1; RUN: opt -S -codegenprepare < %s | FileCheck %s
2
3target datalayout =
4"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
5target triple = "x86_64-unknown-linux-gnu"
6
7@x = external global [1 x [2 x <4 x float>]]
8
9; Can we sink single addressing mode computation to use?
10define void @test1(i1 %cond, i64* %base) {
11; CHECK-LABEL: @test1
12; CHECK: getelementptr i8, {{.+}} 40
13entry:
14  %addr = getelementptr inbounds i64, i64* %base, i64 5
15  %casted = bitcast i64* %addr to i32*
16  br i1 %cond, label %if.then, label %fallthrough
17
18if.then:
19  %v = load i32, i32* %casted, align 4
20  br label %fallthrough
21
22fallthrough:
23  ret void
24}
25
26declare void @foo(i32)
27
28; Make sure sinking two copies of addressing mode into different blocks works
29define void @test2(i1 %cond, i64* %base) {
30; CHECK-LABEL: @test2
31entry:
32  %addr = getelementptr inbounds i64, i64* %base, i64 5
33  %casted = bitcast i64* %addr to i32*
34  br i1 %cond, label %if.then, label %fallthrough
35
36if.then:
37; CHECK-LABEL: if.then:
38; CHECK: getelementptr i8, {{.+}} 40
39  %v1 = load i32, i32* %casted, align 4
40  call void @foo(i32 %v1)
41  %cmp = icmp eq i32 %v1, 0
42  br i1 %cmp, label %next, label %fallthrough
43
44next:
45; CHECK-LABEL: next:
46; CHECK: getelementptr i8, {{.+}} 40
47  %v2 = load i32, i32* %casted, align 4
48  call void @foo(i32 %v2)
49  br label %fallthrough
50
51fallthrough:
52  ret void
53}
54
55; If we have two loads in the same block, only need one copy of addressing mode
56; - instruction selection will duplicate if needed
57define void @test3(i1 %cond, i64* %base) {
58; CHECK-LABEL: @test3
59entry:
60  %addr = getelementptr inbounds i64, i64* %base, i64 5
61  %casted = bitcast i64* %addr to i32*
62  br i1 %cond, label %if.then, label %fallthrough
63
64if.then:
65; CHECK-LABEL: if.then:
66; CHECK: getelementptr i8, {{.+}} 40
67  %v1 = load i32, i32* %casted, align 4
68  call void @foo(i32 %v1)
69; CHECK-NOT: getelementptr i8, {{.+}}, 40
70  %v2 = load i32, i32* %casted, align 4
71  call void @foo(i32 %v2)
72  br label %fallthrough
73
74fallthrough:
75  ret void
76}
77
78; Can we still sink addressing mode if there's a cold use of the
79; address itself?
80define void @test4(i1 %cond, i64* %base) {
81; CHECK-LABEL: @test4
82entry:
83  %addr = getelementptr inbounds i64, i64* %base, i64 5
84  %casted = bitcast i64* %addr to i32*
85  br i1 %cond, label %if.then, label %fallthrough
86
87if.then:
88; CHECK-LABEL: if.then:
89; CHECK: getelementptr i8, {{.+}} 40
90  %v1 = load i32, i32* %casted, align 4
91  call void @foo(i32 %v1)
92  %cmp = icmp eq i32 %v1, 0
93  br i1 %cmp, label %rare.1, label %fallthrough
94
95fallthrough:
96  ret void
97
98rare.1:
99; CHECK-LABEL: rare.1:
100; CHECK: getelementptr i8, {{.+}} 40
101  call void @slowpath(i32 %v1, i32* %casted) cold
102  br label %fallthrough
103}
104
105; Negative test - don't want to duplicate addressing into hot path
106define void @test5(i1 %cond, i64* %base) {
107; CHECK-LABEL: @test5
108entry:
109; CHECK: %addr = getelementptr
110  %addr = getelementptr inbounds i64, i64* %base, i64 5
111  %casted = bitcast i64* %addr to i32*
112  br i1 %cond, label %if.then, label %fallthrough
113
114if.then:
115; CHECK-LABEL: if.then:
116; CHECK-NOT: getelementptr i8, {{.+}} 40
117  %v1 = load i32, i32* %casted, align 4
118  call void @foo(i32 %v1)
119  %cmp = icmp eq i32 %v1, 0
120  br i1 %cmp, label %rare.1, label %fallthrough
121
122fallthrough:
123  ret void
124
125rare.1:
126  call void @slowpath(i32 %v1, i32* %casted) ;; NOT COLD
127  br label %fallthrough
128}
129
130; Negative test - opt for size
131define void @test6(i1 %cond, i64* %base) minsize {
132; CHECK-LABEL: @test6
133entry:
134; CHECK: %addr = getelementptr
135  %addr = getelementptr inbounds i64, i64* %base, i64 5
136  %casted = bitcast i64* %addr to i32*
137  br i1 %cond, label %if.then, label %fallthrough
138
139if.then:
140; CHECK-LABEL: if.then:
141; CHECK-NOT: getelementptr i8, {{.+}} 40
142  %v1 = load i32, i32* %casted, align 4
143  call void @foo(i32 %v1)
144  %cmp = icmp eq i32 %v1, 0
145  br i1 %cmp, label %rare.1, label %fallthrough
146
147fallthrough:
148  ret void
149
150rare.1:
151  call void @slowpath(i32 %v1, i32* %casted) cold
152  br label %fallthrough
153}
154
155
156; Make sure sinking two copies of addressing mode into different blocks works
157; when there are cold paths for each.
158define void @test7(i1 %cond, i64* %base) {
159; CHECK-LABEL: @test7
160entry:
161  %addr = getelementptr inbounds i64, i64* %base, i64 5
162  %casted = bitcast i64* %addr to i32*
163  br i1 %cond, label %if.then, label %fallthrough
164
165if.then:
166; CHECK-LABEL: if.then:
167; CHECK: getelementptr i8, {{.+}} 40
168  %v1 = load i32, i32* %casted, align 4
169  call void @foo(i32 %v1)
170  %cmp = icmp eq i32 %v1, 0
171  br i1 %cmp, label %rare.1, label %next
172
173next:
174; CHECK-LABEL: next:
175; CHECK: getelementptr i8, {{.+}} 40
176  %v2 = load i32, i32* %casted, align 4
177  call void @foo(i32 %v2)
178  %cmp2 = icmp eq i32 %v2, 0
179  br i1 %cmp2, label %rare.1, label %fallthrough
180
181fallthrough:
182  ret void
183
184rare.1:
185; CHECK-LABEL: rare.1:
186; CHECK: getelementptr i8, {{.+}} 40
187  call void @slowpath(i32 %v1, i32* %casted) cold
188  br label %next
189
190rare.2:
191; CHECK-LABEL: rare.2:
192; CHECK: getelementptr i8, {{.+}} 40
193  call void @slowpath(i32 %v2, i32* %casted) cold
194  br label %fallthrough
195}
196
197declare void @slowpath(i32, i32*)
198
199; Make sure we don't end up in an infinite loop after we fail to sink.
200; CHECK-LABEL: define void @test8
201; CHECK: %ptr = getelementptr i8, i8* %aFOO_load_ptr2int_2void, i32 undef
202define void @test8() {
203allocas:
204  %aFOO_load = load float*, float** undef
205  %aFOO_load_ptr2int = ptrtoint float* %aFOO_load to i64
206  %aFOO_load_ptr2int_broadcast_init = insertelement <4 x i64> undef, i64 %aFOO_load_ptr2int, i32 0
207  %aFOO_load_ptr2int_2void = inttoptr i64 %aFOO_load_ptr2int to i8*
208  %ptr = getelementptr i8, i8* %aFOO_load_ptr2int_2void, i32 undef
209  br label %load.i145
210
211load.i145:
212  %ptr.i143 = bitcast i8* %ptr to <4 x float>*
213  %valall.i144 = load <4 x float>, <4 x float>* %ptr.i143, align 4
214  %x_offset = getelementptr [1 x [2 x <4 x float>]], [1 x [2 x <4 x float>]]* @x, i32 0, i64 0
215  br label %pl_loop.i.i122
216
217pl_loop.i.i122:
218  br label %pl_loop.i.i122
219}
220
221; Make sure we can sink address computation even
222; if there is a cycle in phi nodes.
223define void @test9(i1 %cond, i64* %base) {
224; CHECK-LABEL: @test9
225entry:
226  %addr = getelementptr inbounds i64, i64* %base, i64 5
227  %casted = bitcast i64* %addr to i32*
228  br label %header
229
230header:
231  %iv = phi i32 [0, %entry], [%iv.inc, %backedge]
232  %casted.loop = phi i32* [%casted, %entry], [%casted.merged, %backedge]
233  br i1 %cond, label %if.then, label %backedge
234
235if.then:
236  call void @foo(i32 %iv)
237  %addr.1 = getelementptr inbounds i64, i64* %base, i64 5
238  %casted.1 = bitcast i64* %addr.1 to i32*
239  br label %backedge
240
241backedge:
242; CHECK-LABEL: backedge:
243; CHECK: getelementptr i8, {{.+}} 40
244  %casted.merged = phi i32* [%casted.loop, %header], [%casted.1, %if.then]
245  %v = load i32, i32* %casted.merged, align 4
246  call void @foo(i32 %v)
247  %iv.inc = add i32 %iv, 1
248  %cmp = icmp slt i32 %iv.inc, 1000
249  br i1 %cmp, label %header, label %exit
250
251exit:
252  ret void
253}
254
255; Make sure we can eliminate a select when both arguments perform equivalent
256; address computation.
257define void @test10(i1 %cond, i64* %base) {
258; CHECK-LABEL: @test10
259; CHECK: getelementptr i8, {{.+}} 40
260; CHECK-NOT: select
261entry:
262  %gep1 = getelementptr inbounds i64, i64* %base, i64 5
263  %gep1.casted = bitcast i64* %gep1 to i32*
264  %base.casted = bitcast i64* %base to i32*
265  %gep2 = getelementptr inbounds i32, i32* %base.casted, i64 10
266  %casted.merged = select i1 %cond, i32* %gep1.casted, i32* %gep2
267  %v = load i32, i32* %casted.merged, align 4
268  call void @foo(i32 %v)
269  ret void
270}
271
272; Found by fuzzer, getSExtValue of > 64 bit constant
273define void @i96_mul(i1* %base, i96 %offset) {
274BB:
275  ;; RHS = 0x7FFFFFFFFFFFFFFFFFFFFFFF
276  %B84 = mul i96 %offset, 39614081257132168796771975167
277  %G23 = getelementptr i1, i1* %base, i96 %B84
278  store i1 false, i1* %G23
279  ret void
280}
281