Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 1 | ; NOTE: Assertions have been autogenerated by update_test_checks.py |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 2 | ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse | FileCheck %s --check-prefix=ANY --check-prefix=SSE --check-prefix=SSE1 |
| 3 | ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse2 | FileCheck %s --check-prefix=ANY --check-prefix=SSE --check-prefix=SSE2 |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 4 | ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx | FileCheck %s --check-prefix=ANY --check-prefix=AVX --check-prefix=AVX1 |
| 5 | ; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx2 | FileCheck %s --check-prefix=ANY --check-prefix=AVX --check-prefix=AVX2 |
| 6 | |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 7 | ; https://llvm.org/bugs/show_bug.cgi?id=27100 |
| 8 | |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 9 | define void @memset_16_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 10 | ; SSE-LABEL: memset_16_nonzero_bytes: |
| 11 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 12 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 13 | ; SSE-NEXT: movq %rax, (%rdi) |
| 14 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 15 | ; |
Sanjay Patel | 9f41336 | 2016-04-01 17:36:45 +0000 | [diff] [blame] | 16 | ; AVX-LABEL: memset_16_nonzero_bytes: |
| 17 | ; AVX: vmovaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 18 | ; AVX-NEXT: vmovups %xmm0, (%rdi) |
| 19 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 20 | ; |
| 21 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 16, i64 -1) |
| 22 | ret void |
| 23 | } |
| 24 | |
| 25 | define void @memset_32_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 26 | ; SSE-LABEL: memset_32_nonzero_bytes: |
| 27 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 28 | ; SSE-NEXT: movq %rax, 24(%rdi) |
| 29 | ; SSE-NEXT: movq %rax, 16(%rdi) |
| 30 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 31 | ; SSE-NEXT: movq %rax, (%rdi) |
| 32 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 33 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 34 | ; AVX-LABEL: memset_32_nonzero_bytes: |
| 35 | ; AVX: 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] |
| 36 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 37 | ; AVX-NEXT: vzeroupper |
| 38 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 39 | ; |
| 40 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 32, i64 -1) |
| 41 | ret void |
| 42 | } |
| 43 | |
| 44 | define void @memset_64_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 45 | ; SSE-LABEL: memset_64_nonzero_bytes: |
| 46 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 47 | ; SSE-NEXT: movq %rax, 56(%rdi) |
| 48 | ; SSE-NEXT: movq %rax, 48(%rdi) |
| 49 | ; SSE-NEXT: movq %rax, 40(%rdi) |
| 50 | ; SSE-NEXT: movq %rax, 32(%rdi) |
| 51 | ; SSE-NEXT: movq %rax, 24(%rdi) |
| 52 | ; SSE-NEXT: movq %rax, 16(%rdi) |
| 53 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 54 | ; SSE-NEXT: movq %rax, (%rdi) |
| 55 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 56 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 57 | ; AVX-LABEL: memset_64_nonzero_bytes: |
| 58 | ; AVX: 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] |
| 59 | ; AVX-NEXT: vmovups %ymm0, 32(%rdi) |
| 60 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 61 | ; AVX-NEXT: vzeroupper |
| 62 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 63 | ; |
| 64 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 64, i64 -1) |
| 65 | ret void |
| 66 | } |
| 67 | |
| 68 | define void @memset_128_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 69 | ; SSE-LABEL: memset_128_nonzero_bytes: |
| 70 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 71 | ; SSE-NEXT: movq %rax, 120(%rdi) |
| 72 | ; SSE-NEXT: movq %rax, 112(%rdi) |
| 73 | ; SSE-NEXT: movq %rax, 104(%rdi) |
| 74 | ; SSE-NEXT: movq %rax, 96(%rdi) |
| 75 | ; SSE-NEXT: movq %rax, 88(%rdi) |
| 76 | ; SSE-NEXT: movq %rax, 80(%rdi) |
| 77 | ; SSE-NEXT: movq %rax, 72(%rdi) |
| 78 | ; SSE-NEXT: movq %rax, 64(%rdi) |
| 79 | ; SSE-NEXT: movq %rax, 56(%rdi) |
| 80 | ; SSE-NEXT: movq %rax, 48(%rdi) |
| 81 | ; SSE-NEXT: movq %rax, 40(%rdi) |
| 82 | ; SSE-NEXT: movq %rax, 32(%rdi) |
| 83 | ; SSE-NEXT: movq %rax, 24(%rdi) |
| 84 | ; SSE-NEXT: movq %rax, 16(%rdi) |
| 85 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 86 | ; SSE-NEXT: movq %rax, (%rdi) |
| 87 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 88 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 89 | ; AVX-LABEL: memset_128_nonzero_bytes: |
| 90 | ; AVX: 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 | ; AVX-NEXT: vmovups %ymm0, 96(%rdi) |
| 92 | ; AVX-NEXT: vmovups %ymm0, 64(%rdi) |
| 93 | ; AVX-NEXT: vmovups %ymm0, 32(%rdi) |
| 94 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 95 | ; AVX-NEXT: vzeroupper |
| 96 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 97 | ; |
| 98 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 128, i64 -1) |
| 99 | ret void |
| 100 | } |
| 101 | |
| 102 | define void @memset_256_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 103 | ; SSE-LABEL: memset_256_nonzero_bytes: |
| 104 | ; SSE: pushq %rax |
| 105 | ; SSE-NEXT: .Ltmp0: |
| 106 | ; SSE-NEXT: .cfi_def_cfa_offset 16 |
| 107 | ; SSE-NEXT: movl $42, %esi |
| 108 | ; SSE-NEXT: movl $256, %edx # imm = 0x100 |
| 109 | ; SSE-NEXT: callq memset |
| 110 | ; SSE-NEXT: popq %rax |
| 111 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 112 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 113 | ; AVX-LABEL: memset_256_nonzero_bytes: |
| 114 | ; AVX: 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] |
| 115 | ; AVX-NEXT: vmovups %ymm0, 224(%rdi) |
| 116 | ; AVX-NEXT: vmovups %ymm0, 192(%rdi) |
| 117 | ; AVX-NEXT: vmovups %ymm0, 160(%rdi) |
| 118 | ; AVX-NEXT: vmovups %ymm0, 128(%rdi) |
| 119 | ; AVX-NEXT: vmovups %ymm0, 96(%rdi) |
| 120 | ; AVX-NEXT: vmovups %ymm0, 64(%rdi) |
| 121 | ; AVX-NEXT: vmovups %ymm0, 32(%rdi) |
| 122 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 123 | ; AVX-NEXT: vzeroupper |
| 124 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 125 | ; |
| 126 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 256, i64 -1) |
| 127 | ret void |
| 128 | } |
| 129 | |
| 130 | declare i8* @__memset_chk(i8*, i32, i64, i64) |
| 131 | |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 132 | ; Repeat with a non-constant value for the stores. |
| 133 | |
| 134 | define void @memset_16_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 135 | ; SSE-LABEL: memset_16_nonconst_bytes: |
| 136 | ; SSE: movzbl %sil, %eax |
| 137 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 138 | ; SSE-NEXT: imulq %rax, %rcx |
| 139 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 140 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 141 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 142 | ; |
| 143 | ; AVX1-LABEL: memset_16_nonconst_bytes: |
Sanjay Patel | 9f41336 | 2016-04-01 17:36:45 +0000 | [diff] [blame] | 144 | ; AVX1: vmovd %esi, %xmm0 |
| 145 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 146 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 147 | ; AVX1-NEXT: vmovdqu %xmm0, (%rdi) |
| 148 | ; AVX1-NEXT: retq |
| 149 | ; |
| 150 | ; AVX2-LABEL: memset_16_nonconst_bytes: |
Sanjay Patel | 9f41336 | 2016-04-01 17:36:45 +0000 | [diff] [blame] | 151 | ; AVX2: vmovd %esi, %xmm0 |
| 152 | ; AVX2-NEXT: vpbroadcastb %xmm0, %xmm0 |
| 153 | ; AVX2-NEXT: vmovdqu %xmm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 154 | ; AVX2-NEXT: retq |
| 155 | ; |
| 156 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 16, i32 1, i1 false) |
| 157 | ret void |
| 158 | } |
| 159 | |
| 160 | define void @memset_32_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 161 | ; SSE-LABEL: memset_32_nonconst_bytes: |
| 162 | ; SSE: movzbl %sil, %eax |
| 163 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 164 | ; SSE-NEXT: imulq %rax, %rcx |
| 165 | ; SSE-NEXT: movq %rcx, 24(%rdi) |
| 166 | ; SSE-NEXT: movq %rcx, 16(%rdi) |
| 167 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 168 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 169 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 170 | ; |
| 171 | ; AVX1-LABEL: memset_32_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 172 | ; AVX1: vmovd %esi, %xmm0 |
| 173 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 174 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 175 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 176 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 177 | ; AVX1-NEXT: vzeroupper |
| 178 | ; AVX1-NEXT: retq |
| 179 | ; |
| 180 | ; AVX2-LABEL: memset_32_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 181 | ; AVX2: vmovd %esi, %xmm0 |
| 182 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 183 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 184 | ; AVX2-NEXT: vzeroupper |
| 185 | ; AVX2-NEXT: retq |
| 186 | ; |
| 187 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 32, i32 1, i1 false) |
| 188 | ret void |
| 189 | } |
| 190 | |
| 191 | define void @memset_64_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 192 | ; SSE-LABEL: memset_64_nonconst_bytes: |
| 193 | ; SSE: movzbl %sil, %eax |
| 194 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 195 | ; SSE-NEXT: imulq %rax, %rcx |
| 196 | ; SSE-NEXT: movq %rcx, 56(%rdi) |
| 197 | ; SSE-NEXT: movq %rcx, 48(%rdi) |
| 198 | ; SSE-NEXT: movq %rcx, 40(%rdi) |
| 199 | ; SSE-NEXT: movq %rcx, 32(%rdi) |
| 200 | ; SSE-NEXT: movq %rcx, 24(%rdi) |
| 201 | ; SSE-NEXT: movq %rcx, 16(%rdi) |
| 202 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 203 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 204 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 205 | ; |
| 206 | ; AVX1-LABEL: memset_64_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 207 | ; AVX1: vmovd %esi, %xmm0 |
| 208 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 209 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 210 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 211 | ; AVX1-NEXT: vmovups %ymm0, 32(%rdi) |
| 212 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 213 | ; AVX1-NEXT: vzeroupper |
| 214 | ; AVX1-NEXT: retq |
| 215 | ; |
| 216 | ; AVX2-LABEL: memset_64_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 217 | ; AVX2: vmovd %esi, %xmm0 |
| 218 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 219 | ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi) |
| 220 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 221 | ; AVX2-NEXT: vzeroupper |
| 222 | ; AVX2-NEXT: retq |
| 223 | ; |
| 224 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 64, i32 1, i1 false) |
| 225 | ret void |
| 226 | } |
| 227 | |
| 228 | define void @memset_128_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 229 | ; SSE-LABEL: memset_128_nonconst_bytes: |
| 230 | ; SSE: movzbl %sil, %eax |
| 231 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 232 | ; SSE-NEXT: imulq %rax, %rcx |
| 233 | ; SSE-NEXT: movq %rcx, 120(%rdi) |
| 234 | ; SSE-NEXT: movq %rcx, 112(%rdi) |
| 235 | ; SSE-NEXT: movq %rcx, 104(%rdi) |
| 236 | ; SSE-NEXT: movq %rcx, 96(%rdi) |
| 237 | ; SSE-NEXT: movq %rcx, 88(%rdi) |
| 238 | ; SSE-NEXT: movq %rcx, 80(%rdi) |
| 239 | ; SSE-NEXT: movq %rcx, 72(%rdi) |
| 240 | ; SSE-NEXT: movq %rcx, 64(%rdi) |
| 241 | ; SSE-NEXT: movq %rcx, 56(%rdi) |
| 242 | ; SSE-NEXT: movq %rcx, 48(%rdi) |
| 243 | ; SSE-NEXT: movq %rcx, 40(%rdi) |
| 244 | ; SSE-NEXT: movq %rcx, 32(%rdi) |
| 245 | ; SSE-NEXT: movq %rcx, 24(%rdi) |
| 246 | ; SSE-NEXT: movq %rcx, 16(%rdi) |
| 247 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 248 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 249 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 250 | ; |
| 251 | ; AVX1-LABEL: memset_128_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 252 | ; AVX1: vmovd %esi, %xmm0 |
| 253 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 254 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 255 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 256 | ; AVX1-NEXT: vmovups %ymm0, 96(%rdi) |
| 257 | ; AVX1-NEXT: vmovups %ymm0, 64(%rdi) |
| 258 | ; AVX1-NEXT: vmovups %ymm0, 32(%rdi) |
| 259 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 260 | ; AVX1-NEXT: vzeroupper |
| 261 | ; AVX1-NEXT: retq |
| 262 | ; |
| 263 | ; AVX2-LABEL: memset_128_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 264 | ; AVX2: vmovd %esi, %xmm0 |
| 265 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 266 | ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi) |
| 267 | ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi) |
| 268 | ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi) |
| 269 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 270 | ; AVX2-NEXT: vzeroupper |
| 271 | ; AVX2-NEXT: retq |
| 272 | ; |
| 273 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 128, i32 1, i1 false) |
| 274 | ret void |
| 275 | } |
| 276 | |
| 277 | define void @memset_256_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame^] | 278 | ; SSE-LABEL: memset_256_nonconst_bytes: |
| 279 | ; SSE: movl $256, %edx # imm = 0x100 |
| 280 | ; SSE-NEXT: jmp memset # TAILCALL |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 281 | ; |
| 282 | ; AVX1-LABEL: memset_256_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 283 | ; AVX1: vmovd %esi, %xmm0 |
| 284 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 285 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 286 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 287 | ; AVX1-NEXT: vmovups %ymm0, 224(%rdi) |
| 288 | ; AVX1-NEXT: vmovups %ymm0, 192(%rdi) |
| 289 | ; AVX1-NEXT: vmovups %ymm0, 160(%rdi) |
| 290 | ; AVX1-NEXT: vmovups %ymm0, 128(%rdi) |
| 291 | ; AVX1-NEXT: vmovups %ymm0, 96(%rdi) |
| 292 | ; AVX1-NEXT: vmovups %ymm0, 64(%rdi) |
| 293 | ; AVX1-NEXT: vmovups %ymm0, 32(%rdi) |
| 294 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 295 | ; AVX1-NEXT: vzeroupper |
| 296 | ; AVX1-NEXT: retq |
| 297 | ; |
| 298 | ; AVX2-LABEL: memset_256_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 299 | ; AVX2: vmovd %esi, %xmm0 |
| 300 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 301 | ; AVX2-NEXT: vmovdqu %ymm0, 224(%rdi) |
| 302 | ; AVX2-NEXT: vmovdqu %ymm0, 192(%rdi) |
| 303 | ; AVX2-NEXT: vmovdqu %ymm0, 160(%rdi) |
| 304 | ; AVX2-NEXT: vmovdqu %ymm0, 128(%rdi) |
| 305 | ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi) |
| 306 | ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi) |
| 307 | ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi) |
| 308 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 309 | ; AVX2-NEXT: vzeroupper |
| 310 | ; AVX2-NEXT: retq |
| 311 | ; |
| 312 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 256, i32 1, i1 false) |
| 313 | ret void |
| 314 | } |
| 315 | |
| 316 | declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) #1 |
| 317 | |