| blob | 96ac8dff79530da60dc0a7dbaa21ed8342eb2d17 |
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(ptr %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: vbroadcastss {{.*#+}} 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 ptr @__memset_chk(ptr %x, i32 42, i64 16, i64 -1)
35 ret void
36 }
38 define void @memset_32_nonzero_bytes(ptr %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: vbroadcastss {{.*#+}} 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 ptr @__memset_chk(ptr %x, i32 42, i64 32, i64 -1)
62 ret void
63 }
65 define void @memset_64_nonzero_bytes(ptr %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: vbroadcastss {{.*#+}} 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: vbroadcastss {{.*#+}} 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: vbroadcastss {{.*#+}} 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 ptr @__memset_chk(ptr %x, i32 42, i64 64, i64 -1)
119 ret void
120 }
122 define void @memset_128_nonzero_bytes(ptr %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: vbroadcastss {{.*#+}} 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: vbroadcastss {{.*#+}} 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: vbroadcastss {{.*#+}} 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 ptr @__memset_chk(ptr %x, i32 42, i64 128, i64 -1)
193 ret void
194 }
196 define void @memset_256_nonzero_bytes(ptr %x) {
197 ; SSE-LABEL: memset_256_nonzero_bytes:
198 ; SSE: # %bb.0:
199 ; SSE-NEXT: movl $256, %edx # imm = 0x100
200 ; SSE-NEXT: movl $42, %esi
201 ; SSE-NEXT: jmp memset@PLT # TAILCALL
202 ;
203 ; SSE2FAST-LABEL: memset_256_nonzero_bytes:
204 ; SSE2FAST: # %bb.0:
205 ; SSE2FAST-NEXT: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
206 ; SSE2FAST-NEXT: movups %xmm0, 240(%rdi)
207 ; SSE2FAST-NEXT: movups %xmm0, 224(%rdi)
208 ; SSE2FAST-NEXT: movups %xmm0, 208(%rdi)
209 ; SSE2FAST-NEXT: movups %xmm0, 192(%rdi)
210 ; SSE2FAST-NEXT: movups %xmm0, 176(%rdi)
211 ; SSE2FAST-NEXT: movups %xmm0, 160(%rdi)
212 ; SSE2FAST-NEXT: movups %xmm0, 144(%rdi)
213 ; SSE2FAST-NEXT: movups %xmm0, 128(%rdi)
214 ; SSE2FAST-NEXT: movups %xmm0, 112(%rdi)
215 ; SSE2FAST-NEXT: movups %xmm0, 96(%rdi)
216 ; SSE2FAST-NEXT: movups %xmm0, 80(%rdi)
217 ; SSE2FAST-NEXT: movups %xmm0, 64(%rdi)
218 ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi)
219 ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi)
220 ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi)
221 ; SSE2FAST-NEXT: movups %xmm0, (%rdi)
222 ; SSE2FAST-NEXT: retq
223 ;
224 ; AVX1-LABEL: memset_256_nonzero_bytes:
225 ; AVX1: # %bb.0:
226 ; AVX1-NEXT: vbroadcastss {{.*#+}} 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]
227 ; AVX1-NEXT: vmovups %ymm0, 224(%rdi)
228 ; AVX1-NEXT: vmovups %ymm0, 192(%rdi)
229 ; AVX1-NEXT: vmovups %ymm0, 160(%rdi)
230 ; AVX1-NEXT: vmovups %ymm0, 128(%rdi)
231 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
232 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
233 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
234 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
235 ; AVX1-NEXT: vzeroupper
236 ; AVX1-NEXT: retq
237 ;
238 ; AVX2-LABEL: memset_256_nonzero_bytes:
239 ; AVX2: # %bb.0:
240 ; AVX2-NEXT: vbroadcastss {{.*#+}} 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]
241 ; AVX2-NEXT: vmovups %ymm0, 224(%rdi)
242 ; AVX2-NEXT: vmovups %ymm0, 192(%rdi)
243 ; AVX2-NEXT: vmovups %ymm0, 160(%rdi)
244 ; AVX2-NEXT: vmovups %ymm0, 128(%rdi)
245 ; AVX2-NEXT: vmovups %ymm0, 96(%rdi)
246 ; AVX2-NEXT: vmovups %ymm0, 64(%rdi)
247 ; AVX2-NEXT: vmovups %ymm0, 32(%rdi)
248 ; AVX2-NEXT: vmovups %ymm0, (%rdi)
249 ; AVX2-NEXT: vzeroupper
250 ; AVX2-NEXT: retq
251 ;
252 ; AVX512F-LABEL: memset_256_nonzero_bytes:
253 ; AVX512F: # %bb.0:
254 ; AVX512F-NEXT: vbroadcastss {{.*#+}} zmm0 = [707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378,707406378]
255 ; AVX512F-NEXT: vmovups %zmm0, 192(%rdi)
256 ; AVX512F-NEXT: vmovups %zmm0, 128(%rdi)
257 ; AVX512F-NEXT: vmovups %zmm0, 64(%rdi)
258 ; AVX512F-NEXT: vmovups %zmm0, (%rdi)
259 ; AVX512F-NEXT: vzeroupper
260 ; AVX512F-NEXT: retq
261 ;
262 ; AVX512BW-LABEL: memset_256_nonzero_bytes:
263 ; AVX512BW: # %bb.0:
264 ; AVX512BW-NEXT: vbroadcastss {{.*#+}} 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]
265 ; AVX512BW-NEXT: vmovups %zmm0, 192(%rdi)
266 ; AVX512BW-NEXT: vmovups %zmm0, 128(%rdi)
267 ; AVX512BW-NEXT: vmovups %zmm0, 64(%rdi)
268 ; AVX512BW-NEXT: vmovups %zmm0, (%rdi)
269 ; AVX512BW-NEXT: vzeroupper
270 ; AVX512BW-NEXT: retq
271 %call = tail call ptr @__memset_chk(ptr %x, i32 42, i64 256, i64 -1)
272 ret void
273 }
275 declare ptr @__memset_chk(ptr, i32, i64, i64)
277 ; Repeat with a non-constant value for the stores.
279 define void @memset_16_nonconst_bytes(ptr %x, i8 %c) {
280 ; SSE-LABEL: memset_16_nonconst_bytes:
281 ; SSE: # %bb.0:
282 ; SSE-NEXT: movzbl %sil, %eax
283 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
284 ; SSE-NEXT: imulq %rax, %rcx
285 ; SSE-NEXT: movq %rcx, 8(%rdi)
286 ; SSE-NEXT: movq %rcx, (%rdi)
287 ; SSE-NEXT: retq
288 ;
289 ; SSE2FAST-LABEL: memset_16_nonconst_bytes:
290 ; SSE2FAST: # %bb.0:
291 ; SSE2FAST-NEXT: movd %esi, %xmm0
292 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
293 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
294 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
295 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
296 ; SSE2FAST-NEXT: retq
297 ;
298 ; AVX1-LABEL: memset_16_nonconst_bytes:
299 ; AVX1: # %bb.0:
300 ; AVX1-NEXT: vmovd %esi, %xmm0
301 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
302 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
303 ; AVX1-NEXT: vmovdqu %xmm0, (%rdi)
304 ; AVX1-NEXT: retq
305 ;
306 ; AVX2-LABEL: memset_16_nonconst_bytes:
307 ; AVX2: # %bb.0:
308 ; AVX2-NEXT: vmovd %esi, %xmm0
309 ; AVX2-NEXT: vpbroadcastb %xmm0, %xmm0
310 ; AVX2-NEXT: vmovdqu %xmm0, (%rdi)
311 ; AVX2-NEXT: retq
312 ;
313 ; AVX512-LABEL: memset_16_nonconst_bytes:
314 ; AVX512: # %bb.0:
315 ; AVX512-NEXT: vmovd %esi, %xmm0
316 ; AVX512-NEXT: vpbroadcastb %xmm0, %xmm0
317 ; AVX512-NEXT: vmovdqu %xmm0, (%rdi)
318 ; AVX512-NEXT: retq
319 tail call void @llvm.memset.p0.i64(ptr %x, i8 %c, i64 16, i1 false)
320 ret void
321 }
323 define void @memset_32_nonconst_bytes(ptr %x, i8 %c) {
324 ; SSE-LABEL: memset_32_nonconst_bytes:
325 ; SSE: # %bb.0:
326 ; SSE-NEXT: movzbl %sil, %eax
327 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
328 ; SSE-NEXT: imulq %rax, %rcx
329 ; SSE-NEXT: movq %rcx, 24(%rdi)
330 ; SSE-NEXT: movq %rcx, 16(%rdi)
331 ; SSE-NEXT: movq %rcx, 8(%rdi)
332 ; SSE-NEXT: movq %rcx, (%rdi)
333 ; SSE-NEXT: retq
334 ;
335 ; SSE2FAST-LABEL: memset_32_nonconst_bytes:
336 ; SSE2FAST: # %bb.0:
337 ; SSE2FAST-NEXT: movd %esi, %xmm0
338 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
339 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
340 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
341 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
342 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
343 ; SSE2FAST-NEXT: retq
344 ;
345 ; AVX1-LABEL: memset_32_nonconst_bytes:
346 ; AVX1: # %bb.0:
347 ; AVX1-NEXT: vmovd %esi, %xmm0
348 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
349 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
350 ; AVX1-NEXT: vmovdqu %xmm0, 16(%rdi)
351 ; AVX1-NEXT: vmovdqu %xmm0, (%rdi)
352 ; AVX1-NEXT: retq
353 ;
354 ; AVX2-LABEL: memset_32_nonconst_bytes:
355 ; AVX2: # %bb.0:
356 ; AVX2-NEXT: vmovd %esi, %xmm0
357 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
358 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
359 ; AVX2-NEXT: vzeroupper
360 ; AVX2-NEXT: retq
361 ;
362 ; AVX512-LABEL: memset_32_nonconst_bytes:
363 ; AVX512: # %bb.0:
364 ; AVX512-NEXT: vmovd %esi, %xmm0
365 ; AVX512-NEXT: vpbroadcastb %xmm0, %ymm0
366 ; AVX512-NEXT: vmovdqu %ymm0, (%rdi)
367 ; AVX512-NEXT: vzeroupper
368 ; AVX512-NEXT: retq
369 tail call void @llvm.memset.p0.i64(ptr %x, i8 %c, i64 32, i1 false)
370 ret void
371 }
373 define void @memset_64_nonconst_bytes(ptr %x, i8 %c) {
374 ; SSE-LABEL: memset_64_nonconst_bytes:
375 ; SSE: # %bb.0:
376 ; SSE-NEXT: movzbl %sil, %eax
377 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
378 ; SSE-NEXT: imulq %rax, %rcx
379 ; SSE-NEXT: movq %rcx, 56(%rdi)
380 ; SSE-NEXT: movq %rcx, 48(%rdi)
381 ; SSE-NEXT: movq %rcx, 40(%rdi)
382 ; SSE-NEXT: movq %rcx, 32(%rdi)
383 ; SSE-NEXT: movq %rcx, 24(%rdi)
384 ; SSE-NEXT: movq %rcx, 16(%rdi)
385 ; SSE-NEXT: movq %rcx, 8(%rdi)
386 ; SSE-NEXT: movq %rcx, (%rdi)
387 ; SSE-NEXT: retq
388 ;
389 ; SSE2FAST-LABEL: memset_64_nonconst_bytes:
390 ; SSE2FAST: # %bb.0:
391 ; SSE2FAST-NEXT: movd %esi, %xmm0
392 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
393 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
394 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
395 ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi)
396 ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi)
397 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
398 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
399 ; SSE2FAST-NEXT: retq
400 ;
401 ; AVX1-LABEL: memset_64_nonconst_bytes:
402 ; AVX1: # %bb.0:
403 ; AVX1-NEXT: vmovd %esi, %xmm0
404 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
405 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
406 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
407 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
408 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
409 ; AVX1-NEXT: vzeroupper
410 ; AVX1-NEXT: retq
411 ;
412 ; AVX2-LABEL: memset_64_nonconst_bytes:
413 ; AVX2: # %bb.0:
414 ; AVX2-NEXT: vmovd %esi, %xmm0
415 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
416 ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi)
417 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
418 ; AVX2-NEXT: vzeroupper
419 ; AVX2-NEXT: retq
420 ;
421 ; AVX512F-LABEL: memset_64_nonconst_bytes:
422 ; AVX512F: # %bb.0:
423 ; AVX512F-NEXT: movzbl %sil, %eax
424 ; AVX512F-NEXT: imull $16843009, %eax, %eax # imm = 0x1010101
425 ; AVX512F-NEXT: vpbroadcastd %eax, %zmm0
426 ; AVX512F-NEXT: vmovdqu64 %zmm0, (%rdi)
427 ; AVX512F-NEXT: vzeroupper
428 ; AVX512F-NEXT: retq
429 ;
430 ; AVX512BW-LABEL: memset_64_nonconst_bytes:
431 ; AVX512BW: # %bb.0:
432 ; AVX512BW-NEXT: vpbroadcastb %esi, %zmm0
433 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rdi)
434 ; AVX512BW-NEXT: vzeroupper
435 ; AVX512BW-NEXT: retq
436 tail call void @llvm.memset.p0.i64(ptr %x, i8 %c, i64 64, i1 false)
437 ret void
438 }
440 define void @memset_128_nonconst_bytes(ptr %x, i8 %c) {
441 ; SSE-LABEL: memset_128_nonconst_bytes:
442 ; SSE: # %bb.0:
443 ; SSE-NEXT: movzbl %sil, %eax
444 ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101
445 ; SSE-NEXT: imulq %rax, %rcx
446 ; SSE-NEXT: movq %rcx, 120(%rdi)
447 ; SSE-NEXT: movq %rcx, 112(%rdi)
448 ; SSE-NEXT: movq %rcx, 104(%rdi)
449 ; SSE-NEXT: movq %rcx, 96(%rdi)
450 ; SSE-NEXT: movq %rcx, 88(%rdi)
451 ; SSE-NEXT: movq %rcx, 80(%rdi)
452 ; SSE-NEXT: movq %rcx, 72(%rdi)
453 ; SSE-NEXT: movq %rcx, 64(%rdi)
454 ; SSE-NEXT: movq %rcx, 56(%rdi)
455 ; SSE-NEXT: movq %rcx, 48(%rdi)
456 ; SSE-NEXT: movq %rcx, 40(%rdi)
457 ; SSE-NEXT: movq %rcx, 32(%rdi)
458 ; SSE-NEXT: movq %rcx, 24(%rdi)
459 ; SSE-NEXT: movq %rcx, 16(%rdi)
460 ; SSE-NEXT: movq %rcx, 8(%rdi)
461 ; SSE-NEXT: movq %rcx, (%rdi)
462 ; SSE-NEXT: retq
463 ;
464 ; SSE2FAST-LABEL: memset_128_nonconst_bytes:
465 ; SSE2FAST: # %bb.0:
466 ; SSE2FAST-NEXT: movd %esi, %xmm0
467 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
468 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
469 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
470 ; SSE2FAST-NEXT: movdqu %xmm0, 112(%rdi)
471 ; SSE2FAST-NEXT: movdqu %xmm0, 96(%rdi)
472 ; SSE2FAST-NEXT: movdqu %xmm0, 80(%rdi)
473 ; SSE2FAST-NEXT: movdqu %xmm0, 64(%rdi)
474 ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi)
475 ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi)
476 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
477 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
478 ; SSE2FAST-NEXT: retq
479 ;
480 ; AVX1-LABEL: memset_128_nonconst_bytes:
481 ; AVX1: # %bb.0:
482 ; AVX1-NEXT: vmovd %esi, %xmm0
483 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
484 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
485 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
486 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
487 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
488 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
489 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
490 ; AVX1-NEXT: vzeroupper
491 ; AVX1-NEXT: retq
492 ;
493 ; AVX2-LABEL: memset_128_nonconst_bytes:
494 ; AVX2: # %bb.0:
495 ; AVX2-NEXT: vmovd %esi, %xmm0
496 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
497 ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi)
498 ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi)
499 ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi)
500 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
501 ; AVX2-NEXT: vzeroupper
502 ; AVX2-NEXT: retq
503 ;
504 ; AVX512F-LABEL: memset_128_nonconst_bytes:
505 ; AVX512F: # %bb.0:
506 ; AVX512F-NEXT: movzbl %sil, %eax
507 ; AVX512F-NEXT: imull $16843009, %eax, %eax # imm = 0x1010101
508 ; AVX512F-NEXT: vpbroadcastd %eax, %zmm0
509 ; AVX512F-NEXT: vmovdqu64 %zmm0, 64(%rdi)
510 ; AVX512F-NEXT: vmovdqu64 %zmm0, (%rdi)
511 ; AVX512F-NEXT: vzeroupper
512 ; AVX512F-NEXT: retq
513 ;
514 ; AVX512BW-LABEL: memset_128_nonconst_bytes:
515 ; AVX512BW: # %bb.0:
516 ; AVX512BW-NEXT: vpbroadcastb %esi, %zmm0
517 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 64(%rdi)
518 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rdi)
519 ; AVX512BW-NEXT: vzeroupper
520 ; AVX512BW-NEXT: retq
521 tail call void @llvm.memset.p0.i64(ptr %x, i8 %c, i64 128, i1 false)
522 ret void
523 }
525 define void @memset_256_nonconst_bytes(ptr %x, i8 %c) {
526 ; SSE-LABEL: memset_256_nonconst_bytes:
527 ; SSE: # %bb.0:
528 ; SSE-NEXT: movl $256, %edx # imm = 0x100
529 ; SSE-NEXT: jmp memset@PLT # TAILCALL
530 ;
531 ; SSE2FAST-LABEL: memset_256_nonconst_bytes:
532 ; SSE2FAST: # %bb.0:
533 ; SSE2FAST-NEXT: movd %esi, %xmm0
534 ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
535 ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
536 ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
537 ; SSE2FAST-NEXT: movdqu %xmm0, 240(%rdi)
538 ; SSE2FAST-NEXT: movdqu %xmm0, 224(%rdi)
539 ; SSE2FAST-NEXT: movdqu %xmm0, 208(%rdi)
540 ; SSE2FAST-NEXT: movdqu %xmm0, 192(%rdi)
541 ; SSE2FAST-NEXT: movdqu %xmm0, 176(%rdi)
542 ; SSE2FAST-NEXT: movdqu %xmm0, 160(%rdi)
543 ; SSE2FAST-NEXT: movdqu %xmm0, 144(%rdi)
544 ; SSE2FAST-NEXT: movdqu %xmm0, 128(%rdi)
545 ; SSE2FAST-NEXT: movdqu %xmm0, 112(%rdi)
546 ; SSE2FAST-NEXT: movdqu %xmm0, 96(%rdi)
547 ; SSE2FAST-NEXT: movdqu %xmm0, 80(%rdi)
548 ; SSE2FAST-NEXT: movdqu %xmm0, 64(%rdi)
549 ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi)
550 ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi)
551 ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi)
552 ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi)
553 ; SSE2FAST-NEXT: retq
554 ;
555 ; AVX1-LABEL: memset_256_nonconst_bytes:
556 ; AVX1: # %bb.0:
557 ; AVX1-NEXT: vmovd %esi, %xmm0
558 ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
559 ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0
560 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
561 ; AVX1-NEXT: vmovups %ymm0, 224(%rdi)
562 ; AVX1-NEXT: vmovups %ymm0, 192(%rdi)
563 ; AVX1-NEXT: vmovups %ymm0, 160(%rdi)
564 ; AVX1-NEXT: vmovups %ymm0, 128(%rdi)
565 ; AVX1-NEXT: vmovups %ymm0, 96(%rdi)
566 ; AVX1-NEXT: vmovups %ymm0, 64(%rdi)
567 ; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
568 ; AVX1-NEXT: vmovups %ymm0, (%rdi)
569 ; AVX1-NEXT: vzeroupper
570 ; AVX1-NEXT: retq
571 ;
572 ; AVX2-LABEL: memset_256_nonconst_bytes:
573 ; AVX2: # %bb.0:
574 ; AVX2-NEXT: vmovd %esi, %xmm0
575 ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
576 ; AVX2-NEXT: vmovdqu %ymm0, 224(%rdi)
577 ; AVX2-NEXT: vmovdqu %ymm0, 192(%rdi)
578 ; AVX2-NEXT: vmovdqu %ymm0, 160(%rdi)
579 ; AVX2-NEXT: vmovdqu %ymm0, 128(%rdi)
580 ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi)
581 ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi)
582 ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi)
583 ; AVX2-NEXT: vmovdqu %ymm0, (%rdi)
584 ; AVX2-NEXT: vzeroupper
585 ; AVX2-NEXT: retq
586 ;
587 ; AVX512F-LABEL: memset_256_nonconst_bytes:
588 ; AVX512F: # %bb.0:
589 ; AVX512F-NEXT: movzbl %sil, %eax
590 ; AVX512F-NEXT: imull $16843009, %eax, %eax # imm = 0x1010101
591 ; AVX512F-NEXT: vpbroadcastd %eax, %zmm0
592 ; AVX512F-NEXT: vmovdqu64 %zmm0, 192(%rdi)
593 ; AVX512F-NEXT: vmovdqu64 %zmm0, 128(%rdi)
594 ; AVX512F-NEXT: vmovdqu64 %zmm0, 64(%rdi)
595 ; AVX512F-NEXT: vmovdqu64 %zmm0, (%rdi)
596 ; AVX512F-NEXT: vzeroupper
597 ; AVX512F-NEXT: retq
598 ;
599 ; AVX512BW-LABEL: memset_256_nonconst_bytes:
600 ; AVX512BW: # %bb.0:
601 ; AVX512BW-NEXT: vpbroadcastb %esi, %zmm0
602 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 192(%rdi)
603 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 128(%rdi)
604 ; AVX512BW-NEXT: vmovdqu64 %zmm0, 64(%rdi)
605 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rdi)
606 ; AVX512BW-NEXT: vzeroupper
607 ; AVX512BW-NEXT: retq
608 tail call void @llvm.memset.p0.i64(ptr %x, i8 %c, i64 256, i1 false)
609 ret void
610 }
612 declare void @llvm.memset.p0.i64(ptr nocapture, i8, i64, i1) #1