| blob | 6e3dedbf8303964c589a59df30bdd91839fc245a |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse | FileCheck %s --check-prefix=SSE
3 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse2 | FileCheck %s --check-prefix=SSE
4 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse2,-slow-unaligned-mem-16 | FileCheck %s --check-prefix=SSE2FAST
5 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
6 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
7 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx512f -mattr=+prefer-256-bit | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
8 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx512bw -mattr=+prefer-256-bit | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
9 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx512dq -mattr=+prefer-256-bit | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
10 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx512f -mattr=-prefer-256-bit | FileCheck %s --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512F
11 ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx512bw -mattr=-prefer-256-bit | FileCheck %s --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512BW
13 ; https://llvm.org/bugs/show_bug.cgi?id=27100
15 define void @memset_16_nonzero_bytes(i8* %x) {
16 ; SSE-LABEL: memset_16_nonzero_bytes:
17 ; SSE: # %bb.0:
18 ; SSE-NEXT: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
19 ; SSE-NEXT: movq %rax, 8(%rdi)
20 ; SSE-NEXT: movq %rax, (%rdi)
21 ; SSE-NEXT: retq
22 ;
23 ; SSE2FAST-LABEL: memset_16_nonzero_bytes:
24 ; SSE2FAST: # %bb.0:
25 ; SSE2FAST-NEXT: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
26 ; SSE2FAST-NEXT: movups %xmm0, (%rdi)
27 ; SSE2FAST-NEXT: retq
28 ;
29 ; AVX-LABEL: memset_16_nonzero_bytes:
30 ; AVX: # %bb.0:
31 ; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
32 ; AVX-NEXT: vmovups %xmm0, (%rdi)
33 ; AVX-NEXT: retq
34 %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 16, i64 -1)
35 ret void
36 }
38 define void @memset_32_nonzero_bytes(i8* %x) {
39 ; SSE-LABEL: memset_32_nonzero_bytes:
40 ; SSE: # %bb.0:
41 ; SSE-NEXT: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
42 ; SSE-NEXT: movq %rax, 24(%rdi)
43 ; SSE-NEXT: movq %rax, 16(%rdi)
44 ; SSE-NEXT: movq %rax, 8(%rdi)
45 ; SSE-NEXT: movq %rax, (%rdi)
46 ; SSE-NEXT: retq
47 ;
48 ; SSE2FAST-LABEL: memset_32_nonzero_bytes:
49 ; SSE2FAST: # %bb.0:
50 ; SSE2FAST-NEXT: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
51 ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi)
52 ; SSE2FAST-NEXT: movups %xmm0, (%rdi)
53 ; SSE2FAST-NEXT: retq
54 ;
55 ; AVX-LABEL: memset_32_nonzero_bytes:
56 ; AVX: # %bb.0:
57 ; AVX-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
58 ; AVX-NEXT: vmovups %ymm0, (%rdi)
59 ; AVX-NEXT: vzeroupper
60 ; AVX-NEXT: retq
61 %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 32, i64 -1)
62 ret void
63 }
65 define void @memset_64_nonzero_bytes(i8* %x) {
66 ; SSE-LABEL: memset_64_nonzero_bytes:
67 ; SSE: # %bb.0:
68 ; SSE-NEXT: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
69 ; SSE-NEXT: movq %rax, 56(%rdi)
70 ; SSE-NEXT: movq %rax, 48(%rdi)
71 ; SSE-NEXT: movq %rax, 40(%rdi)
72 ; SSE-NEXT: movq %rax, 32(%rdi)
73 ; SSE-NEXT: movq %rax, 24(%rdi)
74 ; SSE-NEXT: movq %rax, 16(%rdi)
75 ; SSE-NEXT: movq %rax, 8(%rdi)
76 ; SSE-NEXT: movq %rax, (%rdi)
77 ; SSE-NEXT: retq
78 ;
79 ; SSE2FAST-LABEL: memset_64_nonzero_bytes:
80 ; SSE2FAST: # %bb.0:
81 ; SSE2FAST-NEXT: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
82 ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi)
83 ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi)
84 ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi)
85 ; SSE2FAST-NEXT: movups %xmm0, (%rdi)
86 ; SSE2FAST-NEXT: retq
87 ;
88 ; AVX1-LABEL: memset_64_nonzero_bytes:
89 ; AVX1: # %bb.0:
90 ; AVX1-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
91 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
92 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
93 ; AVX1-NEXT: vzeroupper
94 ; AVX1-NEXT: retq
95 ;
96 ; AVX2-LABEL: memset_64_nonzero_bytes:
97 ; AVX2: # %bb.0:
98 ; AVX2-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
99 ; AVX2-NEXT: vmovups %ymm0, 32(%rdi)
100 ; AVX2-NEXT: vmovups %ymm0, (%rdi)
101 ; AVX2-NEXT: vzeroupper
102 ; AVX2-NEXT: retq
103 ;
104 ; AVX512F-LABEL: memset_64_nonzero_bytes:
105 ; AVX512F: # %bb.0:
106 ; AVX512F-NEXT: vbroadcastss {{.*#+}} zmm0 = [707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378]
107 ; AVX512F-NEXT: vmovups %zmm0, (%rdi)
108 ; AVX512F-NEXT: vzeroupper
109 ; AVX512F-NEXT: retq
110 ;
111 ; AVX512BW-LABEL: memset_64_nonzero_bytes:
112 ; AVX512BW: # %bb.0:
113 ; AVX512BW-NEXT: vmovaps {{.*#+}} zmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
114 ; AVX512BW-NEXT: vmovups %zmm0, (%rdi)
115 ; AVX512BW-NEXT: vzeroupper
116 ; AVX512BW-NEXT: retq
117 ; AVX512NW-NEXT: retq
118 %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 64, i64 -1)
119 ret void
120 }
122 define void @memset_128_nonzero_bytes(i8* %x) {
123 ; SSE-LABEL: memset_128_nonzero_bytes:
124 ; SSE: # %bb.0:
125 ; SSE-NEXT: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
126 ; SSE-NEXT: movq %rax, 120(%rdi)
127 ; SSE-NEXT: movq %rax, 112(%rdi)
128 ; SSE-NEXT: movq %rax, 104(%rdi)
129 ; SSE-NEXT: movq %rax, 96(%rdi)
130 ; SSE-NEXT: movq %rax, 88(%rdi)
131 ; SSE-NEXT: movq %rax, 80(%rdi)
132 ; SSE-NEXT: movq %rax, 72(%rdi)
133 ; SSE-NEXT: movq %rax, 64(%rdi)
134 ; SSE-NEXT: movq %rax, 56(%rdi)
135 ; SSE-NEXT: movq %rax, 48(%rdi)
136 ; SSE-NEXT: movq %rax, 40(%rdi)
137 ; SSE-NEXT: movq %rax, 32(%rdi)
138 ; SSE-NEXT: movq %rax, 24(%rdi)
139 ; SSE-NEXT: movq %rax, 16(%rdi)
140 ; SSE-NEXT: movq %rax, 8(%rdi)
141 ; SSE-NEXT: movq %rax, (%rdi)
142 ; SSE-NEXT: retq
143 ;
144 ; SSE2FAST-LABEL: memset_128_nonzero_bytes:
145 ; SSE2FAST: # %bb.0:
146 ; SSE2FAST-NEXT: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
147 ; SSE2FAST-NEXT: movups %xmm0, 112(%rdi)
148 ; SSE2FAST-NEXT: movups %xmm0, 96(%rdi)
149 ; SSE2FAST-NEXT: movups %xmm0, 80(%rdi)
150 ; SSE2FAST-NEXT: movups %xmm0, 64(%rdi)
151 ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi)
152 ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi)
153 ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi)
154 ; SSE2FAST-NEXT: movups %xmm0, (%rdi)
155 ; SSE2FAST-NEXT: retq
156 ;
157 ; AVX1-LABEL: memset_128_nonzero_bytes:
158 ; AVX1: # %bb.0:
159 ; AVX1-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
160 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
161 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
162 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
163 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
164 ; AVX1-NEXT: vzeroupper
165 ; AVX1-NEXT: retq
166 ;
167 ; AVX2-LABEL: memset_128_nonzero_bytes:
168 ; AVX2: # %bb.0:
169 ; AVX2-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
170 ; AVX2-NEXT: vmovups %ymm0, 96(%rdi)
171 ; AVX2-NEXT: vmovups %ymm0, 64(%rdi)
172 ; AVX2-NEXT: vmovups %ymm0, 32(%rdi)
173 ; AVX2-NEXT: vmovups %ymm0, (%rdi)
174 ; AVX2-NEXT: vzeroupper
175 ; AVX2-NEXT: retq
176 ;
177 ; AVX512F-LABEL: memset_128_nonzero_bytes:
178 ; AVX512F: # %bb.0:
179 ; AVX512F-NEXT: vbroadcastss {{.*#+}} zmm0 = [707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378]
180 ; AVX512F-NEXT: vmovups %zmm0, 64(%rdi)
181 ; AVX512F-NEXT: vmovups %zmm0, (%rdi)
182 ; AVX512F-NEXT: vzeroupper
183 ; AVX512F-NEXT: retq
184 ;
185 ; AVX512BW-LABEL: memset_128_nonzero_bytes:
186 ; AVX512BW: # %bb.0:
187 ; AVX512BW-NEXT: vmovaps {{.*#+}} zmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
188 ; AVX512BW-NEXT: vmovups %zmm0, 64(%rdi)
189 ; AVX512BW-NEXT: vmovups %zmm0, (%rdi)
190 ; AVX512BW-NEXT: vzeroupper
191 ; AVX512BW-NEXT: retq
192 %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 128, i64 -1)
193 ret void
194 }
196 define void @memset_256_nonzero_bytes(i8* %x) {
197 ; SSE-LABEL: memset_256_nonzero_bytes:
198 ; SSE: # %bb.0:
199 ; SSE-NEXT: pushq %rax
200 ; SSE-NEXT: .cfi_def_cfa_offset 16
201 ; SSE-NEXT: movl $256, %edx # imm = 0x100
202 ; SSE-NEXT: movl $42, %esi
203 ; SSE-NEXT: callq memset
204 ; SSE-NEXT: popq %rax
205 ; SSE-NEXT: .cfi_def_cfa_offset 8
206 ; SSE-NEXT: retq
207 ;
208 ; SSE2FAST-LABEL: memset_256_nonzero_bytes:
209 ; SSE2FAST: # %bb.0:
210 ; SSE2FAST-NEXT: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
211 ; SSE2FAST-NEXT: movups %xmm0, 240(%rdi)
212 ; SSE2FAST-NEXT: movups %xmm0, 224(%rdi)
213 ; SSE2FAST-NEXT: movups %xmm0, 208(%rdi)
214 ; SSE2FAST-NEXT: movups %xmm0, 192(%rdi)
215 ; SSE2FAST-NEXT: movups %xmm0, 176(%rdi)
216 ; SSE2FAST-NEXT: movups %xmm0, 160(%rdi)
217 ; SSE2FAST-NEXT: movups %xmm0, 144(%rdi)
218 ; SSE2FAST-NEXT: movups %xmm0, 128(%rdi)
219 ; SSE2FAST-NEXT: movups %xmm0, 112(%rdi)
220 ; SSE2FAST-NEXT: movups %xmm0, 96(%rdi)
221 ; SSE2FAST-NEXT: movups %xmm0, 80(%rdi)
222 ; SSE2FAST-NEXT: movups %xmm0, 64(%rdi)
223 ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi)
224 ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi)
225 ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi)
226 ; SSE2FAST-NEXT: movups %xmm0, (%rdi)
227 ; SSE2FAST-NEXT: retq
228 ;
229 ; AVX1-LABEL: memset_256_nonzero_bytes:
230 ; AVX1: # %bb.0:
231 ; AVX1-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
232 ; AVX1-NEXT: vmovups %ymm0, 224(%rdi)
233 ; AVX1-NEXT: vmovups %ymm0, 192(%rdi)
234 ; AVX1-NEXT: vmovups %ymm0, 160(%rdi)
235 ; AVX1-NEXT: vmovups %ymm0, 128(%rdi)
236 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
237 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
238 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
239 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
240 ; AVX1-NEXT: vzeroupper
241 ; AVX1-NEXT: retq
242 ;
243 ; AVX2-LABEL: memset_256_nonzero_bytes:
244 ; AVX2: # %bb.0:
245 ; AVX2-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
246 ; AVX2-NEXT: vmovups %ymm0, 224(%rdi)
247 ; AVX2-NEXT: vmovups %ymm0, 192(%rdi)
248 ; AVX2-NEXT: vmovups %ymm0, 160(%rdi)
249 ; AVX2-NEXT: vmovups %ymm0, 128(%rdi)
250 ; AVX2-NEXT: vmovups %ymm0, 96(%rdi)
251 ; AVX2-NEXT: vmovups %ymm0, 64(%rdi)
252 ; AVX2-NEXT: vmovups %ymm0, 32(%rdi)
253 ; AVX2-NEXT: vmovups %ymm0, (%rdi)
254 ; AVX2-NEXT: vzeroupper
255 ; AVX2-NEXT: retq
256 ;
257 ; AVX512F-LABEL: memset_256_nonzero_bytes:
258 ; AVX512F: # %bb.0:
259 ; AVX512F-NEXT: vbroadcastss {{.*#+}} zmm0 = [707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378]
260 ; AVX512F-NEXT: vmovups %zmm0, 192(%rdi)
261 ; AVX512F-NEXT: vmovups %zmm0, 128(%rdi)
262 ; AVX512F-NEXT: vmovups %zmm0, 64(%rdi)
263 ; AVX512F-NEXT: vmovups %zmm0, (%rdi)
264 ; AVX512F-NEXT: vzeroupper
265 ; AVX512F-NEXT: retq
266 ;
267 ; AVX512BW-LABEL: memset_256_nonzero_bytes:
268 ; AVX512BW: # %bb.0:
269 ; AVX512BW-NEXT: vmovaps {{.*#+}} zmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
270 ; AVX512BW-NEXT: vmovups %zmm0, 192(%rdi)
271 ; AVX512BW-NEXT: vmovups %zmm0, 128(%rdi)
272 ; AVX512BW-NEXT: vmovups %zmm0, 64(%rdi)
273 ; AVX512BW-NEXT: vmovups %zmm0, (%rdi)
274 ; AVX512BW-NEXT: vzeroupper
275 ; AVX512BW-NEXT: retq
276 %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 256, i64 -1)
277 ret void
278 }
280 declare i8* @__memset_chk(i8*, i32, i64, i64)
282 ; Repeat with a non-constant value for the stores.
284 define void @memset_16_nonconst_bytes(i8* %x, i8 %c) {
285 ; SSE-LABEL: memset_16_nonconst_bytes:
286 ; SSE: # %bb.0:
287 ; SSE-NEXT: # kill: def $esi killed $esi def $rsi
288 ; SSE-NEXT: movzbl %sil, %eax
289 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
290 ; SSE-NEXT: imulq %rax, %rcx
291 ; SSE-NEXT: movq %rcx, 8(%rdi)
292 ; SSE-NEXT: movq %rcx, (%rdi)
293 ; SSE-NEXT: retq
294 ;
295 ; SSE2FAST-LABEL: memset_16_nonconst_bytes:
296 ; SSE2FAST: # %bb.0:
297 ; SSE2FAST-NEXT: movd %esi, %xmm0
298 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
299 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
300 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
301 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
302 ; SSE2FAST-NEXT: retq
303 ;
304 ; AVX1-LABEL: memset_16_nonconst_bytes:
305 ; AVX1: # %bb.0:
306 ; AVX1-NEXT: vmovd %esi, %xmm0
307 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
308 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
309 ; AVX1-NEXT: vmovdqu %xmm0, (%rdi)
310 ; AVX1-NEXT: retq
311 ;
312 ; AVX2-LABEL: memset_16_nonconst_bytes:
313 ; AVX2: # %bb.0:
314 ; AVX2-NEXT: vmovd %esi, %xmm0
315 ; AVX2-NEXT: vpbroadcastb %xmm0, %xmm0
316 ; AVX2-NEXT: vmovdqu %xmm0, (%rdi)
317 ; AVX2-NEXT: retq
318 ;
319 ; AVX512-LABEL: memset_16_nonconst_bytes:
320 ; AVX512: # %bb.0:
321 ; AVX512-NEXT: vmovd %esi, %xmm0
322 ; AVX512-NEXT: vpbroadcastb %xmm0, %xmm0
323 ; AVX512-NEXT: vmovdqu %xmm0, (%rdi)
324 ; AVX512-NEXT: retq
325 tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 16, i1 false)
326 ret void
327 }
329 define void @memset_32_nonconst_bytes(i8* %x, i8 %c) {
330 ; SSE-LABEL: memset_32_nonconst_bytes:
331 ; SSE: # %bb.0:
332 ; SSE-NEXT: # kill: def $esi killed $esi def $rsi
333 ; SSE-NEXT: movzbl %sil, %eax
334 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
335 ; SSE-NEXT: imulq %rax, %rcx
336 ; SSE-NEXT: movq %rcx, 24(%rdi)
337 ; SSE-NEXT: movq %rcx, 16(%rdi)
338 ; SSE-NEXT: movq %rcx, 8(%rdi)
339 ; SSE-NEXT: movq %rcx, (%rdi)
340 ; SSE-NEXT: retq
341 ;
342 ; SSE2FAST-LABEL: memset_32_nonconst_bytes:
343 ; SSE2FAST: # %bb.0:
344 ; SSE2FAST-NEXT: movd %esi, %xmm0
345 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
346 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
347 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
348 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
349 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
350 ; SSE2FAST-NEXT: retq
351 ;
352 ; AVX1-LABEL: memset_32_nonconst_bytes:
353 ; AVX1: # %bb.0:
354 ; AVX1-NEXT: vmovd %esi, %xmm0
355 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
356 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
357 ; AVX1-NEXT: vmovdqu %xmm0, 16(%rdi)
358 ; AVX1-NEXT: vmovdqu %xmm0, (%rdi)
359 ; AVX1-NEXT: retq
360 ;
361 ; AVX2-LABEL: memset_32_nonconst_bytes:
362 ; AVX2: # %bb.0:
363 ; AVX2-NEXT: vmovd %esi, %xmm0
364 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
365 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
366 ; AVX2-NEXT: vzeroupper
367 ; AVX2-NEXT: retq
368 ;
369 ; AVX512-LABEL: memset_32_nonconst_bytes:
370 ; AVX512: # %bb.0:
371 ; AVX512-NEXT: vmovd %esi, %xmm0
372 ; AVX512-NEXT: vpbroadcastb %xmm0, %ymm0
373 ; AVX512-NEXT: vmovdqu %ymm0, (%rdi)
374 ; AVX512-NEXT: vzeroupper
375 ; AVX512-NEXT: retq
376 tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 32, i1 false)
377 ret void
378 }
380 define void @memset_64_nonconst_bytes(i8* %x, i8 %c) {
381 ; SSE-LABEL: memset_64_nonconst_bytes:
382 ; SSE: # %bb.0:
383 ; SSE-NEXT: # kill: def $esi killed $esi def $rsi
384 ; SSE-NEXT: movzbl %sil, %eax
385 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
386 ; SSE-NEXT: imulq %rax, %rcx
387 ; SSE-NEXT: movq %rcx, 56(%rdi)
388 ; SSE-NEXT: movq %rcx, 48(%rdi)
389 ; SSE-NEXT: movq %rcx, 40(%rdi)
390 ; SSE-NEXT: movq %rcx, 32(%rdi)
391 ; SSE-NEXT: movq %rcx, 24(%rdi)
392 ; SSE-NEXT: movq %rcx, 16(%rdi)
393 ; SSE-NEXT: movq %rcx, 8(%rdi)
394 ; SSE-NEXT: movq %rcx, (%rdi)
395 ; SSE-NEXT: retq
396 ;
397 ; SSE2FAST-LABEL: memset_64_nonconst_bytes:
398 ; SSE2FAST: # %bb.0:
399 ; SSE2FAST-NEXT: movd %esi, %xmm0
400 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
401 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
402 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
403 ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi)
404 ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi)
405 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
406 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
407 ; SSE2FAST-NEXT: retq
408 ;
409 ; AVX1-LABEL: memset_64_nonconst_bytes:
410 ; AVX1: # %bb.0:
411 ; AVX1-NEXT: vmovd %esi, %xmm0
412 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
413 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
414 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
415 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
416 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
417 ; AVX1-NEXT: vzeroupper
418 ; AVX1-NEXT: retq
419 ;
420 ; AVX2-LABEL: memset_64_nonconst_bytes:
421 ; AVX2: # %bb.0:
422 ; AVX2-NEXT: vmovd %esi, %xmm0
423 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
424 ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi)
425 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
426 ; AVX2-NEXT: vzeroupper
427 ; AVX2-NEXT: retq
428 ;
429 ; AVX512F-LABEL: memset_64_nonconst_bytes:
430 ; AVX512F: # %bb.0:
431 ; AVX512F-NEXT: movzbl %sil, %eax
432 ; AVX512F-NEXT: imull $16843009, %eax, %eax # imm = 0x1010101
433 ; AVX512F-NEXT: vpbroadcastd %eax, %zmm0
434 ; AVX512F-NEXT: vmovdqu64 %zmm0, (%rdi)
435 ; AVX512F-NEXT: vzeroupper
436 ; AVX512F-NEXT: retq
437 ;
438 ; AVX512BW-LABEL: memset_64_nonconst_bytes:
439 ; AVX512BW: # %bb.0:
440 ; AVX512BW-NEXT: vpbroadcastb %esi, %zmm0
441 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rdi)
442 ; AVX512BW-NEXT: vzeroupper
443 ; AVX512BW-NEXT: retq
444 tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 64, i1 false)
445 ret void
446 }
448 define void @memset_128_nonconst_bytes(i8* %x, i8 %c) {
449 ; SSE-LABEL: memset_128_nonconst_bytes:
450 ; SSE: # %bb.0:
451 ; SSE-NEXT: # kill: def $esi killed $esi def $rsi
452 ; SSE-NEXT: movzbl %sil, %eax
453 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
454 ; SSE-NEXT: imulq %rax, %rcx
455 ; SSE-NEXT: movq %rcx, 120(%rdi)
456 ; SSE-NEXT: movq %rcx, 112(%rdi)
457 ; SSE-NEXT: movq %rcx, 104(%rdi)
458 ; SSE-NEXT: movq %rcx, 96(%rdi)
459 ; SSE-NEXT: movq %rcx, 88(%rdi)
460 ; SSE-NEXT: movq %rcx, 80(%rdi)
461 ; SSE-NEXT: movq %rcx, 72(%rdi)
462 ; SSE-NEXT: movq %rcx, 64(%rdi)
463 ; SSE-NEXT: movq %rcx, 56(%rdi)
464 ; SSE-NEXT: movq %rcx, 48(%rdi)
465 ; SSE-NEXT: movq %rcx, 40(%rdi)
466 ; SSE-NEXT: movq %rcx, 32(%rdi)
467 ; SSE-NEXT: movq %rcx, 24(%rdi)
468 ; SSE-NEXT: movq %rcx, 16(%rdi)
469 ; SSE-NEXT: movq %rcx, 8(%rdi)
470 ; SSE-NEXT: movq %rcx, (%rdi)
471 ; SSE-NEXT: retq
472 ;
473 ; SSE2FAST-LABEL: memset_128_nonconst_bytes:
474 ; SSE2FAST: # %bb.0:
475 ; SSE2FAST-NEXT: movd %esi, %xmm0
476 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
477 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
478 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
479 ; SSE2FAST-NEXT: movdqu %xmm0, 112(%rdi)
480 ; SSE2FAST-NEXT: movdqu %xmm0, 96(%rdi)
481 ; SSE2FAST-NEXT: movdqu %xmm0, 80(%rdi)
482 ; SSE2FAST-NEXT: movdqu %xmm0, 64(%rdi)
483 ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi)
484 ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi)
485 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
486 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
487 ; SSE2FAST-NEXT: retq
488 ;
489 ; AVX1-LABEL: memset_128_nonconst_bytes:
490 ; AVX1: # %bb.0:
491 ; AVX1-NEXT: vmovd %esi, %xmm0
492 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
493 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
494 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
495 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
496 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
497 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
498 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
499 ; AVX1-NEXT: vzeroupper
500 ; AVX1-NEXT: retq
501 ;
502 ; AVX2-LABEL: memset_128_nonconst_bytes:
503 ; AVX2: # %bb.0:
504 ; AVX2-NEXT: vmovd %esi, %xmm0
505 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
506 ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi)
507 ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi)
508 ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi)
509 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
510 ; AVX2-NEXT: vzeroupper
511 ; AVX2-NEXT: retq
512 ;
513 ; AVX512F-LABEL: memset_128_nonconst_bytes:
514 ; AVX512F: # %bb.0:
515 ; AVX512F-NEXT: movzbl %sil, %eax
516 ; AVX512F-NEXT: imull $16843009, %eax, %eax # imm = 0x1010101
517 ; AVX512F-NEXT: vpbroadcastd %eax, %zmm0
518 ; AVX512F-NEXT: vmovdqu64 %zmm0, 64(%rdi)
519 ; AVX512F-NEXT: vmovdqu64 %zmm0, (%rdi)
520 ; AVX512F-NEXT: vzeroupper
521 ; AVX512F-NEXT: retq
522 ;
523 ; AVX512BW-LABEL: memset_128_nonconst_bytes:
524 ; AVX512BW: # %bb.0:
525 ; AVX512BW-NEXT: vpbroadcastb %esi, %zmm0
526 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 64(%rdi)
527 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rdi)
528 ; AVX512BW-NEXT: vzeroupper
529 ; AVX512BW-NEXT: retq
530 tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 128, i1 false)
531 ret void
532 }
534 define void @memset_256_nonconst_bytes(i8* %x, i8 %c) {
535 ; SSE-LABEL: memset_256_nonconst_bytes:
536 ; SSE: # %bb.0:
537 ; SSE-NEXT: movl $256, %edx # imm = 0x100
538 ; SSE-NEXT: jmp memset # TAILCALL
539 ;
540 ; SSE2FAST-LABEL: memset_256_nonconst_bytes:
541 ; SSE2FAST: # %bb.0:
542 ; SSE2FAST-NEXT: movd %esi, %xmm0
543 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
544 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
545 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
546 ; SSE2FAST-NEXT: movdqu %xmm0, 240(%rdi)
547 ; SSE2FAST-NEXT: movdqu %xmm0, 224(%rdi)
548 ; SSE2FAST-NEXT: movdqu %xmm0, 208(%rdi)
549 ; SSE2FAST-NEXT: movdqu %xmm0, 192(%rdi)
550 ; SSE2FAST-NEXT: movdqu %xmm0, 176(%rdi)
551 ; SSE2FAST-NEXT: movdqu %xmm0, 160(%rdi)
552 ; SSE2FAST-NEXT: movdqu %xmm0, 144(%rdi)
553 ; SSE2FAST-NEXT: movdqu %xmm0, 128(%rdi)
554 ; SSE2FAST-NEXT: movdqu %xmm0, 112(%rdi)
555 ; SSE2FAST-NEXT: movdqu %xmm0, 96(%rdi)
556 ; SSE2FAST-NEXT: movdqu %xmm0, 80(%rdi)
557 ; SSE2FAST-NEXT: movdqu %xmm0, 64(%rdi)
558 ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi)
559 ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi)
560 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
561 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
562 ; SSE2FAST-NEXT: retq
563 ;
564 ; AVX1-LABEL: memset_256_nonconst_bytes:
565 ; AVX1: # %bb.0:
566 ; AVX1-NEXT: vmovd %esi, %xmm0
567 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
568 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
569 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
570 ; AVX1-NEXT: vmovups %ymm0, 224(%rdi)
571 ; AVX1-NEXT: vmovups %ymm0, 192(%rdi)
572 ; AVX1-NEXT: vmovups %ymm0, 160(%rdi)
573 ; AVX1-NEXT: vmovups %ymm0, 128(%rdi)
574 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
575 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
576 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
577 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
578 ; AVX1-NEXT: vzeroupper
579 ; AVX1-NEXT: retq
580 ;
581 ; AVX2-LABEL: memset_256_nonconst_bytes:
582 ; AVX2: # %bb.0:
583 ; AVX2-NEXT: vmovd %esi, %xmm0
584 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
585 ; AVX2-NEXT: vmovdqu %ymm0, 224(%rdi)
586 ; AVX2-NEXT: vmovdqu %ymm0, 192(%rdi)
587 ; AVX2-NEXT: vmovdqu %ymm0, 160(%rdi)
588 ; AVX2-NEXT: vmovdqu %ymm0, 128(%rdi)
589 ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi)
590 ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi)
591 ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi)
592 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
593 ; AVX2-NEXT: vzeroupper
594 ; AVX2-NEXT: retq
595 ;
596 ; AVX512F-LABEL: memset_256_nonconst_bytes:
597 ; AVX512F: # %bb.0:
598 ; AVX512F-NEXT: movzbl %sil, %eax
599 ; AVX512F-NEXT: imull $16843009, %eax, %eax # imm = 0x1010101
600 ; AVX512F-NEXT: vpbroadcastd %eax, %zmm0
601 ; AVX512F-NEXT: vmovdqu64 %zmm0, 192(%rdi)
602 ; AVX512F-NEXT: vmovdqu64 %zmm0, 128(%rdi)
603 ; AVX512F-NEXT: vmovdqu64 %zmm0, 64(%rdi)
604 ; AVX512F-NEXT: vmovdqu64 %zmm0, (%rdi)
605 ; AVX512F-NEXT: vzeroupper
606 ; AVX512F-NEXT: retq
607 ;
608 ; AVX512BW-LABEL: memset_256_nonconst_bytes:
609 ; AVX512BW: # %bb.0:
610 ; AVX512BW-NEXT: vpbroadcastb %esi, %zmm0
611 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 192(%rdi)
612 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 128(%rdi)
613 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 64(%rdi)
614 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rdi)
615 ; AVX512BW-NEXT: vzeroupper
616 ; AVX512BW-NEXT: retq
617 tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 256, i1 false)
618 ret void
619 }
621 declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i1) #1