2-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
2662 // CHECK2-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
2663 // CHECK2-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
2664 // CHECK2-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
2665 // CHECK2-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
2666 // CHECK2-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
2667 // CHECK2-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
2668 // CHECK2-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
2669 // CHECK2-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
2670 // CHECK2-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2671 // CHECK2-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
2672 // CHECK2-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
2673 // CHECK2:       cond.true:
2674 // CHECK2-NEXT:    br label [[COND_END:%.*]]
2675 // CHECK2:       cond.false:
2676 // CHECK2-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2677 // CHECK2-NEXT:    br label [[COND_END]]
2678 // CHECK2:       cond.end:
2679 // CHECK2-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
2680 // CHECK2-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
2681 // CHECK2-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
2682 // CHECK2-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
2683 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
2684 // CHECK2:       omp.inner.for.cond:
2685 // CHECK2-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2686 // CHECK2-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
2687 // CHECK2-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
2688 // CHECK2-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
2689 // CHECK2:       omp.inner.for.body:
2690 // CHECK2-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2691 // CHECK2-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
2692 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
2693 // CHECK2-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
2694 // CHECK2-NEXT:    call void @_Z3fn3v()
2695 // CHECK2-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
2696 // CHECK2:       omp.body.continue:
2697 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
2698 // CHECK2:       omp.inner.for.inc:
2699 // CHECK2-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
2700 // CHECK2-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
2701 // CHECK2-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
2702 // CHECK2-NEXT:    br label [[OMP_INNER_FOR_COND]]
2703 // CHECK2:       omp.inner.for.end:
2704 // CHECK2-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
2705 // CHECK2:       omp.loop.exit:
2706 // CHECK2-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
2707 // CHECK2-NEXT:    ret void
2708 //
2709 //
2710 // CHECK2-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
2711 // CHECK2-SAME: () #[[ATTR5:[0-9]+]] {
2712 // CHECK2-NEXT:  entry:
2713 // CHECK2-NEXT:    call void @__tgt_register_requires(i64 1)
2714 // CHECK2-NEXT:    ret void
2715 //
2716 //
2717 // CHECK3-LABEL: define {{[^@]+}}@_Z9gtid_testv
2718 // CHECK3-SAME: () #[[ATTR0:[0-9]+]] {
2719 // CHECK3-NEXT:  entry:
2720 // CHECK3-NEXT:    [[I:%.*]] = alloca i32, align 4
2721 // CHECK3-NEXT:    [[I1:%.*]] = alloca i32, align 4
2722 // CHECK3-NEXT:    store i32 0, i32* [[I]], align 4
2723 // CHECK3-NEXT:    br label [[FOR_COND:%.*]]
2724 // CHECK3:       for.cond:
2725 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
2726 // CHECK3-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
2727 // CHECK3-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
2728 // CHECK3:       for.body:
2729 // CHECK3-NEXT:    br label [[FOR_INC:%.*]]
2730 // CHECK3:       for.inc:
2731 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
2732 // CHECK3-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
2733 // CHECK3-NEXT:    store i32 [[INC]], i32* [[I]], align 4
2734 // CHECK3-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
2735 // CHECK3:       for.end:
2736 // CHECK3-NEXT:    store i32 0, i32* [[I1]], align 4
2737 // CHECK3-NEXT:    br label [[FOR_COND2:%.*]]
2738 // CHECK3:       for.cond2:
2739 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
2740 // CHECK3-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
2741 // CHECK3-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
2742 // CHECK3:       for.body4:
2743 // CHECK3-NEXT:    call void @_Z9gtid_testv()
2744 // CHECK3-NEXT:    br label [[FOR_INC5:%.*]]
2745 // CHECK3:       for.inc5:
2746 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
2747 // CHECK3-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
2748 // CHECK3-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
2749 // CHECK3-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP4:![0-9]+]]
2750 // CHECK3:       for.end7:
2751 // CHECK3-NEXT:    ret void
2752 //
2753 //
2754 // CHECK3-LABEL: define {{[^@]+}}@main
2755 // CHECK3-SAME: () #[[ATTR1:[0-9]+]] {
2756 // CHECK3-NEXT:  entry:
2757 // CHECK3-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
2758 // CHECK3-NEXT:    [[I:%.*]] = alloca i32, align 4
2759 // CHECK3-NEXT:    [[I1:%.*]] = alloca i32, align 4
2760 // CHECK3-NEXT:    [[I8:%.*]] = alloca i32, align 4
2761 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
2762 // CHECK3-NEXT:    store i32 0, i32* [[I]], align 4
2763 // CHECK3-NEXT:    br label [[FOR_COND:%.*]]
2764 // CHECK3:       for.cond:
2765 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
2766 // CHECK3-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
2767 // CHECK3-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
2768 // CHECK3:       for.body:
2769 // CHECK3-NEXT:    call void @_Z3fn4v()
2770 // CHECK3-NEXT:    br label [[FOR_INC:%.*]]
2771 // CHECK3:       for.inc:
2772 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
2773 // CHECK3-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
2774 // CHECK3-NEXT:    store i32 [[INC]], i32* [[I]], align 4
2775 // CHECK3-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
2776 // CHECK3:       for.end:
2777 // CHECK3-NEXT:    store i32 0, i32* [[I1]], align 4
2778 // CHECK3-NEXT:    br label [[FOR_COND2:%.*]]
2779 // CHECK3:       for.cond2:
2780 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
2781 // CHECK3-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
2782 // CHECK3-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
2783 // CHECK3:       for.body4:
2784 // CHECK3-NEXT:    call void @_Z3fn5v()
2785 // CHECK3-NEXT:    br label [[FOR_INC5:%.*]]
2786 // CHECK3:       for.inc5:
2787 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
2788 // CHECK3-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
2789 // CHECK3-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
2790 // CHECK3-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP6:![0-9]+]]
2791 // CHECK3:       for.end7:
2792 // CHECK3-NEXT:    store i32 0, i32* [[I8]], align 4
2793 // CHECK3-NEXT:    br label [[FOR_COND9:%.*]]
2794 // CHECK3:       for.cond9:
2795 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
2796 // CHECK3-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
2797 // CHECK3-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
2798 // CHECK3:       for.body11:
2799 // CHECK3-NEXT:    call void @_Z3fn6v()
2800 // CHECK3-NEXT:    br label [[FOR_INC12:%.*]]
2801 // CHECK3:       for.inc12:
2802 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
2803 // CHECK3-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
2804 // CHECK3-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
2805 // CHECK3-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP7:![0-9]+]]
2806 // CHECK3:       for.end14:
2807 // CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
2808 // CHECK3-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
2809 // CHECK3-NEXT:    ret i32 [[CALL]]
2810 //
2811 //
2812 // CHECK3-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
2813 // CHECK3-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
2814 // CHECK3-NEXT:  entry:
2815 // CHECK3-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
2816 // CHECK3-NEXT:    [[I:%.*]] = alloca i32, align 4
2817 // CHECK3-NEXT:    [[I1:%.*]] = alloca i32, align 4
2818 // CHECK3-NEXT:    [[I8:%.*]] = alloca i32, align 4
2819 // CHECK3-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
2820 // CHECK3-NEXT:    store i32 0, i32* [[I]], align 4
2821 // CHECK3-NEXT:    br label [[FOR_COND:%.*]]
2822 // CHECK3:       for.cond:
2823 // CHECK3-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
2824 // CHECK3-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
2825 // CHECK3-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
2826 // CHECK3:       for.body:
2827 // CHECK3-NEXT:    call void @_Z3fn1v()
2828 // CHECK3-NEXT:    br label [[FOR_INC:%.*]]
2829 // CHECK3:       for.inc:
2830 // CHECK3-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
2831 // CHECK3-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
2832 // CHECK3-NEXT:    store i32 [[INC]], i32* [[I]], align 4
2833 // CHECK3-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
2834 // CHECK3:       for.end:
2835 // CHECK3-NEXT:    store i32 0, i32* [[I1]], align 4
2836 // CHECK3-NEXT:    br label [[FOR_COND2:%.*]]
2837 // CHECK3:       for.cond2:
2838 // CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
2839 // CHECK3-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
2840 // CHECK3-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
2841 // CHECK3:       for.body4:
2842 // CHECK3-NEXT:    call void @_Z3fn2v()
2843 // CHECK3-NEXT:    br label [[FOR_INC5:%.*]]
2844 // CHECK3:       for.inc5:
2845 // CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
2846 // CHECK3-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
2847 // CHECK3-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
2848 // CHECK3-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP9:![0-9]+]]
2849 // CHECK3:       for.end7:
2850 // CHECK3-NEXT:    store i32 0, i32* [[I8]], align 4
2851 // CHECK3-NEXT:    br label [[FOR_COND9:%.*]]
2852 // CHECK3:       for.cond9:
2853 // CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
2854 // CHECK3-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
2855 // CHECK3-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
2856 // CHECK3:       for.body11:
2857 // CHECK3-NEXT:    call void @_Z3fn3v()
2858 // CHECK3-NEXT:    br label [[FOR_INC12:%.*]]
2859 // CHECK3:       for.inc12:
2860 // CHECK3-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
2861 // CHECK3-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
2862 // CHECK3-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
2863 // CHECK3-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP10:![0-9]+]]
2864 // CHECK3:       for.end14:
2865 // CHECK3-NEXT:    ret i32 0
2866 //
2867 //
2868 // CHECK4-LABEL: define {{[^@]+}}@_Z9gtid_testv
2869 // CHECK4-SAME: () #[[ATTR0:[0-9]+]] {
2870 // CHECK4-NEXT:  entry:
2871 // CHECK4-NEXT:    [[I:%.*]] = alloca i32, align 4
2872 // CHECK4-NEXT:    [[I1:%.*]] = alloca i32, align 4
2873 // CHECK4-NEXT:    store i32 0, i32* [[I]], align 4
2874 // CHECK4-NEXT:    br label [[FOR_COND:%.*]]
2875 // CHECK4:       for.cond:
2876 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
2877 // CHECK4-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
2878 // CHECK4-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
2879 // CHECK4:       for.body:
2880 // CHECK4-NEXT:    br label [[FOR_INC:%.*]]
2881 // CHECK4:       for.inc:
2882 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
2883 // CHECK4-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
2884 // CHECK4-NEXT:    store i32 [[INC]], i32* [[I]], align 4
2885 // CHECK4-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
2886 // CHECK4:       for.end:
2887 // CHECK4-NEXT:    store i32 0, i32* [[I1]], align 4
2888 // CHECK4-NEXT:    br label [[FOR_COND2:%.*]]
2889 // CHECK4:       for.cond2:
2890 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
2891 // CHECK4-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
2892 // CHECK4-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
2893 // CHECK4:       for.body4:
2894 // CHECK4-NEXT:    call void @_Z9gtid_testv()
2895 // CHECK4-NEXT:    br label [[FOR_INC5:%.*]]
2896 // CHECK4:       for.inc5:
2897 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
2898 // CHECK4-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
2899 // CHECK4-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
2900 // CHECK4-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP4:![0-9]+]]
2901 // CHECK4:       for.end7:
2902 // CHECK4-NEXT:    ret void
2903 //
2904 //
2905 // CHECK4-LABEL: define {{[^@]+}}@main
2906 // CHECK4-SAME: () #[[ATTR1:[0-9]+]] {
2907 // CHECK4-NEXT:  entry:
2908 // CHECK4-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
2909 // CHECK4-NEXT:    [[I:%.*]] = alloca i32, align 4
2910 // CHECK4-NEXT:    [[I1:%.*]] = alloca i32, align 4
2911 // CHECK4-NEXT:    [[I8:%.*]] = alloca i32, align 4
2912 // CHECK4-NEXT:    store i32 0, i32* [[RETVAL]], align 4
2913 // CHECK4-NEXT:    store i32 0, i32* [[I]], align 4
2914 // CHECK4-NEXT:    br label [[FOR_COND:%.*]]
2915 // CHECK4:       for.cond:
2916 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
2917 // CHECK4-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
2918 // CHECK4-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
2919 // CHECK4:       for.body:
2920 // CHECK4-NEXT:    call void @_Z3fn4v()
2921 // CHECK4-NEXT:    br label [[FOR_INC:%.*]]
2922 // CHECK4:       for.inc:
2923 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
2924 // CHECK4-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
2925 // CHECK4-NEXT:    store i32 [[INC]], i32* [[I]], align 4
2926 // CHECK4-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
2927 // CHECK4:       for.end:
2928 // CHECK4-NEXT:    store i32 0, i32* [[I1]], align 4
2929 // CHECK4-NEXT:    br label [[FOR_COND2:%.*]]
2930 // CHECK4:       for.cond2:
2931 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
2932 // CHECK4-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
2933 // CHECK4-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
2934 // CHECK4:       for.body4:
2935 // CHECK4-NEXT:    call void @_Z3fn5v()
2936 // CHECK4-NEXT:    br label [[FOR_INC5:%.*]]
2937 // CHECK4:       for.inc5:
2938 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
2939 // CHECK4-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
2940 // CHECK4-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
2941 // CHECK4-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP6:![0-9]+]]
2942 // CHECK4:       for.end7:
2943 // CHECK4-NEXT:    store i32 0, i32* [[I8]], align 4
2944 // CHECK4-NEXT:    br label [[FOR_COND9:%.*]]
2945 // CHECK4:       for.cond9:
2946 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
2947 // CHECK4-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
2948 // CHECK4-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
2949 // CHECK4:       for.body11:
2950 // CHECK4-NEXT:    call void @_Z3fn6v()
2951 // CHECK4-NEXT:    br label [[FOR_INC12:%.*]]
2952 // CHECK4:       for.inc12:
2953 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
2954 // CHECK4-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
2955 // CHECK4-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
2956 // CHECK4-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP7:![0-9]+]]
2957 // CHECK4:       for.end14:
2958 // CHECK4-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
2959 // CHECK4-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
2960 // CHECK4-NEXT:    ret i32 [[CALL]]
2961 //
2962 //
2963 // CHECK4-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
2964 // CHECK4-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
2965 // CHECK4-NEXT:  entry:
2966 // CHECK4-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
2967 // CHECK4-NEXT:    [[I:%.*]] = alloca i32, align 4
2968 // CHECK4-NEXT:    [[I1:%.*]] = alloca i32, align 4
2969 // CHECK4-NEXT:    [[I8:%.*]] = alloca i32, align 4
2970 // CHECK4-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
2971 // CHECK4-NEXT:    store i32 0, i32* [[I]], align 4
2972 // CHECK4-NEXT:    br label [[FOR_COND:%.*]]
2973 // CHECK4:       for.cond:
2974 // CHECK4-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
2975 // CHECK4-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
2976 // CHECK4-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
2977 // CHECK4:       for.body:
2978 // CHECK4-NEXT:    call void @_Z3fn1v()
2979 // CHECK4-NEXT:    br label [[FOR_INC:%.*]]
2980 // CHECK4:       for.inc:
2981 // CHECK4-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
2982 // CHECK4-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
2983 // CHECK4-NEXT:    store i32 [[INC]], i32* [[I]], align 4
2984 // CHECK4-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
2985 // CHECK4:       for.end:
2986 // CHECK4-NEXT:    store i32 0, i32* [[I1]], align 4
2987 // CHECK4-NEXT:    br label [[FOR_COND2:%.*]]
2988 // CHECK4:       for.cond2:
2989 // CHECK4-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
2990 // CHECK4-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
2991 // CHECK4-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
2992 // CHECK4:       for.body4:
2993 // CHECK4-NEXT:    call void @_Z3fn2v()
2994 // CHECK4-NEXT:    br label [[FOR_INC5:%.*]]
2995 // CHECK4:       for.inc5:
2996 // CHECK4-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
2997 // CHECK4-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
2998 // CHECK4-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
2999 // CHECK4-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP9:![0-9]+]]
3000 // CHECK4:       for.end7:
3001 // CHECK4-NEXT:    store i32 0, i32* [[I8]], align 4
3002 // CHECK4-NEXT:    br label [[FOR_COND9:%.*]]
3003 // CHECK4:       for.cond9:
3004 // CHECK4-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
3005 // CHECK4-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
3006 // CHECK4-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
3007 // CHECK4:       for.body11:
3008 // CHECK4-NEXT:    call void @_Z3fn3v()
3009 // CHECK4-NEXT:    br label [[FOR_INC12:%.*]]
3010 // CHECK4:       for.inc12:
3011 // CHECK4-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
3012 // CHECK4-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
3013 // CHECK4-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
3014 // CHECK4-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP10:![0-9]+]]
3015 // CHECK4:       for.end14:
3016 // CHECK4-NEXT:    ret i32 0
3017 //
3018 //
3019 // CHECK5-LABEL: define {{[^@]+}}@_Z9gtid_testv
3020 // CHECK5-SAME: () #[[ATTR0:[0-9]+]] {
3021 // CHECK5-NEXT:  entry:
3022 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3023 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
3024 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
3025 // CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
3026 // CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
3027 // CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
3028 // CHECK5:       omp_offload.failed:
3029 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
3030 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
3031 // CHECK5:       omp_offload.cont:
3032 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3033 // CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
3034 // CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
3035 // CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
3036 // CHECK5:       omp_offload.failed2:
3037 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
3038 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
3039 // CHECK5:       omp_offload.cont3:
3040 // CHECK5-NEXT:    ret void
3041 //
3042 //
3043 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
3044 // CHECK5-SAME: () #[[ATTR1:[0-9]+]] {
3045 // CHECK5-NEXT:  entry:
3046 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
3047 // CHECK5-NEXT:    ret void
3048 //
3049 //
3050 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined.
3051 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
3052 // CHECK5-NEXT:  entry:
3053 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3054 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3055 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3056 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3057 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
3058 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
3059 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3060 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3061 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3062 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3063 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3064 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
3065 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
3066 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3067 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3068 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3069 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
3070 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3071 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3072 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
3073 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3074 // CHECK5:       cond.true:
3075 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3076 // CHECK5:       cond.false:
3077 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3078 // CHECK5-NEXT:    br label [[COND_END]]
3079 // CHECK5:       cond.end:
3080 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
3081 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
3082 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3083 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
3084 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3085 // CHECK5:       omp.inner.for.cond:
3086 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3087 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3088 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
3089 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3090 // CHECK5:       omp.inner.for.body:
3091 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3092 // CHECK5-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
3093 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3094 // CHECK5-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
3095 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
3096 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3097 // CHECK5:       omp.inner.for.inc:
3098 // CHECK5-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3099 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
3100 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
3101 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
3102 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3103 // CHECK5:       omp.inner.for.end:
3104 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3105 // CHECK5:       omp.loop.exit:
3106 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
3107 // CHECK5-NEXT:    ret void
3108 //
3109 //
3110 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..1
3111 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
3112 // CHECK5-NEXT:  entry:
3113 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3114 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3115 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
3116 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
3117 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3118 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3119 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
3120 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
3121 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3122 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3123 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3124 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3125 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3126 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3127 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3128 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
3129 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
3130 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3131 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
3132 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3133 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
3134 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
3135 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
3136 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3137 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3138 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3139 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
3140 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3141 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3142 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
3143 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3144 // CHECK5:       cond.true:
3145 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3146 // CHECK5:       cond.false:
3147 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3148 // CHECK5-NEXT:    br label [[COND_END]]
3149 // CHECK5:       cond.end:
3150 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
3151 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
3152 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
3153 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
3154 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3155 // CHECK5:       omp.inner.for.cond:
3156 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3157 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3158 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
3159 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3160 // CHECK5:       omp.inner.for.body:
3161 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3162 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
3163 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
3164 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
3165 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
3166 // CHECK5:       omp.body.continue:
3167 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3168 // CHECK5:       omp.inner.for.inc:
3169 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3170 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
3171 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
3172 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3173 // CHECK5:       omp.inner.for.end:
3174 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3175 // CHECK5:       omp.loop.exit:
3176 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
3177 // CHECK5-NEXT:    ret void
3178 //
3179 //
3180 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
3181 // CHECK5-SAME: () #[[ATTR1]] {
3182 // CHECK5-NEXT:  entry:
3183 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
3184 // CHECK5-NEXT:    ret void
3185 //
3186 //
3187 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..2
3188 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
3189 // CHECK5-NEXT:  entry:
3190 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3191 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3192 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3193 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3194 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
3195 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
3196 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3197 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3198 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3199 // CHECK5-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
3200 // CHECK5-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
3201 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3202 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3203 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
3204 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
3205 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3206 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3207 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3208 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
3209 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3210 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3211 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
3212 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3213 // CHECK5:       cond.true:
3214 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3215 // CHECK5:       cond.false:
3216 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3217 // CHECK5-NEXT:    br label [[COND_END]]
3218 // CHECK5:       cond.end:
3219 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
3220 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
3221 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3222 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
3223 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3224 // CHECK5:       omp.inner.for.cond:
3225 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3226 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3227 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
3228 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3229 // CHECK5:       omp.inner.for.body:
3230 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3231 // CHECK5-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
3232 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3233 // CHECK5-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
3234 // CHECK5-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
3235 // CHECK5-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3236 // CHECK5-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
3237 // CHECK5-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
3238 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3239 // CHECK5:       omp.inner.for.inc:
3240 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3241 // CHECK5-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
3242 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
3243 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
3244 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3245 // CHECK5:       omp.inner.for.end:
3246 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3247 // CHECK5:       omp.loop.exit:
3248 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
3249 // CHECK5-NEXT:    ret void
3250 //
3251 //
3252 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..3
3253 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
3254 // CHECK5-NEXT:  entry:
3255 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3256 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3257 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
3258 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
3259 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3260 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3261 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
3262 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
3263 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3264 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3265 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3266 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3267 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3268 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3269 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3270 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
3271 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
3272 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3273 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
3274 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3275 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
3276 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
3277 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
3278 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3279 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3280 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3281 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
3282 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3283 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3284 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
3285 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3286 // CHECK5:       cond.true:
3287 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3288 // CHECK5:       cond.false:
3289 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3290 // CHECK5-NEXT:    br label [[COND_END]]
3291 // CHECK5:       cond.end:
3292 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
3293 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
3294 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
3295 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
3296 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3297 // CHECK5:       omp.inner.for.cond:
3298 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3299 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3300 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
3301 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3302 // CHECK5:       omp.inner.for.body:
3303 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3304 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
3305 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
3306 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
3307 // CHECK5-NEXT:    call void @_Z9gtid_testv()
3308 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
3309 // CHECK5:       omp.body.continue:
3310 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3311 // CHECK5:       omp.inner.for.inc:
3312 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3313 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
3314 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
3315 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3316 // CHECK5:       omp.inner.for.end:
3317 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3318 // CHECK5:       omp.loop.exit:
3319 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
3320 // CHECK5-NEXT:    ret void
3321 //
3322 //
3323 // CHECK5-LABEL: define {{[^@]+}}@main
3324 // CHECK5-SAME: () #[[ATTR3:[0-9]+]] {
3325 // CHECK5-NEXT:  entry:
3326 // CHECK5-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
3327 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3328 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
3329 // CHECK5-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
3330 // CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
3331 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
3332 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
3333 // CHECK5-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
3334 // CHECK5-NEXT:    store i32 0, i32* [[RETVAL]], align 4
3335 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3336 // CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
3337 // CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
3338 // CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
3339 // CHECK5:       omp_offload.failed:
3340 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
3341 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
3342 // CHECK5:       omp_offload.cont:
3343 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3344 // CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
3345 // CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
3346 // CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
3347 // CHECK5:       omp_offload.failed2:
3348 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
3349 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
3350 // CHECK5:       omp_offload.cont3:
3351 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
3352 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
3353 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
3354 // CHECK5-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
3355 // CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3356 // CHECK5-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
3357 // CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
3358 // CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3359 // CHECK5-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
3360 // CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
3361 // CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
3362 // CHECK5-NEXT:    store i8* null, i8** [[TMP10]], align 8
3363 // CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3364 // CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3365 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3366 // CHECK5-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3367 // CHECK5-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
3368 // CHECK5-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
3369 // CHECK5:       omp_offload.failed5:
3370 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
3371 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
3372 // CHECK5:       omp_offload.cont6:
3373 // CHECK5-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
3374 // CHECK5-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
3375 // CHECK5-NEXT:    ret i32 [[CALL]]
3376 //
3377 //
3378 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
3379 // CHECK5-SAME: () #[[ATTR1]] {
3380 // CHECK5-NEXT:  entry:
3381 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
3382 // CHECK5-NEXT:    ret void
3383 //
3384 //
3385 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..4
3386 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
3387 // CHECK5-NEXT:  entry:
3388 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3389 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3390 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3391 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3392 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
3393 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
3394 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3395 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3396 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3397 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3398 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3399 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
3400 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
3401 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3402 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3403 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3404 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
3405 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3406 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3407 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
3408 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3409 // CHECK5:       cond.true:
3410 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3411 // CHECK5:       cond.false:
3412 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3413 // CHECK5-NEXT:    br label [[COND_END]]
3414 // CHECK5:       cond.end:
3415 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
3416 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
3417 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3418 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
3419 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3420 // CHECK5:       omp.inner.for.cond:
3421 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3422 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3423 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
3424 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3425 // CHECK5:       omp.inner.for.body:
3426 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3427 // CHECK5-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
3428 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3429 // CHECK5-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
3430 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
3431 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3432 // CHECK5:       omp.inner.for.inc:
3433 // CHECK5-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3434 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
3435 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
3436 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
3437 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3438 // CHECK5:       omp.inner.for.end:
3439 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3440 // CHECK5:       omp.loop.exit:
3441 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
3442 // CHECK5-NEXT:    ret void
3443 //
3444 //
3445 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..5
3446 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
3447 // CHECK5-NEXT:  entry:
3448 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3449 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3450 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
3451 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
3452 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3453 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3454 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
3455 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
3456 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3457 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3458 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3459 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3460 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3461 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3462 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3463 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
3464 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
3465 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3466 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
3467 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3468 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
3469 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
3470 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
3471 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3472 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3473 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3474 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
3475 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3476 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3477 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
3478 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3479 // CHECK5:       cond.true:
3480 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3481 // CHECK5:       cond.false:
3482 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3483 // CHECK5-NEXT:    br label [[COND_END]]
3484 // CHECK5:       cond.end:
3485 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
3486 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
3487 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
3488 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
3489 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3490 // CHECK5:       omp.inner.for.cond:
3491 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3492 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3493 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
3494 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3495 // CHECK5:       omp.inner.for.body:
3496 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3497 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
3498 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
3499 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
3500 // CHECK5-NEXT:    call void @_Z3fn4v()
3501 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
3502 // CHECK5:       omp.body.continue:
3503 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3504 // CHECK5:       omp.inner.for.inc:
3505 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3506 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
3507 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
3508 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3509 // CHECK5:       omp.inner.for.end:
3510 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3511 // CHECK5:       omp.loop.exit:
3512 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
3513 // CHECK5-NEXT:    ret void
3514 //
3515 //
3516 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
3517 // CHECK5-SAME: () #[[ATTR1]] {
3518 // CHECK5-NEXT:  entry:
3519 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
3520 // CHECK5-NEXT:    ret void
3521 //
3522 //
3523 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..6
3524 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
3525 // CHECK5-NEXT:  entry:
3526 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3527 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3528 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3529 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3530 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
3531 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
3532 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3533 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3534 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3535 // CHECK5-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
3536 // CHECK5-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
3537 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3538 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3539 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
3540 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
3541 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3542 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3543 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3544 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
3545 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3546 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3547 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
3548 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3549 // CHECK5:       cond.true:
3550 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3551 // CHECK5:       cond.false:
3552 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3553 // CHECK5-NEXT:    br label [[COND_END]]
3554 // CHECK5:       cond.end:
3555 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
3556 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
3557 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3558 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
3559 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3560 // CHECK5:       omp.inner.for.cond:
3561 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3562 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3563 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
3564 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3565 // CHECK5:       omp.inner.for.body:
3566 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3567 // CHECK5-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
3568 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3569 // CHECK5-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
3570 // CHECK5-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
3571 // CHECK5-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3572 // CHECK5-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
3573 // CHECK5-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
3574 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3575 // CHECK5:       omp.inner.for.inc:
3576 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3577 // CHECK5-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
3578 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
3579 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
3580 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3581 // CHECK5:       omp.inner.for.end:
3582 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3583 // CHECK5:       omp.loop.exit:
3584 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
3585 // CHECK5-NEXT:    ret void
3586 //
3587 //
3588 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..7
3589 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
3590 // CHECK5-NEXT:  entry:
3591 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3592 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3593 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
3594 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
3595 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3596 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3597 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
3598 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
3599 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3600 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3601 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3602 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3603 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3604 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3605 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3606 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
3607 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
3608 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3609 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
3610 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3611 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
3612 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
3613 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
3614 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3615 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3616 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3617 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
3618 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3619 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3620 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
3621 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3622 // CHECK5:       cond.true:
3623 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3624 // CHECK5:       cond.false:
3625 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3626 // CHECK5-NEXT:    br label [[COND_END]]
3627 // CHECK5:       cond.end:
3628 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
3629 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
3630 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
3631 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
3632 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3633 // CHECK5:       omp.inner.for.cond:
3634 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3635 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3636 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
3637 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3638 // CHECK5:       omp.inner.for.body:
3639 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3640 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
3641 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
3642 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
3643 // CHECK5-NEXT:    call void @_Z3fn5v()
3644 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
3645 // CHECK5:       omp.body.continue:
3646 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3647 // CHECK5:       omp.inner.for.inc:
3648 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3649 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
3650 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
3651 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3652 // CHECK5:       omp.inner.for.end:
3653 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3654 // CHECK5:       omp.loop.exit:
3655 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
3656 // CHECK5-NEXT:    ret void
3657 //
3658 //
3659 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
3660 // CHECK5-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
3661 // CHECK5-NEXT:  entry:
3662 // CHECK5-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
3663 // CHECK5-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
3664 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
3665 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
3666 // CHECK5-NEXT:    ret void
3667 //
3668 //
3669 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..8
3670 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
3671 // CHECK5-NEXT:  entry:
3672 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3673 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3674 // CHECK5-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
3675 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3676 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3677 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
3678 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
3679 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3680 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3681 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3682 // CHECK5-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
3683 // CHECK5-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
3684 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3685 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3686 // CHECK5-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
3687 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
3688 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
3689 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
3690 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3691 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3692 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3693 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
3694 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3695 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3696 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
3697 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3698 // CHECK5:       cond.true:
3699 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3700 // CHECK5:       cond.false:
3701 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3702 // CHECK5-NEXT:    br label [[COND_END]]
3703 // CHECK5:       cond.end:
3704 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
3705 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
3706 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3707 // CHECK5-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
3708 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3709 // CHECK5:       omp.inner.for.cond:
3710 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3711 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3712 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
3713 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3714 // CHECK5:       omp.inner.for.body:
3715 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3716 // CHECK5-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
3717 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3718 // CHECK5-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
3719 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
3720 // CHECK5-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
3721 // CHECK5-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
3722 // CHECK5:       omp_if.then:
3723 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
3724 // CHECK5-NEXT:    br label [[OMP_IF_END:%.*]]
3725 // CHECK5:       omp_if.else:
3726 // CHECK5-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
3727 // CHECK5-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3728 // CHECK5-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
3729 // CHECK5-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
3730 // CHECK5-NEXT:    br label [[OMP_IF_END]]
3731 // CHECK5:       omp_if.end:
3732 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3733 // CHECK5:       omp.inner.for.inc:
3734 // CHECK5-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3735 // CHECK5-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
3736 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
3737 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
3738 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3739 // CHECK5:       omp.inner.for.end:
3740 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3741 // CHECK5:       omp.loop.exit:
3742 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
3743 // CHECK5-NEXT:    ret void
3744 //
3745 //
3746 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..9
3747 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
3748 // CHECK5-NEXT:  entry:
3749 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3750 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3751 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
3752 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
3753 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3754 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3755 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
3756 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
3757 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3758 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3759 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3760 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3761 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3762 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3763 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3764 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
3765 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
3766 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3767 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
3768 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3769 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
3770 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
3771 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
3772 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3773 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3774 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3775 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
3776 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3777 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3778 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
3779 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3780 // CHECK5:       cond.true:
3781 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3782 // CHECK5:       cond.false:
3783 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3784 // CHECK5-NEXT:    br label [[COND_END]]
3785 // CHECK5:       cond.end:
3786 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
3787 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
3788 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
3789 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
3790 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3791 // CHECK5:       omp.inner.for.cond:
3792 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3793 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3794 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
3795 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3796 // CHECK5:       omp.inner.for.body:
3797 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3798 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
3799 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
3800 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
3801 // CHECK5-NEXT:    call void @_Z3fn6v()
3802 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
3803 // CHECK5:       omp.body.continue:
3804 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3805 // CHECK5:       omp.inner.for.inc:
3806 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3807 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
3808 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
3809 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3810 // CHECK5:       omp.inner.for.end:
3811 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3812 // CHECK5:       omp.loop.exit:
3813 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
3814 // CHECK5-NEXT:    ret void
3815 //
3816 //
3817 // CHECK5-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
3818 // CHECK5-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
3819 // CHECK5-NEXT:  entry:
3820 // CHECK5-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
3821 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3822 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
3823 // CHECK5-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
3824 // CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
3825 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
3826 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
3827 // CHECK5-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
3828 // CHECK5-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
3829 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3830 // CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
3831 // CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
3832 // CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
3833 // CHECK5:       omp_offload.failed:
3834 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
3835 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
3836 // CHECK5:       omp_offload.cont:
3837 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3838 // CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
3839 // CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
3840 // CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
3841 // CHECK5:       omp_offload.failed2:
3842 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
3843 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
3844 // CHECK5:       omp_offload.cont3:
3845 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
3846 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
3847 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
3848 // CHECK5-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
3849 // CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3850 // CHECK5-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
3851 // CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
3852 // CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3853 // CHECK5-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
3854 // CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
3855 // CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
3856 // CHECK5-NEXT:    store i8* null, i8** [[TMP10]], align 8
3857 // CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
3858 // CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
3859 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
3860 // CHECK5-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
3861 // CHECK5-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
3862 // CHECK5-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
3863 // CHECK5:       omp_offload.failed5:
3864 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
3865 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
3866 // CHECK5:       omp_offload.cont6:
3867 // CHECK5-NEXT:    ret i32 0
3868 //
3869 //
3870 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
3871 // CHECK5-SAME: () #[[ATTR1]] {
3872 // CHECK5-NEXT:  entry:
3873 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
3874 // CHECK5-NEXT:    ret void
3875 //
3876 //
3877 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..10
3878 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
3879 // CHECK5-NEXT:  entry:
3880 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3881 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3882 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3883 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3884 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
3885 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
3886 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3887 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3888 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3889 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3890 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3891 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
3892 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
3893 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3894 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3895 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3896 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
3897 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3898 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3899 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
3900 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3901 // CHECK5:       cond.true:
3902 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3903 // CHECK5:       cond.false:
3904 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3905 // CHECK5-NEXT:    br label [[COND_END]]
3906 // CHECK5:       cond.end:
3907 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
3908 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
3909 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3910 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
3911 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3912 // CHECK5:       omp.inner.for.cond:
3913 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3914 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3915 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
3916 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3917 // CHECK5:       omp.inner.for.body:
3918 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
3919 // CHECK5-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
3920 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
3921 // CHECK5-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
3922 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
3923 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3924 // CHECK5:       omp.inner.for.inc:
3925 // CHECK5-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3926 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
3927 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
3928 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
3929 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
3930 // CHECK5:       omp.inner.for.end:
3931 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
3932 // CHECK5:       omp.loop.exit:
3933 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
3934 // CHECK5-NEXT:    ret void
3935 //
3936 //
3937 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..11
3938 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
3939 // CHECK5-NEXT:  entry:
3940 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
3941 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
3942 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
3943 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
3944 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
3945 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
3946 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
3947 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
3948 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
3949 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
3950 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
3951 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
3952 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
3953 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3954 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3955 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
3956 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
3957 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
3958 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
3959 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
3960 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
3961 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
3962 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
3963 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
3964 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
3965 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
3966 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
3967 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
3968 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3969 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
3970 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
3971 // CHECK5:       cond.true:
3972 // CHECK5-NEXT:    br label [[COND_END:%.*]]
3973 // CHECK5:       cond.false:
3974 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3975 // CHECK5-NEXT:    br label [[COND_END]]
3976 // CHECK5:       cond.end:
3977 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
3978 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
3979 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
3980 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
3981 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
3982 // CHECK5:       omp.inner.for.cond:
3983 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3984 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
3985 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
3986 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
3987 // CHECK5:       omp.inner.for.body:
3988 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3989 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
3990 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
3991 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
3992 // CHECK5-NEXT:    call void @_Z3fn1v()
3993 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
3994 // CHECK5:       omp.body.continue:
3995 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
3996 // CHECK5:       omp.inner.for.inc:
3997 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
3998 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
3999 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4000 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
4001 // CHECK5:       omp.inner.for.end:
4002 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4003 // CHECK5:       omp.loop.exit:
4004 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4005 // CHECK5-NEXT:    ret void
4006 //
4007 //
4008 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
4009 // CHECK5-SAME: () #[[ATTR1]] {
4010 // CHECK5-NEXT:  entry:
4011 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
4012 // CHECK5-NEXT:    ret void
4013 //
4014 //
4015 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..12
4016 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
4017 // CHECK5-NEXT:  entry:
4018 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4019 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4020 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4021 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4022 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4023 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4024 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4025 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4026 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
4027 // CHECK5-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
4028 // CHECK5-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
4029 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4030 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4031 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4032 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4033 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4034 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4035 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4036 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
4037 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4038 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4039 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
4040 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4041 // CHECK5:       cond.true:
4042 // CHECK5-NEXT:    br label [[COND_END:%.*]]
4043 // CHECK5:       cond.false:
4044 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4045 // CHECK5-NEXT:    br label [[COND_END]]
4046 // CHECK5:       cond.end:
4047 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
4048 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
4049 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4050 // CHECK5-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
4051 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4052 // CHECK5:       omp.inner.for.cond:
4053 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4054 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4055 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
4056 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4057 // CHECK5:       omp.inner.for.body:
4058 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4059 // CHECK5-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
4060 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4061 // CHECK5-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
4062 // CHECK5-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
4063 // CHECK5-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4064 // CHECK5-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
4065 // CHECK5-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
4066 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4067 // CHECK5:       omp.inner.for.inc:
4068 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4069 // CHECK5-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
4070 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
4071 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
4072 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
4073 // CHECK5:       omp.inner.for.end:
4074 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4075 // CHECK5:       omp.loop.exit:
4076 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
4077 // CHECK5-NEXT:    ret void
4078 //
4079 //
4080 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..13
4081 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
4082 // CHECK5-NEXT:  entry:
4083 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4084 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4085 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
4086 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
4087 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4088 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4089 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
4090 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
4091 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4092 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4093 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
4094 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4095 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4096 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4097 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4098 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
4099 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
4100 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4101 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
4102 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4103 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
4104 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
4105 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
4106 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4107 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4108 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4109 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
4110 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4111 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4112 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
4113 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4114 // CHECK5:       cond.true:
4115 // CHECK5-NEXT:    br label [[COND_END:%.*]]
4116 // CHECK5:       cond.false:
4117 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4118 // CHECK5-NEXT:    br label [[COND_END]]
4119 // CHECK5:       cond.end:
4120 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
4121 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
4122 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
4123 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
4124 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4125 // CHECK5:       omp.inner.for.cond:
4126 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4127 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4128 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
4129 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4130 // CHECK5:       omp.inner.for.body:
4131 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4132 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
4133 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
4134 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
4135 // CHECK5-NEXT:    call void @_Z3fn2v()
4136 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
4137 // CHECK5:       omp.body.continue:
4138 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4139 // CHECK5:       omp.inner.for.inc:
4140 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4141 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
4142 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4143 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
4144 // CHECK5:       omp.inner.for.end:
4145 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4146 // CHECK5:       omp.loop.exit:
4147 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4148 // CHECK5-NEXT:    ret void
4149 //
4150 //
4151 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
4152 // CHECK5-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
4153 // CHECK5-NEXT:  entry:
4154 // CHECK5-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
4155 // CHECK5-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
4156 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
4157 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
4158 // CHECK5-NEXT:    ret void
4159 //
4160 //
4161 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..14
4162 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
4163 // CHECK5-NEXT:  entry:
4164 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4165 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4166 // CHECK5-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
4167 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4168 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4169 // CHECK5-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4170 // CHECK5-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4171 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4172 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4173 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
4174 // CHECK5-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
4175 // CHECK5-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
4176 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4177 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4178 // CHECK5-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
4179 // CHECK5-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
4180 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4181 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4182 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4183 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4184 // CHECK5-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4185 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
4186 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4187 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4188 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
4189 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4190 // CHECK5:       cond.true:
4191 // CHECK5-NEXT:    br label [[COND_END:%.*]]
4192 // CHECK5:       cond.false:
4193 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4194 // CHECK5-NEXT:    br label [[COND_END]]
4195 // CHECK5:       cond.end:
4196 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
4197 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
4198 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4199 // CHECK5-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
4200 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4201 // CHECK5:       omp.inner.for.cond:
4202 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4203 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4204 // CHECK5-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
4205 // CHECK5-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4206 // CHECK5:       omp.inner.for.body:
4207 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4208 // CHECK5-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
4209 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4210 // CHECK5-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
4211 // CHECK5-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
4212 // CHECK5-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
4213 // CHECK5-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
4214 // CHECK5:       omp_if.then:
4215 // CHECK5-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
4216 // CHECK5-NEXT:    br label [[OMP_IF_END:%.*]]
4217 // CHECK5:       omp_if.else:
4218 // CHECK5-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
4219 // CHECK5-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4220 // CHECK5-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
4221 // CHECK5-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
4222 // CHECK5-NEXT:    br label [[OMP_IF_END]]
4223 // CHECK5:       omp_if.end:
4224 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4225 // CHECK5:       omp.inner.for.inc:
4226 // CHECK5-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4227 // CHECK5-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
4228 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
4229 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
4230 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
4231 // CHECK5:       omp.inner.for.end:
4232 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4233 // CHECK5:       omp.loop.exit:
4234 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
4235 // CHECK5-NEXT:    ret void
4236 //
4237 //
4238 // CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..15
4239 // CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
4240 // CHECK5-NEXT:  entry:
4241 // CHECK5-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4242 // CHECK5-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4243 // CHECK5-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
4244 // CHECK5-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
4245 // CHECK5-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4246 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4247 // CHECK5-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
4248 // CHECK5-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
4249 // CHECK5-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4250 // CHECK5-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4251 // CHECK5-NEXT:    [[I:%.*]] = alloca i32, align 4
4252 // CHECK5-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4253 // CHECK5-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4254 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4255 // CHECK5-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4256 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
4257 // CHECK5-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
4258 // CHECK5-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4259 // CHECK5-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
4260 // CHECK5-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4261 // CHECK5-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
4262 // CHECK5-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
4263 // CHECK5-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
4264 // CHECK5-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4265 // CHECK5-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4266 // CHECK5-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4267 // CHECK5-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
4268 // CHECK5-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4269 // CHECK5-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4270 // CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
4271 // CHECK5-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4272 // CHECK5:       cond.true:
4273 // CHECK5-NEXT:    br label [[COND_END:%.*]]
4274 // CHECK5:       cond.false:
4275 // CHECK5-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4276 // CHECK5-NEXT:    br label [[COND_END]]
4277 // CHECK5:       cond.end:
4278 // CHECK5-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
4279 // CHECK5-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
4280 // CHECK5-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
4281 // CHECK5-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
4282 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4283 // CHECK5:       omp.inner.for.cond:
4284 // CHECK5-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4285 // CHECK5-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4286 // CHECK5-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
4287 // CHECK5-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4288 // CHECK5:       omp.inner.for.body:
4289 // CHECK5-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4290 // CHECK5-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
4291 // CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
4292 // CHECK5-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
4293 // CHECK5-NEXT:    call void @_Z3fn3v()
4294 // CHECK5-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
4295 // CHECK5:       omp.body.continue:
4296 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4297 // CHECK5:       omp.inner.for.inc:
4298 // CHECK5-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4299 // CHECK5-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
4300 // CHECK5-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4301 // CHECK5-NEXT:    br label [[OMP_INNER_FOR_COND]]
4302 // CHECK5:       omp.inner.for.end:
4303 // CHECK5-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4304 // CHECK5:       omp.loop.exit:
4305 // CHECK5-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4306 // CHECK5-NEXT:    ret void
4307 //
4308 //
4309 // CHECK5-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
4310 // CHECK5-SAME: () #[[ATTR5:[0-9]+]] {
4311 // CHECK5-NEXT:  entry:
4312 // CHECK5-NEXT:    call void @__tgt_register_requires(i64 1)
4313 // CHECK5-NEXT:    ret void
4314 //
4315 //
4316 // CHECK6-LABEL: define {{[^@]+}}@_Z9gtid_testv
4317 // CHECK6-SAME: () #[[ATTR0:[0-9]+]] {
4318 // CHECK6-NEXT:  entry:
4319 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4320 // CHECK6-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
4321 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
4322 // CHECK6-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
4323 // CHECK6-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
4324 // CHECK6-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
4325 // CHECK6:       omp_offload.failed:
4326 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
4327 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT]]
4328 // CHECK6:       omp_offload.cont:
4329 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
4330 // CHECK6-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
4331 // CHECK6-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
4332 // CHECK6-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
4333 // CHECK6:       omp_offload.failed2:
4334 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
4335 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
4336 // CHECK6:       omp_offload.cont3:
4337 // CHECK6-NEXT:    ret void
4338 //
4339 //
4340 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
4341 // CHECK6-SAME: () #[[ATTR1:[0-9]+]] {
4342 // CHECK6-NEXT:  entry:
4343 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
4344 // CHECK6-NEXT:    ret void
4345 //
4346 //
4347 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined.
4348 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
4349 // CHECK6-NEXT:  entry:
4350 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4351 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4352 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4353 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4354 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4355 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4356 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4357 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4358 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4359 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4360 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4361 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4362 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4363 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4364 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4365 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4366 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
4367 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4368 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4369 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
4370 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4371 // CHECK6:       cond.true:
4372 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4373 // CHECK6:       cond.false:
4374 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4375 // CHECK6-NEXT:    br label [[COND_END]]
4376 // CHECK6:       cond.end:
4377 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
4378 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
4379 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4380 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
4381 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4382 // CHECK6:       omp.inner.for.cond:
4383 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4384 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4385 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
4386 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4387 // CHECK6:       omp.inner.for.body:
4388 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4389 // CHECK6-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
4390 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4391 // CHECK6-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
4392 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
4393 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4394 // CHECK6:       omp.inner.for.inc:
4395 // CHECK6-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4396 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
4397 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
4398 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
4399 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4400 // CHECK6:       omp.inner.for.end:
4401 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4402 // CHECK6:       omp.loop.exit:
4403 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
4404 // CHECK6-NEXT:    ret void
4405 //
4406 //
4407 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..1
4408 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
4409 // CHECK6-NEXT:  entry:
4410 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4411 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4412 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
4413 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
4414 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4415 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4416 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
4417 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
4418 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4419 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4420 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4421 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4422 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4423 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4424 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4425 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
4426 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
4427 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4428 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
4429 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4430 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
4431 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
4432 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
4433 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4434 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4435 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4436 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
4437 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4438 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4439 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
4440 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4441 // CHECK6:       cond.true:
4442 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4443 // CHECK6:       cond.false:
4444 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4445 // CHECK6-NEXT:    br label [[COND_END]]
4446 // CHECK6:       cond.end:
4447 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
4448 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
4449 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
4450 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
4451 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4452 // CHECK6:       omp.inner.for.cond:
4453 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4454 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4455 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
4456 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4457 // CHECK6:       omp.inner.for.body:
4458 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4459 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
4460 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
4461 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
4462 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
4463 // CHECK6:       omp.body.continue:
4464 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4465 // CHECK6:       omp.inner.for.inc:
4466 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4467 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
4468 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4469 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4470 // CHECK6:       omp.inner.for.end:
4471 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4472 // CHECK6:       omp.loop.exit:
4473 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4474 // CHECK6-NEXT:    ret void
4475 //
4476 //
4477 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
4478 // CHECK6-SAME: () #[[ATTR1]] {
4479 // CHECK6-NEXT:  entry:
4480 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
4481 // CHECK6-NEXT:    ret void
4482 //
4483 //
4484 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..2
4485 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
4486 // CHECK6-NEXT:  entry:
4487 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4488 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4489 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4490 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4491 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4492 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4493 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4494 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4495 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4496 // CHECK6-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
4497 // CHECK6-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
4498 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4499 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4500 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4501 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4502 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4503 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4504 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4505 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
4506 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4507 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4508 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
4509 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4510 // CHECK6:       cond.true:
4511 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4512 // CHECK6:       cond.false:
4513 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4514 // CHECK6-NEXT:    br label [[COND_END]]
4515 // CHECK6:       cond.end:
4516 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
4517 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
4518 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4519 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
4520 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4521 // CHECK6:       omp.inner.for.cond:
4522 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4523 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4524 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
4525 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4526 // CHECK6:       omp.inner.for.body:
4527 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4528 // CHECK6-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
4529 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4530 // CHECK6-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
4531 // CHECK6-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
4532 // CHECK6-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4533 // CHECK6-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
4534 // CHECK6-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
4535 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4536 // CHECK6:       omp.inner.for.inc:
4537 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4538 // CHECK6-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
4539 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
4540 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
4541 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4542 // CHECK6:       omp.inner.for.end:
4543 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4544 // CHECK6:       omp.loop.exit:
4545 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
4546 // CHECK6-NEXT:    ret void
4547 //
4548 //
4549 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..3
4550 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
4551 // CHECK6-NEXT:  entry:
4552 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4553 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4554 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
4555 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
4556 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4557 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4558 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
4559 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
4560 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4561 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4562 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4563 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4564 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4565 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4566 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4567 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
4568 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
4569 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4570 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
4571 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4572 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
4573 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
4574 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
4575 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4576 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4577 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4578 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
4579 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4580 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4581 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
4582 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4583 // CHECK6:       cond.true:
4584 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4585 // CHECK6:       cond.false:
4586 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4587 // CHECK6-NEXT:    br label [[COND_END]]
4588 // CHECK6:       cond.end:
4589 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
4590 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
4591 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
4592 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
4593 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4594 // CHECK6:       omp.inner.for.cond:
4595 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4596 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4597 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
4598 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4599 // CHECK6:       omp.inner.for.body:
4600 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4601 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
4602 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
4603 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
4604 // CHECK6-NEXT:    call void @_Z9gtid_testv()
4605 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
4606 // CHECK6:       omp.body.continue:
4607 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4608 // CHECK6:       omp.inner.for.inc:
4609 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4610 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
4611 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4612 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4613 // CHECK6:       omp.inner.for.end:
4614 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4615 // CHECK6:       omp.loop.exit:
4616 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4617 // CHECK6-NEXT:    ret void
4618 //
4619 //
4620 // CHECK6-LABEL: define {{[^@]+}}@main
4621 // CHECK6-SAME: () #[[ATTR3:[0-9]+]] {
4622 // CHECK6-NEXT:  entry:
4623 // CHECK6-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
4624 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4625 // CHECK6-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
4626 // CHECK6-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
4627 // CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
4628 // CHECK6-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
4629 // CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
4630 // CHECK6-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
4631 // CHECK6-NEXT:    store i32 0, i32* [[RETVAL]], align 4
4632 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
4633 // CHECK6-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
4634 // CHECK6-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
4635 // CHECK6-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
4636 // CHECK6:       omp_offload.failed:
4637 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
4638 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT]]
4639 // CHECK6:       omp_offload.cont:
4640 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
4641 // CHECK6-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
4642 // CHECK6-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
4643 // CHECK6-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
4644 // CHECK6:       omp_offload.failed2:
4645 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
4646 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
4647 // CHECK6:       omp_offload.cont3:
4648 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
4649 // CHECK6-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
4650 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
4651 // CHECK6-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
4652 // CHECK6-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4653 // CHECK6-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
4654 // CHECK6-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
4655 // CHECK6-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4656 // CHECK6-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
4657 // CHECK6-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
4658 // CHECK6-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
4659 // CHECK6-NEXT:    store i8* null, i8** [[TMP10]], align 8
4660 // CHECK6-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
4661 // CHECK6-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
4662 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
4663 // CHECK6-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
4664 // CHECK6-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
4665 // CHECK6-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
4666 // CHECK6:       omp_offload.failed5:
4667 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
4668 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
4669 // CHECK6:       omp_offload.cont6:
4670 // CHECK6-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
4671 // CHECK6-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
4672 // CHECK6-NEXT:    ret i32 [[CALL]]
4673 //
4674 //
4675 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
4676 // CHECK6-SAME: () #[[ATTR1]] {
4677 // CHECK6-NEXT:  entry:
4678 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
4679 // CHECK6-NEXT:    ret void
4680 //
4681 //
4682 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..4
4683 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
4684 // CHECK6-NEXT:  entry:
4685 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4686 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4687 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4688 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4689 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4690 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4691 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4692 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4693 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4694 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4695 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4696 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4697 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4698 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4699 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4700 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4701 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
4702 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4703 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4704 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
4705 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4706 // CHECK6:       cond.true:
4707 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4708 // CHECK6:       cond.false:
4709 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4710 // CHECK6-NEXT:    br label [[COND_END]]
4711 // CHECK6:       cond.end:
4712 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
4713 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
4714 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4715 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
4716 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4717 // CHECK6:       omp.inner.for.cond:
4718 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4719 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4720 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
4721 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4722 // CHECK6:       omp.inner.for.body:
4723 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4724 // CHECK6-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
4725 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4726 // CHECK6-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
4727 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
4728 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4729 // CHECK6:       omp.inner.for.inc:
4730 // CHECK6-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4731 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
4732 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
4733 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
4734 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4735 // CHECK6:       omp.inner.for.end:
4736 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4737 // CHECK6:       omp.loop.exit:
4738 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
4739 // CHECK6-NEXT:    ret void
4740 //
4741 //
4742 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..5
4743 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
4744 // CHECK6-NEXT:  entry:
4745 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4746 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4747 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
4748 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
4749 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4750 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4751 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
4752 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
4753 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4754 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4755 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4756 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4757 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4758 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4759 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4760 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
4761 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
4762 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4763 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
4764 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4765 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
4766 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
4767 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
4768 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4769 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4770 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4771 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
4772 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4773 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4774 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
4775 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4776 // CHECK6:       cond.true:
4777 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4778 // CHECK6:       cond.false:
4779 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4780 // CHECK6-NEXT:    br label [[COND_END]]
4781 // CHECK6:       cond.end:
4782 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
4783 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
4784 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
4785 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
4786 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4787 // CHECK6:       omp.inner.for.cond:
4788 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4789 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4790 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
4791 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4792 // CHECK6:       omp.inner.for.body:
4793 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4794 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
4795 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
4796 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
4797 // CHECK6-NEXT:    call void @_Z3fn4v()
4798 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
4799 // CHECK6:       omp.body.continue:
4800 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4801 // CHECK6:       omp.inner.for.inc:
4802 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4803 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
4804 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4805 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4806 // CHECK6:       omp.inner.for.end:
4807 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4808 // CHECK6:       omp.loop.exit:
4809 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4810 // CHECK6-NEXT:    ret void
4811 //
4812 //
4813 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
4814 // CHECK6-SAME: () #[[ATTR1]] {
4815 // CHECK6-NEXT:  entry:
4816 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
4817 // CHECK6-NEXT:    ret void
4818 //
4819 //
4820 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..6
4821 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
4822 // CHECK6-NEXT:  entry:
4823 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4824 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4825 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4826 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4827 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4828 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4829 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4830 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4831 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4832 // CHECK6-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
4833 // CHECK6-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
4834 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4835 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4836 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4837 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4838 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4839 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4840 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4841 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
4842 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4843 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4844 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
4845 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4846 // CHECK6:       cond.true:
4847 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4848 // CHECK6:       cond.false:
4849 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4850 // CHECK6-NEXT:    br label [[COND_END]]
4851 // CHECK6:       cond.end:
4852 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
4853 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
4854 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4855 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
4856 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4857 // CHECK6:       omp.inner.for.cond:
4858 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4859 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4860 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
4861 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4862 // CHECK6:       omp.inner.for.body:
4863 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
4864 // CHECK6-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
4865 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4866 // CHECK6-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
4867 // CHECK6-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
4868 // CHECK6-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4869 // CHECK6-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
4870 // CHECK6-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
4871 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4872 // CHECK6:       omp.inner.for.inc:
4873 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4874 // CHECK6-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
4875 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
4876 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
4877 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4878 // CHECK6:       omp.inner.for.end:
4879 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4880 // CHECK6:       omp.loop.exit:
4881 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
4882 // CHECK6-NEXT:    ret void
4883 //
4884 //
4885 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..7
4886 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
4887 // CHECK6-NEXT:  entry:
4888 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4889 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4890 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
4891 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
4892 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4893 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4894 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
4895 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
4896 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4897 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4898 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4899 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4900 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4901 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4902 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4903 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
4904 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
4905 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
4906 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
4907 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
4908 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
4909 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
4910 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
4911 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4912 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4913 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4914 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
4915 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4916 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4917 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
4918 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4919 // CHECK6:       cond.true:
4920 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4921 // CHECK6:       cond.false:
4922 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4923 // CHECK6-NEXT:    br label [[COND_END]]
4924 // CHECK6:       cond.end:
4925 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
4926 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
4927 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
4928 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
4929 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
4930 // CHECK6:       omp.inner.for.cond:
4931 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4932 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
4933 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
4934 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
4935 // CHECK6:       omp.inner.for.body:
4936 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4937 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
4938 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
4939 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
4940 // CHECK6-NEXT:    call void @_Z3fn5v()
4941 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
4942 // CHECK6:       omp.body.continue:
4943 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
4944 // CHECK6:       omp.inner.for.inc:
4945 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
4946 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
4947 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
4948 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
4949 // CHECK6:       omp.inner.for.end:
4950 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
4951 // CHECK6:       omp.loop.exit:
4952 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
4953 // CHECK6-NEXT:    ret void
4954 //
4955 //
4956 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
4957 // CHECK6-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
4958 // CHECK6-NEXT:  entry:
4959 // CHECK6-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
4960 // CHECK6-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
4961 // CHECK6-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
4962 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
4963 // CHECK6-NEXT:    ret void
4964 //
4965 //
4966 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..8
4967 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
4968 // CHECK6-NEXT:  entry:
4969 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
4970 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
4971 // CHECK6-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
4972 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
4973 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
4974 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
4975 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
4976 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
4977 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
4978 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
4979 // CHECK6-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
4980 // CHECK6-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
4981 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
4982 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
4983 // CHECK6-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
4984 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
4985 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
4986 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
4987 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
4988 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
4989 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
4990 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
4991 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
4992 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4993 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
4994 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
4995 // CHECK6:       cond.true:
4996 // CHECK6-NEXT:    br label [[COND_END:%.*]]
4997 // CHECK6:       cond.false:
4998 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
4999 // CHECK6-NEXT:    br label [[COND_END]]
5000 // CHECK6:       cond.end:
5001 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
5002 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
5003 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5004 // CHECK6-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
5005 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5006 // CHECK6:       omp.inner.for.cond:
5007 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5008 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5009 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
5010 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5011 // CHECK6:       omp.inner.for.body:
5012 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5013 // CHECK6-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
5014 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5015 // CHECK6-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
5016 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
5017 // CHECK6-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
5018 // CHECK6-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
5019 // CHECK6:       omp_if.then:
5020 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
5021 // CHECK6-NEXT:    br label [[OMP_IF_END:%.*]]
5022 // CHECK6:       omp_if.else:
5023 // CHECK6-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
5024 // CHECK6-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5025 // CHECK6-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
5026 // CHECK6-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
5027 // CHECK6-NEXT:    br label [[OMP_IF_END]]
5028 // CHECK6:       omp_if.end:
5029 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5030 // CHECK6:       omp.inner.for.inc:
5031 // CHECK6-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5032 // CHECK6-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
5033 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
5034 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
5035 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5036 // CHECK6:       omp.inner.for.end:
5037 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5038 // CHECK6:       omp.loop.exit:
5039 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
5040 // CHECK6-NEXT:    ret void
5041 //
5042 //
5043 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..9
5044 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
5045 // CHECK6-NEXT:  entry:
5046 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5047 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5048 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
5049 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
5050 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5051 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5052 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
5053 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
5054 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5055 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5056 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5057 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5058 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5059 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5060 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5061 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
5062 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
5063 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5064 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
5065 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5066 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
5067 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
5068 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
5069 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5070 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5071 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5072 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
5073 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5074 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5075 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
5076 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5077 // CHECK6:       cond.true:
5078 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5079 // CHECK6:       cond.false:
5080 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5081 // CHECK6-NEXT:    br label [[COND_END]]
5082 // CHECK6:       cond.end:
5083 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
5084 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
5085 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
5086 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
5087 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5088 // CHECK6:       omp.inner.for.cond:
5089 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5090 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5091 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
5092 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5093 // CHECK6:       omp.inner.for.body:
5094 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5095 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
5096 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
5097 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
5098 // CHECK6-NEXT:    call void @_Z3fn6v()
5099 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
5100 // CHECK6:       omp.body.continue:
5101 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5102 // CHECK6:       omp.inner.for.inc:
5103 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5104 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
5105 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
5106 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5107 // CHECK6:       omp.inner.for.end:
5108 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5109 // CHECK6:       omp.loop.exit:
5110 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
5111 // CHECK6-NEXT:    ret void
5112 //
5113 //
5114 // CHECK6-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
5115 // CHECK6-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
5116 // CHECK6-NEXT:  entry:
5117 // CHECK6-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
5118 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5119 // CHECK6-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
5120 // CHECK6-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
5121 // CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
5122 // CHECK6-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
5123 // CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
5124 // CHECK6-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
5125 // CHECK6-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
5126 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
5127 // CHECK6-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
5128 // CHECK6-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
5129 // CHECK6-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
5130 // CHECK6:       omp_offload.failed:
5131 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
5132 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT]]
5133 // CHECK6:       omp_offload.cont:
5134 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
5135 // CHECK6-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
5136 // CHECK6-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
5137 // CHECK6-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
5138 // CHECK6:       omp_offload.failed2:
5139 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
5140 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
5141 // CHECK6:       omp_offload.cont3:
5142 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
5143 // CHECK6-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
5144 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
5145 // CHECK6-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
5146 // CHECK6-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5147 // CHECK6-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
5148 // CHECK6-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
5149 // CHECK6-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5150 // CHECK6-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
5151 // CHECK6-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
5152 // CHECK6-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
5153 // CHECK6-NEXT:    store i8* null, i8** [[TMP10]], align 8
5154 // CHECK6-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
5155 // CHECK6-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
5156 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
5157 // CHECK6-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
5158 // CHECK6-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
5159 // CHECK6-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
5160 // CHECK6:       omp_offload.failed5:
5161 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
5162 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
5163 // CHECK6:       omp_offload.cont6:
5164 // CHECK6-NEXT:    ret i32 0
5165 //
5166 //
5167 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
5168 // CHECK6-SAME: () #[[ATTR1]] {
5169 // CHECK6-NEXT:  entry:
5170 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
5171 // CHECK6-NEXT:    ret void
5172 //
5173 //
5174 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..10
5175 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
5176 // CHECK6-NEXT:  entry:
5177 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5178 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5179 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5180 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5181 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
5182 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
5183 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5184 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5185 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5186 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5187 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5188 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
5189 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
5190 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5191 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5192 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5193 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
5194 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5195 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5196 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
5197 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5198 // CHECK6:       cond.true:
5199 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5200 // CHECK6:       cond.false:
5201 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5202 // CHECK6-NEXT:    br label [[COND_END]]
5203 // CHECK6:       cond.end:
5204 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
5205 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
5206 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5207 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
5208 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5209 // CHECK6:       omp.inner.for.cond:
5210 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5211 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5212 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
5213 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5214 // CHECK6:       omp.inner.for.body:
5215 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5216 // CHECK6-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
5217 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5218 // CHECK6-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
5219 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
5220 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5221 // CHECK6:       omp.inner.for.inc:
5222 // CHECK6-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5223 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
5224 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
5225 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
5226 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5227 // CHECK6:       omp.inner.for.end:
5228 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5229 // CHECK6:       omp.loop.exit:
5230 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
5231 // CHECK6-NEXT:    ret void
5232 //
5233 //
5234 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..11
5235 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
5236 // CHECK6-NEXT:  entry:
5237 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5238 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5239 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
5240 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
5241 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5242 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5243 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
5244 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
5245 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5246 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5247 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5248 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5249 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5250 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5251 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5252 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
5253 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
5254 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5255 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
5256 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5257 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
5258 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
5259 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
5260 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5261 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5262 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5263 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
5264 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5265 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5266 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
5267 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5268 // CHECK6:       cond.true:
5269 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5270 // CHECK6:       cond.false:
5271 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5272 // CHECK6-NEXT:    br label [[COND_END]]
5273 // CHECK6:       cond.end:
5274 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
5275 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
5276 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
5277 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
5278 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5279 // CHECK6:       omp.inner.for.cond:
5280 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5281 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5282 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
5283 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5284 // CHECK6:       omp.inner.for.body:
5285 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5286 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
5287 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
5288 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
5289 // CHECK6-NEXT:    call void @_Z3fn1v()
5290 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
5291 // CHECK6:       omp.body.continue:
5292 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5293 // CHECK6:       omp.inner.for.inc:
5294 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5295 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
5296 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
5297 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5298 // CHECK6:       omp.inner.for.end:
5299 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5300 // CHECK6:       omp.loop.exit:
5301 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
5302 // CHECK6-NEXT:    ret void
5303 //
5304 //
5305 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
5306 // CHECK6-SAME: () #[[ATTR1]] {
5307 // CHECK6-NEXT:  entry:
5308 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
5309 // CHECK6-NEXT:    ret void
5310 //
5311 //
5312 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..12
5313 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
5314 // CHECK6-NEXT:  entry:
5315 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5316 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5317 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5318 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5319 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
5320 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
5321 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5322 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5323 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5324 // CHECK6-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
5325 // CHECK6-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
5326 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5327 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5328 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
5329 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
5330 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5331 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5332 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5333 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
5334 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5335 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5336 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
5337 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5338 // CHECK6:       cond.true:
5339 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5340 // CHECK6:       cond.false:
5341 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5342 // CHECK6-NEXT:    br label [[COND_END]]
5343 // CHECK6:       cond.end:
5344 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
5345 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
5346 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5347 // CHECK6-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
5348 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5349 // CHECK6:       omp.inner.for.cond:
5350 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5351 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5352 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
5353 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5354 // CHECK6:       omp.inner.for.body:
5355 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5356 // CHECK6-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
5357 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5358 // CHECK6-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
5359 // CHECK6-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
5360 // CHECK6-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5361 // CHECK6-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
5362 // CHECK6-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
5363 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5364 // CHECK6:       omp.inner.for.inc:
5365 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5366 // CHECK6-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
5367 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
5368 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
5369 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5370 // CHECK6:       omp.inner.for.end:
5371 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5372 // CHECK6:       omp.loop.exit:
5373 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
5374 // CHECK6-NEXT:    ret void
5375 //
5376 //
5377 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..13
5378 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
5379 // CHECK6-NEXT:  entry:
5380 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5381 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5382 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
5383 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
5384 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5385 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5386 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
5387 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
5388 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5389 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5390 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5391 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5392 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5393 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5394 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5395 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
5396 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
5397 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5398 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
5399 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5400 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
5401 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
5402 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
5403 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5404 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5405 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5406 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
5407 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5408 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5409 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
5410 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5411 // CHECK6:       cond.true:
5412 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5413 // CHECK6:       cond.false:
5414 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5415 // CHECK6-NEXT:    br label [[COND_END]]
5416 // CHECK6:       cond.end:
5417 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
5418 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
5419 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
5420 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
5421 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5422 // CHECK6:       omp.inner.for.cond:
5423 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5424 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5425 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
5426 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5427 // CHECK6:       omp.inner.for.body:
5428 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5429 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
5430 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
5431 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
5432 // CHECK6-NEXT:    call void @_Z3fn2v()
5433 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
5434 // CHECK6:       omp.body.continue:
5435 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5436 // CHECK6:       omp.inner.for.inc:
5437 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5438 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
5439 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
5440 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5441 // CHECK6:       omp.inner.for.end:
5442 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5443 // CHECK6:       omp.loop.exit:
5444 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
5445 // CHECK6-NEXT:    ret void
5446 //
5447 //
5448 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
5449 // CHECK6-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
5450 // CHECK6-NEXT:  entry:
5451 // CHECK6-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
5452 // CHECK6-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
5453 // CHECK6-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
5454 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
5455 // CHECK6-NEXT:    ret void
5456 //
5457 //
5458 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..14
5459 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
5460 // CHECK6-NEXT:  entry:
5461 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5462 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5463 // CHECK6-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
5464 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5465 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5466 // CHECK6-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
5467 // CHECK6-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
5468 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5469 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5470 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5471 // CHECK6-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
5472 // CHECK6-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
5473 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5474 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5475 // CHECK6-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
5476 // CHECK6-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
5477 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
5478 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
5479 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5480 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5481 // CHECK6-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5482 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
5483 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5484 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5485 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
5486 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5487 // CHECK6:       cond.true:
5488 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5489 // CHECK6:       cond.false:
5490 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5491 // CHECK6-NEXT:    br label [[COND_END]]
5492 // CHECK6:       cond.end:
5493 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
5494 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
5495 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5496 // CHECK6-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
5497 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5498 // CHECK6:       omp.inner.for.cond:
5499 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5500 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5501 // CHECK6-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
5502 // CHECK6-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5503 // CHECK6:       omp.inner.for.body:
5504 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5505 // CHECK6-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
5506 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5507 // CHECK6-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
5508 // CHECK6-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
5509 // CHECK6-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
5510 // CHECK6-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
5511 // CHECK6:       omp_if.then:
5512 // CHECK6-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
5513 // CHECK6-NEXT:    br label [[OMP_IF_END:%.*]]
5514 // CHECK6:       omp_if.else:
5515 // CHECK6-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
5516 // CHECK6-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5517 // CHECK6-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
5518 // CHECK6-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
5519 // CHECK6-NEXT:    br label [[OMP_IF_END]]
5520 // CHECK6:       omp_if.end:
5521 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5522 // CHECK6:       omp.inner.for.inc:
5523 // CHECK6-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5524 // CHECK6-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
5525 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
5526 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
5527 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5528 // CHECK6:       omp.inner.for.end:
5529 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5530 // CHECK6:       omp.loop.exit:
5531 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
5532 // CHECK6-NEXT:    ret void
5533 //
5534 //
5535 // CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..15
5536 // CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
5537 // CHECK6-NEXT:  entry:
5538 // CHECK6-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5539 // CHECK6-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5540 // CHECK6-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
5541 // CHECK6-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
5542 // CHECK6-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5543 // CHECK6-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5544 // CHECK6-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
5545 // CHECK6-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
5546 // CHECK6-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5547 // CHECK6-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5548 // CHECK6-NEXT:    [[I:%.*]] = alloca i32, align 4
5549 // CHECK6-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5550 // CHECK6-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5551 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5552 // CHECK6-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5553 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
5554 // CHECK6-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
5555 // CHECK6-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
5556 // CHECK6-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
5557 // CHECK6-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
5558 // CHECK6-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
5559 // CHECK6-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
5560 // CHECK6-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
5561 // CHECK6-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5562 // CHECK6-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5563 // CHECK6-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5564 // CHECK6-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
5565 // CHECK6-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5566 // CHECK6-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5567 // CHECK6-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
5568 // CHECK6-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5569 // CHECK6:       cond.true:
5570 // CHECK6-NEXT:    br label [[COND_END:%.*]]
5571 // CHECK6:       cond.false:
5572 // CHECK6-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5573 // CHECK6-NEXT:    br label [[COND_END]]
5574 // CHECK6:       cond.end:
5575 // CHECK6-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
5576 // CHECK6-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
5577 // CHECK6-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
5578 // CHECK6-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
5579 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5580 // CHECK6:       omp.inner.for.cond:
5581 // CHECK6-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5582 // CHECK6-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
5583 // CHECK6-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
5584 // CHECK6-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5585 // CHECK6:       omp.inner.for.body:
5586 // CHECK6-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5587 // CHECK6-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
5588 // CHECK6-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
5589 // CHECK6-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
5590 // CHECK6-NEXT:    call void @_Z3fn3v()
5591 // CHECK6-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
5592 // CHECK6:       omp.body.continue:
5593 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5594 // CHECK6:       omp.inner.for.inc:
5595 // CHECK6-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5596 // CHECK6-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
5597 // CHECK6-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
5598 // CHECK6-NEXT:    br label [[OMP_INNER_FOR_COND]]
5599 // CHECK6:       omp.inner.for.end:
5600 // CHECK6-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
5601 // CHECK6:       omp.loop.exit:
5602 // CHECK6-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
5603 // CHECK6-NEXT:    ret void
5604 //
5605 //
5606 // CHECK6-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
5607 // CHECK6-SAME: () #[[ATTR5:[0-9]+]] {
5608 // CHECK6-NEXT:  entry:
5609 // CHECK6-NEXT:    call void @__tgt_register_requires(i64 1)
5610 // CHECK6-NEXT:    ret void
5611 //
5612 //
5613 // CHECK7-LABEL: define {{[^@]+}}@_Z9gtid_testv
5614 // CHECK7-SAME: () #[[ATTR0:[0-9]+]] {
5615 // CHECK7-NEXT:  entry:
5616 // CHECK7-NEXT:    [[I:%.*]] = alloca i32, align 4
5617 // CHECK7-NEXT:    [[I1:%.*]] = alloca i32, align 4
5618 // CHECK7-NEXT:    store i32 0, i32* [[I]], align 4
5619 // CHECK7-NEXT:    br label [[FOR_COND:%.*]]
5620 // CHECK7:       for.cond:
5621 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
5622 // CHECK7-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
5623 // CHECK7-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
5624 // CHECK7:       for.body:
5625 // CHECK7-NEXT:    br label [[FOR_INC:%.*]]
5626 // CHECK7:       for.inc:
5627 // CHECK7-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
5628 // CHECK7-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
5629 // CHECK7-NEXT:    store i32 [[INC]], i32* [[I]], align 4
5630 // CHECK7-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
5631 // CHECK7:       for.end:
5632 // CHECK7-NEXT:    store i32 0, i32* [[I1]], align 4
5633 // CHECK7-NEXT:    br label [[FOR_COND2:%.*]]
5634 // CHECK7:       for.cond2:
5635 // CHECK7-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
5636 // CHECK7-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
5637 // CHECK7-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
5638 // CHECK7:       for.body4:
5639 // CHECK7-NEXT:    call void @_Z9gtid_testv()
5640 // CHECK7-NEXT:    br label [[FOR_INC5:%.*]]
5641 // CHECK7:       for.inc5:
5642 // CHECK7-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
5643 // CHECK7-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
5644 // CHECK7-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
5645 // CHECK7-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP4:![0-9]+]]
5646 // CHECK7:       for.end7:
5647 // CHECK7-NEXT:    ret void
5648 //
5649 //
5650 // CHECK7-LABEL: define {{[^@]+}}@main
5651 // CHECK7-SAME: () #[[ATTR1:[0-9]+]] {
5652 // CHECK7-NEXT:  entry:
5653 // CHECK7-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
5654 // CHECK7-NEXT:    [[I:%.*]] = alloca i32, align 4
5655 // CHECK7-NEXT:    [[I1:%.*]] = alloca i32, align 4
5656 // CHECK7-NEXT:    [[I8:%.*]] = alloca i32, align 4
5657 // CHECK7-NEXT:    store i32 0, i32* [[RETVAL]], align 4
5658 // CHECK7-NEXT:    store i32 0, i32* [[I]], align 4
5659 // CHECK7-NEXT:    br label [[FOR_COND:%.*]]
5660 // CHECK7:       for.cond:
5661 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
5662 // CHECK7-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
5663 // CHECK7-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
5664 // CHECK7:       for.body:
5665 // CHECK7-NEXT:    call void @_Z3fn4v()
5666 // CHECK7-NEXT:    br label [[FOR_INC:%.*]]
5667 // CHECK7:       for.inc:
5668 // CHECK7-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
5669 // CHECK7-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
5670 // CHECK7-NEXT:    store i32 [[INC]], i32* [[I]], align 4
5671 // CHECK7-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
5672 // CHECK7:       for.end:
5673 // CHECK7-NEXT:    store i32 0, i32* [[I1]], align 4
5674 // CHECK7-NEXT:    br label [[FOR_COND2:%.*]]
5675 // CHECK7:       for.cond2:
5676 // CHECK7-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
5677 // CHECK7-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
5678 // CHECK7-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
5679 // CHECK7:       for.body4:
5680 // CHECK7-NEXT:    call void @_Z3fn5v()
5681 // CHECK7-NEXT:    br label [[FOR_INC5:%.*]]
5682 // CHECK7:       for.inc5:
5683 // CHECK7-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
5684 // CHECK7-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
5685 // CHECK7-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
5686 // CHECK7-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP6:![0-9]+]]
5687 // CHECK7:       for.end7:
5688 // CHECK7-NEXT:    store i32 0, i32* [[I8]], align 4
5689 // CHECK7-NEXT:    br label [[FOR_COND9:%.*]]
5690 // CHECK7:       for.cond9:
5691 // CHECK7-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
5692 // CHECK7-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
5693 // CHECK7-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
5694 // CHECK7:       for.body11:
5695 // CHECK7-NEXT:    call void @_Z3fn6v()
5696 // CHECK7-NEXT:    br label [[FOR_INC12:%.*]]
5697 // CHECK7:       for.inc12:
5698 // CHECK7-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
5699 // CHECK7-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
5700 // CHECK7-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
5701 // CHECK7-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP7:![0-9]+]]
5702 // CHECK7:       for.end14:
5703 // CHECK7-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
5704 // CHECK7-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
5705 // CHECK7-NEXT:    ret i32 [[CALL]]
5706 //
5707 //
5708 // CHECK7-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
5709 // CHECK7-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
5710 // CHECK7-NEXT:  entry:
5711 // CHECK7-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
5712 // CHECK7-NEXT:    [[I:%.*]] = alloca i32, align 4
5713 // CHECK7-NEXT:    [[I1:%.*]] = alloca i32, align 4
5714 // CHECK7-NEXT:    [[I8:%.*]] = alloca i32, align 4
5715 // CHECK7-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
5716 // CHECK7-NEXT:    store i32 0, i32* [[I]], align 4
5717 // CHECK7-NEXT:    br label [[FOR_COND:%.*]]
5718 // CHECK7:       for.cond:
5719 // CHECK7-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
5720 // CHECK7-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
5721 // CHECK7-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
5722 // CHECK7:       for.body:
5723 // CHECK7-NEXT:    call void @_Z3fn1v()
5724 // CHECK7-NEXT:    br label [[FOR_INC:%.*]]
5725 // CHECK7:       for.inc:
5726 // CHECK7-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
5727 // CHECK7-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
5728 // CHECK7-NEXT:    store i32 [[INC]], i32* [[I]], align 4
5729 // CHECK7-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
5730 // CHECK7:       for.end:
5731 // CHECK7-NEXT:    store i32 0, i32* [[I1]], align 4
5732 // CHECK7-NEXT:    br label [[FOR_COND2:%.*]]
5733 // CHECK7:       for.cond2:
5734 // CHECK7-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
5735 // CHECK7-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
5736 // CHECK7-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
5737 // CHECK7:       for.body4:
5738 // CHECK7-NEXT:    call void @_Z3fn2v()
5739 // CHECK7-NEXT:    br label [[FOR_INC5:%.*]]
5740 // CHECK7:       for.inc5:
5741 // CHECK7-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
5742 // CHECK7-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
5743 // CHECK7-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
5744 // CHECK7-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP9:![0-9]+]]
5745 // CHECK7:       for.end7:
5746 // CHECK7-NEXT:    store i32 0, i32* [[I8]], align 4
5747 // CHECK7-NEXT:    br label [[FOR_COND9:%.*]]
5748 // CHECK7:       for.cond9:
5749 // CHECK7-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
5750 // CHECK7-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
5751 // CHECK7-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
5752 // CHECK7:       for.body11:
5753 // CHECK7-NEXT:    call void @_Z3fn3v()
5754 // CHECK7-NEXT:    br label [[FOR_INC12:%.*]]
5755 // CHECK7:       for.inc12:
5756 // CHECK7-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
5757 // CHECK7-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
5758 // CHECK7-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
5759 // CHECK7-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP10:![0-9]+]]
5760 // CHECK7:       for.end14:
5761 // CHECK7-NEXT:    ret i32 0
5762 //
5763 //
5764 // CHECK8-LABEL: define {{[^@]+}}@_Z9gtid_testv
5765 // CHECK8-SAME: () #[[ATTR0:[0-9]+]] {
5766 // CHECK8-NEXT:  entry:
5767 // CHECK8-NEXT:    [[I:%.*]] = alloca i32, align 4
5768 // CHECK8-NEXT:    [[I1:%.*]] = alloca i32, align 4
5769 // CHECK8-NEXT:    store i32 0, i32* [[I]], align 4
5770 // CHECK8-NEXT:    br label [[FOR_COND:%.*]]
5771 // CHECK8:       for.cond:
5772 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
5773 // CHECK8-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
5774 // CHECK8-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
5775 // CHECK8:       for.body:
5776 // CHECK8-NEXT:    br label [[FOR_INC:%.*]]
5777 // CHECK8:       for.inc:
5778 // CHECK8-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
5779 // CHECK8-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
5780 // CHECK8-NEXT:    store i32 [[INC]], i32* [[I]], align 4
5781 // CHECK8-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
5782 // CHECK8:       for.end:
5783 // CHECK8-NEXT:    store i32 0, i32* [[I1]], align 4
5784 // CHECK8-NEXT:    br label [[FOR_COND2:%.*]]
5785 // CHECK8:       for.cond2:
5786 // CHECK8-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
5787 // CHECK8-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
5788 // CHECK8-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
5789 // CHECK8:       for.body4:
5790 // CHECK8-NEXT:    call void @_Z9gtid_testv()
5791 // CHECK8-NEXT:    br label [[FOR_INC5:%.*]]
5792 // CHECK8:       for.inc5:
5793 // CHECK8-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
5794 // CHECK8-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
5795 // CHECK8-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
5796 // CHECK8-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP4:![0-9]+]]
5797 // CHECK8:       for.end7:
5798 // CHECK8-NEXT:    ret void
5799 //
5800 //
5801 // CHECK8-LABEL: define {{[^@]+}}@main
5802 // CHECK8-SAME: () #[[ATTR1:[0-9]+]] {
5803 // CHECK8-NEXT:  entry:
5804 // CHECK8-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
5805 // CHECK8-NEXT:    [[I:%.*]] = alloca i32, align 4
5806 // CHECK8-NEXT:    [[I1:%.*]] = alloca i32, align 4
5807 // CHECK8-NEXT:    [[I8:%.*]] = alloca i32, align 4
5808 // CHECK8-NEXT:    store i32 0, i32* [[RETVAL]], align 4
5809 // CHECK8-NEXT:    store i32 0, i32* [[I]], align 4
5810 // CHECK8-NEXT:    br label [[FOR_COND:%.*]]
5811 // CHECK8:       for.cond:
5812 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
5813 // CHECK8-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
5814 // CHECK8-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
5815 // CHECK8:       for.body:
5816 // CHECK8-NEXT:    call void @_Z3fn4v()
5817 // CHECK8-NEXT:    br label [[FOR_INC:%.*]]
5818 // CHECK8:       for.inc:
5819 // CHECK8-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
5820 // CHECK8-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
5821 // CHECK8-NEXT:    store i32 [[INC]], i32* [[I]], align 4
5822 // CHECK8-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
5823 // CHECK8:       for.end:
5824 // CHECK8-NEXT:    store i32 0, i32* [[I1]], align 4
5825 // CHECK8-NEXT:    br label [[FOR_COND2:%.*]]
5826 // CHECK8:       for.cond2:
5827 // CHECK8-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
5828 // CHECK8-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
5829 // CHECK8-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
5830 // CHECK8:       for.body4:
5831 // CHECK8-NEXT:    call void @_Z3fn5v()
5832 // CHECK8-NEXT:    br label [[FOR_INC5:%.*]]
5833 // CHECK8:       for.inc5:
5834 // CHECK8-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
5835 // CHECK8-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
5836 // CHECK8-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
5837 // CHECK8-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP6:![0-9]+]]
5838 // CHECK8:       for.end7:
5839 // CHECK8-NEXT:    store i32 0, i32* [[I8]], align 4
5840 // CHECK8-NEXT:    br label [[FOR_COND9:%.*]]
5841 // CHECK8:       for.cond9:
5842 // CHECK8-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
5843 // CHECK8-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
5844 // CHECK8-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
5845 // CHECK8:       for.body11:
5846 // CHECK8-NEXT:    call void @_Z3fn6v()
5847 // CHECK8-NEXT:    br label [[FOR_INC12:%.*]]
5848 // CHECK8:       for.inc12:
5849 // CHECK8-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
5850 // CHECK8-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
5851 // CHECK8-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
5852 // CHECK8-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP7:![0-9]+]]
5853 // CHECK8:       for.end14:
5854 // CHECK8-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
5855 // CHECK8-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
5856 // CHECK8-NEXT:    ret i32 [[CALL]]
5857 //
5858 //
5859 // CHECK8-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
5860 // CHECK8-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
5861 // CHECK8-NEXT:  entry:
5862 // CHECK8-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
5863 // CHECK8-NEXT:    [[I:%.*]] = alloca i32, align 4
5864 // CHECK8-NEXT:    [[I1:%.*]] = alloca i32, align 4
5865 // CHECK8-NEXT:    [[I8:%.*]] = alloca i32, align 4
5866 // CHECK8-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
5867 // CHECK8-NEXT:    store i32 0, i32* [[I]], align 4
5868 // CHECK8-NEXT:    br label [[FOR_COND:%.*]]
5869 // CHECK8:       for.cond:
5870 // CHECK8-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
5871 // CHECK8-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
5872 // CHECK8-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
5873 // CHECK8:       for.body:
5874 // CHECK8-NEXT:    call void @_Z3fn1v()
5875 // CHECK8-NEXT:    br label [[FOR_INC:%.*]]
5876 // CHECK8:       for.inc:
5877 // CHECK8-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
5878 // CHECK8-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
5879 // CHECK8-NEXT:    store i32 [[INC]], i32* [[I]], align 4
5880 // CHECK8-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
5881 // CHECK8:       for.end:
5882 // CHECK8-NEXT:    store i32 0, i32* [[I1]], align 4
5883 // CHECK8-NEXT:    br label [[FOR_COND2:%.*]]
5884 // CHECK8:       for.cond2:
5885 // CHECK8-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
5886 // CHECK8-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
5887 // CHECK8-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
5888 // CHECK8:       for.body4:
5889 // CHECK8-NEXT:    call void @_Z3fn2v()
5890 // CHECK8-NEXT:    br label [[FOR_INC5:%.*]]
5891 // CHECK8:       for.inc5:
5892 // CHECK8-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
5893 // CHECK8-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
5894 // CHECK8-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
5895 // CHECK8-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP9:![0-9]+]]
5896 // CHECK8:       for.end7:
5897 // CHECK8-NEXT:    store i32 0, i32* [[I8]], align 4
5898 // CHECK8-NEXT:    br label [[FOR_COND9:%.*]]
5899 // CHECK8:       for.cond9:
5900 // CHECK8-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
5901 // CHECK8-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
5902 // CHECK8-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
5903 // CHECK8:       for.body11:
5904 // CHECK8-NEXT:    call void @_Z3fn3v()
5905 // CHECK8-NEXT:    br label [[FOR_INC12:%.*]]
5906 // CHECK8:       for.inc12:
5907 // CHECK8-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
5908 // CHECK8-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
5909 // CHECK8-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
5910 // CHECK8-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP10:![0-9]+]]
5911 // CHECK8:       for.end14:
5912 // CHECK8-NEXT:    ret i32 0
5913 //
5914 //
5915 // CHECK9-LABEL: define {{[^@]+}}@_Z9gtid_testv
5916 // CHECK9-SAME: () #[[ATTR0:[0-9]+]] {
5917 // CHECK9-NEXT:  entry:
5918 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5919 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
5920 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
5921 // CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
5922 // CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
5923 // CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
5924 // CHECK9:       omp_offload.failed:
5925 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
5926 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
5927 // CHECK9:       omp_offload.cont:
5928 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
5929 // CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
5930 // CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
5931 // CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
5932 // CHECK9:       omp_offload.failed2:
5933 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
5934 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
5935 // CHECK9:       omp_offload.cont3:
5936 // CHECK9-NEXT:    ret void
5937 //
5938 //
5939 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
5940 // CHECK9-SAME: () #[[ATTR1:[0-9]+]] {
5941 // CHECK9-NEXT:  entry:
5942 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
5943 // CHECK9-NEXT:    ret void
5944 //
5945 //
5946 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined.
5947 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
5948 // CHECK9-NEXT:  entry:
5949 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
5950 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
5951 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
5952 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
5953 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
5954 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
5955 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
5956 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
5957 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
5958 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
5959 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
5960 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
5961 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
5962 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
5963 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
5964 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
5965 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
5966 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
5967 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5968 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
5969 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
5970 // CHECK9:       cond.true:
5971 // CHECK9-NEXT:    br label [[COND_END:%.*]]
5972 // CHECK9:       cond.false:
5973 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5974 // CHECK9-NEXT:    br label [[COND_END]]
5975 // CHECK9:       cond.end:
5976 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
5977 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
5978 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5979 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
5980 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
5981 // CHECK9:       omp.inner.for.cond:
5982 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5983 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5984 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
5985 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
5986 // CHECK9:       omp.inner.for.body:
5987 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
5988 // CHECK9-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
5989 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
5990 // CHECK9-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
5991 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
5992 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
5993 // CHECK9:       omp.inner.for.inc:
5994 // CHECK9-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
5995 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
5996 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
5997 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
5998 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
5999 // CHECK9:       omp.inner.for.end:
6000 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6001 // CHECK9:       omp.loop.exit:
6002 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
6003 // CHECK9-NEXT:    ret void
6004 //
6005 //
6006 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..1
6007 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6008 // CHECK9-NEXT:  entry:
6009 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6010 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6011 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6012 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6013 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6014 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6015 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6016 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6017 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6018 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6019 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6020 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6021 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6022 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6023 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6024 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6025 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6026 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6027 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6028 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6029 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
6030 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
6031 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
6032 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6033 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6034 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6035 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
6036 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6037 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6038 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
6039 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6040 // CHECK9:       cond.true:
6041 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6042 // CHECK9:       cond.false:
6043 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6044 // CHECK9-NEXT:    br label [[COND_END]]
6045 // CHECK9:       cond.end:
6046 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
6047 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
6048 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
6049 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
6050 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6051 // CHECK9:       omp.inner.for.cond:
6052 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6053 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6054 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
6055 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6056 // CHECK9:       omp.inner.for.body:
6057 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6058 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
6059 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
6060 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
6061 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
6062 // CHECK9:       omp.body.continue:
6063 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6064 // CHECK9:       omp.inner.for.inc:
6065 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6066 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
6067 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
6068 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6069 // CHECK9:       omp.inner.for.end:
6070 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6071 // CHECK9:       omp.loop.exit:
6072 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
6073 // CHECK9-NEXT:    ret void
6074 //
6075 //
6076 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
6077 // CHECK9-SAME: () #[[ATTR1]] {
6078 // CHECK9-NEXT:  entry:
6079 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
6080 // CHECK9-NEXT:    ret void
6081 //
6082 //
6083 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..2
6084 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
6085 // CHECK9-NEXT:  entry:
6086 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6087 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6088 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6089 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6090 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
6091 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
6092 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6093 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6094 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6095 // CHECK9-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
6096 // CHECK9-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
6097 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6098 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6099 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
6100 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
6101 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6102 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6103 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6104 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
6105 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6106 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6107 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
6108 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6109 // CHECK9:       cond.true:
6110 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6111 // CHECK9:       cond.false:
6112 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6113 // CHECK9-NEXT:    br label [[COND_END]]
6114 // CHECK9:       cond.end:
6115 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
6116 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
6117 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6118 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
6119 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6120 // CHECK9:       omp.inner.for.cond:
6121 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6122 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6123 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
6124 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6125 // CHECK9:       omp.inner.for.body:
6126 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6127 // CHECK9-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
6128 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6129 // CHECK9-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
6130 // CHECK9-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
6131 // CHECK9-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6132 // CHECK9-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
6133 // CHECK9-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
6134 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6135 // CHECK9:       omp.inner.for.inc:
6136 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6137 // CHECK9-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
6138 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
6139 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
6140 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6141 // CHECK9:       omp.inner.for.end:
6142 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6143 // CHECK9:       omp.loop.exit:
6144 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
6145 // CHECK9-NEXT:    ret void
6146 //
6147 //
6148 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..3
6149 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6150 // CHECK9-NEXT:  entry:
6151 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6152 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6153 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6154 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6155 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6156 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6157 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6158 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6159 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6160 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6161 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6162 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6163 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6164 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6165 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6166 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6167 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6168 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6169 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6170 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6171 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
6172 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
6173 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
6174 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6175 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6176 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6177 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
6178 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6179 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6180 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
6181 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6182 // CHECK9:       cond.true:
6183 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6184 // CHECK9:       cond.false:
6185 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6186 // CHECK9-NEXT:    br label [[COND_END]]
6187 // CHECK9:       cond.end:
6188 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
6189 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
6190 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
6191 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
6192 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6193 // CHECK9:       omp.inner.for.cond:
6194 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6195 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6196 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
6197 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6198 // CHECK9:       omp.inner.for.body:
6199 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6200 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
6201 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
6202 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
6203 // CHECK9-NEXT:    call void @_Z9gtid_testv()
6204 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
6205 // CHECK9:       omp.body.continue:
6206 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6207 // CHECK9:       omp.inner.for.inc:
6208 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6209 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
6210 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
6211 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6212 // CHECK9:       omp.inner.for.end:
6213 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6214 // CHECK9:       omp.loop.exit:
6215 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
6216 // CHECK9-NEXT:    ret void
6217 //
6218 //
6219 // CHECK9-LABEL: define {{[^@]+}}@main
6220 // CHECK9-SAME: () #[[ATTR3:[0-9]+]] {
6221 // CHECK9-NEXT:  entry:
6222 // CHECK9-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
6223 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6224 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
6225 // CHECK9-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
6226 // CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
6227 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
6228 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
6229 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
6230 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
6231 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
6232 // CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
6233 // CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
6234 // CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
6235 // CHECK9:       omp_offload.failed:
6236 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
6237 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
6238 // CHECK9:       omp_offload.cont:
6239 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
6240 // CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
6241 // CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
6242 // CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
6243 // CHECK9:       omp_offload.failed2:
6244 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
6245 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
6246 // CHECK9:       omp_offload.cont3:
6247 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
6248 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
6249 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
6250 // CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
6251 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
6252 // CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
6253 // CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
6254 // CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
6255 // CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
6256 // CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
6257 // CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
6258 // CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
6259 // CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
6260 // CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
6261 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
6262 // CHECK9-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
6263 // CHECK9-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
6264 // CHECK9-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
6265 // CHECK9:       omp_offload.failed5:
6266 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
6267 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
6268 // CHECK9:       omp_offload.cont6:
6269 // CHECK9-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
6270 // CHECK9-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
6271 // CHECK9-NEXT:    ret i32 [[CALL]]
6272 //
6273 //
6274 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
6275 // CHECK9-SAME: () #[[ATTR1]] {
6276 // CHECK9-NEXT:  entry:
6277 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
6278 // CHECK9-NEXT:    ret void
6279 //
6280 //
6281 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..4
6282 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
6283 // CHECK9-NEXT:  entry:
6284 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6285 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6286 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6287 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6288 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
6289 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
6290 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6291 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6292 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6293 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6294 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6295 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
6296 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
6297 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6298 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6299 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6300 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
6301 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6302 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6303 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
6304 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6305 // CHECK9:       cond.true:
6306 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6307 // CHECK9:       cond.false:
6308 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6309 // CHECK9-NEXT:    br label [[COND_END]]
6310 // CHECK9:       cond.end:
6311 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
6312 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
6313 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6314 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
6315 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6316 // CHECK9:       omp.inner.for.cond:
6317 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6318 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6319 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
6320 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6321 // CHECK9:       omp.inner.for.body:
6322 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6323 // CHECK9-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
6324 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6325 // CHECK9-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
6326 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
6327 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6328 // CHECK9:       omp.inner.for.inc:
6329 // CHECK9-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6330 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
6331 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
6332 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
6333 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6334 // CHECK9:       omp.inner.for.end:
6335 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6336 // CHECK9:       omp.loop.exit:
6337 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
6338 // CHECK9-NEXT:    ret void
6339 //
6340 //
6341 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..5
6342 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6343 // CHECK9-NEXT:  entry:
6344 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6345 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6346 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6347 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6348 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6349 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6350 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6351 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6352 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6353 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6354 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6355 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6356 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6357 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6358 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6359 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6360 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6361 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6362 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6363 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6364 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
6365 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
6366 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
6367 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6368 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6369 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6370 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
6371 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6372 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6373 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
6374 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6375 // CHECK9:       cond.true:
6376 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6377 // CHECK9:       cond.false:
6378 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6379 // CHECK9-NEXT:    br label [[COND_END]]
6380 // CHECK9:       cond.end:
6381 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
6382 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
6383 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
6384 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
6385 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6386 // CHECK9:       omp.inner.for.cond:
6387 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6388 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6389 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
6390 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6391 // CHECK9:       omp.inner.for.body:
6392 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6393 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
6394 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
6395 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
6396 // CHECK9-NEXT:    call void @_Z3fn4v()
6397 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
6398 // CHECK9:       omp.body.continue:
6399 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6400 // CHECK9:       omp.inner.for.inc:
6401 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6402 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
6403 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
6404 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6405 // CHECK9:       omp.inner.for.end:
6406 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6407 // CHECK9:       omp.loop.exit:
6408 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
6409 // CHECK9-NEXT:    ret void
6410 //
6411 //
6412 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
6413 // CHECK9-SAME: () #[[ATTR1]] {
6414 // CHECK9-NEXT:  entry:
6415 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
6416 // CHECK9-NEXT:    ret void
6417 //
6418 //
6419 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..6
6420 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
6421 // CHECK9-NEXT:  entry:
6422 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6423 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6424 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6425 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6426 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
6427 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
6428 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6429 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6430 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6431 // CHECK9-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
6432 // CHECK9-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
6433 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6434 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6435 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
6436 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
6437 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6438 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6439 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6440 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
6441 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6442 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6443 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
6444 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6445 // CHECK9:       cond.true:
6446 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6447 // CHECK9:       cond.false:
6448 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6449 // CHECK9-NEXT:    br label [[COND_END]]
6450 // CHECK9:       cond.end:
6451 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
6452 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
6453 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6454 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
6455 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6456 // CHECK9:       omp.inner.for.cond:
6457 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6458 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6459 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
6460 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6461 // CHECK9:       omp.inner.for.body:
6462 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6463 // CHECK9-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
6464 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6465 // CHECK9-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
6466 // CHECK9-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
6467 // CHECK9-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6468 // CHECK9-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
6469 // CHECK9-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
6470 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6471 // CHECK9:       omp.inner.for.inc:
6472 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6473 // CHECK9-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
6474 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
6475 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
6476 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6477 // CHECK9:       omp.inner.for.end:
6478 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6479 // CHECK9:       omp.loop.exit:
6480 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
6481 // CHECK9-NEXT:    ret void
6482 //
6483 //
6484 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..7
6485 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6486 // CHECK9-NEXT:  entry:
6487 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6488 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6489 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6490 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6491 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6492 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6493 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6494 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6495 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6496 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6497 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6498 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6499 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6500 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6501 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6502 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6503 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6504 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6505 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6506 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6507 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
6508 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
6509 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
6510 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6511 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6512 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6513 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
6514 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6515 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6516 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
6517 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6518 // CHECK9:       cond.true:
6519 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6520 // CHECK9:       cond.false:
6521 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6522 // CHECK9-NEXT:    br label [[COND_END]]
6523 // CHECK9:       cond.end:
6524 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
6525 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
6526 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
6527 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
6528 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6529 // CHECK9:       omp.inner.for.cond:
6530 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6531 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6532 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
6533 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6534 // CHECK9:       omp.inner.for.body:
6535 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6536 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
6537 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
6538 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
6539 // CHECK9-NEXT:    call void @_Z3fn5v()
6540 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
6541 // CHECK9:       omp.body.continue:
6542 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6543 // CHECK9:       omp.inner.for.inc:
6544 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6545 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
6546 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
6547 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6548 // CHECK9:       omp.inner.for.end:
6549 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6550 // CHECK9:       omp.loop.exit:
6551 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
6552 // CHECK9-NEXT:    ret void
6553 //
6554 //
6555 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
6556 // CHECK9-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
6557 // CHECK9-NEXT:  entry:
6558 // CHECK9-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
6559 // CHECK9-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
6560 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
6561 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
6562 // CHECK9-NEXT:    ret void
6563 //
6564 //
6565 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..8
6566 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
6567 // CHECK9-NEXT:  entry:
6568 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6569 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6570 // CHECK9-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
6571 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6572 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6573 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
6574 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
6575 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6576 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6577 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6578 // CHECK9-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
6579 // CHECK9-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
6580 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6581 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6582 // CHECK9-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
6583 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
6584 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
6585 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
6586 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6587 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6588 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6589 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
6590 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6591 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6592 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
6593 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6594 // CHECK9:       cond.true:
6595 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6596 // CHECK9:       cond.false:
6597 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6598 // CHECK9-NEXT:    br label [[COND_END]]
6599 // CHECK9:       cond.end:
6600 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
6601 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
6602 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6603 // CHECK9-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
6604 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6605 // CHECK9:       omp.inner.for.cond:
6606 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6607 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6608 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
6609 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6610 // CHECK9:       omp.inner.for.body:
6611 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6612 // CHECK9-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
6613 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6614 // CHECK9-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
6615 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
6616 // CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
6617 // CHECK9-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
6618 // CHECK9:       omp_if.then:
6619 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
6620 // CHECK9-NEXT:    br label [[OMP_IF_END:%.*]]
6621 // CHECK9:       omp_if.else:
6622 // CHECK9-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
6623 // CHECK9-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6624 // CHECK9-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
6625 // CHECK9-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
6626 // CHECK9-NEXT:    br label [[OMP_IF_END]]
6627 // CHECK9:       omp_if.end:
6628 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6629 // CHECK9:       omp.inner.for.inc:
6630 // CHECK9-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6631 // CHECK9-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
6632 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
6633 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
6634 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6635 // CHECK9:       omp.inner.for.end:
6636 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6637 // CHECK9:       omp.loop.exit:
6638 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
6639 // CHECK9-NEXT:    ret void
6640 //
6641 //
6642 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..9
6643 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6644 // CHECK9-NEXT:  entry:
6645 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6646 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6647 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6648 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6649 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6650 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6651 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6652 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6653 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6654 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6655 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6656 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6657 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6658 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6659 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6660 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6661 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6662 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6663 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6664 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6665 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
6666 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
6667 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
6668 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6669 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6670 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6671 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
6672 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6673 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6674 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
6675 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6676 // CHECK9:       cond.true:
6677 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6678 // CHECK9:       cond.false:
6679 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6680 // CHECK9-NEXT:    br label [[COND_END]]
6681 // CHECK9:       cond.end:
6682 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
6683 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
6684 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
6685 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
6686 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6687 // CHECK9:       omp.inner.for.cond:
6688 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6689 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6690 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
6691 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6692 // CHECK9:       omp.inner.for.body:
6693 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6694 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
6695 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
6696 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
6697 // CHECK9-NEXT:    call void @_Z3fn6v()
6698 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
6699 // CHECK9:       omp.body.continue:
6700 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6701 // CHECK9:       omp.inner.for.inc:
6702 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6703 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
6704 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
6705 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6706 // CHECK9:       omp.inner.for.end:
6707 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6708 // CHECK9:       omp.loop.exit:
6709 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
6710 // CHECK9-NEXT:    ret void
6711 //
6712 //
6713 // CHECK9-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
6714 // CHECK9-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
6715 // CHECK9-NEXT:  entry:
6716 // CHECK9-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
6717 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6718 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
6719 // CHECK9-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
6720 // CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
6721 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
6722 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
6723 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
6724 // CHECK9-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
6725 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
6726 // CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
6727 // CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
6728 // CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
6729 // CHECK9:       omp_offload.failed:
6730 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
6731 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
6732 // CHECK9:       omp_offload.cont:
6733 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
6734 // CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
6735 // CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
6736 // CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
6737 // CHECK9:       omp_offload.failed2:
6738 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
6739 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
6740 // CHECK9:       omp_offload.cont3:
6741 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
6742 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
6743 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
6744 // CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
6745 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
6746 // CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
6747 // CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
6748 // CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
6749 // CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
6750 // CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
6751 // CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
6752 // CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
6753 // CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
6754 // CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
6755 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
6756 // CHECK9-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
6757 // CHECK9-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
6758 // CHECK9-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
6759 // CHECK9:       omp_offload.failed5:
6760 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
6761 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
6762 // CHECK9:       omp_offload.cont6:
6763 // CHECK9-NEXT:    ret i32 0
6764 //
6765 //
6766 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
6767 // CHECK9-SAME: () #[[ATTR1]] {
6768 // CHECK9-NEXT:  entry:
6769 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
6770 // CHECK9-NEXT:    ret void
6771 //
6772 //
6773 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..10
6774 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
6775 // CHECK9-NEXT:  entry:
6776 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6777 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6778 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6779 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6780 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
6781 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
6782 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6783 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6784 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6785 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6786 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6787 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
6788 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
6789 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6790 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6791 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6792 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
6793 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6794 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6795 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
6796 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6797 // CHECK9:       cond.true:
6798 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6799 // CHECK9:       cond.false:
6800 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6801 // CHECK9-NEXT:    br label [[COND_END]]
6802 // CHECK9:       cond.end:
6803 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
6804 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
6805 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6806 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
6807 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6808 // CHECK9:       omp.inner.for.cond:
6809 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6810 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6811 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
6812 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6813 // CHECK9:       omp.inner.for.body:
6814 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6815 // CHECK9-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
6816 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6817 // CHECK9-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
6818 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
6819 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6820 // CHECK9:       omp.inner.for.inc:
6821 // CHECK9-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6822 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
6823 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
6824 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
6825 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6826 // CHECK9:       omp.inner.for.end:
6827 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6828 // CHECK9:       omp.loop.exit:
6829 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
6830 // CHECK9-NEXT:    ret void
6831 //
6832 //
6833 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..11
6834 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6835 // CHECK9-NEXT:  entry:
6836 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6837 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6838 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6839 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6840 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6841 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6842 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6843 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6844 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6845 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6846 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6847 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6848 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6849 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6850 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6851 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6852 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6853 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6854 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6855 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6856 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
6857 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
6858 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
6859 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6860 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6861 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6862 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
6863 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6864 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6865 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
6866 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6867 // CHECK9:       cond.true:
6868 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6869 // CHECK9:       cond.false:
6870 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6871 // CHECK9-NEXT:    br label [[COND_END]]
6872 // CHECK9:       cond.end:
6873 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
6874 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
6875 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
6876 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
6877 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6878 // CHECK9:       omp.inner.for.cond:
6879 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6880 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
6881 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
6882 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6883 // CHECK9:       omp.inner.for.body:
6884 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6885 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
6886 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
6887 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
6888 // CHECK9-NEXT:    call void @_Z3fn1v()
6889 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
6890 // CHECK9:       omp.body.continue:
6891 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6892 // CHECK9:       omp.inner.for.inc:
6893 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6894 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
6895 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
6896 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6897 // CHECK9:       omp.inner.for.end:
6898 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6899 // CHECK9:       omp.loop.exit:
6900 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
6901 // CHECK9-NEXT:    ret void
6902 //
6903 //
6904 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
6905 // CHECK9-SAME: () #[[ATTR1]] {
6906 // CHECK9-NEXT:  entry:
6907 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
6908 // CHECK9-NEXT:    ret void
6909 //
6910 //
6911 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..12
6912 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
6913 // CHECK9-NEXT:  entry:
6914 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6915 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6916 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6917 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6918 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
6919 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
6920 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6921 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6922 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6923 // CHECK9-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
6924 // CHECK9-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
6925 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6926 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6927 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
6928 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
6929 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
6930 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
6931 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6932 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
6933 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
6934 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6935 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
6936 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
6937 // CHECK9:       cond.true:
6938 // CHECK9-NEXT:    br label [[COND_END:%.*]]
6939 // CHECK9:       cond.false:
6940 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6941 // CHECK9-NEXT:    br label [[COND_END]]
6942 // CHECK9:       cond.end:
6943 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
6944 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
6945 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6946 // CHECK9-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
6947 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
6948 // CHECK9:       omp.inner.for.cond:
6949 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6950 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6951 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
6952 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
6953 // CHECK9:       omp.inner.for.body:
6954 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
6955 // CHECK9-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
6956 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
6957 // CHECK9-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
6958 // CHECK9-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
6959 // CHECK9-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
6960 // CHECK9-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
6961 // CHECK9-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
6962 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
6963 // CHECK9:       omp.inner.for.inc:
6964 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
6965 // CHECK9-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
6966 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
6967 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
6968 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
6969 // CHECK9:       omp.inner.for.end:
6970 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
6971 // CHECK9:       omp.loop.exit:
6972 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
6973 // CHECK9-NEXT:    ret void
6974 //
6975 //
6976 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..13
6977 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
6978 // CHECK9-NEXT:  entry:
6979 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
6980 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
6981 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
6982 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
6983 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
6984 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
6985 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
6986 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
6987 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
6988 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
6989 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
6990 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
6991 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
6992 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6993 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6994 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
6995 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
6996 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
6997 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
6998 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
6999 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7000 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7001 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7002 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7003 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7004 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7005 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7006 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7007 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7008 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7009 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7010 // CHECK9:       cond.true:
7011 // CHECK9-NEXT:    br label [[COND_END:%.*]]
7012 // CHECK9:       cond.false:
7013 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7014 // CHECK9-NEXT:    br label [[COND_END]]
7015 // CHECK9:       cond.end:
7016 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7017 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7018 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7019 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7020 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7021 // CHECK9:       omp.inner.for.cond:
7022 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7023 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7024 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7025 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7026 // CHECK9:       omp.inner.for.body:
7027 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7028 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7029 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7030 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7031 // CHECK9-NEXT:    call void @_Z3fn2v()
7032 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7033 // CHECK9:       omp.body.continue:
7034 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7035 // CHECK9:       omp.inner.for.inc:
7036 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7037 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
7038 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
7039 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
7040 // CHECK9:       omp.inner.for.end:
7041 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7042 // CHECK9:       omp.loop.exit:
7043 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
7044 // CHECK9-NEXT:    ret void
7045 //
7046 //
7047 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
7048 // CHECK9-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
7049 // CHECK9-NEXT:  entry:
7050 // CHECK9-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
7051 // CHECK9-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
7052 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
7053 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
7054 // CHECK9-NEXT:    ret void
7055 //
7056 //
7057 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..14
7058 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
7059 // CHECK9-NEXT:  entry:
7060 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7061 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7062 // CHECK9-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
7063 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7064 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7065 // CHECK9-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
7066 // CHECK9-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
7067 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7068 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7069 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
7070 // CHECK9-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
7071 // CHECK9-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
7072 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7073 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7074 // CHECK9-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
7075 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
7076 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
7077 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
7078 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7079 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7080 // CHECK9-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7081 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
7082 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7083 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7084 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
7085 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7086 // CHECK9:       cond.true:
7087 // CHECK9-NEXT:    br label [[COND_END:%.*]]
7088 // CHECK9:       cond.false:
7089 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7090 // CHECK9-NEXT:    br label [[COND_END]]
7091 // CHECK9:       cond.end:
7092 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
7093 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
7094 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7095 // CHECK9-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
7096 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7097 // CHECK9:       omp.inner.for.cond:
7098 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7099 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7100 // CHECK9-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
7101 // CHECK9-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7102 // CHECK9:       omp.inner.for.body:
7103 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7104 // CHECK9-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
7105 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7106 // CHECK9-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
7107 // CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
7108 // CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
7109 // CHECK9-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
7110 // CHECK9:       omp_if.then:
7111 // CHECK9-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
7112 // CHECK9-NEXT:    br label [[OMP_IF_END:%.*]]
7113 // CHECK9:       omp_if.else:
7114 // CHECK9-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
7115 // CHECK9-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7116 // CHECK9-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
7117 // CHECK9-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
7118 // CHECK9-NEXT:    br label [[OMP_IF_END]]
7119 // CHECK9:       omp_if.end:
7120 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7121 // CHECK9:       omp.inner.for.inc:
7122 // CHECK9-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7123 // CHECK9-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
7124 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
7125 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
7126 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
7127 // CHECK9:       omp.inner.for.end:
7128 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7129 // CHECK9:       omp.loop.exit:
7130 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
7131 // CHECK9-NEXT:    ret void
7132 //
7133 //
7134 // CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..15
7135 // CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
7136 // CHECK9-NEXT:  entry:
7137 // CHECK9-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7138 // CHECK9-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7139 // CHECK9-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
7140 // CHECK9-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
7141 // CHECK9-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7142 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7143 // CHECK9-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
7144 // CHECK9-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
7145 // CHECK9-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7146 // CHECK9-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7147 // CHECK9-NEXT:    [[I:%.*]] = alloca i32, align 4
7148 // CHECK9-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7149 // CHECK9-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7150 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7151 // CHECK9-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7152 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
7153 // CHECK9-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
7154 // CHECK9-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7155 // CHECK9-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
7156 // CHECK9-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7157 // CHECK9-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7158 // CHECK9-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7159 // CHECK9-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7160 // CHECK9-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7161 // CHECK9-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7162 // CHECK9-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7163 // CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7164 // CHECK9-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7165 // CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7166 // CHECK9-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7167 // CHECK9-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7168 // CHECK9:       cond.true:
7169 // CHECK9-NEXT:    br label [[COND_END:%.*]]
7170 // CHECK9:       cond.false:
7171 // CHECK9-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7172 // CHECK9-NEXT:    br label [[COND_END]]
7173 // CHECK9:       cond.end:
7174 // CHECK9-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7175 // CHECK9-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7176 // CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7177 // CHECK9-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7178 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7179 // CHECK9:       omp.inner.for.cond:
7180 // CHECK9-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7181 // CHECK9-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7182 // CHECK9-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7183 // CHECK9-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7184 // CHECK9:       omp.inner.for.body:
7185 // CHECK9-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7186 // CHECK9-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7187 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7188 // CHECK9-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7189 // CHECK9-NEXT:    call void @_Z3fn3v()
7190 // CHECK9-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7191 // CHECK9:       omp.body.continue:
7192 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7193 // CHECK9:       omp.inner.for.inc:
7194 // CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7195 // CHECK9-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
7196 // CHECK9-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
7197 // CHECK9-NEXT:    br label [[OMP_INNER_FOR_COND]]
7198 // CHECK9:       omp.inner.for.end:
7199 // CHECK9-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7200 // CHECK9:       omp.loop.exit:
7201 // CHECK9-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
7202 // CHECK9-NEXT:    ret void
7203 //
7204 //
7205 // CHECK9-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
7206 // CHECK9-SAME: () #[[ATTR5:[0-9]+]] {
7207 // CHECK9-NEXT:  entry:
7208 // CHECK9-NEXT:    call void @__tgt_register_requires(i64 1)
7209 // CHECK9-NEXT:    ret void
7210 //
7211 //
7212 // CHECK10-LABEL: define {{[^@]+}}@_Z9gtid_testv
7213 // CHECK10-SAME: () #[[ATTR0:[0-9]+]] {
7214 // CHECK10-NEXT:  entry:
7215 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7216 // CHECK10-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
7217 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
7218 // CHECK10-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
7219 // CHECK10-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
7220 // CHECK10-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
7221 // CHECK10:       omp_offload.failed:
7222 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
7223 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
7224 // CHECK10:       omp_offload.cont:
7225 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
7226 // CHECK10-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
7227 // CHECK10-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
7228 // CHECK10-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
7229 // CHECK10:       omp_offload.failed2:
7230 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
7231 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
7232 // CHECK10:       omp_offload.cont3:
7233 // CHECK10-NEXT:    ret void
7234 //
7235 //
7236 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
7237 // CHECK10-SAME: () #[[ATTR1:[0-9]+]] {
7238 // CHECK10-NEXT:  entry:
7239 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
7240 // CHECK10-NEXT:    ret void
7241 //
7242 //
7243 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined.
7244 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
7245 // CHECK10-NEXT:  entry:
7246 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7247 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7248 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7249 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7250 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
7251 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
7252 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7253 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7254 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7255 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7256 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7257 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
7258 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
7259 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7260 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7261 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7262 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
7263 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7264 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7265 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
7266 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7267 // CHECK10:       cond.true:
7268 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7269 // CHECK10:       cond.false:
7270 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7271 // CHECK10-NEXT:    br label [[COND_END]]
7272 // CHECK10:       cond.end:
7273 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
7274 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
7275 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7276 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
7277 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7278 // CHECK10:       omp.inner.for.cond:
7279 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7280 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7281 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
7282 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7283 // CHECK10:       omp.inner.for.body:
7284 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7285 // CHECK10-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
7286 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7287 // CHECK10-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
7288 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
7289 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7290 // CHECK10:       omp.inner.for.inc:
7291 // CHECK10-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7292 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
7293 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
7294 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
7295 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7296 // CHECK10:       omp.inner.for.end:
7297 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7298 // CHECK10:       omp.loop.exit:
7299 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
7300 // CHECK10-NEXT:    ret void
7301 //
7302 //
7303 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..1
7304 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
7305 // CHECK10-NEXT:  entry:
7306 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7307 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7308 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
7309 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
7310 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7311 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7312 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
7313 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
7314 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7315 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7316 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
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 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7320 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7321 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
7322 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
7323 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7324 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
7325 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7326 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7327 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7328 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7329 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7330 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7331 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7332 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7333 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7334 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7335 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7336 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7337 // CHECK10:       cond.true:
7338 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7339 // CHECK10:       cond.false:
7340 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7341 // CHECK10-NEXT:    br label [[COND_END]]
7342 // CHECK10:       cond.end:
7343 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7344 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7345 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7346 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7347 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7348 // CHECK10:       omp.inner.for.cond:
7349 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7350 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7351 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7352 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7353 // CHECK10:       omp.inner.for.body:
7354 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7355 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7356 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7357 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7358 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7359 // CHECK10:       omp.body.continue:
7360 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7361 // CHECK10:       omp.inner.for.inc:
7362 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7363 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
7364 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
7365 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7366 // CHECK10:       omp.inner.for.end:
7367 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7368 // CHECK10:       omp.loop.exit:
7369 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
7370 // CHECK10-NEXT:    ret void
7371 //
7372 //
7373 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
7374 // CHECK10-SAME: () #[[ATTR1]] {
7375 // CHECK10-NEXT:  entry:
7376 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
7377 // CHECK10-NEXT:    ret void
7378 //
7379 //
7380 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..2
7381 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
7382 // CHECK10-NEXT:  entry:
7383 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7384 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7385 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7386 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7387 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
7388 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
7389 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7390 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7391 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7392 // CHECK10-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
7393 // CHECK10-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
7394 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7395 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7396 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
7397 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
7398 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7399 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7400 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7401 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
7402 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7403 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7404 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
7405 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7406 // CHECK10:       cond.true:
7407 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7408 // CHECK10:       cond.false:
7409 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7410 // CHECK10-NEXT:    br label [[COND_END]]
7411 // CHECK10:       cond.end:
7412 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
7413 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
7414 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7415 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
7416 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7417 // CHECK10:       omp.inner.for.cond:
7418 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7419 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7420 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
7421 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7422 // CHECK10:       omp.inner.for.body:
7423 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7424 // CHECK10-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
7425 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7426 // CHECK10-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
7427 // CHECK10-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
7428 // CHECK10-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7429 // CHECK10-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
7430 // CHECK10-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
7431 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7432 // CHECK10:       omp.inner.for.inc:
7433 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7434 // CHECK10-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
7435 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
7436 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
7437 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7438 // CHECK10:       omp.inner.for.end:
7439 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7440 // CHECK10:       omp.loop.exit:
7441 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
7442 // CHECK10-NEXT:    ret void
7443 //
7444 //
7445 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..3
7446 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
7447 // CHECK10-NEXT:  entry:
7448 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7449 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7450 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
7451 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
7452 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7453 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7454 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
7455 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
7456 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7457 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7458 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7459 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7460 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7461 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7462 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7463 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
7464 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
7465 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7466 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
7467 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7468 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7469 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7470 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7471 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7472 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7473 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7474 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7475 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7476 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7477 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7478 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7479 // CHECK10:       cond.true:
7480 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7481 // CHECK10:       cond.false:
7482 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7483 // CHECK10-NEXT:    br label [[COND_END]]
7484 // CHECK10:       cond.end:
7485 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7486 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7487 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7488 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7489 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7490 // CHECK10:       omp.inner.for.cond:
7491 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7492 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7493 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7494 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7495 // CHECK10:       omp.inner.for.body:
7496 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7497 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7498 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7499 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7500 // CHECK10-NEXT:    call void @_Z9gtid_testv()
7501 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7502 // CHECK10:       omp.body.continue:
7503 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7504 // CHECK10:       omp.inner.for.inc:
7505 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7506 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
7507 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
7508 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7509 // CHECK10:       omp.inner.for.end:
7510 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7511 // CHECK10:       omp.loop.exit:
7512 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
7513 // CHECK10-NEXT:    ret void
7514 //
7515 //
7516 // CHECK10-LABEL: define {{[^@]+}}@main
7517 // CHECK10-SAME: () #[[ATTR3:[0-9]+]] {
7518 // CHECK10-NEXT:  entry:
7519 // CHECK10-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
7520 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7521 // CHECK10-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
7522 // CHECK10-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
7523 // CHECK10-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
7524 // CHECK10-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
7525 // CHECK10-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
7526 // CHECK10-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
7527 // CHECK10-NEXT:    store i32 0, i32* [[RETVAL]], align 4
7528 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
7529 // CHECK10-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
7530 // CHECK10-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
7531 // CHECK10-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
7532 // CHECK10:       omp_offload.failed:
7533 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
7534 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
7535 // CHECK10:       omp_offload.cont:
7536 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
7537 // CHECK10-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
7538 // CHECK10-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
7539 // CHECK10-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
7540 // CHECK10:       omp_offload.failed2:
7541 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
7542 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
7543 // CHECK10:       omp_offload.cont3:
7544 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
7545 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
7546 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
7547 // CHECK10-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
7548 // CHECK10-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
7549 // CHECK10-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
7550 // CHECK10-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
7551 // CHECK10-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
7552 // CHECK10-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
7553 // CHECK10-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
7554 // CHECK10-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
7555 // CHECK10-NEXT:    store i8* null, i8** [[TMP10]], align 8
7556 // CHECK10-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
7557 // CHECK10-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
7558 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
7559 // CHECK10-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
7560 // CHECK10-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
7561 // CHECK10-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
7562 // CHECK10:       omp_offload.failed5:
7563 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
7564 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
7565 // CHECK10:       omp_offload.cont6:
7566 // CHECK10-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
7567 // CHECK10-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
7568 // CHECK10-NEXT:    ret i32 [[CALL]]
7569 //
7570 //
7571 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
7572 // CHECK10-SAME: () #[[ATTR1]] {
7573 // CHECK10-NEXT:  entry:
7574 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
7575 // CHECK10-NEXT:    ret void
7576 //
7577 //
7578 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..4
7579 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
7580 // CHECK10-NEXT:  entry:
7581 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7582 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7583 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7584 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7585 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
7586 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
7587 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7588 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7589 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7590 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7591 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7592 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
7593 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
7594 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7595 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7596 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7597 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
7598 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7599 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7600 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
7601 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7602 // CHECK10:       cond.true:
7603 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7604 // CHECK10:       cond.false:
7605 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7606 // CHECK10-NEXT:    br label [[COND_END]]
7607 // CHECK10:       cond.end:
7608 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
7609 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
7610 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7611 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
7612 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7613 // CHECK10:       omp.inner.for.cond:
7614 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7615 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7616 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
7617 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7618 // CHECK10:       omp.inner.for.body:
7619 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7620 // CHECK10-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
7621 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7622 // CHECK10-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
7623 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
7624 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7625 // CHECK10:       omp.inner.for.inc:
7626 // CHECK10-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7627 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
7628 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
7629 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
7630 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7631 // CHECK10:       omp.inner.for.end:
7632 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7633 // CHECK10:       omp.loop.exit:
7634 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
7635 // CHECK10-NEXT:    ret void
7636 //
7637 //
7638 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..5
7639 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
7640 // CHECK10-NEXT:  entry:
7641 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7642 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7643 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
7644 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
7645 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7646 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7647 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
7648 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
7649 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7650 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7651 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7652 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7653 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7654 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7655 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7656 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
7657 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
7658 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7659 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
7660 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7661 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7662 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7663 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7664 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7665 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7666 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7667 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7668 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7669 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7670 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7671 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7672 // CHECK10:       cond.true:
7673 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7674 // CHECK10:       cond.false:
7675 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7676 // CHECK10-NEXT:    br label [[COND_END]]
7677 // CHECK10:       cond.end:
7678 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7679 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7680 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7681 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7682 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7683 // CHECK10:       omp.inner.for.cond:
7684 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7685 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7686 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7687 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7688 // CHECK10:       omp.inner.for.body:
7689 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7690 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7691 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7692 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7693 // CHECK10-NEXT:    call void @_Z3fn4v()
7694 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7695 // CHECK10:       omp.body.continue:
7696 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7697 // CHECK10:       omp.inner.for.inc:
7698 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7699 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
7700 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
7701 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7702 // CHECK10:       omp.inner.for.end:
7703 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7704 // CHECK10:       omp.loop.exit:
7705 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
7706 // CHECK10-NEXT:    ret void
7707 //
7708 //
7709 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
7710 // CHECK10-SAME: () #[[ATTR1]] {
7711 // CHECK10-NEXT:  entry:
7712 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
7713 // CHECK10-NEXT:    ret void
7714 //
7715 //
7716 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..6
7717 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
7718 // CHECK10-NEXT:  entry:
7719 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7720 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7721 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7722 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7723 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
7724 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
7725 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7726 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7727 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7728 // CHECK10-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
7729 // CHECK10-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
7730 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7731 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7732 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
7733 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
7734 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7735 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7736 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7737 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
7738 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7739 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7740 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
7741 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7742 // CHECK10:       cond.true:
7743 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7744 // CHECK10:       cond.false:
7745 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7746 // CHECK10-NEXT:    br label [[COND_END]]
7747 // CHECK10:       cond.end:
7748 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
7749 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
7750 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7751 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
7752 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7753 // CHECK10:       omp.inner.for.cond:
7754 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7755 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7756 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
7757 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7758 // CHECK10:       omp.inner.for.body:
7759 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7760 // CHECK10-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
7761 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7762 // CHECK10-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
7763 // CHECK10-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
7764 // CHECK10-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7765 // CHECK10-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
7766 // CHECK10-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
7767 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7768 // CHECK10:       omp.inner.for.inc:
7769 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7770 // CHECK10-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
7771 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
7772 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
7773 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7774 // CHECK10:       omp.inner.for.end:
7775 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7776 // CHECK10:       omp.loop.exit:
7777 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
7778 // CHECK10-NEXT:    ret void
7779 //
7780 //
7781 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..7
7782 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
7783 // CHECK10-NEXT:  entry:
7784 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7785 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7786 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
7787 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
7788 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7789 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7790 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
7791 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
7792 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7793 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7794 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7795 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7796 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7797 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7798 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7799 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
7800 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
7801 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7802 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
7803 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7804 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7805 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7806 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7807 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7808 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7809 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7810 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7811 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7812 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7813 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7814 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7815 // CHECK10:       cond.true:
7816 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7817 // CHECK10:       cond.false:
7818 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7819 // CHECK10-NEXT:    br label [[COND_END]]
7820 // CHECK10:       cond.end:
7821 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7822 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7823 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7824 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7825 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7826 // CHECK10:       omp.inner.for.cond:
7827 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7828 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7829 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7830 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7831 // CHECK10:       omp.inner.for.body:
7832 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7833 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7834 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7835 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7836 // CHECK10-NEXT:    call void @_Z3fn5v()
7837 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7838 // CHECK10:       omp.body.continue:
7839 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7840 // CHECK10:       omp.inner.for.inc:
7841 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7842 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
7843 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
7844 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7845 // CHECK10:       omp.inner.for.end:
7846 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7847 // CHECK10:       omp.loop.exit:
7848 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
7849 // CHECK10-NEXT:    ret void
7850 //
7851 //
7852 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
7853 // CHECK10-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
7854 // CHECK10-NEXT:  entry:
7855 // CHECK10-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
7856 // CHECK10-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
7857 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
7858 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
7859 // CHECK10-NEXT:    ret void
7860 //
7861 //
7862 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..8
7863 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
7864 // CHECK10-NEXT:  entry:
7865 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7866 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7867 // CHECK10-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
7868 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7869 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7870 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
7871 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
7872 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7873 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7874 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7875 // CHECK10-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
7876 // CHECK10-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
7877 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7878 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7879 // CHECK10-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
7880 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
7881 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
7882 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
7883 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7884 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7885 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7886 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
7887 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7888 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7889 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
7890 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7891 // CHECK10:       cond.true:
7892 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7893 // CHECK10:       cond.false:
7894 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7895 // CHECK10-NEXT:    br label [[COND_END]]
7896 // CHECK10:       cond.end:
7897 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
7898 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
7899 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7900 // CHECK10-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
7901 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7902 // CHECK10:       omp.inner.for.cond:
7903 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7904 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7905 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
7906 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7907 // CHECK10:       omp.inner.for.body:
7908 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
7909 // CHECK10-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
7910 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
7911 // CHECK10-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
7912 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
7913 // CHECK10-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
7914 // CHECK10-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
7915 // CHECK10:       omp_if.then:
7916 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
7917 // CHECK10-NEXT:    br label [[OMP_IF_END:%.*]]
7918 // CHECK10:       omp_if.else:
7919 // CHECK10-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
7920 // CHECK10-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7921 // CHECK10-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
7922 // CHECK10-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
7923 // CHECK10-NEXT:    br label [[OMP_IF_END]]
7924 // CHECK10:       omp_if.end:
7925 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7926 // CHECK10:       omp.inner.for.inc:
7927 // CHECK10-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7928 // CHECK10-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
7929 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
7930 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
7931 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
7932 // CHECK10:       omp.inner.for.end:
7933 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
7934 // CHECK10:       omp.loop.exit:
7935 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
7936 // CHECK10-NEXT:    ret void
7937 //
7938 //
7939 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..9
7940 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
7941 // CHECK10-NEXT:  entry:
7942 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
7943 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
7944 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
7945 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
7946 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
7947 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
7948 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
7949 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
7950 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
7951 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
7952 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
7953 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
7954 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
7955 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7956 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7957 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
7958 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
7959 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
7960 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
7961 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
7962 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
7963 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
7964 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
7965 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
7966 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
7967 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
7968 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
7969 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
7970 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7971 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
7972 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
7973 // CHECK10:       cond.true:
7974 // CHECK10-NEXT:    br label [[COND_END:%.*]]
7975 // CHECK10:       cond.false:
7976 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7977 // CHECK10-NEXT:    br label [[COND_END]]
7978 // CHECK10:       cond.end:
7979 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
7980 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
7981 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
7982 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
7983 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
7984 // CHECK10:       omp.inner.for.cond:
7985 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7986 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
7987 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
7988 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
7989 // CHECK10:       omp.inner.for.body:
7990 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
7991 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
7992 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
7993 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
7994 // CHECK10-NEXT:    call void @_Z3fn6v()
7995 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
7996 // CHECK10:       omp.body.continue:
7997 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
7998 // CHECK10:       omp.inner.for.inc:
7999 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8000 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
8001 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
8002 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8003 // CHECK10:       omp.inner.for.end:
8004 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8005 // CHECK10:       omp.loop.exit:
8006 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
8007 // CHECK10-NEXT:    ret void
8008 //
8009 //
8010 // CHECK10-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
8011 // CHECK10-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
8012 // CHECK10-NEXT:  entry:
8013 // CHECK10-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
8014 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8015 // CHECK10-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
8016 // CHECK10-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
8017 // CHECK10-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
8018 // CHECK10-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
8019 // CHECK10-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
8020 // CHECK10-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
8021 // CHECK10-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
8022 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
8023 // CHECK10-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
8024 // CHECK10-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
8025 // CHECK10-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
8026 // CHECK10:       omp_offload.failed:
8027 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
8028 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
8029 // CHECK10:       omp_offload.cont:
8030 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
8031 // CHECK10-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
8032 // CHECK10-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
8033 // CHECK10-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
8034 // CHECK10:       omp_offload.failed2:
8035 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
8036 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
8037 // CHECK10:       omp_offload.cont3:
8038 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
8039 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
8040 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
8041 // CHECK10-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
8042 // CHECK10-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
8043 // CHECK10-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
8044 // CHECK10-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
8045 // CHECK10-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
8046 // CHECK10-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
8047 // CHECK10-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
8048 // CHECK10-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
8049 // CHECK10-NEXT:    store i8* null, i8** [[TMP10]], align 8
8050 // CHECK10-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
8051 // CHECK10-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
8052 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
8053 // CHECK10-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
8054 // CHECK10-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
8055 // CHECK10-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
8056 // CHECK10:       omp_offload.failed5:
8057 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
8058 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
8059 // CHECK10:       omp_offload.cont6:
8060 // CHECK10-NEXT:    ret i32 0
8061 //
8062 //
8063 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
8064 // CHECK10-SAME: () #[[ATTR1]] {
8065 // CHECK10-NEXT:  entry:
8066 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
8067 // CHECK10-NEXT:    ret void
8068 //
8069 //
8070 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..10
8071 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
8072 // CHECK10-NEXT:  entry:
8073 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8074 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8075 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8076 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8077 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
8078 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
8079 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8080 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8081 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
8082 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8083 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8084 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
8085 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
8086 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8087 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8088 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8089 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
8090 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8091 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8092 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
8093 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8094 // CHECK10:       cond.true:
8095 // CHECK10-NEXT:    br label [[COND_END:%.*]]
8096 // CHECK10:       cond.false:
8097 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8098 // CHECK10-NEXT:    br label [[COND_END]]
8099 // CHECK10:       cond.end:
8100 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
8101 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
8102 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8103 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
8104 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8105 // CHECK10:       omp.inner.for.cond:
8106 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8107 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8108 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
8109 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8110 // CHECK10:       omp.inner.for.body:
8111 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8112 // CHECK10-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
8113 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8114 // CHECK10-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
8115 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
8116 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8117 // CHECK10:       omp.inner.for.inc:
8118 // CHECK10-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8119 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
8120 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
8121 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
8122 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8123 // CHECK10:       omp.inner.for.end:
8124 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8125 // CHECK10:       omp.loop.exit:
8126 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
8127 // CHECK10-NEXT:    ret void
8128 //
8129 //
8130 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..11
8131 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
8132 // CHECK10-NEXT:  entry:
8133 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8134 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8135 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
8136 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
8137 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8138 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8139 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
8140 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
8141 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8142 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8143 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
8144 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8145 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8146 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8147 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8148 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
8149 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
8150 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8151 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
8152 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8153 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
8154 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
8155 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
8156 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8157 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8158 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8159 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
8160 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8161 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8162 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
8163 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8164 // CHECK10:       cond.true:
8165 // CHECK10-NEXT:    br label [[COND_END:%.*]]
8166 // CHECK10:       cond.false:
8167 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8168 // CHECK10-NEXT:    br label [[COND_END]]
8169 // CHECK10:       cond.end:
8170 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
8171 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
8172 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
8173 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
8174 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8175 // CHECK10:       omp.inner.for.cond:
8176 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8177 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8178 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
8179 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8180 // CHECK10:       omp.inner.for.body:
8181 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8182 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
8183 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
8184 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
8185 // CHECK10-NEXT:    call void @_Z3fn1v()
8186 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
8187 // CHECK10:       omp.body.continue:
8188 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8189 // CHECK10:       omp.inner.for.inc:
8190 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8191 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
8192 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
8193 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8194 // CHECK10:       omp.inner.for.end:
8195 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8196 // CHECK10:       omp.loop.exit:
8197 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
8198 // CHECK10-NEXT:    ret void
8199 //
8200 //
8201 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
8202 // CHECK10-SAME: () #[[ATTR1]] {
8203 // CHECK10-NEXT:  entry:
8204 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
8205 // CHECK10-NEXT:    ret void
8206 //
8207 //
8208 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..12
8209 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
8210 // CHECK10-NEXT:  entry:
8211 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8212 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8213 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8214 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8215 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
8216 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
8217 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8218 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8219 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
8220 // CHECK10-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
8221 // CHECK10-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
8222 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8223 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8224 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
8225 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
8226 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8227 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8228 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8229 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
8230 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8231 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8232 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
8233 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8234 // CHECK10:       cond.true:
8235 // CHECK10-NEXT:    br label [[COND_END:%.*]]
8236 // CHECK10:       cond.false:
8237 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8238 // CHECK10-NEXT:    br label [[COND_END]]
8239 // CHECK10:       cond.end:
8240 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
8241 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
8242 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8243 // CHECK10-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
8244 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8245 // CHECK10:       omp.inner.for.cond:
8246 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8247 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8248 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
8249 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8250 // CHECK10:       omp.inner.for.body:
8251 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8252 // CHECK10-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
8253 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8254 // CHECK10-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
8255 // CHECK10-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
8256 // CHECK10-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8257 // CHECK10-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
8258 // CHECK10-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
8259 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8260 // CHECK10:       omp.inner.for.inc:
8261 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8262 // CHECK10-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
8263 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
8264 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
8265 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8266 // CHECK10:       omp.inner.for.end:
8267 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8268 // CHECK10:       omp.loop.exit:
8269 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
8270 // CHECK10-NEXT:    ret void
8271 //
8272 //
8273 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..13
8274 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
8275 // CHECK10-NEXT:  entry:
8276 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8277 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8278 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
8279 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
8280 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8281 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8282 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
8283 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
8284 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8285 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8286 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
8287 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8288 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8289 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8290 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8291 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
8292 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
8293 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8294 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
8295 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8296 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
8297 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
8298 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
8299 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8300 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8301 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8302 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
8303 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8304 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8305 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
8306 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8307 // CHECK10:       cond.true:
8308 // CHECK10-NEXT:    br label [[COND_END:%.*]]
8309 // CHECK10:       cond.false:
8310 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8311 // CHECK10-NEXT:    br label [[COND_END]]
8312 // CHECK10:       cond.end:
8313 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
8314 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
8315 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
8316 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
8317 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8318 // CHECK10:       omp.inner.for.cond:
8319 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8320 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8321 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
8322 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8323 // CHECK10:       omp.inner.for.body:
8324 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8325 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
8326 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
8327 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
8328 // CHECK10-NEXT:    call void @_Z3fn2v()
8329 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
8330 // CHECK10:       omp.body.continue:
8331 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8332 // CHECK10:       omp.inner.for.inc:
8333 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8334 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
8335 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
8336 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8337 // CHECK10:       omp.inner.for.end:
8338 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8339 // CHECK10:       omp.loop.exit:
8340 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
8341 // CHECK10-NEXT:    ret void
8342 //
8343 //
8344 // CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
8345 // CHECK10-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
8346 // CHECK10-NEXT:  entry:
8347 // CHECK10-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
8348 // CHECK10-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
8349 // CHECK10-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
8350 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
8351 // CHECK10-NEXT:    ret void
8352 //
8353 //
8354 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..14
8355 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
8356 // CHECK10-NEXT:  entry:
8357 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8358 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8359 // CHECK10-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
8360 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8361 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8362 // CHECK10-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
8363 // CHECK10-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
8364 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8365 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8366 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
8367 // CHECK10-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
8368 // CHECK10-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
8369 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8370 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8371 // CHECK10-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
8372 // CHECK10-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
8373 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
8374 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
8375 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8376 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8377 // CHECK10-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8378 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
8379 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8380 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8381 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
8382 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8383 // CHECK10:       cond.true:
8384 // CHECK10-NEXT:    br label [[COND_END:%.*]]
8385 // CHECK10:       cond.false:
8386 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8387 // CHECK10-NEXT:    br label [[COND_END]]
8388 // CHECK10:       cond.end:
8389 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
8390 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
8391 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8392 // CHECK10-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
8393 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8394 // CHECK10:       omp.inner.for.cond:
8395 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8396 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8397 // CHECK10-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
8398 // CHECK10-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8399 // CHECK10:       omp.inner.for.body:
8400 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8401 // CHECK10-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
8402 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8403 // CHECK10-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
8404 // CHECK10-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
8405 // CHECK10-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
8406 // CHECK10-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
8407 // CHECK10:       omp_if.then:
8408 // CHECK10-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
8409 // CHECK10-NEXT:    br label [[OMP_IF_END:%.*]]
8410 // CHECK10:       omp_if.else:
8411 // CHECK10-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
8412 // CHECK10-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8413 // CHECK10-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
8414 // CHECK10-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
8415 // CHECK10-NEXT:    br label [[OMP_IF_END]]
8416 // CHECK10:       omp_if.end:
8417 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8418 // CHECK10:       omp.inner.for.inc:
8419 // CHECK10-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8420 // CHECK10-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
8421 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
8422 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
8423 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8424 // CHECK10:       omp.inner.for.end:
8425 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8426 // CHECK10:       omp.loop.exit:
8427 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
8428 // CHECK10-NEXT:    ret void
8429 //
8430 //
8431 // CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..15
8432 // CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
8433 // CHECK10-NEXT:  entry:
8434 // CHECK10-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8435 // CHECK10-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8436 // CHECK10-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
8437 // CHECK10-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
8438 // CHECK10-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8439 // CHECK10-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8440 // CHECK10-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
8441 // CHECK10-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
8442 // CHECK10-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8443 // CHECK10-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8444 // CHECK10-NEXT:    [[I:%.*]] = alloca i32, align 4
8445 // CHECK10-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8446 // CHECK10-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8447 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8448 // CHECK10-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8449 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
8450 // CHECK10-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
8451 // CHECK10-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8452 // CHECK10-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
8453 // CHECK10-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8454 // CHECK10-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
8455 // CHECK10-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
8456 // CHECK10-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
8457 // CHECK10-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8458 // CHECK10-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8459 // CHECK10-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8460 // CHECK10-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
8461 // CHECK10-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8462 // CHECK10-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8463 // CHECK10-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
8464 // CHECK10-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8465 // CHECK10:       cond.true:
8466 // CHECK10-NEXT:    br label [[COND_END:%.*]]
8467 // CHECK10:       cond.false:
8468 // CHECK10-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8469 // CHECK10-NEXT:    br label [[COND_END]]
8470 // CHECK10:       cond.end:
8471 // CHECK10-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
8472 // CHECK10-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
8473 // CHECK10-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
8474 // CHECK10-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
8475 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8476 // CHECK10:       omp.inner.for.cond:
8477 // CHECK10-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8478 // CHECK10-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8479 // CHECK10-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
8480 // CHECK10-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8481 // CHECK10:       omp.inner.for.body:
8482 // CHECK10-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8483 // CHECK10-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
8484 // CHECK10-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
8485 // CHECK10-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
8486 // CHECK10-NEXT:    call void @_Z3fn3v()
8487 // CHECK10-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
8488 // CHECK10:       omp.body.continue:
8489 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8490 // CHECK10:       omp.inner.for.inc:
8491 // CHECK10-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8492 // CHECK10-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
8493 // CHECK10-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
8494 // CHECK10-NEXT:    br label [[OMP_INNER_FOR_COND]]
8495 // CHECK10:       omp.inner.for.end:
8496 // CHECK10-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8497 // CHECK10:       omp.loop.exit:
8498 // CHECK10-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
8499 // CHECK10-NEXT:    ret void
8500 //
8501 //
8502 // CHECK10-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
8503 // CHECK10-SAME: () #[[ATTR5:[0-9]+]] {
8504 // CHECK10-NEXT:  entry:
8505 // CHECK10-NEXT:    call void @__tgt_register_requires(i64 1)
8506 // CHECK10-NEXT:    ret void
8507 //
8508 //
8509 // CHECK11-LABEL: define {{[^@]+}}@_Z9gtid_testv
8510 // CHECK11-SAME: () #[[ATTR0:[0-9]+]] {
8511 // CHECK11-NEXT:  entry:
8512 // CHECK11-NEXT:    [[I:%.*]] = alloca i32, align 4
8513 // CHECK11-NEXT:    [[I1:%.*]] = alloca i32, align 4
8514 // CHECK11-NEXT:    store i32 0, i32* [[I]], align 4
8515 // CHECK11-NEXT:    br label [[FOR_COND:%.*]]
8516 // CHECK11:       for.cond:
8517 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
8518 // CHECK11-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
8519 // CHECK11-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
8520 // CHECK11:       for.body:
8521 // CHECK11-NEXT:    br label [[FOR_INC:%.*]]
8522 // CHECK11:       for.inc:
8523 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
8524 // CHECK11-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
8525 // CHECK11-NEXT:    store i32 [[INC]], i32* [[I]], align 4
8526 // CHECK11-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
8527 // CHECK11:       for.end:
8528 // CHECK11-NEXT:    store i32 0, i32* [[I1]], align 4
8529 // CHECK11-NEXT:    br label [[FOR_COND2:%.*]]
8530 // CHECK11:       for.cond2:
8531 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
8532 // CHECK11-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
8533 // CHECK11-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
8534 // CHECK11:       for.body4:
8535 // CHECK11-NEXT:    call void @_Z9gtid_testv()
8536 // CHECK11-NEXT:    br label [[FOR_INC5:%.*]]
8537 // CHECK11:       for.inc5:
8538 // CHECK11-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
8539 // CHECK11-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
8540 // CHECK11-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
8541 // CHECK11-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP8:![0-9]+]]
8542 // CHECK11:       for.end7:
8543 // CHECK11-NEXT:    ret void
8544 //
8545 //
8546 // CHECK11-LABEL: define {{[^@]+}}@main
8547 // CHECK11-SAME: () #[[ATTR1:[0-9]+]] {
8548 // CHECK11-NEXT:  entry:
8549 // CHECK11-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
8550 // CHECK11-NEXT:    [[I:%.*]] = alloca i32, align 4
8551 // CHECK11-NEXT:    [[I1:%.*]] = alloca i32, align 4
8552 // CHECK11-NEXT:    [[I8:%.*]] = alloca i32, align 4
8553 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
8554 // CHECK11-NEXT:    store i32 0, i32* [[I]], align 4
8555 // CHECK11-NEXT:    br label [[FOR_COND:%.*]]
8556 // CHECK11:       for.cond:
8557 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
8558 // CHECK11-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
8559 // CHECK11-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
8560 // CHECK11:       for.body:
8561 // CHECK11-NEXT:    call void @_Z3fn4v()
8562 // CHECK11-NEXT:    br label [[FOR_INC:%.*]]
8563 // CHECK11:       for.inc:
8564 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
8565 // CHECK11-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
8566 // CHECK11-NEXT:    store i32 [[INC]], i32* [[I]], align 4
8567 // CHECK11-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
8568 // CHECK11:       for.end:
8569 // CHECK11-NEXT:    store i32 0, i32* [[I1]], align 4
8570 // CHECK11-NEXT:    br label [[FOR_COND2:%.*]]
8571 // CHECK11:       for.cond2:
8572 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
8573 // CHECK11-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
8574 // CHECK11-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
8575 // CHECK11:       for.body4:
8576 // CHECK11-NEXT:    call void @_Z3fn5v()
8577 // CHECK11-NEXT:    br label [[FOR_INC5:%.*]]
8578 // CHECK11:       for.inc5:
8579 // CHECK11-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
8580 // CHECK11-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
8581 // CHECK11-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
8582 // CHECK11-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP10:![0-9]+]]
8583 // CHECK11:       for.end7:
8584 // CHECK11-NEXT:    store i32 0, i32* [[I8]], align 4
8585 // CHECK11-NEXT:    br label [[FOR_COND9:%.*]]
8586 // CHECK11:       for.cond9:
8587 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
8588 // CHECK11-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
8589 // CHECK11-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
8590 // CHECK11:       for.body11:
8591 // CHECK11-NEXT:    call void @_Z3fn6v()
8592 // CHECK11-NEXT:    br label [[FOR_INC12:%.*]]
8593 // CHECK11:       for.inc12:
8594 // CHECK11-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
8595 // CHECK11-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
8596 // CHECK11-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
8597 // CHECK11-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP11:![0-9]+]]
8598 // CHECK11:       for.end14:
8599 // CHECK11-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
8600 // CHECK11-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
8601 // CHECK11-NEXT:    ret i32 [[CALL]]
8602 //
8603 //
8604 // CHECK11-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
8605 // CHECK11-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
8606 // CHECK11-NEXT:  entry:
8607 // CHECK11-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
8608 // CHECK11-NEXT:    [[I:%.*]] = alloca i32, align 4
8609 // CHECK11-NEXT:    [[I1:%.*]] = alloca i32, align 4
8610 // CHECK11-NEXT:    [[I8:%.*]] = alloca i32, align 4
8611 // CHECK11-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
8612 // CHECK11-NEXT:    store i32 0, i32* [[I]], align 4
8613 // CHECK11-NEXT:    br label [[FOR_COND:%.*]]
8614 // CHECK11:       for.cond:
8615 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
8616 // CHECK11-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
8617 // CHECK11-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
8618 // CHECK11:       for.body:
8619 // CHECK11-NEXT:    call void @_Z3fn1v()
8620 // CHECK11-NEXT:    br label [[FOR_INC:%.*]]
8621 // CHECK11:       for.inc:
8622 // CHECK11-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
8623 // CHECK11-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
8624 // CHECK11-NEXT:    store i32 [[INC]], i32* [[I]], align 4
8625 // CHECK11-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
8626 // CHECK11:       for.end:
8627 // CHECK11-NEXT:    store i32 0, i32* [[I1]], align 4
8628 // CHECK11-NEXT:    br label [[FOR_COND2:%.*]]
8629 // CHECK11:       for.cond2:
8630 // CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
8631 // CHECK11-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
8632 // CHECK11-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
8633 // CHECK11:       for.body4:
8634 // CHECK11-NEXT:    call void @_Z3fn2v()
8635 // CHECK11-NEXT:    br label [[FOR_INC5:%.*]]
8636 // CHECK11:       for.inc5:
8637 // CHECK11-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
8638 // CHECK11-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
8639 // CHECK11-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
8640 // CHECK11-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP13:![0-9]+]]
8641 // CHECK11:       for.end7:
8642 // CHECK11-NEXT:    store i32 0, i32* [[I8]], align 4
8643 // CHECK11-NEXT:    br label [[FOR_COND9:%.*]]
8644 // CHECK11:       for.cond9:
8645 // CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
8646 // CHECK11-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
8647 // CHECK11-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
8648 // CHECK11:       for.body11:
8649 // CHECK11-NEXT:    call void @_Z3fn3v()
8650 // CHECK11-NEXT:    br label [[FOR_INC12:%.*]]
8651 // CHECK11:       for.inc12:
8652 // CHECK11-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
8653 // CHECK11-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
8654 // CHECK11-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
8655 // CHECK11-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP14:![0-9]+]]
8656 // CHECK11:       for.end14:
8657 // CHECK11-NEXT:    ret i32 0
8658 //
8659 //
8660 // CHECK12-LABEL: define {{[^@]+}}@_Z9gtid_testv
8661 // CHECK12-SAME: () #[[ATTR0:[0-9]+]] {
8662 // CHECK12-NEXT:  entry:
8663 // CHECK12-NEXT:    [[I:%.*]] = alloca i32, align 4
8664 // CHECK12-NEXT:    [[I1:%.*]] = alloca i32, align 4
8665 // CHECK12-NEXT:    store i32 0, i32* [[I]], align 4
8666 // CHECK12-NEXT:    br label [[FOR_COND:%.*]]
8667 // CHECK12:       for.cond:
8668 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
8669 // CHECK12-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
8670 // CHECK12-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
8671 // CHECK12:       for.body:
8672 // CHECK12-NEXT:    br label [[FOR_INC:%.*]]
8673 // CHECK12:       for.inc:
8674 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
8675 // CHECK12-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
8676 // CHECK12-NEXT:    store i32 [[INC]], i32* [[I]], align 4
8677 // CHECK12-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
8678 // CHECK12:       for.end:
8679 // CHECK12-NEXT:    store i32 0, i32* [[I1]], align 4
8680 // CHECK12-NEXT:    br label [[FOR_COND2:%.*]]
8681 // CHECK12:       for.cond2:
8682 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
8683 // CHECK12-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
8684 // CHECK12-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
8685 // CHECK12:       for.body4:
8686 // CHECK12-NEXT:    call void @_Z9gtid_testv()
8687 // CHECK12-NEXT:    br label [[FOR_INC5:%.*]]
8688 // CHECK12:       for.inc5:
8689 // CHECK12-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
8690 // CHECK12-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
8691 // CHECK12-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
8692 // CHECK12-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP8:![0-9]+]]
8693 // CHECK12:       for.end7:
8694 // CHECK12-NEXT:    ret void
8695 //
8696 //
8697 // CHECK12-LABEL: define {{[^@]+}}@main
8698 // CHECK12-SAME: () #[[ATTR1:[0-9]+]] {
8699 // CHECK12-NEXT:  entry:
8700 // CHECK12-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
8701 // CHECK12-NEXT:    [[I:%.*]] = alloca i32, align 4
8702 // CHECK12-NEXT:    [[I1:%.*]] = alloca i32, align 4
8703 // CHECK12-NEXT:    [[I8:%.*]] = alloca i32, align 4
8704 // CHECK12-NEXT:    store i32 0, i32* [[RETVAL]], align 4
8705 // CHECK12-NEXT:    store i32 0, i32* [[I]], align 4
8706 // CHECK12-NEXT:    br label [[FOR_COND:%.*]]
8707 // CHECK12:       for.cond:
8708 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
8709 // CHECK12-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
8710 // CHECK12-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
8711 // CHECK12:       for.body:
8712 // CHECK12-NEXT:    call void @_Z3fn4v()
8713 // CHECK12-NEXT:    br label [[FOR_INC:%.*]]
8714 // CHECK12:       for.inc:
8715 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
8716 // CHECK12-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
8717 // CHECK12-NEXT:    store i32 [[INC]], i32* [[I]], align 4
8718 // CHECK12-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
8719 // CHECK12:       for.end:
8720 // CHECK12-NEXT:    store i32 0, i32* [[I1]], align 4
8721 // CHECK12-NEXT:    br label [[FOR_COND2:%.*]]
8722 // CHECK12:       for.cond2:
8723 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
8724 // CHECK12-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
8725 // CHECK12-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
8726 // CHECK12:       for.body4:
8727 // CHECK12-NEXT:    call void @_Z3fn5v()
8728 // CHECK12-NEXT:    br label [[FOR_INC5:%.*]]
8729 // CHECK12:       for.inc5:
8730 // CHECK12-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
8731 // CHECK12-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
8732 // CHECK12-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
8733 // CHECK12-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP10:![0-9]+]]
8734 // CHECK12:       for.end7:
8735 // CHECK12-NEXT:    store i32 0, i32* [[I8]], align 4
8736 // CHECK12-NEXT:    br label [[FOR_COND9:%.*]]
8737 // CHECK12:       for.cond9:
8738 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
8739 // CHECK12-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
8740 // CHECK12-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
8741 // CHECK12:       for.body11:
8742 // CHECK12-NEXT:    call void @_Z3fn6v()
8743 // CHECK12-NEXT:    br label [[FOR_INC12:%.*]]
8744 // CHECK12:       for.inc12:
8745 // CHECK12-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
8746 // CHECK12-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
8747 // CHECK12-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
8748 // CHECK12-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP11:![0-9]+]]
8749 // CHECK12:       for.end14:
8750 // CHECK12-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
8751 // CHECK12-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
8752 // CHECK12-NEXT:    ret i32 [[CALL]]
8753 //
8754 //
8755 // CHECK12-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
8756 // CHECK12-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
8757 // CHECK12-NEXT:  entry:
8758 // CHECK12-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
8759 // CHECK12-NEXT:    [[I:%.*]] = alloca i32, align 4
8760 // CHECK12-NEXT:    [[I1:%.*]] = alloca i32, align 4
8761 // CHECK12-NEXT:    [[I8:%.*]] = alloca i32, align 4
8762 // CHECK12-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
8763 // CHECK12-NEXT:    store i32 0, i32* [[I]], align 4
8764 // CHECK12-NEXT:    br label [[FOR_COND:%.*]]
8765 // CHECK12:       for.cond:
8766 // CHECK12-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
8767 // CHECK12-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
8768 // CHECK12-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
8769 // CHECK12:       for.body:
8770 // CHECK12-NEXT:    call void @_Z3fn1v()
8771 // CHECK12-NEXT:    br label [[FOR_INC:%.*]]
8772 // CHECK12:       for.inc:
8773 // CHECK12-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
8774 // CHECK12-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
8775 // CHECK12-NEXT:    store i32 [[INC]], i32* [[I]], align 4
8776 // CHECK12-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
8777 // CHECK12:       for.end:
8778 // CHECK12-NEXT:    store i32 0, i32* [[I1]], align 4
8779 // CHECK12-NEXT:    br label [[FOR_COND2:%.*]]
8780 // CHECK12:       for.cond2:
8781 // CHECK12-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
8782 // CHECK12-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
8783 // CHECK12-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
8784 // CHECK12:       for.body4:
8785 // CHECK12-NEXT:    call void @_Z3fn2v()
8786 // CHECK12-NEXT:    br label [[FOR_INC5:%.*]]
8787 // CHECK12:       for.inc5:
8788 // CHECK12-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
8789 // CHECK12-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
8790 // CHECK12-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
8791 // CHECK12-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP13:![0-9]+]]
8792 // CHECK12:       for.end7:
8793 // CHECK12-NEXT:    store i32 0, i32* [[I8]], align 4
8794 // CHECK12-NEXT:    br label [[FOR_COND9:%.*]]
8795 // CHECK12:       for.cond9:
8796 // CHECK12-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
8797 // CHECK12-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
8798 // CHECK12-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
8799 // CHECK12:       for.body11:
8800 // CHECK12-NEXT:    call void @_Z3fn3v()
8801 // CHECK12-NEXT:    br label [[FOR_INC12:%.*]]
8802 // CHECK12:       for.inc12:
8803 // CHECK12-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
8804 // CHECK12-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
8805 // CHECK12-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
8806 // CHECK12-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP14:![0-9]+]]
8807 // CHECK12:       for.end14:
8808 // CHECK12-NEXT:    ret i32 0
8809 //
8810 //
8811 // CHECK13-LABEL: define {{[^@]+}}@_Z9gtid_testv
8812 // CHECK13-SAME: () #[[ATTR0:[0-9]+]] {
8813 // CHECK13-NEXT:  entry:
8814 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8815 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
8816 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
8817 // CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
8818 // CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
8819 // CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
8820 // CHECK13:       omp_offload.failed:
8821 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
8822 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
8823 // CHECK13:       omp_offload.cont:
8824 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
8825 // CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
8826 // CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
8827 // CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
8828 // CHECK13:       omp_offload.failed2:
8829 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
8830 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
8831 // CHECK13:       omp_offload.cont3:
8832 // CHECK13-NEXT:    ret void
8833 //
8834 //
8835 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
8836 // CHECK13-SAME: () #[[ATTR1:[0-9]+]] {
8837 // CHECK13-NEXT:  entry:
8838 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
8839 // CHECK13-NEXT:    ret void
8840 //
8841 //
8842 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined.
8843 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
8844 // CHECK13-NEXT:  entry:
8845 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8846 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8847 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8848 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8849 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
8850 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
8851 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8852 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8853 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
8854 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8855 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8856 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
8857 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
8858 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8859 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8860 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8861 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
8862 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8863 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8864 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
8865 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8866 // CHECK13:       cond.true:
8867 // CHECK13-NEXT:    br label [[COND_END:%.*]]
8868 // CHECK13:       cond.false:
8869 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8870 // CHECK13-NEXT:    br label [[COND_END]]
8871 // CHECK13:       cond.end:
8872 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
8873 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
8874 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8875 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
8876 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8877 // CHECK13:       omp.inner.for.cond:
8878 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8879 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8880 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
8881 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8882 // CHECK13:       omp.inner.for.body:
8883 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
8884 // CHECK13-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
8885 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
8886 // CHECK13-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
8887 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
8888 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8889 // CHECK13:       omp.inner.for.inc:
8890 // CHECK13-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8891 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
8892 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
8893 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
8894 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
8895 // CHECK13:       omp.inner.for.end:
8896 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8897 // CHECK13:       omp.loop.exit:
8898 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
8899 // CHECK13-NEXT:    ret void
8900 //
8901 //
8902 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..1
8903 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
8904 // CHECK13-NEXT:  entry:
8905 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8906 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8907 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
8908 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
8909 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8910 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8911 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
8912 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
8913 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8914 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8915 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
8916 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8917 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8918 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8919 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8920 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
8921 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
8922 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
8923 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
8924 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
8925 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
8926 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
8927 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
8928 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8929 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8930 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
8931 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
8932 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
8933 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8934 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
8935 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
8936 // CHECK13:       cond.true:
8937 // CHECK13-NEXT:    br label [[COND_END:%.*]]
8938 // CHECK13:       cond.false:
8939 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8940 // CHECK13-NEXT:    br label [[COND_END]]
8941 // CHECK13:       cond.end:
8942 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
8943 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
8944 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
8945 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
8946 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
8947 // CHECK13:       omp.inner.for.cond:
8948 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8949 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
8950 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
8951 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
8952 // CHECK13:       omp.inner.for.body:
8953 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8954 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
8955 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
8956 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
8957 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
8958 // CHECK13:       omp.body.continue:
8959 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
8960 // CHECK13:       omp.inner.for.inc:
8961 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
8962 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
8963 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
8964 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
8965 // CHECK13:       omp.inner.for.end:
8966 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
8967 // CHECK13:       omp.loop.exit:
8968 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
8969 // CHECK13-NEXT:    ret void
8970 //
8971 //
8972 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
8973 // CHECK13-SAME: () #[[ATTR1]] {
8974 // CHECK13-NEXT:  entry:
8975 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
8976 // CHECK13-NEXT:    ret void
8977 //
8978 //
8979 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..2
8980 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
8981 // CHECK13-NEXT:  entry:
8982 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
8983 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
8984 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
8985 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
8986 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
8987 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
8988 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
8989 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
8990 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
8991 // CHECK13-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
8992 // CHECK13-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
8993 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
8994 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
8995 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
8996 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
8997 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
8998 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
8999 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9000 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
9001 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9002 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9003 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
9004 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9005 // CHECK13:       cond.true:
9006 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9007 // CHECK13:       cond.false:
9008 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9009 // CHECK13-NEXT:    br label [[COND_END]]
9010 // CHECK13:       cond.end:
9011 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
9012 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9013 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9014 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
9015 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9016 // CHECK13:       omp.inner.for.cond:
9017 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9018 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9019 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
9020 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9021 // CHECK13:       omp.inner.for.body:
9022 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9023 // CHECK13-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
9024 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9025 // CHECK13-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
9026 // CHECK13-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
9027 // CHECK13-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9028 // CHECK13-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
9029 // CHECK13-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
9030 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9031 // CHECK13:       omp.inner.for.inc:
9032 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9033 // CHECK13-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
9034 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
9035 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
9036 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9037 // CHECK13:       omp.inner.for.end:
9038 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9039 // CHECK13:       omp.loop.exit:
9040 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
9041 // CHECK13-NEXT:    ret void
9042 //
9043 //
9044 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..3
9045 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
9046 // CHECK13-NEXT:  entry:
9047 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9048 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9049 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
9050 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
9051 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9052 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9053 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
9054 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
9055 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9056 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9057 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9058 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9059 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9060 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9061 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9062 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
9063 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
9064 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9065 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
9066 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9067 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
9068 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
9069 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
9070 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9071 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9072 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9073 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
9074 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9075 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9076 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
9077 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9078 // CHECK13:       cond.true:
9079 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9080 // CHECK13:       cond.false:
9081 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9082 // CHECK13-NEXT:    br label [[COND_END]]
9083 // CHECK13:       cond.end:
9084 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
9085 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
9086 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
9087 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
9088 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9089 // CHECK13:       omp.inner.for.cond:
9090 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9091 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9092 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
9093 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9094 // CHECK13:       omp.inner.for.body:
9095 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9096 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
9097 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
9098 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
9099 // CHECK13-NEXT:    call void @_Z9gtid_testv()
9100 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
9101 // CHECK13:       omp.body.continue:
9102 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9103 // CHECK13:       omp.inner.for.inc:
9104 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9105 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
9106 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
9107 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9108 // CHECK13:       omp.inner.for.end:
9109 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9110 // CHECK13:       omp.loop.exit:
9111 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
9112 // CHECK13-NEXT:    ret void
9113 //
9114 //
9115 // CHECK13-LABEL: define {{[^@]+}}@main
9116 // CHECK13-SAME: () #[[ATTR3:[0-9]+]] {
9117 // CHECK13-NEXT:  entry:
9118 // CHECK13-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
9119 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9120 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
9121 // CHECK13-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
9122 // CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
9123 // CHECK13-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
9124 // CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
9125 // CHECK13-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
9126 // CHECK13-NEXT:    store i32 0, i32* [[RETVAL]], align 4
9127 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
9128 // CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
9129 // CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
9130 // CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
9131 // CHECK13:       omp_offload.failed:
9132 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
9133 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
9134 // CHECK13:       omp_offload.cont:
9135 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
9136 // CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
9137 // CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
9138 // CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
9139 // CHECK13:       omp_offload.failed2:
9140 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
9141 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
9142 // CHECK13:       omp_offload.cont3:
9143 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
9144 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
9145 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
9146 // CHECK13-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
9147 // CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
9148 // CHECK13-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
9149 // CHECK13-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
9150 // CHECK13-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
9151 // CHECK13-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
9152 // CHECK13-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
9153 // CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
9154 // CHECK13-NEXT:    store i8* null, i8** [[TMP10]], align 8
9155 // CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
9156 // CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
9157 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
9158 // CHECK13-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
9159 // CHECK13-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
9160 // CHECK13-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
9161 // CHECK13:       omp_offload.failed5:
9162 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
9163 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
9164 // CHECK13:       omp_offload.cont6:
9165 // CHECK13-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
9166 // CHECK13-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
9167 // CHECK13-NEXT:    ret i32 [[CALL]]
9168 //
9169 //
9170 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
9171 // CHECK13-SAME: () #[[ATTR1]] {
9172 // CHECK13-NEXT:  entry:
9173 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
9174 // CHECK13-NEXT:    ret void
9175 //
9176 //
9177 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..4
9178 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
9179 // CHECK13-NEXT:  entry:
9180 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9181 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9182 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9183 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9184 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
9185 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
9186 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9187 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9188 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9189 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9190 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9191 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
9192 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
9193 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9194 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9195 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9196 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
9197 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9198 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9199 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
9200 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9201 // CHECK13:       cond.true:
9202 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9203 // CHECK13:       cond.false:
9204 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9205 // CHECK13-NEXT:    br label [[COND_END]]
9206 // CHECK13:       cond.end:
9207 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
9208 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9209 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9210 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
9211 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9212 // CHECK13:       omp.inner.for.cond:
9213 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9214 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9215 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
9216 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9217 // CHECK13:       omp.inner.for.body:
9218 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9219 // CHECK13-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
9220 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9221 // CHECK13-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
9222 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
9223 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9224 // CHECK13:       omp.inner.for.inc:
9225 // CHECK13-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9226 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
9227 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
9228 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
9229 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9230 // CHECK13:       omp.inner.for.end:
9231 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9232 // CHECK13:       omp.loop.exit:
9233 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
9234 // CHECK13-NEXT:    ret void
9235 //
9236 //
9237 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..5
9238 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
9239 // CHECK13-NEXT:  entry:
9240 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9241 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9242 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
9243 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
9244 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9245 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9246 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
9247 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
9248 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9249 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9250 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9251 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9252 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9253 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9254 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9255 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
9256 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
9257 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9258 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
9259 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9260 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
9261 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
9262 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
9263 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9264 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9265 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9266 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
9267 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9268 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9269 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
9270 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9271 // CHECK13:       cond.true:
9272 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9273 // CHECK13:       cond.false:
9274 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9275 // CHECK13-NEXT:    br label [[COND_END]]
9276 // CHECK13:       cond.end:
9277 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
9278 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
9279 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
9280 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
9281 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9282 // CHECK13:       omp.inner.for.cond:
9283 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9284 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9285 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
9286 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9287 // CHECK13:       omp.inner.for.body:
9288 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9289 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
9290 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
9291 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
9292 // CHECK13-NEXT:    call void @_Z3fn4v()
9293 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
9294 // CHECK13:       omp.body.continue:
9295 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9296 // CHECK13:       omp.inner.for.inc:
9297 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9298 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
9299 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
9300 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9301 // CHECK13:       omp.inner.for.end:
9302 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9303 // CHECK13:       omp.loop.exit:
9304 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
9305 // CHECK13-NEXT:    ret void
9306 //
9307 //
9308 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
9309 // CHECK13-SAME: () #[[ATTR1]] {
9310 // CHECK13-NEXT:  entry:
9311 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
9312 // CHECK13-NEXT:    ret void
9313 //
9314 //
9315 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..6
9316 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
9317 // CHECK13-NEXT:  entry:
9318 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9319 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9320 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9321 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9322 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
9323 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
9324 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9325 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9326 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9327 // CHECK13-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
9328 // CHECK13-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
9329 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9330 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9331 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
9332 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
9333 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9334 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9335 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9336 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
9337 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9338 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9339 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
9340 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9341 // CHECK13:       cond.true:
9342 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9343 // CHECK13:       cond.false:
9344 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9345 // CHECK13-NEXT:    br label [[COND_END]]
9346 // CHECK13:       cond.end:
9347 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
9348 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9349 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9350 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
9351 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9352 // CHECK13:       omp.inner.for.cond:
9353 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9354 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9355 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
9356 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9357 // CHECK13:       omp.inner.for.body:
9358 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9359 // CHECK13-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
9360 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9361 // CHECK13-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
9362 // CHECK13-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
9363 // CHECK13-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9364 // CHECK13-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
9365 // CHECK13-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
9366 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9367 // CHECK13:       omp.inner.for.inc:
9368 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9369 // CHECK13-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
9370 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
9371 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
9372 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9373 // CHECK13:       omp.inner.for.end:
9374 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9375 // CHECK13:       omp.loop.exit:
9376 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
9377 // CHECK13-NEXT:    ret void
9378 //
9379 //
9380 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..7
9381 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
9382 // CHECK13-NEXT:  entry:
9383 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9384 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9385 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
9386 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
9387 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9388 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9389 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
9390 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
9391 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9392 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9393 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9394 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9395 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9396 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9397 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9398 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
9399 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
9400 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9401 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
9402 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9403 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
9404 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
9405 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
9406 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9407 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9408 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9409 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
9410 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9411 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9412 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
9413 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9414 // CHECK13:       cond.true:
9415 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9416 // CHECK13:       cond.false:
9417 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9418 // CHECK13-NEXT:    br label [[COND_END]]
9419 // CHECK13:       cond.end:
9420 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
9421 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
9422 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
9423 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
9424 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9425 // CHECK13:       omp.inner.for.cond:
9426 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9427 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9428 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
9429 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9430 // CHECK13:       omp.inner.for.body:
9431 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9432 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
9433 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
9434 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
9435 // CHECK13-NEXT:    call void @_Z3fn5v()
9436 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
9437 // CHECK13:       omp.body.continue:
9438 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9439 // CHECK13:       omp.inner.for.inc:
9440 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9441 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
9442 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
9443 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9444 // CHECK13:       omp.inner.for.end:
9445 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9446 // CHECK13:       omp.loop.exit:
9447 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
9448 // CHECK13-NEXT:    ret void
9449 //
9450 //
9451 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
9452 // CHECK13-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
9453 // CHECK13-NEXT:  entry:
9454 // CHECK13-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
9455 // CHECK13-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
9456 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
9457 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
9458 // CHECK13-NEXT:    ret void
9459 //
9460 //
9461 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..8
9462 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
9463 // CHECK13-NEXT:  entry:
9464 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9465 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9466 // CHECK13-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
9467 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9468 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9469 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
9470 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
9471 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9472 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9473 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9474 // CHECK13-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
9475 // CHECK13-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
9476 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9477 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9478 // CHECK13-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
9479 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
9480 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
9481 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
9482 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9483 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9484 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9485 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
9486 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9487 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9488 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
9489 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9490 // CHECK13:       cond.true:
9491 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9492 // CHECK13:       cond.false:
9493 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9494 // CHECK13-NEXT:    br label [[COND_END]]
9495 // CHECK13:       cond.end:
9496 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
9497 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9498 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9499 // CHECK13-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
9500 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9501 // CHECK13:       omp.inner.for.cond:
9502 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9503 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9504 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
9505 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9506 // CHECK13:       omp.inner.for.body:
9507 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9508 // CHECK13-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
9509 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9510 // CHECK13-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
9511 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
9512 // CHECK13-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
9513 // CHECK13-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
9514 // CHECK13:       omp_if.then:
9515 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
9516 // CHECK13-NEXT:    br label [[OMP_IF_END:%.*]]
9517 // CHECK13:       omp_if.else:
9518 // CHECK13-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
9519 // CHECK13-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9520 // CHECK13-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
9521 // CHECK13-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
9522 // CHECK13-NEXT:    br label [[OMP_IF_END]]
9523 // CHECK13:       omp_if.end:
9524 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9525 // CHECK13:       omp.inner.for.inc:
9526 // CHECK13-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9527 // CHECK13-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
9528 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
9529 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
9530 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9531 // CHECK13:       omp.inner.for.end:
9532 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9533 // CHECK13:       omp.loop.exit:
9534 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
9535 // CHECK13-NEXT:    ret void
9536 //
9537 //
9538 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..9
9539 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
9540 // CHECK13-NEXT:  entry:
9541 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9542 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9543 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
9544 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
9545 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9546 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9547 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
9548 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
9549 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9550 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9551 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9552 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9553 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9554 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9555 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9556 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
9557 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
9558 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9559 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
9560 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9561 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
9562 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
9563 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
9564 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9565 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9566 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9567 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
9568 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9569 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9570 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
9571 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9572 // CHECK13:       cond.true:
9573 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9574 // CHECK13:       cond.false:
9575 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9576 // CHECK13-NEXT:    br label [[COND_END]]
9577 // CHECK13:       cond.end:
9578 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
9579 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
9580 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
9581 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
9582 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9583 // CHECK13:       omp.inner.for.cond:
9584 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9585 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9586 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
9587 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9588 // CHECK13:       omp.inner.for.body:
9589 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9590 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
9591 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
9592 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
9593 // CHECK13-NEXT:    call void @_Z3fn6v()
9594 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
9595 // CHECK13:       omp.body.continue:
9596 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9597 // CHECK13:       omp.inner.for.inc:
9598 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9599 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
9600 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
9601 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9602 // CHECK13:       omp.inner.for.end:
9603 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9604 // CHECK13:       omp.loop.exit:
9605 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
9606 // CHECK13-NEXT:    ret void
9607 //
9608 //
9609 // CHECK13-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
9610 // CHECK13-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
9611 // CHECK13-NEXT:  entry:
9612 // CHECK13-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
9613 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9614 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
9615 // CHECK13-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
9616 // CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
9617 // CHECK13-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
9618 // CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
9619 // CHECK13-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
9620 // CHECK13-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
9621 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
9622 // CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
9623 // CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
9624 // CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
9625 // CHECK13:       omp_offload.failed:
9626 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
9627 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
9628 // CHECK13:       omp_offload.cont:
9629 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
9630 // CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
9631 // CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
9632 // CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
9633 // CHECK13:       omp_offload.failed2:
9634 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
9635 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
9636 // CHECK13:       omp_offload.cont3:
9637 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
9638 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
9639 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
9640 // CHECK13-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
9641 // CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
9642 // CHECK13-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
9643 // CHECK13-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
9644 // CHECK13-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
9645 // CHECK13-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
9646 // CHECK13-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
9647 // CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
9648 // CHECK13-NEXT:    store i8* null, i8** [[TMP10]], align 8
9649 // CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
9650 // CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
9651 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
9652 // CHECK13-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
9653 // CHECK13-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
9654 // CHECK13-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
9655 // CHECK13:       omp_offload.failed5:
9656 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
9657 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
9658 // CHECK13:       omp_offload.cont6:
9659 // CHECK13-NEXT:    ret i32 0
9660 //
9661 //
9662 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
9663 // CHECK13-SAME: () #[[ATTR1]] {
9664 // CHECK13-NEXT:  entry:
9665 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
9666 // CHECK13-NEXT:    ret void
9667 //
9668 //
9669 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..10
9670 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
9671 // CHECK13-NEXT:  entry:
9672 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9673 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9674 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9675 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9676 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
9677 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
9678 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9679 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9680 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9681 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9682 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9683 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
9684 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
9685 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9686 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9687 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9688 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
9689 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9690 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9691 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
9692 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9693 // CHECK13:       cond.true:
9694 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9695 // CHECK13:       cond.false:
9696 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9697 // CHECK13-NEXT:    br label [[COND_END]]
9698 // CHECK13:       cond.end:
9699 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
9700 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9701 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9702 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
9703 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9704 // CHECK13:       omp.inner.for.cond:
9705 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9706 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9707 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
9708 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9709 // CHECK13:       omp.inner.for.body:
9710 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9711 // CHECK13-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
9712 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9713 // CHECK13-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
9714 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
9715 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9716 // CHECK13:       omp.inner.for.inc:
9717 // CHECK13-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9718 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
9719 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
9720 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
9721 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9722 // CHECK13:       omp.inner.for.end:
9723 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9724 // CHECK13:       omp.loop.exit:
9725 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
9726 // CHECK13-NEXT:    ret void
9727 //
9728 //
9729 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..11
9730 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
9731 // CHECK13-NEXT:  entry:
9732 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9733 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9734 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
9735 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
9736 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9737 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9738 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
9739 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
9740 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9741 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9742 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9743 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9744 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9745 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9746 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9747 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
9748 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
9749 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9750 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
9751 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9752 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
9753 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
9754 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
9755 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9756 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9757 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9758 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
9759 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9760 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9761 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
9762 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9763 // CHECK13:       cond.true:
9764 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9765 // CHECK13:       cond.false:
9766 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9767 // CHECK13-NEXT:    br label [[COND_END]]
9768 // CHECK13:       cond.end:
9769 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
9770 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
9771 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
9772 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
9773 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9774 // CHECK13:       omp.inner.for.cond:
9775 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9776 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9777 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
9778 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9779 // CHECK13:       omp.inner.for.body:
9780 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9781 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
9782 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
9783 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
9784 // CHECK13-NEXT:    call void @_Z3fn1v()
9785 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
9786 // CHECK13:       omp.body.continue:
9787 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9788 // CHECK13:       omp.inner.for.inc:
9789 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9790 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
9791 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
9792 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9793 // CHECK13:       omp.inner.for.end:
9794 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9795 // CHECK13:       omp.loop.exit:
9796 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
9797 // CHECK13-NEXT:    ret void
9798 //
9799 //
9800 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
9801 // CHECK13-SAME: () #[[ATTR1]] {
9802 // CHECK13-NEXT:  entry:
9803 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
9804 // CHECK13-NEXT:    ret void
9805 //
9806 //
9807 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..12
9808 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
9809 // CHECK13-NEXT:  entry:
9810 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9811 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9812 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9813 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9814 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
9815 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
9816 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9817 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9818 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9819 // CHECK13-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
9820 // CHECK13-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
9821 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9822 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9823 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
9824 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
9825 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9826 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9827 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9828 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
9829 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9830 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9831 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
9832 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9833 // CHECK13:       cond.true:
9834 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9835 // CHECK13:       cond.false:
9836 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9837 // CHECK13-NEXT:    br label [[COND_END]]
9838 // CHECK13:       cond.end:
9839 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
9840 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9841 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9842 // CHECK13-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
9843 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9844 // CHECK13:       omp.inner.for.cond:
9845 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9846 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9847 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
9848 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9849 // CHECK13:       omp.inner.for.body:
9850 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9851 // CHECK13-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
9852 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9853 // CHECK13-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
9854 // CHECK13-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
9855 // CHECK13-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9856 // CHECK13-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
9857 // CHECK13-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
9858 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9859 // CHECK13:       omp.inner.for.inc:
9860 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9861 // CHECK13-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
9862 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
9863 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
9864 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9865 // CHECK13:       omp.inner.for.end:
9866 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9867 // CHECK13:       omp.loop.exit:
9868 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
9869 // CHECK13-NEXT:    ret void
9870 //
9871 //
9872 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..13
9873 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
9874 // CHECK13-NEXT:  entry:
9875 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9876 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9877 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
9878 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
9879 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9880 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9881 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
9882 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
9883 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9884 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9885 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9886 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9887 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9888 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9889 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9890 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
9891 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
9892 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
9893 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
9894 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
9895 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
9896 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
9897 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
9898 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9899 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9900 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9901 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
9902 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9903 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9904 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
9905 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9906 // CHECK13:       cond.true:
9907 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9908 // CHECK13:       cond.false:
9909 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9910 // CHECK13-NEXT:    br label [[COND_END]]
9911 // CHECK13:       cond.end:
9912 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
9913 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
9914 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
9915 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
9916 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9917 // CHECK13:       omp.inner.for.cond:
9918 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9919 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
9920 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
9921 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9922 // CHECK13:       omp.inner.for.body:
9923 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9924 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
9925 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
9926 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
9927 // CHECK13-NEXT:    call void @_Z3fn2v()
9928 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
9929 // CHECK13:       omp.body.continue:
9930 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
9931 // CHECK13:       omp.inner.for.inc:
9932 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9933 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
9934 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
9935 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
9936 // CHECK13:       omp.inner.for.end:
9937 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
9938 // CHECK13:       omp.loop.exit:
9939 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
9940 // CHECK13-NEXT:    ret void
9941 //
9942 //
9943 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
9944 // CHECK13-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
9945 // CHECK13-NEXT:  entry:
9946 // CHECK13-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
9947 // CHECK13-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
9948 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
9949 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
9950 // CHECK13-NEXT:    ret void
9951 //
9952 //
9953 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..14
9954 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
9955 // CHECK13-NEXT:  entry:
9956 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
9957 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
9958 // CHECK13-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
9959 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
9960 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
9961 // CHECK13-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
9962 // CHECK13-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
9963 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
9964 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
9965 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
9966 // CHECK13-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
9967 // CHECK13-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
9968 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
9969 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
9970 // CHECK13-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
9971 // CHECK13-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
9972 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
9973 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
9974 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
9975 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
9976 // CHECK13-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
9977 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
9978 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
9979 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9980 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
9981 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
9982 // CHECK13:       cond.true:
9983 // CHECK13-NEXT:    br label [[COND_END:%.*]]
9984 // CHECK13:       cond.false:
9985 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9986 // CHECK13-NEXT:    br label [[COND_END]]
9987 // CHECK13:       cond.end:
9988 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
9989 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
9990 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
9991 // CHECK13-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
9992 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
9993 // CHECK13:       omp.inner.for.cond:
9994 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
9995 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
9996 // CHECK13-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
9997 // CHECK13-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
9998 // CHECK13:       omp.inner.for.body:
9999 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10000 // CHECK13-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
10001 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10002 // CHECK13-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
10003 // CHECK13-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
10004 // CHECK13-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
10005 // CHECK13-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
10006 // CHECK13:       omp_if.then:
10007 // CHECK13-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
10008 // CHECK13-NEXT:    br label [[OMP_IF_END:%.*]]
10009 // CHECK13:       omp_if.else:
10010 // CHECK13-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
10011 // CHECK13-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10012 // CHECK13-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
10013 // CHECK13-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
10014 // CHECK13-NEXT:    br label [[OMP_IF_END]]
10015 // CHECK13:       omp_if.end:
10016 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10017 // CHECK13:       omp.inner.for.inc:
10018 // CHECK13-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10019 // CHECK13-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
10020 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
10021 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
10022 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
10023 // CHECK13:       omp.inner.for.end:
10024 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10025 // CHECK13:       omp.loop.exit:
10026 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
10027 // CHECK13-NEXT:    ret void
10028 //
10029 //
10030 // CHECK13-LABEL: define {{[^@]+}}@.omp_outlined..15
10031 // CHECK13-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
10032 // CHECK13-NEXT:  entry:
10033 // CHECK13-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10034 // CHECK13-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10035 // CHECK13-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
10036 // CHECK13-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
10037 // CHECK13-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10038 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10039 // CHECK13-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
10040 // CHECK13-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
10041 // CHECK13-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10042 // CHECK13-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10043 // CHECK13-NEXT:    [[I:%.*]] = alloca i32, align 4
10044 // CHECK13-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10045 // CHECK13-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10046 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10047 // CHECK13-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10048 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
10049 // CHECK13-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
10050 // CHECK13-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10051 // CHECK13-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
10052 // CHECK13-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10053 // CHECK13-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
10054 // CHECK13-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
10055 // CHECK13-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
10056 // CHECK13-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10057 // CHECK13-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10058 // CHECK13-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10059 // CHECK13-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
10060 // CHECK13-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10061 // CHECK13-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10062 // CHECK13-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
10063 // CHECK13-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10064 // CHECK13:       cond.true:
10065 // CHECK13-NEXT:    br label [[COND_END:%.*]]
10066 // CHECK13:       cond.false:
10067 // CHECK13-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10068 // CHECK13-NEXT:    br label [[COND_END]]
10069 // CHECK13:       cond.end:
10070 // CHECK13-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
10071 // CHECK13-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
10072 // CHECK13-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
10073 // CHECK13-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
10074 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10075 // CHECK13:       omp.inner.for.cond:
10076 // CHECK13-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10077 // CHECK13-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10078 // CHECK13-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
10079 // CHECK13-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10080 // CHECK13:       omp.inner.for.body:
10081 // CHECK13-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10082 // CHECK13-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
10083 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
10084 // CHECK13-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
10085 // CHECK13-NEXT:    call void @_Z3fn3v()
10086 // CHECK13-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
10087 // CHECK13:       omp.body.continue:
10088 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10089 // CHECK13:       omp.inner.for.inc:
10090 // CHECK13-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10091 // CHECK13-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
10092 // CHECK13-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
10093 // CHECK13-NEXT:    br label [[OMP_INNER_FOR_COND]]
10094 // CHECK13:       omp.inner.for.end:
10095 // CHECK13-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10096 // CHECK13:       omp.loop.exit:
10097 // CHECK13-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
10098 // CHECK13-NEXT:    ret void
10099 //
10100 //
10101 // CHECK13-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
10102 // CHECK13-SAME: () #[[ATTR5:[0-9]+]] {
10103 // CHECK13-NEXT:  entry:
10104 // CHECK13-NEXT:    call void @__tgt_register_requires(i64 1)
10105 // CHECK13-NEXT:    ret void
10106 //
10107 //
10108 // CHECK14-LABEL: define {{[^@]+}}@_Z9gtid_testv
10109 // CHECK14-SAME: () #[[ATTR0:[0-9]+]] {
10110 // CHECK14-NEXT:  entry:
10111 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10112 // CHECK14-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
10113 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
10114 // CHECK14-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
10115 // CHECK14-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
10116 // CHECK14-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10117 // CHECK14:       omp_offload.failed:
10118 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
10119 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10120 // CHECK14:       omp_offload.cont:
10121 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10122 // CHECK14-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
10123 // CHECK14-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
10124 // CHECK14-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
10125 // CHECK14:       omp_offload.failed2:
10126 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
10127 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
10128 // CHECK14:       omp_offload.cont3:
10129 // CHECK14-NEXT:    ret void
10130 //
10131 //
10132 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47
10133 // CHECK14-SAME: () #[[ATTR1:[0-9]+]] {
10134 // CHECK14-NEXT:  entry:
10135 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
10136 // CHECK14-NEXT:    ret void
10137 //
10138 //
10139 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined.
10140 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
10141 // CHECK14-NEXT:  entry:
10142 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10143 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10144 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10145 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10146 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
10147 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
10148 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10149 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10150 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10151 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10152 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10153 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
10154 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
10155 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10156 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10157 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10158 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
10159 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10160 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10161 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
10162 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10163 // CHECK14:       cond.true:
10164 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10165 // CHECK14:       cond.false:
10166 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10167 // CHECK14-NEXT:    br label [[COND_END]]
10168 // CHECK14:       cond.end:
10169 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
10170 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
10171 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10172 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
10173 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10174 // CHECK14:       omp.inner.for.cond:
10175 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10176 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10177 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
10178 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10179 // CHECK14:       omp.inner.for.body:
10180 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10181 // CHECK14-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
10182 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10183 // CHECK14-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
10184 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..1 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
10185 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10186 // CHECK14:       omp.inner.for.inc:
10187 // CHECK14-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10188 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
10189 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
10190 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
10191 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10192 // CHECK14:       omp.inner.for.end:
10193 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10194 // CHECK14:       omp.loop.exit:
10195 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
10196 // CHECK14-NEXT:    ret void
10197 //
10198 //
10199 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..1
10200 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
10201 // CHECK14-NEXT:  entry:
10202 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10203 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10204 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
10205 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
10206 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10207 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10208 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
10209 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
10210 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10211 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10212 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10213 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10214 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10215 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10216 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10217 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
10218 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
10219 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10220 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
10221 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10222 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
10223 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
10224 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
10225 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10226 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10227 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10228 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
10229 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10230 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10231 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
10232 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10233 // CHECK14:       cond.true:
10234 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10235 // CHECK14:       cond.false:
10236 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10237 // CHECK14-NEXT:    br label [[COND_END]]
10238 // CHECK14:       cond.end:
10239 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
10240 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
10241 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
10242 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
10243 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10244 // CHECK14:       omp.inner.for.cond:
10245 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10246 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10247 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
10248 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10249 // CHECK14:       omp.inner.for.body:
10250 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10251 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
10252 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
10253 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
10254 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
10255 // CHECK14:       omp.body.continue:
10256 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10257 // CHECK14:       omp.inner.for.inc:
10258 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10259 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
10260 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
10261 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10262 // CHECK14:       omp.inner.for.end:
10263 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10264 // CHECK14:       omp.loop.exit:
10265 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
10266 // CHECK14-NEXT:    ret void
10267 //
10268 //
10269 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52
10270 // CHECK14-SAME: () #[[ATTR1]] {
10271 // CHECK14-NEXT:  entry:
10272 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..2 to void (i32*, i32*, ...)*))
10273 // CHECK14-NEXT:    ret void
10274 //
10275 //
10276 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..2
10277 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
10278 // CHECK14-NEXT:  entry:
10279 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10280 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10281 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10282 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10283 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
10284 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
10285 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10286 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10287 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10288 // CHECK14-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
10289 // CHECK14-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
10290 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10291 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10292 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
10293 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
10294 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10295 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10296 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10297 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
10298 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10299 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10300 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
10301 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10302 // CHECK14:       cond.true:
10303 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10304 // CHECK14:       cond.false:
10305 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10306 // CHECK14-NEXT:    br label [[COND_END]]
10307 // CHECK14:       cond.end:
10308 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
10309 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
10310 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10311 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
10312 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10313 // CHECK14:       omp.inner.for.cond:
10314 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10315 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10316 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
10317 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10318 // CHECK14:       omp.inner.for.body:
10319 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10320 // CHECK14-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
10321 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10322 // CHECK14-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
10323 // CHECK14-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
10324 // CHECK14-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10325 // CHECK14-NEXT:    call void @.omp_outlined..3(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
10326 // CHECK14-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
10327 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10328 // CHECK14:       omp.inner.for.inc:
10329 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10330 // CHECK14-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
10331 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
10332 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
10333 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10334 // CHECK14:       omp.inner.for.end:
10335 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10336 // CHECK14:       omp.loop.exit:
10337 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
10338 // CHECK14-NEXT:    ret void
10339 //
10340 //
10341 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..3
10342 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
10343 // CHECK14-NEXT:  entry:
10344 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10345 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10346 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
10347 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
10348 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10349 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10350 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
10351 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
10352 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10353 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10354 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10355 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10356 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10357 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10358 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10359 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
10360 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
10361 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10362 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
10363 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10364 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
10365 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
10366 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
10367 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10368 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10369 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10370 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
10371 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10372 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10373 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
10374 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10375 // CHECK14:       cond.true:
10376 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10377 // CHECK14:       cond.false:
10378 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10379 // CHECK14-NEXT:    br label [[COND_END]]
10380 // CHECK14:       cond.end:
10381 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
10382 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
10383 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
10384 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
10385 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10386 // CHECK14:       omp.inner.for.cond:
10387 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10388 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10389 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
10390 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10391 // CHECK14:       omp.inner.for.body:
10392 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10393 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
10394 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
10395 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
10396 // CHECK14-NEXT:    call void @_Z9gtid_testv()
10397 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
10398 // CHECK14:       omp.body.continue:
10399 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10400 // CHECK14:       omp.inner.for.inc:
10401 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10402 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
10403 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
10404 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10405 // CHECK14:       omp.inner.for.end:
10406 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10407 // CHECK14:       omp.loop.exit:
10408 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
10409 // CHECK14-NEXT:    ret void
10410 //
10411 //
10412 // CHECK14-LABEL: define {{[^@]+}}@main
10413 // CHECK14-SAME: () #[[ATTR3:[0-9]+]] {
10414 // CHECK14-NEXT:  entry:
10415 // CHECK14-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
10416 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10417 // CHECK14-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
10418 // CHECK14-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
10419 // CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
10420 // CHECK14-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
10421 // CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
10422 // CHECK14-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
10423 // CHECK14-NEXT:    store i32 0, i32* [[RETVAL]], align 4
10424 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10425 // CHECK14-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
10426 // CHECK14-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
10427 // CHECK14-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10428 // CHECK14:       omp_offload.failed:
10429 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
10430 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10431 // CHECK14:       omp_offload.cont:
10432 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10433 // CHECK14-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
10434 // CHECK14-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
10435 // CHECK14-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
10436 // CHECK14:       omp_offload.failed2:
10437 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
10438 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
10439 // CHECK14:       omp_offload.cont3:
10440 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
10441 // CHECK14-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
10442 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
10443 // CHECK14-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
10444 // CHECK14-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10445 // CHECK14-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
10446 // CHECK14-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
10447 // CHECK14-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10448 // CHECK14-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
10449 // CHECK14-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
10450 // CHECK14-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
10451 // CHECK14-NEXT:    store i8* null, i8** [[TMP10]], align 8
10452 // CHECK14-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10453 // CHECK14-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10454 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10455 // CHECK14-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10456 // CHECK14-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
10457 // CHECK14-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
10458 // CHECK14:       omp_offload.failed5:
10459 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
10460 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
10461 // CHECK14:       omp_offload.cont6:
10462 // CHECK14-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
10463 // CHECK14-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP15]])
10464 // CHECK14-NEXT:    ret i32 [[CALL]]
10465 //
10466 //
10467 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85
10468 // CHECK14-SAME: () #[[ATTR1]] {
10469 // CHECK14-NEXT:  entry:
10470 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*))
10471 // CHECK14-NEXT:    ret void
10472 //
10473 //
10474 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..4
10475 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
10476 // CHECK14-NEXT:  entry:
10477 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10478 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10479 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10480 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10481 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
10482 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
10483 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10484 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10485 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10486 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10487 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10488 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
10489 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
10490 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10491 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10492 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10493 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
10494 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10495 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10496 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
10497 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10498 // CHECK14:       cond.true:
10499 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10500 // CHECK14:       cond.false:
10501 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10502 // CHECK14-NEXT:    br label [[COND_END]]
10503 // CHECK14:       cond.end:
10504 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
10505 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
10506 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10507 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
10508 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10509 // CHECK14:       omp.inner.for.cond:
10510 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10511 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10512 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
10513 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10514 // CHECK14:       omp.inner.for.body:
10515 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10516 // CHECK14-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
10517 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10518 // CHECK14-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
10519 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
10520 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10521 // CHECK14:       omp.inner.for.inc:
10522 // CHECK14-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10523 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
10524 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
10525 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
10526 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10527 // CHECK14:       omp.inner.for.end:
10528 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10529 // CHECK14:       omp.loop.exit:
10530 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
10531 // CHECK14-NEXT:    ret void
10532 //
10533 //
10534 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..5
10535 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
10536 // CHECK14-NEXT:  entry:
10537 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10538 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10539 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
10540 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
10541 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10542 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10543 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
10544 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
10545 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10546 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10547 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10548 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10549 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10550 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10551 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10552 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
10553 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
10554 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10555 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
10556 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10557 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
10558 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
10559 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
10560 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10561 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10562 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10563 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
10564 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10565 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10566 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
10567 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10568 // CHECK14:       cond.true:
10569 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10570 // CHECK14:       cond.false:
10571 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10572 // CHECK14-NEXT:    br label [[COND_END]]
10573 // CHECK14:       cond.end:
10574 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
10575 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
10576 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
10577 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
10578 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10579 // CHECK14:       omp.inner.for.cond:
10580 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10581 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10582 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
10583 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10584 // CHECK14:       omp.inner.for.body:
10585 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10586 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
10587 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
10588 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
10589 // CHECK14-NEXT:    call void @_Z3fn4v()
10590 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
10591 // CHECK14:       omp.body.continue:
10592 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10593 // CHECK14:       omp.inner.for.inc:
10594 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10595 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
10596 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
10597 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10598 // CHECK14:       omp.inner.for.end:
10599 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10600 // CHECK14:       omp.loop.exit:
10601 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
10602 // CHECK14-NEXT:    ret void
10603 //
10604 //
10605 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
10606 // CHECK14-SAME: () #[[ATTR1]] {
10607 // CHECK14-NEXT:  entry:
10608 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*))
10609 // CHECK14-NEXT:    ret void
10610 //
10611 //
10612 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..6
10613 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
10614 // CHECK14-NEXT:  entry:
10615 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10616 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10617 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10618 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10619 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
10620 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
10621 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10622 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10623 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10624 // CHECK14-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
10625 // CHECK14-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
10626 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10627 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10628 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
10629 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
10630 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10631 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10632 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10633 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
10634 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10635 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10636 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
10637 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10638 // CHECK14:       cond.true:
10639 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10640 // CHECK14:       cond.false:
10641 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10642 // CHECK14-NEXT:    br label [[COND_END]]
10643 // CHECK14:       cond.end:
10644 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
10645 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
10646 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10647 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
10648 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10649 // CHECK14:       omp.inner.for.cond:
10650 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10651 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10652 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
10653 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10654 // CHECK14:       omp.inner.for.body:
10655 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10656 // CHECK14-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
10657 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10658 // CHECK14-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
10659 // CHECK14-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
10660 // CHECK14-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10661 // CHECK14-NEXT:    call void @.omp_outlined..7(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
10662 // CHECK14-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
10663 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10664 // CHECK14:       omp.inner.for.inc:
10665 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10666 // CHECK14-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
10667 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
10668 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
10669 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10670 // CHECK14:       omp.inner.for.end:
10671 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10672 // CHECK14:       omp.loop.exit:
10673 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
10674 // CHECK14-NEXT:    ret void
10675 //
10676 //
10677 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..7
10678 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
10679 // CHECK14-NEXT:  entry:
10680 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10681 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10682 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
10683 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
10684 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10685 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10686 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
10687 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
10688 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10689 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10690 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10691 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10692 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10693 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10694 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10695 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
10696 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
10697 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10698 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
10699 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10700 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
10701 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
10702 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
10703 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10704 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10705 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10706 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
10707 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10708 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10709 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
10710 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10711 // CHECK14:       cond.true:
10712 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10713 // CHECK14:       cond.false:
10714 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10715 // CHECK14-NEXT:    br label [[COND_END]]
10716 // CHECK14:       cond.end:
10717 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
10718 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
10719 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
10720 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
10721 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10722 // CHECK14:       omp.inner.for.cond:
10723 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10724 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10725 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
10726 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10727 // CHECK14:       omp.inner.for.body:
10728 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10729 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
10730 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
10731 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
10732 // CHECK14-NEXT:    call void @_Z3fn5v()
10733 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
10734 // CHECK14:       omp.body.continue:
10735 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10736 // CHECK14:       omp.inner.for.inc:
10737 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10738 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
10739 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
10740 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10741 // CHECK14:       omp.inner.for.end:
10742 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10743 // CHECK14:       omp.loop.exit:
10744 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
10745 // CHECK14-NEXT:    ret void
10746 //
10747 //
10748 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103
10749 // CHECK14-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
10750 // CHECK14-NEXT:  entry:
10751 // CHECK14-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
10752 // CHECK14-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
10753 // CHECK14-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
10754 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..8 to void (i32*, i32*, ...)*), i32* [[CONV]])
10755 // CHECK14-NEXT:    ret void
10756 //
10757 //
10758 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..8
10759 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
10760 // CHECK14-NEXT:  entry:
10761 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10762 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10763 // CHECK14-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
10764 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10765 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10766 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
10767 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
10768 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10769 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10770 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10771 // CHECK14-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
10772 // CHECK14-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
10773 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10774 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10775 // CHECK14-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
10776 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
10777 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
10778 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
10779 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10780 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10781 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10782 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
10783 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10784 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10785 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
10786 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10787 // CHECK14:       cond.true:
10788 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10789 // CHECK14:       cond.false:
10790 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10791 // CHECK14-NEXT:    br label [[COND_END]]
10792 // CHECK14:       cond.end:
10793 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
10794 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
10795 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10796 // CHECK14-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
10797 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10798 // CHECK14:       omp.inner.for.cond:
10799 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10800 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10801 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
10802 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10803 // CHECK14:       omp.inner.for.body:
10804 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10805 // CHECK14-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
10806 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10807 // CHECK14-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
10808 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
10809 // CHECK14-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
10810 // CHECK14-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
10811 // CHECK14:       omp_if.then:
10812 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..9 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
10813 // CHECK14-NEXT:    br label [[OMP_IF_END:%.*]]
10814 // CHECK14:       omp_if.else:
10815 // CHECK14-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
10816 // CHECK14-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10817 // CHECK14-NEXT:    call void @.omp_outlined..9(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
10818 // CHECK14-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
10819 // CHECK14-NEXT:    br label [[OMP_IF_END]]
10820 // CHECK14:       omp_if.end:
10821 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10822 // CHECK14:       omp.inner.for.inc:
10823 // CHECK14-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10824 // CHECK14-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
10825 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
10826 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
10827 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10828 // CHECK14:       omp.inner.for.end:
10829 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10830 // CHECK14:       omp.loop.exit:
10831 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
10832 // CHECK14-NEXT:    ret void
10833 //
10834 //
10835 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..9
10836 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
10837 // CHECK14-NEXT:  entry:
10838 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10839 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10840 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
10841 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
10842 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10843 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10844 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
10845 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
10846 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10847 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10848 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10849 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10850 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10851 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10852 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10853 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
10854 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
10855 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
10856 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
10857 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
10858 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
10859 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
10860 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
10861 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10862 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10863 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10864 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
10865 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10866 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10867 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
10868 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10869 // CHECK14:       cond.true:
10870 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10871 // CHECK14:       cond.false:
10872 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10873 // CHECK14-NEXT:    br label [[COND_END]]
10874 // CHECK14:       cond.end:
10875 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
10876 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
10877 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
10878 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
10879 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
10880 // CHECK14:       omp.inner.for.cond:
10881 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10882 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
10883 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
10884 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
10885 // CHECK14:       omp.inner.for.body:
10886 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10887 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
10888 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
10889 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
10890 // CHECK14-NEXT:    call void @_Z3fn6v()
10891 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
10892 // CHECK14:       omp.body.continue:
10893 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
10894 // CHECK14:       omp.inner.for.inc:
10895 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
10896 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
10897 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
10898 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
10899 // CHECK14:       omp.inner.for.end:
10900 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
10901 // CHECK14:       omp.loop.exit:
10902 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
10903 // CHECK14-NEXT:    ret void
10904 //
10905 //
10906 // CHECK14-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
10907 // CHECK14-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
10908 // CHECK14-NEXT:  entry:
10909 // CHECK14-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
10910 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10911 // CHECK14-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
10912 // CHECK14-NEXT:    [[ARG_CASTED:%.*]] = alloca i64, align 8
10913 // CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
10914 // CHECK14-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
10915 // CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
10916 // CHECK14-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
10917 // CHECK14-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
10918 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10919 // CHECK14-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
10920 // CHECK14-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
10921 // CHECK14-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
10922 // CHECK14:       omp_offload.failed:
10923 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
10924 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT]]
10925 // CHECK14:       omp_offload.cont:
10926 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10927 // CHECK14-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
10928 // CHECK14-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
10929 // CHECK14-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
10930 // CHECK14:       omp_offload.failed2:
10931 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
10932 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
10933 // CHECK14:       omp_offload.cont3:
10934 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
10935 // CHECK14-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
10936 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
10937 // CHECK14-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
10938 // CHECK14-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10939 // CHECK14-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
10940 // CHECK14-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
10941 // CHECK14-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10942 // CHECK14-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
10943 // CHECK14-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
10944 // CHECK14-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
10945 // CHECK14-NEXT:    store i8* null, i8** [[TMP10]], align 8
10946 // CHECK14-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
10947 // CHECK14-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
10948 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
10949 // CHECK14-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
10950 // CHECK14-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
10951 // CHECK14-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
10952 // CHECK14:       omp_offload.failed5:
10953 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
10954 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
10955 // CHECK14:       omp_offload.cont6:
10956 // CHECK14-NEXT:    ret i32 0
10957 //
10958 //
10959 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63
10960 // CHECK14-SAME: () #[[ATTR1]] {
10961 // CHECK14-NEXT:  entry:
10962 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..10 to void (i32*, i32*, ...)*))
10963 // CHECK14-NEXT:    ret void
10964 //
10965 //
10966 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..10
10967 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
10968 // CHECK14-NEXT:  entry:
10969 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
10970 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
10971 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
10972 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
10973 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
10974 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
10975 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
10976 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
10977 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
10978 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
10979 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
10980 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
10981 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
10982 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
10983 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
10984 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
10985 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
10986 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
10987 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10988 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
10989 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
10990 // CHECK14:       cond.true:
10991 // CHECK14-NEXT:    br label [[COND_END:%.*]]
10992 // CHECK14:       cond.false:
10993 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
10994 // CHECK14-NEXT:    br label [[COND_END]]
10995 // CHECK14:       cond.end:
10996 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
10997 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
10998 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
10999 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
11000 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
11001 // CHECK14:       omp.inner.for.cond:
11002 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11003 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11004 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
11005 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
11006 // CHECK14:       omp.inner.for.body:
11007 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
11008 // CHECK14-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
11009 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11010 // CHECK14-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
11011 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..11 to void (i32*, i32*, ...)*), i64 [[TMP8]], i64 [[TMP10]])
11012 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
11013 // CHECK14:       omp.inner.for.inc:
11014 // CHECK14-NEXT:    [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11015 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
11016 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP11]], [[TMP12]]
11017 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
11018 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
11019 // CHECK14:       omp.inner.for.end:
11020 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
11021 // CHECK14:       omp.loop.exit:
11022 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
11023 // CHECK14-NEXT:    ret void
11024 //
11025 //
11026 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..11
11027 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
11028 // CHECK14-NEXT:  entry:
11029 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11030 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11031 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
11032 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
11033 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
11034 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
11035 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
11036 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
11037 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
11038 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
11039 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
11040 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11041 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11042 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
11043 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
11044 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
11045 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
11046 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
11047 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
11048 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
11049 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
11050 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
11051 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
11052 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
11053 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
11054 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11055 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
11056 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
11057 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11058 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
11059 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
11060 // CHECK14:       cond.true:
11061 // CHECK14-NEXT:    br label [[COND_END:%.*]]
11062 // CHECK14:       cond.false:
11063 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11064 // CHECK14-NEXT:    br label [[COND_END]]
11065 // CHECK14:       cond.end:
11066 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
11067 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
11068 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
11069 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
11070 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
11071 // CHECK14:       omp.inner.for.cond:
11072 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11073 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11074 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
11075 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
11076 // CHECK14:       omp.inner.for.body:
11077 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11078 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
11079 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
11080 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
11081 // CHECK14-NEXT:    call void @_Z3fn1v()
11082 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
11083 // CHECK14:       omp.body.continue:
11084 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
11085 // CHECK14:       omp.inner.for.inc:
11086 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11087 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
11088 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
11089 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
11090 // CHECK14:       omp.inner.for.end:
11091 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
11092 // CHECK14:       omp.loop.exit:
11093 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
11094 // CHECK14-NEXT:    ret void
11095 //
11096 //
11097 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69
11098 // CHECK14-SAME: () #[[ATTR1]] {
11099 // CHECK14-NEXT:  entry:
11100 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined..12 to void (i32*, i32*, ...)*))
11101 // CHECK14-NEXT:    ret void
11102 //
11103 //
11104 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..12
11105 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR1]] {
11106 // CHECK14-NEXT:  entry:
11107 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11108 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11109 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
11110 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
11111 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
11112 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
11113 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
11114 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
11115 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
11116 // CHECK14-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
11117 // CHECK14-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
11118 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11119 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11120 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
11121 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
11122 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
11123 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
11124 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11125 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
11126 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
11127 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11128 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP2]], 99
11129 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
11130 // CHECK14:       cond.true:
11131 // CHECK14-NEXT:    br label [[COND_END:%.*]]
11132 // CHECK14:       cond.false:
11133 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11134 // CHECK14-NEXT:    br label [[COND_END]]
11135 // CHECK14:       cond.end:
11136 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP3]], [[COND_FALSE]] ]
11137 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
11138 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
11139 // CHECK14-NEXT:    store i32 [[TMP4]], i32* [[DOTOMP_IV]], align 4
11140 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
11141 // CHECK14:       omp.inner.for.cond:
11142 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11143 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11144 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP5]], [[TMP6]]
11145 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
11146 // CHECK14:       omp.inner.for.body:
11147 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
11148 // CHECK14-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
11149 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11150 // CHECK14-NEXT:    [[TMP10:%.*]] = zext i32 [[TMP9]] to i64
11151 // CHECK14-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
11152 // CHECK14-NEXT:    [[TMP11:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11153 // CHECK14-NEXT:    call void @.omp_outlined..13(i32* [[TMP11]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR2]]
11154 // CHECK14-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP1]])
11155 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
11156 // CHECK14:       omp.inner.for.inc:
11157 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11158 // CHECK14-NEXT:    [[TMP13:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
11159 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP12]], [[TMP13]]
11160 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
11161 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
11162 // CHECK14:       omp.inner.for.end:
11163 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
11164 // CHECK14:       omp.loop.exit:
11165 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP1]])
11166 // CHECK14-NEXT:    ret void
11167 //
11168 //
11169 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..13
11170 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
11171 // CHECK14-NEXT:  entry:
11172 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11173 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11174 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
11175 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
11176 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
11177 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
11178 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
11179 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
11180 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
11181 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
11182 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
11183 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11184 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11185 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
11186 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
11187 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
11188 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
11189 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
11190 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
11191 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
11192 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
11193 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
11194 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
11195 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
11196 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
11197 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11198 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
11199 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
11200 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11201 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
11202 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
11203 // CHECK14:       cond.true:
11204 // CHECK14-NEXT:    br label [[COND_END:%.*]]
11205 // CHECK14:       cond.false:
11206 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11207 // CHECK14-NEXT:    br label [[COND_END]]
11208 // CHECK14:       cond.end:
11209 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
11210 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
11211 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
11212 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
11213 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
11214 // CHECK14:       omp.inner.for.cond:
11215 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11216 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11217 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
11218 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
11219 // CHECK14:       omp.inner.for.body:
11220 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11221 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
11222 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
11223 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
11224 // CHECK14-NEXT:    call void @_Z3fn2v()
11225 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
11226 // CHECK14:       omp.body.continue:
11227 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
11228 // CHECK14:       omp.inner.for.inc:
11229 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11230 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
11231 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
11232 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
11233 // CHECK14:       omp.inner.for.end:
11234 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
11235 // CHECK14:       omp.loop.exit:
11236 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
11237 // CHECK14-NEXT:    ret void
11238 //
11239 //
11240 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75
11241 // CHECK14-SAME: (i64 [[ARG:%.*]]) #[[ATTR1]] {
11242 // CHECK14-NEXT:  entry:
11243 // CHECK14-NEXT:    [[ARG_ADDR:%.*]] = alloca i64, align 8
11244 // CHECK14-NEXT:    store i64 [[ARG]], i64* [[ARG_ADDR]], align 8
11245 // CHECK14-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_ADDR]] to i32*
11246 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_teams(%struct.ident_t* @[[GLOB3]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i32*)* @.omp_outlined..14 to void (i32*, i32*, ...)*), i32* [[CONV]])
11247 // CHECK14-NEXT:    ret void
11248 //
11249 //
11250 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..14
11251 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ARG:%.*]]) #[[ATTR1]] {
11252 // CHECK14-NEXT:  entry:
11253 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11254 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11255 // CHECK14-NEXT:    [[ARG_ADDR:%.*]] = alloca i32*, align 8
11256 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
11257 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
11258 // CHECK14-NEXT:    [[DOTOMP_COMB_LB:%.*]] = alloca i32, align 4
11259 // CHECK14-NEXT:    [[DOTOMP_COMB_UB:%.*]] = alloca i32, align 4
11260 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
11261 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
11262 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
11263 // CHECK14-NEXT:    [[DOTBOUND_ZERO_ADDR:%.*]] = alloca i32, align 4
11264 // CHECK14-NEXT:    store i32 0, i32* [[DOTBOUND_ZERO_ADDR]], align 4
11265 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11266 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11267 // CHECK14-NEXT:    store i32* [[ARG]], i32** [[ARG_ADDR]], align 8
11268 // CHECK14-NEXT:    [[TMP0:%.*]] = load i32*, i32** [[ARG_ADDR]], align 8
11269 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_COMB_LB]], align 4
11270 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_COMB_UB]], align 4
11271 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
11272 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
11273 // CHECK14-NEXT:    [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11274 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
11275 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_COMB_LB]], i32* [[DOTOMP_COMB_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
11276 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11277 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP3]], 99
11278 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
11279 // CHECK14:       cond.true:
11280 // CHECK14-NEXT:    br label [[COND_END:%.*]]
11281 // CHECK14:       cond.false:
11282 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11283 // CHECK14-NEXT:    br label [[COND_END]]
11284 // CHECK14:       cond.end:
11285 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP4]], [[COND_FALSE]] ]
11286 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_COMB_UB]], align 4
11287 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
11288 // CHECK14-NEXT:    store i32 [[TMP5]], i32* [[DOTOMP_IV]], align 4
11289 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
11290 // CHECK14:       omp.inner.for.cond:
11291 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11292 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11293 // CHECK14-NEXT:    [[CMP1:%.*]] = icmp sle i32 [[TMP6]], [[TMP7]]
11294 // CHECK14-NEXT:    br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
11295 // CHECK14:       omp.inner.for.body:
11296 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_COMB_LB]], align 4
11297 // CHECK14-NEXT:    [[TMP9:%.*]] = zext i32 [[TMP8]] to i64
11298 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_COMB_UB]], align 4
11299 // CHECK14-NEXT:    [[TMP11:%.*]] = zext i32 [[TMP10]] to i64
11300 // CHECK14-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP0]], align 4
11301 // CHECK14-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP12]], 0
11302 // CHECK14-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
11303 // CHECK14:       omp_if.then:
11304 // CHECK14-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB3]], i32 2, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, i64, i64)* @.omp_outlined..15 to void (i32*, i32*, ...)*), i64 [[TMP9]], i64 [[TMP11]])
11305 // CHECK14-NEXT:    br label [[OMP_IF_END:%.*]]
11306 // CHECK14:       omp_if.else:
11307 // CHECK14-NEXT:    call void @__kmpc_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
11308 // CHECK14-NEXT:    [[TMP13:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11309 // CHECK14-NEXT:    call void @.omp_outlined..15(i32* [[TMP13]], i32* [[DOTBOUND_ZERO_ADDR]], i64 [[TMP9]], i64 [[TMP11]]) #[[ATTR2]]
11310 // CHECK14-NEXT:    call void @__kmpc_end_serialized_parallel(%struct.ident_t* @[[GLOB3]], i32 [[TMP2]])
11311 // CHECK14-NEXT:    br label [[OMP_IF_END]]
11312 // CHECK14:       omp_if.end:
11313 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
11314 // CHECK14:       omp.inner.for.inc:
11315 // CHECK14-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11316 // CHECK14-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4
11317 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP14]], [[TMP15]]
11318 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[DOTOMP_IV]], align 4
11319 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
11320 // CHECK14:       omp.inner.for.end:
11321 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
11322 // CHECK14:       omp.loop.exit:
11323 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
11324 // CHECK14-NEXT:    ret void
11325 //
11326 //
11327 // CHECK14-LABEL: define {{[^@]+}}@.omp_outlined..15
11328 // CHECK14-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i64 [[DOTPREVIOUS_LB_:%.*]], i64 [[DOTPREVIOUS_UB_:%.*]]) #[[ATTR1]] {
11329 // CHECK14-NEXT:  entry:
11330 // CHECK14-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
11331 // CHECK14-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
11332 // CHECK14-NEXT:    [[DOTPREVIOUS_LB__ADDR:%.*]] = alloca i64, align 8
11333 // CHECK14-NEXT:    [[DOTPREVIOUS_UB__ADDR:%.*]] = alloca i64, align 8
11334 // CHECK14-NEXT:    [[DOTOMP_IV:%.*]] = alloca i32, align 4
11335 // CHECK14-NEXT:    [[TMP:%.*]] = alloca i32, align 4
11336 // CHECK14-NEXT:    [[DOTOMP_LB:%.*]] = alloca i32, align 4
11337 // CHECK14-NEXT:    [[DOTOMP_UB:%.*]] = alloca i32, align 4
11338 // CHECK14-NEXT:    [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
11339 // CHECK14-NEXT:    [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
11340 // CHECK14-NEXT:    [[I:%.*]] = alloca i32, align 4
11341 // CHECK14-NEXT:    store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
11342 // CHECK14-NEXT:    store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
11343 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_LB_]], i64* [[DOTPREVIOUS_LB__ADDR]], align 8
11344 // CHECK14-NEXT:    store i64 [[DOTPREVIOUS_UB_]], i64* [[DOTPREVIOUS_UB__ADDR]], align 8
11345 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_LB]], align 4
11346 // CHECK14-NEXT:    store i32 99, i32* [[DOTOMP_UB]], align 4
11347 // CHECK14-NEXT:    [[TMP0:%.*]] = load i64, i64* [[DOTPREVIOUS_LB__ADDR]], align 8
11348 // CHECK14-NEXT:    [[CONV:%.*]] = trunc i64 [[TMP0]] to i32
11349 // CHECK14-NEXT:    [[TMP1:%.*]] = load i64, i64* [[DOTPREVIOUS_UB__ADDR]], align 8
11350 // CHECK14-NEXT:    [[CONV1:%.*]] = trunc i64 [[TMP1]] to i32
11351 // CHECK14-NEXT:    store i32 [[CONV]], i32* [[DOTOMP_LB]], align 4
11352 // CHECK14-NEXT:    store i32 [[CONV1]], i32* [[DOTOMP_UB]], align 4
11353 // CHECK14-NEXT:    store i32 1, i32* [[DOTOMP_STRIDE]], align 4
11354 // CHECK14-NEXT:    store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
11355 // CHECK14-NEXT:    [[TMP2:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
11356 // CHECK14-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
11357 // CHECK14-NEXT:    call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP3]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
11358 // CHECK14-NEXT:    [[TMP4:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11359 // CHECK14-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP4]], 99
11360 // CHECK14-NEXT:    br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
11361 // CHECK14:       cond.true:
11362 // CHECK14-NEXT:    br label [[COND_END:%.*]]
11363 // CHECK14:       cond.false:
11364 // CHECK14-NEXT:    [[TMP5:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11365 // CHECK14-NEXT:    br label [[COND_END]]
11366 // CHECK14:       cond.end:
11367 // CHECK14-NEXT:    [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP5]], [[COND_FALSE]] ]
11368 // CHECK14-NEXT:    store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
11369 // CHECK14-NEXT:    [[TMP6:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
11370 // CHECK14-NEXT:    store i32 [[TMP6]], i32* [[DOTOMP_IV]], align 4
11371 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND:%.*]]
11372 // CHECK14:       omp.inner.for.cond:
11373 // CHECK14-NEXT:    [[TMP7:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11374 // CHECK14-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
11375 // CHECK14-NEXT:    [[CMP2:%.*]] = icmp sle i32 [[TMP7]], [[TMP8]]
11376 // CHECK14-NEXT:    br i1 [[CMP2]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
11377 // CHECK14:       omp.inner.for.body:
11378 // CHECK14-NEXT:    [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11379 // CHECK14-NEXT:    [[MUL:%.*]] = mul nsw i32 [[TMP9]], 1
11380 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 0, [[MUL]]
11381 // CHECK14-NEXT:    store i32 [[ADD]], i32* [[I]], align 4
11382 // CHECK14-NEXT:    call void @_Z3fn3v()
11383 // CHECK14-NEXT:    br label [[OMP_BODY_CONTINUE:%.*]]
11384 // CHECK14:       omp.body.continue:
11385 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_INC:%.*]]
11386 // CHECK14:       omp.inner.for.inc:
11387 // CHECK14-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4
11388 // CHECK14-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP10]], 1
11389 // CHECK14-NEXT:    store i32 [[ADD3]], i32* [[DOTOMP_IV]], align 4
11390 // CHECK14-NEXT:    br label [[OMP_INNER_FOR_COND]]
11391 // CHECK14:       omp.inner.for.end:
11392 // CHECK14-NEXT:    br label [[OMP_LOOP_EXIT:%.*]]
11393 // CHECK14:       omp.loop.exit:
11394 // CHECK14-NEXT:    call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP3]])
11395 // CHECK14-NEXT:    ret void
11396 //
11397 //
11398 // CHECK14-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
11399 // CHECK14-SAME: () #[[ATTR5:[0-9]+]] {
11400 // CHECK14-NEXT:  entry:
11401 // CHECK14-NEXT:    call void @__tgt_register_requires(i64 1)
11402 // CHECK14-NEXT:    ret void
11403 //
11404 //
11405 // CHECK15-LABEL: define {{[^@]+}}@_Z9gtid_testv
11406 // CHECK15-SAME: () #[[ATTR0:[0-9]+]] {
11407 // CHECK15-NEXT:  entry:
11408 // CHECK15-NEXT:    [[I:%.*]] = alloca i32, align 4
11409 // CHECK15-NEXT:    [[I1:%.*]] = alloca i32, align 4
11410 // CHECK15-NEXT:    store i32 0, i32* [[I]], align 4
11411 // CHECK15-NEXT:    br label [[FOR_COND:%.*]]
11412 // CHECK15:       for.cond:
11413 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
11414 // CHECK15-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
11415 // CHECK15-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
11416 // CHECK15:       for.body:
11417 // CHECK15-NEXT:    br label [[FOR_INC:%.*]]
11418 // CHECK15:       for.inc:
11419 // CHECK15-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
11420 // CHECK15-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
11421 // CHECK15-NEXT:    store i32 [[INC]], i32* [[I]], align 4
11422 // CHECK15-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
11423 // CHECK15:       for.end:
11424 // CHECK15-NEXT:    store i32 0, i32* [[I1]], align 4
11425 // CHECK15-NEXT:    br label [[FOR_COND2:%.*]]
11426 // CHECK15:       for.cond2:
11427 // CHECK15-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
11428 // CHECK15-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
11429 // CHECK15-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
11430 // CHECK15:       for.body4:
11431 // CHECK15-NEXT:    call void @_Z9gtid_testv()
11432 // CHECK15-NEXT:    br label [[FOR_INC5:%.*]]
11433 // CHECK15:       for.inc5:
11434 // CHECK15-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
11435 // CHECK15-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
11436 // CHECK15-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
11437 // CHECK15-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP8:![0-9]+]]
11438 // CHECK15:       for.end7:
11439 // CHECK15-NEXT:    ret void
11440 //
11441 //
11442 // CHECK15-LABEL: define {{[^@]+}}@main
11443 // CHECK15-SAME: () #[[ATTR1:[0-9]+]] {
11444 // CHECK15-NEXT:  entry:
11445 // CHECK15-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
11446 // CHECK15-NEXT:    [[I:%.*]] = alloca i32, align 4
11447 // CHECK15-NEXT:    [[I1:%.*]] = alloca i32, align 4
11448 // CHECK15-NEXT:    [[I8:%.*]] = alloca i32, align 4
11449 // CHECK15-NEXT:    store i32 0, i32* [[RETVAL]], align 4
11450 // CHECK15-NEXT:    store i32 0, i32* [[I]], align 4
11451 // CHECK15-NEXT:    br label [[FOR_COND:%.*]]
11452 // CHECK15:       for.cond:
11453 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
11454 // CHECK15-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
11455 // CHECK15-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
11456 // CHECK15:       for.body:
11457 // CHECK15-NEXT:    call void @_Z3fn4v()
11458 // CHECK15-NEXT:    br label [[FOR_INC:%.*]]
11459 // CHECK15:       for.inc:
11460 // CHECK15-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
11461 // CHECK15-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
11462 // CHECK15-NEXT:    store i32 [[INC]], i32* [[I]], align 4
11463 // CHECK15-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
11464 // CHECK15:       for.end:
11465 // CHECK15-NEXT:    store i32 0, i32* [[I1]], align 4
11466 // CHECK15-NEXT:    br label [[FOR_COND2:%.*]]
11467 // CHECK15:       for.cond2:
11468 // CHECK15-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
11469 // CHECK15-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
11470 // CHECK15-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
11471 // CHECK15:       for.body4:
11472 // CHECK15-NEXT:    call void @_Z3fn5v()
11473 // CHECK15-NEXT:    br label [[FOR_INC5:%.*]]
11474 // CHECK15:       for.inc5:
11475 // CHECK15-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
11476 // CHECK15-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
11477 // CHECK15-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
11478 // CHECK15-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP10:![0-9]+]]
11479 // CHECK15:       for.end7:
11480 // CHECK15-NEXT:    store i32 0, i32* [[I8]], align 4
11481 // CHECK15-NEXT:    br label [[FOR_COND9:%.*]]
11482 // CHECK15:       for.cond9:
11483 // CHECK15-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
11484 // CHECK15-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
11485 // CHECK15-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
11486 // CHECK15:       for.body11:
11487 // CHECK15-NEXT:    call void @_Z3fn6v()
11488 // CHECK15-NEXT:    br label [[FOR_INC12:%.*]]
11489 // CHECK15:       for.inc12:
11490 // CHECK15-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
11491 // CHECK15-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
11492 // CHECK15-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
11493 // CHECK15-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP11:![0-9]+]]
11494 // CHECK15:       for.end14:
11495 // CHECK15-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
11496 // CHECK15-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
11497 // CHECK15-NEXT:    ret i32 [[CALL]]
11498 //
11499 //
11500 // CHECK15-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
11501 // CHECK15-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
11502 // CHECK15-NEXT:  entry:
11503 // CHECK15-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
11504 // CHECK15-NEXT:    [[I:%.*]] = alloca i32, align 4
11505 // CHECK15-NEXT:    [[I1:%.*]] = alloca i32, align 4
11506 // CHECK15-NEXT:    [[I8:%.*]] = alloca i32, align 4
11507 // CHECK15-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
11508 // CHECK15-NEXT:    store i32 0, i32* [[I]], align 4
11509 // CHECK15-NEXT:    br label [[FOR_COND:%.*]]
11510 // CHECK15:       for.cond:
11511 // CHECK15-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
11512 // CHECK15-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
11513 // CHECK15-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
11514 // CHECK15:       for.body:
11515 // CHECK15-NEXT:    call void @_Z3fn1v()
11516 // CHECK15-NEXT:    br label [[FOR_INC:%.*]]
11517 // CHECK15:       for.inc:
11518 // CHECK15-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
11519 // CHECK15-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
11520 // CHECK15-NEXT:    store i32 [[INC]], i32* [[I]], align 4
11521 // CHECK15-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
11522 // CHECK15:       for.end:
11523 // CHECK15-NEXT:    store i32 0, i32* [[I1]], align 4
11524 // CHECK15-NEXT:    br label [[FOR_COND2:%.*]]
11525 // CHECK15:       for.cond2:
11526 // CHECK15-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
11527 // CHECK15-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
11528 // CHECK15-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
11529 // CHECK15:       for.body4:
11530 // CHECK15-NEXT:    call void @_Z3fn2v()
11531 // CHECK15-NEXT:    br label [[FOR_INC5:%.*]]
11532 // CHECK15:       for.inc5:
11533 // CHECK15-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
11534 // CHECK15-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
11535 // CHECK15-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
11536 // CHECK15-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP13:![0-9]+]]
11537 // CHECK15:       for.end7:
11538 // CHECK15-NEXT:    store i32 0, i32* [[I8]], align 4
11539 // CHECK15-NEXT:    br label [[FOR_COND9:%.*]]
11540 // CHECK15:       for.cond9:
11541 // CHECK15-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
11542 // CHECK15-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
11543 // CHECK15-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
11544 // CHECK15:       for.body11:
11545 // CHECK15-NEXT:    call void @_Z3fn3v()
11546 // CHECK15-NEXT:    br label [[FOR_INC12:%.*]]
11547 // CHECK15:       for.inc12:
11548 // CHECK15-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
11549 // CHECK15-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
11550 // CHECK15-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
11551 // CHECK15-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP14:![0-9]+]]
11552 // CHECK15:       for.end14:
11553 // CHECK15-NEXT:    ret i32 0
11554 //
11555 //
11556 // CHECK16-LABEL: define {{[^@]+}}@_Z9gtid_testv
11557 // CHECK16-SAME: () #[[ATTR0:[0-9]+]] {
11558 // CHECK16-NEXT:  entry:
11559 // CHECK16-NEXT:    [[I:%.*]] = alloca i32, align 4
11560 // CHECK16-NEXT:    [[I1:%.*]] = alloca i32, align 4
11561 // CHECK16-NEXT:    store i32 0, i32* [[I]], align 4
11562 // CHECK16-NEXT:    br label [[FOR_COND:%.*]]
11563 // CHECK16:       for.cond:
11564 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
11565 // CHECK16-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
11566 // CHECK16-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
11567 // CHECK16:       for.body:
11568 // CHECK16-NEXT:    br label [[FOR_INC:%.*]]
11569 // CHECK16:       for.inc:
11570 // CHECK16-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
11571 // CHECK16-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
11572 // CHECK16-NEXT:    store i32 [[INC]], i32* [[I]], align 4
11573 // CHECK16-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
11574 // CHECK16:       for.end:
11575 // CHECK16-NEXT:    store i32 0, i32* [[I1]], align 4
11576 // CHECK16-NEXT:    br label [[FOR_COND2:%.*]]
11577 // CHECK16:       for.cond2:
11578 // CHECK16-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
11579 // CHECK16-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
11580 // CHECK16-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
11581 // CHECK16:       for.body4:
11582 // CHECK16-NEXT:    call void @_Z9gtid_testv()
11583 // CHECK16-NEXT:    br label [[FOR_INC5:%.*]]
11584 // CHECK16:       for.inc5:
11585 // CHECK16-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
11586 // CHECK16-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
11587 // CHECK16-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
11588 // CHECK16-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP8:![0-9]+]]
11589 // CHECK16:       for.end7:
11590 // CHECK16-NEXT:    ret void
11591 //
11592 //
11593 // CHECK16-LABEL: define {{[^@]+}}@main
11594 // CHECK16-SAME: () #[[ATTR1:[0-9]+]] {
11595 // CHECK16-NEXT:  entry:
11596 // CHECK16-NEXT:    [[RETVAL:%.*]] = alloca i32, align 4
11597 // CHECK16-NEXT:    [[I:%.*]] = alloca i32, align 4
11598 // CHECK16-NEXT:    [[I1:%.*]] = alloca i32, align 4
11599 // CHECK16-NEXT:    [[I8:%.*]] = alloca i32, align 4
11600 // CHECK16-NEXT:    store i32 0, i32* [[RETVAL]], align 4
11601 // CHECK16-NEXT:    store i32 0, i32* [[I]], align 4
11602 // CHECK16-NEXT:    br label [[FOR_COND:%.*]]
11603 // CHECK16:       for.cond:
11604 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
11605 // CHECK16-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
11606 // CHECK16-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
11607 // CHECK16:       for.body:
11608 // CHECK16-NEXT:    call void @_Z3fn4v()
11609 // CHECK16-NEXT:    br label [[FOR_INC:%.*]]
11610 // CHECK16:       for.inc:
11611 // CHECK16-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
11612 // CHECK16-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
11613 // CHECK16-NEXT:    store i32 [[INC]], i32* [[I]], align 4
11614 // CHECK16-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
11615 // CHECK16:       for.end:
11616 // CHECK16-NEXT:    store i32 0, i32* [[I1]], align 4
11617 // CHECK16-NEXT:    br label [[FOR_COND2:%.*]]
11618 // CHECK16:       for.cond2:
11619 // CHECK16-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
11620 // CHECK16-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
11621 // CHECK16-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
11622 // CHECK16:       for.body4:
11623 // CHECK16-NEXT:    call void @_Z3fn5v()
11624 // CHECK16-NEXT:    br label [[FOR_INC5:%.*]]
11625 // CHECK16:       for.inc5:
11626 // CHECK16-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
11627 // CHECK16-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
11628 // CHECK16-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
11629 // CHECK16-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP10:![0-9]+]]
11630 // CHECK16:       for.end7:
11631 // CHECK16-NEXT:    store i32 0, i32* [[I8]], align 4
11632 // CHECK16-NEXT:    br label [[FOR_COND9:%.*]]
11633 // CHECK16:       for.cond9:
11634 // CHECK16-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
11635 // CHECK16-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
11636 // CHECK16-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
11637 // CHECK16:       for.body11:
11638 // CHECK16-NEXT:    call void @_Z3fn6v()
11639 // CHECK16-NEXT:    br label [[FOR_INC12:%.*]]
11640 // CHECK16:       for.inc12:
11641 // CHECK16-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
11642 // CHECK16-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
11643 // CHECK16-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
11644 // CHECK16-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP11:![0-9]+]]
11645 // CHECK16:       for.end14:
11646 // CHECK16-NEXT:    [[TMP6:%.*]] = load i32, i32* @Arg, align 4
11647 // CHECK16-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiEiT_(i32 [[TMP6]])
11648 // CHECK16-NEXT:    ret i32 [[CALL]]
11649 //
11650 //
11651 // CHECK16-LABEL: define {{[^@]+}}@_Z5tmainIiEiT_
11652 // CHECK16-SAME: (i32 [[ARG:%.*]]) #[[ATTR0]] comdat {
11653 // CHECK16-NEXT:  entry:
11654 // CHECK16-NEXT:    [[ARG_ADDR:%.*]] = alloca i32, align 4
11655 // CHECK16-NEXT:    [[I:%.*]] = alloca i32, align 4
11656 // CHECK16-NEXT:    [[I1:%.*]] = alloca i32, align 4
11657 // CHECK16-NEXT:    [[I8:%.*]] = alloca i32, align 4
11658 // CHECK16-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
11659 // CHECK16-NEXT:    store i32 0, i32* [[I]], align 4
11660 // CHECK16-NEXT:    br label [[FOR_COND:%.*]]
11661 // CHECK16:       for.cond:
11662 // CHECK16-NEXT:    [[TMP0:%.*]] = load i32, i32* [[I]], align 4
11663 // CHECK16-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP0]], 100
11664 // CHECK16-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
11665 // CHECK16:       for.body:
11666 // CHECK16-NEXT:    call void @_Z3fn1v()
11667 // CHECK16-NEXT:    br label [[FOR_INC:%.*]]
11668 // CHECK16:       for.inc:
11669 // CHECK16-NEXT:    [[TMP1:%.*]] = load i32, i32* [[I]], align 4
11670 // CHECK16-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP1]], 1
11671 // CHECK16-NEXT:    store i32 [[INC]], i32* [[I]], align 4
11672 // CHECK16-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
11673 // CHECK16:       for.end:
11674 // CHECK16-NEXT:    store i32 0, i32* [[I1]], align 4
11675 // CHECK16-NEXT:    br label [[FOR_COND2:%.*]]
11676 // CHECK16:       for.cond2:
11677 // CHECK16-NEXT:    [[TMP2:%.*]] = load i32, i32* [[I1]], align 4
11678 // CHECK16-NEXT:    [[CMP3:%.*]] = icmp slt i32 [[TMP2]], 100
11679 // CHECK16-NEXT:    br i1 [[CMP3]], label [[FOR_BODY4:%.*]], label [[FOR_END7:%.*]]
11680 // CHECK16:       for.body4:
11681 // CHECK16-NEXT:    call void @_Z3fn2v()
11682 // CHECK16-NEXT:    br label [[FOR_INC5:%.*]]
11683 // CHECK16:       for.inc5:
11684 // CHECK16-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I1]], align 4
11685 // CHECK16-NEXT:    [[INC6:%.*]] = add nsw i32 [[TMP3]], 1
11686 // CHECK16-NEXT:    store i32 [[INC6]], i32* [[I1]], align 4
11687 // CHECK16-NEXT:    br label [[FOR_COND2]], !llvm.loop [[LOOP13:![0-9]+]]
11688 // CHECK16:       for.end7:
11689 // CHECK16-NEXT:    store i32 0, i32* [[I8]], align 4
11690 // CHECK16-NEXT:    br label [[FOR_COND9:%.*]]
11691 // CHECK16:       for.cond9:
11692 // CHECK16-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I8]], align 4
11693 // CHECK16-NEXT:    [[CMP10:%.*]] = icmp slt i32 [[TMP4]], 100
11694 // CHECK16-NEXT:    br i1 [[CMP10]], label [[FOR_BODY11:%.*]], label [[FOR_END14:%.*]]
11695 // CHECK16:       for.body11:
11696 // CHECK16-NEXT:    call void @_Z3fn3v()
11697 // CHECK16-NEXT:    br label [[FOR_INC12:%.*]]
11698 // CHECK16:       for.inc12:
11699 // CHECK16-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I8]], align 4
11700 // CHECK16-NEXT:    [[INC13:%.*]] = add nsw i32 [[TMP5]], 1
11701 // CHECK16-NEXT:    store i32 [[INC13]], i32* [[I8]], align 4
11702 // CHECK16-NEXT:    br label [[FOR_COND9]], !llvm.loop [[LOOP14:![0-9]+]]
11703 // CHECK16:       for.end14:
11704 // CHECK16-NEXT:    ret i32 0
11705 //
11706