| blob | da9d16ca9cf726e4298dcfc20ef4a7e93b6b9ca7 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-linux -mcpu=x86-64 -verify-machineinstrs | FileCheck %s -check-prefix=CHECK
3 ; RUN: llc < %s -mtriple=x86_64-linux -mcpu=x86-64 --x86-disable-avoid-SFB -verify-machineinstrs | FileCheck %s --check-prefix=DISABLED
4 ; RUN: llc < %s -mtriple=x86_64-linux -mcpu=core-avx2 -verify-machineinstrs | FileCheck %s -check-prefix=CHECK-AVX2
5 ; RUN: llc < %s -mtriple=x86_64-linux -mcpu=skx -verify-machineinstrs | FileCheck %s -check-prefix=CHECK-AVX512
7 ; ModuleID = '../testSFB/testOverlapBlocks.c'
8 source_filename = "../testSFB/testOverlapBlocks.c"
9 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
10 target triple = "x86_64-unknown-linux-gnu"
12 ; Function Attrs: nounwind uwtable
13 define dso_local void @test_overlap_1(ptr nocapture %A, i32 %x) local_unnamed_addr #0 {
14 ; CHECK-LABEL: test_overlap_1:
15 ; CHECK: # %bb.0: # %entry
16 ; CHECK-NEXT: movl $7, -8(%rdi)
17 ; CHECK-NEXT: movq -16(%rdi), %rax
18 ; CHECK-NEXT: movq %rax, (%rdi)
19 ; CHECK-NEXT: movl -8(%rdi), %eax
20 ; CHECK-NEXT: movl %eax, 8(%rdi)
21 ; CHECK-NEXT: movl -4(%rdi), %eax
22 ; CHECK-NEXT: movl %eax, 12(%rdi)
23 ; CHECK-NEXT: movslq %esi, %rax
24 ; CHECK-NEXT: movq %rax, -9(%rdi)
25 ; CHECK-NEXT: movq %rax, -16(%rdi)
26 ; CHECK-NEXT: movb $0, -1(%rdi)
27 ; CHECK-NEXT: movq -16(%rdi), %rax
28 ; CHECK-NEXT: movq %rax, 16(%rdi)
29 ; CHECK-NEXT: movl -8(%rdi), %eax
30 ; CHECK-NEXT: movl %eax, 24(%rdi)
31 ; CHECK-NEXT: movzwl -4(%rdi), %eax
32 ; CHECK-NEXT: movw %ax, 28(%rdi)
33 ; CHECK-NEXT: movzbl -2(%rdi), %eax
34 ; CHECK-NEXT: movb %al, 30(%rdi)
35 ; CHECK-NEXT: movzbl -1(%rdi), %eax
36 ; CHECK-NEXT: movb %al, 31(%rdi)
37 ; CHECK-NEXT: retq
38 ;
39 ; DISABLED-LABEL: test_overlap_1:
40 ; DISABLED: # %bb.0: # %entry
41 ; DISABLED-NEXT: movl $7, -8(%rdi)
42 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
43 ; DISABLED-NEXT: movups %xmm0, (%rdi)
44 ; DISABLED-NEXT: movslq %esi, %rax
45 ; DISABLED-NEXT: movq %rax, -9(%rdi)
46 ; DISABLED-NEXT: movq %rax, -16(%rdi)
47 ; DISABLED-NEXT: movb $0, -1(%rdi)
48 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
49 ; DISABLED-NEXT: movups %xmm0, 16(%rdi)
50 ; DISABLED-NEXT: retq
51 ;
52 ; CHECK-AVX2-LABEL: test_overlap_1:
53 ; CHECK-AVX2: # %bb.0: # %entry
54 ; CHECK-AVX2-NEXT: movl $7, -8(%rdi)
55 ; CHECK-AVX2-NEXT: movq -16(%rdi), %rax
56 ; CHECK-AVX2-NEXT: movq %rax, (%rdi)
57 ; CHECK-AVX2-NEXT: movl -8(%rdi), %eax
58 ; CHECK-AVX2-NEXT: movl %eax, 8(%rdi)
59 ; CHECK-AVX2-NEXT: movl -4(%rdi), %eax
60 ; CHECK-AVX2-NEXT: movl %eax, 12(%rdi)
61 ; CHECK-AVX2-NEXT: movslq %esi, %rax
62 ; CHECK-AVX2-NEXT: movq %rax, -9(%rdi)
63 ; CHECK-AVX2-NEXT: movq %rax, -16(%rdi)
64 ; CHECK-AVX2-NEXT: movb $0, -1(%rdi)
65 ; CHECK-AVX2-NEXT: movq -16(%rdi), %rax
66 ; CHECK-AVX2-NEXT: movq %rax, 16(%rdi)
67 ; CHECK-AVX2-NEXT: movl -8(%rdi), %eax
68 ; CHECK-AVX2-NEXT: movl %eax, 24(%rdi)
69 ; CHECK-AVX2-NEXT: movzwl -4(%rdi), %eax
70 ; CHECK-AVX2-NEXT: movw %ax, 28(%rdi)
71 ; CHECK-AVX2-NEXT: movzbl -2(%rdi), %eax
72 ; CHECK-AVX2-NEXT: movb %al, 30(%rdi)
73 ; CHECK-AVX2-NEXT: movzbl -1(%rdi), %eax
74 ; CHECK-AVX2-NEXT: movb %al, 31(%rdi)
75 ; CHECK-AVX2-NEXT: retq
76 ;
77 ; CHECK-AVX512-LABEL: test_overlap_1:
78 ; CHECK-AVX512: # %bb.0: # %entry
79 ; CHECK-AVX512-NEXT: movl $7, -8(%rdi)
80 ; CHECK-AVX512-NEXT: movq -16(%rdi), %rax
81 ; CHECK-AVX512-NEXT: movq %rax, (%rdi)
82 ; CHECK-AVX512-NEXT: movl -8(%rdi), %eax
83 ; CHECK-AVX512-NEXT: movl %eax, 8(%rdi)
84 ; CHECK-AVX512-NEXT: movl -4(%rdi), %eax
85 ; CHECK-AVX512-NEXT: movl %eax, 12(%rdi)
86 ; CHECK-AVX512-NEXT: movslq %esi, %rax
87 ; CHECK-AVX512-NEXT: movq %rax, -9(%rdi)
88 ; CHECK-AVX512-NEXT: movq %rax, -16(%rdi)
89 ; CHECK-AVX512-NEXT: movb $0, -1(%rdi)
90 ; CHECK-AVX512-NEXT: movq -16(%rdi), %rax
91 ; CHECK-AVX512-NEXT: movq %rax, 16(%rdi)
92 ; CHECK-AVX512-NEXT: movl -8(%rdi), %eax
93 ; CHECK-AVX512-NEXT: movl %eax, 24(%rdi)
94 ; CHECK-AVX512-NEXT: movzwl -4(%rdi), %eax
95 ; CHECK-AVX512-NEXT: movw %ax, 28(%rdi)
96 ; CHECK-AVX512-NEXT: movzbl -2(%rdi), %eax
97 ; CHECK-AVX512-NEXT: movb %al, 30(%rdi)
98 ; CHECK-AVX512-NEXT: movzbl -1(%rdi), %eax
99 ; CHECK-AVX512-NEXT: movb %al, 31(%rdi)
100 ; CHECK-AVX512-NEXT: retq
101 entry:
102 %add.ptr = getelementptr inbounds i8, ptr %A, i64 -16
103 %add.ptr1 = getelementptr inbounds i8, ptr %A, i64 -8
104 store i32 7, ptr %add.ptr1, align 4
105 tail call void @llvm.memcpy.p0.p0.i64(ptr align 4 %A, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
106 %conv = sext i32 %x to i64
107 %add.ptr2 = getelementptr inbounds i8, ptr %A, i64 -9
108 store i64 %conv, ptr %add.ptr2, align 8
109 store i64 %conv, ptr %add.ptr, align 8
110 %add.ptr5 = getelementptr inbounds i8, ptr %A, i64 -1
111 store i8 0, ptr %add.ptr5, align 1
112 %add.ptr6 = getelementptr inbounds i8, ptr %A, i64 16
113 tail call void @llvm.memcpy.p0.p0.i64(ptr nonnull align 4 %add.ptr6, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
114 ret void
115 }
117 ; Function Attrs: argmemonly nounwind
118 declare void @llvm.memcpy.p0.p0.i64(ptr nocapture writeonly, ptr nocapture readonly, i64, i1) #1
120 ; Function Attrs: nounwind uwtable
121 define dso_local void @test_overlap_2(ptr nocapture %A, i32 %x) local_unnamed_addr #0 {
122 ; CHECK-LABEL: test_overlap_2:
123 ; CHECK: # %bb.0: # %entry
124 ; CHECK-NEXT: movslq %esi, %rax
125 ; CHECK-NEXT: movq %rax, -16(%rdi)
126 ; CHECK-NEXT: movq -16(%rdi), %rcx
127 ; CHECK-NEXT: movq %rcx, (%rdi)
128 ; CHECK-NEXT: movq -8(%rdi), %rcx
129 ; CHECK-NEXT: movq %rcx, 8(%rdi)
130 ; CHECK-NEXT: movq %rax, -8(%rdi)
131 ; CHECK-NEXT: movl $7, -12(%rdi)
132 ; CHECK-NEXT: movl -16(%rdi), %eax
133 ; CHECK-NEXT: movl %eax, 16(%rdi)
134 ; CHECK-NEXT: movl -12(%rdi), %eax
135 ; CHECK-NEXT: movl %eax, 20(%rdi)
136 ; CHECK-NEXT: movq -8(%rdi), %rax
137 ; CHECK-NEXT: movq %rax, 24(%rdi)
138 ; CHECK-NEXT: retq
139 ;
140 ; DISABLED-LABEL: test_overlap_2:
141 ; DISABLED: # %bb.0: # %entry
142 ; DISABLED-NEXT: movslq %esi, %rax
143 ; DISABLED-NEXT: movq %rax, -16(%rdi)
144 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
145 ; DISABLED-NEXT: movups %xmm0, (%rdi)
146 ; DISABLED-NEXT: movq %rax, -8(%rdi)
147 ; DISABLED-NEXT: movl $7, -12(%rdi)
148 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
149 ; DISABLED-NEXT: movups %xmm0, 16(%rdi)
150 ; DISABLED-NEXT: retq
151 ;
152 ; CHECK-AVX2-LABEL: test_overlap_2:
153 ; CHECK-AVX2: # %bb.0: # %entry
154 ; CHECK-AVX2-NEXT: movslq %esi, %rax
155 ; CHECK-AVX2-NEXT: movq %rax, -16(%rdi)
156 ; CHECK-AVX2-NEXT: movq -16(%rdi), %rcx
157 ; CHECK-AVX2-NEXT: movq %rcx, (%rdi)
158 ; CHECK-AVX2-NEXT: movq -8(%rdi), %rcx
159 ; CHECK-AVX2-NEXT: movq %rcx, 8(%rdi)
160 ; CHECK-AVX2-NEXT: movq %rax, -8(%rdi)
161 ; CHECK-AVX2-NEXT: movl $7, -12(%rdi)
162 ; CHECK-AVX2-NEXT: movl -16(%rdi), %eax
163 ; CHECK-AVX2-NEXT: movl %eax, 16(%rdi)
164 ; CHECK-AVX2-NEXT: movl -12(%rdi), %eax
165 ; CHECK-AVX2-NEXT: movl %eax, 20(%rdi)
166 ; CHECK-AVX2-NEXT: movq -8(%rdi), %rax
167 ; CHECK-AVX2-NEXT: movq %rax, 24(%rdi)
168 ; CHECK-AVX2-NEXT: retq
169 ;
170 ; CHECK-AVX512-LABEL: test_overlap_2:
171 ; CHECK-AVX512: # %bb.0: # %entry
172 ; CHECK-AVX512-NEXT: movslq %esi, %rax
173 ; CHECK-AVX512-NEXT: movq %rax, -16(%rdi)
174 ; CHECK-AVX512-NEXT: movq -16(%rdi), %rcx
175 ; CHECK-AVX512-NEXT: movq %rcx, (%rdi)
176 ; CHECK-AVX512-NEXT: movq -8(%rdi), %rcx
177 ; CHECK-AVX512-NEXT: movq %rcx, 8(%rdi)
178 ; CHECK-AVX512-NEXT: movq %rax, -8(%rdi)
179 ; CHECK-AVX512-NEXT: movl $7, -12(%rdi)
180 ; CHECK-AVX512-NEXT: movl -16(%rdi), %eax
181 ; CHECK-AVX512-NEXT: movl %eax, 16(%rdi)
182 ; CHECK-AVX512-NEXT: movl -12(%rdi), %eax
183 ; CHECK-AVX512-NEXT: movl %eax, 20(%rdi)
184 ; CHECK-AVX512-NEXT: movq -8(%rdi), %rax
185 ; CHECK-AVX512-NEXT: movq %rax, 24(%rdi)
186 ; CHECK-AVX512-NEXT: retq
187 entry:
188 %add.ptr = getelementptr inbounds i8, ptr %A, i64 -16
189 %conv = sext i32 %x to i64
190 store i64 %conv, ptr %add.ptr, align 8
191 tail call void @llvm.memcpy.p0.p0.i64(ptr align 4 %A, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
192 %add.ptr3 = getelementptr inbounds i8, ptr %A, i64 -8
193 store i64 %conv, ptr %add.ptr3, align 8
194 %add.ptr4 = getelementptr inbounds i8, ptr %A, i64 -12
195 store i32 7, ptr %add.ptr4, align 4
196 %add.ptr5 = getelementptr inbounds i8, ptr %A, i64 16
197 tail call void @llvm.memcpy.p0.p0.i64(ptr nonnull align 4 %add.ptr5, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
198 ret void
199 }
201 ; Function Attrs: nounwind uwtable
202 define dso_local void @test_overlap_3(ptr nocapture %A, i32 %x) local_unnamed_addr #0 {
203 ; CHECK-LABEL: test_overlap_3:
204 ; CHECK: # %bb.0: # %entry
205 ; CHECK-NEXT: movl $7, -10(%rdi)
206 ; CHECK-NEXT: movl -16(%rdi), %eax
207 ; CHECK-NEXT: movl %eax, (%rdi)
208 ; CHECK-NEXT: movzwl -12(%rdi), %eax
209 ; CHECK-NEXT: movw %ax, 4(%rdi)
210 ; CHECK-NEXT: movl -10(%rdi), %eax
211 ; CHECK-NEXT: movl %eax, 6(%rdi)
212 ; CHECK-NEXT: movl -6(%rdi), %eax
213 ; CHECK-NEXT: movl %eax, 10(%rdi)
214 ; CHECK-NEXT: movzwl -2(%rdi), %eax
215 ; CHECK-NEXT: movw %ax, 14(%rdi)
216 ; CHECK-NEXT: movslq %esi, %rax
217 ; CHECK-NEXT: movq %rax, -9(%rdi)
218 ; CHECK-NEXT: movq %rax, -16(%rdi)
219 ; CHECK-NEXT: movb $0, -1(%rdi)
220 ; CHECK-NEXT: movq -16(%rdi), %rax
221 ; CHECK-NEXT: movq %rax, 16(%rdi)
222 ; CHECK-NEXT: movzwl -8(%rdi), %eax
223 ; CHECK-NEXT: movw %ax, 24(%rdi)
224 ; CHECK-NEXT: movl -6(%rdi), %eax
225 ; CHECK-NEXT: movl %eax, 26(%rdi)
226 ; CHECK-NEXT: movzbl -2(%rdi), %eax
227 ; CHECK-NEXT: movb %al, 30(%rdi)
228 ; CHECK-NEXT: movzbl -1(%rdi), %eax
229 ; CHECK-NEXT: movb %al, 31(%rdi)
230 ; CHECK-NEXT: retq
231 ;
232 ; DISABLED-LABEL: test_overlap_3:
233 ; DISABLED: # %bb.0: # %entry
234 ; DISABLED-NEXT: movl $7, -10(%rdi)
235 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
236 ; DISABLED-NEXT: movups %xmm0, (%rdi)
237 ; DISABLED-NEXT: movslq %esi, %rax
238 ; DISABLED-NEXT: movq %rax, -9(%rdi)
239 ; DISABLED-NEXT: movq %rax, -16(%rdi)
240 ; DISABLED-NEXT: movb $0, -1(%rdi)
241 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
242 ; DISABLED-NEXT: movups %xmm0, 16(%rdi)
243 ; DISABLED-NEXT: retq
244 ;
245 ; CHECK-AVX2-LABEL: test_overlap_3:
246 ; CHECK-AVX2: # %bb.0: # %entry
247 ; CHECK-AVX2-NEXT: movl $7, -10(%rdi)
248 ; CHECK-AVX2-NEXT: movl -16(%rdi), %eax
249 ; CHECK-AVX2-NEXT: movl %eax, (%rdi)
250 ; CHECK-AVX2-NEXT: movzwl -12(%rdi), %eax
251 ; CHECK-AVX2-NEXT: movw %ax, 4(%rdi)
252 ; CHECK-AVX2-NEXT: movl -10(%rdi), %eax
253 ; CHECK-AVX2-NEXT: movl %eax, 6(%rdi)
254 ; CHECK-AVX2-NEXT: movl -6(%rdi), %eax
255 ; CHECK-AVX2-NEXT: movl %eax, 10(%rdi)
256 ; CHECK-AVX2-NEXT: movzwl -2(%rdi), %eax
257 ; CHECK-AVX2-NEXT: movw %ax, 14(%rdi)
258 ; CHECK-AVX2-NEXT: movslq %esi, %rax
259 ; CHECK-AVX2-NEXT: movq %rax, -9(%rdi)
260 ; CHECK-AVX2-NEXT: movq %rax, -16(%rdi)
261 ; CHECK-AVX2-NEXT: movb $0, -1(%rdi)
262 ; CHECK-AVX2-NEXT: movq -16(%rdi), %rax
263 ; CHECK-AVX2-NEXT: movq %rax, 16(%rdi)
264 ; CHECK-AVX2-NEXT: movzwl -8(%rdi), %eax
265 ; CHECK-AVX2-NEXT: movw %ax, 24(%rdi)
266 ; CHECK-AVX2-NEXT: movl -6(%rdi), %eax
267 ; CHECK-AVX2-NEXT: movl %eax, 26(%rdi)
268 ; CHECK-AVX2-NEXT: movzbl -2(%rdi), %eax
269 ; CHECK-AVX2-NEXT: movb %al, 30(%rdi)
270 ; CHECK-AVX2-NEXT: movzbl -1(%rdi), %eax
271 ; CHECK-AVX2-NEXT: movb %al, 31(%rdi)
272 ; CHECK-AVX2-NEXT: retq
273 ;
274 ; CHECK-AVX512-LABEL: test_overlap_3:
275 ; CHECK-AVX512: # %bb.0: # %entry
276 ; CHECK-AVX512-NEXT: movl $7, -10(%rdi)
277 ; CHECK-AVX512-NEXT: movl -16(%rdi), %eax
278 ; CHECK-AVX512-NEXT: movl %eax, (%rdi)
279 ; CHECK-AVX512-NEXT: movzwl -12(%rdi), %eax
280 ; CHECK-AVX512-NEXT: movw %ax, 4(%rdi)
281 ; CHECK-AVX512-NEXT: movl -10(%rdi), %eax
282 ; CHECK-AVX512-NEXT: movl %eax, 6(%rdi)
283 ; CHECK-AVX512-NEXT: movl -6(%rdi), %eax
284 ; CHECK-AVX512-NEXT: movl %eax, 10(%rdi)
285 ; CHECK-AVX512-NEXT: movzwl -2(%rdi), %eax
286 ; CHECK-AVX512-NEXT: movw %ax, 14(%rdi)
287 ; CHECK-AVX512-NEXT: movslq %esi, %rax
288 ; CHECK-AVX512-NEXT: movq %rax, -9(%rdi)
289 ; CHECK-AVX512-NEXT: movq %rax, -16(%rdi)
290 ; CHECK-AVX512-NEXT: movb $0, -1(%rdi)
291 ; CHECK-AVX512-NEXT: movq -16(%rdi), %rax
292 ; CHECK-AVX512-NEXT: movq %rax, 16(%rdi)
293 ; CHECK-AVX512-NEXT: movzwl -8(%rdi), %eax
294 ; CHECK-AVX512-NEXT: movw %ax, 24(%rdi)
295 ; CHECK-AVX512-NEXT: movl -6(%rdi), %eax
296 ; CHECK-AVX512-NEXT: movl %eax, 26(%rdi)
297 ; CHECK-AVX512-NEXT: movzbl -2(%rdi), %eax
298 ; CHECK-AVX512-NEXT: movb %al, 30(%rdi)
299 ; CHECK-AVX512-NEXT: movzbl -1(%rdi), %eax
300 ; CHECK-AVX512-NEXT: movb %al, 31(%rdi)
301 ; CHECK-AVX512-NEXT: retq
302 entry:
303 %add.ptr = getelementptr inbounds i8, ptr %A, i64 -16
304 %add.ptr1 = getelementptr inbounds i8, ptr %A, i64 -10
305 store i32 7, ptr %add.ptr1, align 4
306 tail call void @llvm.memcpy.p0.p0.i64(ptr align 4 %A, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
307 %conv = sext i32 %x to i64
308 %add.ptr2 = getelementptr inbounds i8, ptr %A, i64 -9
309 store i64 %conv, ptr %add.ptr2, align 8
310 store i64 %conv, ptr %add.ptr, align 8
311 %add.ptr5 = getelementptr inbounds i8, ptr %A, i64 -1
312 store i8 0, ptr %add.ptr5, align 1
313 %add.ptr6 = getelementptr inbounds i8, ptr %A, i64 16
314 tail call void @llvm.memcpy.p0.p0.i64(ptr nonnull align 4 %add.ptr6, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
315 ret void
316 }
318 ; Function Attrs: nounwind uwtable
319 define dso_local void @test_overlap_4(ptr nocapture %A, i32 %x) local_unnamed_addr #0 {
320 ; CHECK-LABEL: test_overlap_4:
321 ; CHECK: # %bb.0: # %entry
322 ; CHECK-NEXT: movups -16(%rdi), %xmm0
323 ; CHECK-NEXT: movups %xmm0, (%rdi)
324 ; CHECK-NEXT: movslq %esi, %rax
325 ; CHECK-NEXT: movq %rax, -8(%rdi)
326 ; CHECK-NEXT: movl %eax, -16(%rdi)
327 ; CHECK-NEXT: movl $0, -11(%rdi)
328 ; CHECK-NEXT: movl -16(%rdi), %eax
329 ; CHECK-NEXT: movl %eax, 16(%rdi)
330 ; CHECK-NEXT: movzbl -12(%rdi), %eax
331 ; CHECK-NEXT: movb %al, 20(%rdi)
332 ; CHECK-NEXT: movl -11(%rdi), %eax
333 ; CHECK-NEXT: movl %eax, 21(%rdi)
334 ; CHECK-NEXT: movl -7(%rdi), %eax
335 ; CHECK-NEXT: movl %eax, 25(%rdi)
336 ; CHECK-NEXT: movzwl -3(%rdi), %eax
337 ; CHECK-NEXT: movw %ax, 29(%rdi)
338 ; CHECK-NEXT: movzbl -1(%rdi), %eax
339 ; CHECK-NEXT: movb %al, 31(%rdi)
340 ; CHECK-NEXT: retq
341 ;
342 ; DISABLED-LABEL: test_overlap_4:
343 ; DISABLED: # %bb.0: # %entry
344 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
345 ; DISABLED-NEXT: movups %xmm0, (%rdi)
346 ; DISABLED-NEXT: movslq %esi, %rax
347 ; DISABLED-NEXT: movq %rax, -8(%rdi)
348 ; DISABLED-NEXT: movl %eax, -16(%rdi)
349 ; DISABLED-NEXT: movl $0, -11(%rdi)
350 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
351 ; DISABLED-NEXT: movups %xmm0, 16(%rdi)
352 ; DISABLED-NEXT: retq
353 ;
354 ; CHECK-AVX2-LABEL: test_overlap_4:
355 ; CHECK-AVX2: # %bb.0: # %entry
356 ; CHECK-AVX2-NEXT: vmovups -16(%rdi), %xmm0
357 ; CHECK-AVX2-NEXT: vmovups %xmm0, (%rdi)
358 ; CHECK-AVX2-NEXT: movslq %esi, %rax
359 ; CHECK-AVX2-NEXT: movq %rax, -8(%rdi)
360 ; CHECK-AVX2-NEXT: movl %eax, -16(%rdi)
361 ; CHECK-AVX2-NEXT: movl $0, -11(%rdi)
362 ; CHECK-AVX2-NEXT: movl -16(%rdi), %eax
363 ; CHECK-AVX2-NEXT: movl %eax, 16(%rdi)
364 ; CHECK-AVX2-NEXT: movzbl -12(%rdi), %eax
365 ; CHECK-AVX2-NEXT: movb %al, 20(%rdi)
366 ; CHECK-AVX2-NEXT: movl -11(%rdi), %eax
367 ; CHECK-AVX2-NEXT: movl %eax, 21(%rdi)
368 ; CHECK-AVX2-NEXT: movl -7(%rdi), %eax
369 ; CHECK-AVX2-NEXT: movl %eax, 25(%rdi)
370 ; CHECK-AVX2-NEXT: movzwl -3(%rdi), %eax
371 ; CHECK-AVX2-NEXT: movw %ax, 29(%rdi)
372 ; CHECK-AVX2-NEXT: movzbl -1(%rdi), %eax
373 ; CHECK-AVX2-NEXT: movb %al, 31(%rdi)
374 ; CHECK-AVX2-NEXT: retq
375 ;
376 ; CHECK-AVX512-LABEL: test_overlap_4:
377 ; CHECK-AVX512: # %bb.0: # %entry
378 ; CHECK-AVX512-NEXT: vmovups -16(%rdi), %xmm0
379 ; CHECK-AVX512-NEXT: vmovups %xmm0, (%rdi)
380 ; CHECK-AVX512-NEXT: movslq %esi, %rax
381 ; CHECK-AVX512-NEXT: movq %rax, -8(%rdi)
382 ; CHECK-AVX512-NEXT: movl %eax, -16(%rdi)
383 ; CHECK-AVX512-NEXT: movl $0, -11(%rdi)
384 ; CHECK-AVX512-NEXT: movl -16(%rdi), %eax
385 ; CHECK-AVX512-NEXT: movl %eax, 16(%rdi)
386 ; CHECK-AVX512-NEXT: movzbl -12(%rdi), %eax
387 ; CHECK-AVX512-NEXT: movb %al, 20(%rdi)
388 ; CHECK-AVX512-NEXT: movl -11(%rdi), %eax
389 ; CHECK-AVX512-NEXT: movl %eax, 21(%rdi)
390 ; CHECK-AVX512-NEXT: movl -7(%rdi), %eax
391 ; CHECK-AVX512-NEXT: movl %eax, 25(%rdi)
392 ; CHECK-AVX512-NEXT: movzwl -3(%rdi), %eax
393 ; CHECK-AVX512-NEXT: movw %ax, 29(%rdi)
394 ; CHECK-AVX512-NEXT: movzbl -1(%rdi), %eax
395 ; CHECK-AVX512-NEXT: movb %al, 31(%rdi)
396 ; CHECK-AVX512-NEXT: retq
397 entry:
398 %add.ptr = getelementptr inbounds i8, ptr %A, i64 -16
399 tail call void @llvm.memcpy.p0.p0.i64(ptr align 4 %A, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
400 %conv = sext i32 %x to i64
401 %add.ptr1 = getelementptr inbounds i8, ptr %A, i64 -8
402 store i64 %conv, ptr %add.ptr1, align 8
403 store i32 %x, ptr %add.ptr, align 4
404 %add.ptr3 = getelementptr inbounds i8, ptr %A, i64 -11
405 store i32 0, ptr %add.ptr3, align 4
406 %add.ptr4 = getelementptr inbounds i8, ptr %A, i64 16
407 tail call void @llvm.memcpy.p0.p0.i64(ptr nonnull align 4 %add.ptr4, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
408 ret void
409 }
411 ; Function Attrs: nounwind uwtable
412 define dso_local void @test_overlap_5(ptr nocapture %A, i32 %x) local_unnamed_addr #0 {
413 ; CHECK-LABEL: test_overlap_5:
414 ; CHECK: # %bb.0: # %entry
415 ; CHECK-NEXT: movups -16(%rdi), %xmm0
416 ; CHECK-NEXT: movups %xmm0, (%rdi)
417 ; CHECK-NEXT: movslq %esi, %rax
418 ; CHECK-NEXT: movq %rax, -16(%rdi)
419 ; CHECK-NEXT: movb %al, -14(%rdi)
420 ; CHECK-NEXT: movb $0, -11(%rdi)
421 ; CHECK-NEXT: movzwl -16(%rdi), %eax
422 ; CHECK-NEXT: movw %ax, 16(%rdi)
423 ; CHECK-NEXT: movzbl -14(%rdi), %eax
424 ; CHECK-NEXT: movb %al, 18(%rdi)
425 ; CHECK-NEXT: movzwl -13(%rdi), %eax
426 ; CHECK-NEXT: movw %ax, 19(%rdi)
427 ; CHECK-NEXT: movzbl -11(%rdi), %eax
428 ; CHECK-NEXT: movb %al, 21(%rdi)
429 ; CHECK-NEXT: movq -10(%rdi), %rax
430 ; CHECK-NEXT: movq %rax, 22(%rdi)
431 ; CHECK-NEXT: movzwl -2(%rdi), %eax
432 ; CHECK-NEXT: movw %ax, 30(%rdi)
433 ; CHECK-NEXT: retq
434 ;
435 ; DISABLED-LABEL: test_overlap_5:
436 ; DISABLED: # %bb.0: # %entry
437 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
438 ; DISABLED-NEXT: movups %xmm0, (%rdi)
439 ; DISABLED-NEXT: movslq %esi, %rax
440 ; DISABLED-NEXT: movq %rax, -16(%rdi)
441 ; DISABLED-NEXT: movb %al, -14(%rdi)
442 ; DISABLED-NEXT: movb $0, -11(%rdi)
443 ; DISABLED-NEXT: movups -16(%rdi), %xmm0
444 ; DISABLED-NEXT: movups %xmm0, 16(%rdi)
445 ; DISABLED-NEXT: retq
446 ;
447 ; CHECK-AVX2-LABEL: test_overlap_5:
448 ; CHECK-AVX2: # %bb.0: # %entry
449 ; CHECK-AVX2-NEXT: vmovups -16(%rdi), %xmm0
450 ; CHECK-AVX2-NEXT: vmovups %xmm0, (%rdi)
451 ; CHECK-AVX2-NEXT: movslq %esi, %rax
452 ; CHECK-AVX2-NEXT: movq %rax, -16(%rdi)
453 ; CHECK-AVX2-NEXT: movb %al, -14(%rdi)
454 ; CHECK-AVX2-NEXT: movb $0, -11(%rdi)
455 ; CHECK-AVX2-NEXT: movzwl -16(%rdi), %eax
456 ; CHECK-AVX2-NEXT: movw %ax, 16(%rdi)
457 ; CHECK-AVX2-NEXT: movzbl -14(%rdi), %eax
458 ; CHECK-AVX2-NEXT: movb %al, 18(%rdi)
459 ; CHECK-AVX2-NEXT: movzwl -13(%rdi), %eax
460 ; CHECK-AVX2-NEXT: movw %ax, 19(%rdi)
461 ; CHECK-AVX2-NEXT: movzbl -11(%rdi), %eax
462 ; CHECK-AVX2-NEXT: movb %al, 21(%rdi)
463 ; CHECK-AVX2-NEXT: movq -10(%rdi), %rax
464 ; CHECK-AVX2-NEXT: movq %rax, 22(%rdi)
465 ; CHECK-AVX2-NEXT: movzwl -2(%rdi), %eax
466 ; CHECK-AVX2-NEXT: movw %ax, 30(%rdi)
467 ; CHECK-AVX2-NEXT: retq
468 ;
469 ; CHECK-AVX512-LABEL: test_overlap_5:
470 ; CHECK-AVX512: # %bb.0: # %entry
471 ; CHECK-AVX512-NEXT: vmovups -16(%rdi), %xmm0
472 ; CHECK-AVX512-NEXT: vmovups %xmm0, (%rdi)
473 ; CHECK-AVX512-NEXT: movslq %esi, %rax
474 ; CHECK-AVX512-NEXT: movq %rax, -16(%rdi)
475 ; CHECK-AVX512-NEXT: movb %al, -14(%rdi)
476 ; CHECK-AVX512-NEXT: movb $0, -11(%rdi)
477 ; CHECK-AVX512-NEXT: movzwl -16(%rdi), %eax
478 ; CHECK-AVX512-NEXT: movw %ax, 16(%rdi)
479 ; CHECK-AVX512-NEXT: movzbl -14(%rdi), %eax
480 ; CHECK-AVX512-NEXT: movb %al, 18(%rdi)
481 ; CHECK-AVX512-NEXT: movzwl -13(%rdi), %eax
482 ; CHECK-AVX512-NEXT: movw %ax, 19(%rdi)
483 ; CHECK-AVX512-NEXT: movzbl -11(%rdi), %eax
484 ; CHECK-AVX512-NEXT: movb %al, 21(%rdi)
485 ; CHECK-AVX512-NEXT: movq -10(%rdi), %rax
486 ; CHECK-AVX512-NEXT: movq %rax, 22(%rdi)
487 ; CHECK-AVX512-NEXT: movzwl -2(%rdi), %eax
488 ; CHECK-AVX512-NEXT: movw %ax, 30(%rdi)
489 ; CHECK-AVX512-NEXT: retq
490 entry:
491 %add.ptr = getelementptr inbounds i8, ptr %A, i64 -16
492 tail call void @llvm.memcpy.p0.p0.i64(ptr align 4 %A, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
493 %conv = sext i32 %x to i64
494 store i64 %conv, ptr %add.ptr, align 8
495 %conv2 = trunc i32 %x to i8
496 %add.ptr3 = getelementptr inbounds i8, ptr %A, i64 -14
497 store i8 %conv2, ptr %add.ptr3, align 1
498 %add.ptr4 = getelementptr inbounds i8, ptr %A, i64 -11
499 store i8 0, ptr %add.ptr4, align 1
500 %add.ptr5 = getelementptr inbounds i8, ptr %A, i64 16
501 tail call void @llvm.memcpy.p0.p0.i64(ptr nonnull align 4 %add.ptr5, ptr nonnull align 4 %add.ptr, i64 16, i1 false)
502 ret void
503 }
505 attributes #0 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "frame-pointer"="none" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-features"="+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
507 attributes #1 = { argmemonly nounwind }