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 | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 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 |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 7 | |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 8 | ; https://llvm.org/bugs/show_bug.cgi?id=27100 |
| 9 | |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 10 | define void @memset_16_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 11 | ; SSE-LABEL: memset_16_nonzero_bytes: |
| 12 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 13 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 14 | ; SSE-NEXT: movq %rax, (%rdi) |
| 15 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 16 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 17 | ; SSE2FAST-LABEL: memset_16_nonzero_bytes: |
| 18 | ; SSE2FAST: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 19 | ; SSE2FAST-NEXT: movups %xmm0, (%rdi) |
| 20 | ; SSE2FAST-NEXT: retq |
| 21 | ; |
Sanjay Patel | 9f41336 | 2016-04-01 17:36:45 +0000 | [diff] [blame] | 22 | ; AVX-LABEL: memset_16_nonzero_bytes: |
| 23 | ; AVX: vmovaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 24 | ; AVX-NEXT: vmovups %xmm0, (%rdi) |
| 25 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 26 | ; |
| 27 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 16, i64 -1) |
| 28 | ret void |
| 29 | } |
| 30 | |
| 31 | define void @memset_32_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 32 | ; SSE-LABEL: memset_32_nonzero_bytes: |
| 33 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 34 | ; SSE-NEXT: movq %rax, 24(%rdi) |
| 35 | ; SSE-NEXT: movq %rax, 16(%rdi) |
| 36 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 37 | ; SSE-NEXT: movq %rax, (%rdi) |
| 38 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 39 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 40 | ; SSE2FAST-LABEL: memset_32_nonzero_bytes: |
| 41 | ; SSE2FAST: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 42 | ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi) |
| 43 | ; SSE2FAST-NEXT: movups %xmm0, (%rdi) |
| 44 | ; SSE2FAST-NEXT: retq |
| 45 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 46 | ; AVX-LABEL: memset_32_nonzero_bytes: |
| 47 | ; 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] |
| 48 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 49 | ; AVX-NEXT: vzeroupper |
| 50 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 51 | ; |
| 52 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 32, i64 -1) |
| 53 | ret void |
| 54 | } |
| 55 | |
| 56 | define void @memset_64_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 57 | ; SSE-LABEL: memset_64_nonzero_bytes: |
| 58 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 59 | ; SSE-NEXT: movq %rax, 56(%rdi) |
| 60 | ; SSE-NEXT: movq %rax, 48(%rdi) |
| 61 | ; SSE-NEXT: movq %rax, 40(%rdi) |
| 62 | ; SSE-NEXT: movq %rax, 32(%rdi) |
| 63 | ; SSE-NEXT: movq %rax, 24(%rdi) |
| 64 | ; SSE-NEXT: movq %rax, 16(%rdi) |
| 65 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 66 | ; SSE-NEXT: movq %rax, (%rdi) |
| 67 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 68 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 69 | ; SSE2FAST-LABEL: memset_64_nonzero_bytes: |
| 70 | ; SSE2FAST: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 71 | ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi) |
| 72 | ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi) |
| 73 | ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi) |
| 74 | ; SSE2FAST-NEXT: movups %xmm0, (%rdi) |
| 75 | ; SSE2FAST-NEXT: retq |
| 76 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 77 | ; AVX-LABEL: memset_64_nonzero_bytes: |
| 78 | ; 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] |
| 79 | ; AVX-NEXT: vmovups %ymm0, 32(%rdi) |
| 80 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 81 | ; AVX-NEXT: vzeroupper |
| 82 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 83 | ; |
| 84 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 64, i64 -1) |
| 85 | ret void |
| 86 | } |
| 87 | |
| 88 | define void @memset_128_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 89 | ; SSE-LABEL: memset_128_nonzero_bytes: |
| 90 | ; SSE: movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A |
| 91 | ; SSE-NEXT: movq %rax, 120(%rdi) |
| 92 | ; SSE-NEXT: movq %rax, 112(%rdi) |
| 93 | ; SSE-NEXT: movq %rax, 104(%rdi) |
| 94 | ; SSE-NEXT: movq %rax, 96(%rdi) |
| 95 | ; SSE-NEXT: movq %rax, 88(%rdi) |
| 96 | ; SSE-NEXT: movq %rax, 80(%rdi) |
| 97 | ; SSE-NEXT: movq %rax, 72(%rdi) |
| 98 | ; SSE-NEXT: movq %rax, 64(%rdi) |
| 99 | ; SSE-NEXT: movq %rax, 56(%rdi) |
| 100 | ; SSE-NEXT: movq %rax, 48(%rdi) |
| 101 | ; SSE-NEXT: movq %rax, 40(%rdi) |
| 102 | ; SSE-NEXT: movq %rax, 32(%rdi) |
| 103 | ; SSE-NEXT: movq %rax, 24(%rdi) |
| 104 | ; SSE-NEXT: movq %rax, 16(%rdi) |
| 105 | ; SSE-NEXT: movq %rax, 8(%rdi) |
| 106 | ; SSE-NEXT: movq %rax, (%rdi) |
| 107 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 108 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 109 | ; SSE2FAST-LABEL: memset_128_nonzero_bytes: |
| 110 | ; SSE2FAST: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 111 | ; SSE2FAST-NEXT: movups %xmm0, 112(%rdi) |
| 112 | ; SSE2FAST-NEXT: movups %xmm0, 96(%rdi) |
| 113 | ; SSE2FAST-NEXT: movups %xmm0, 80(%rdi) |
| 114 | ; SSE2FAST-NEXT: movups %xmm0, 64(%rdi) |
| 115 | ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi) |
| 116 | ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi) |
| 117 | ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi) |
| 118 | ; SSE2FAST-NEXT: movups %xmm0, (%rdi) |
| 119 | ; SSE2FAST-NEXT: retq |
| 120 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 121 | ; AVX-LABEL: memset_128_nonzero_bytes: |
| 122 | ; 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] |
| 123 | ; AVX-NEXT: vmovups %ymm0, 96(%rdi) |
| 124 | ; AVX-NEXT: vmovups %ymm0, 64(%rdi) |
| 125 | ; AVX-NEXT: vmovups %ymm0, 32(%rdi) |
| 126 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 127 | ; AVX-NEXT: vzeroupper |
| 128 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 129 | ; |
| 130 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 128, i64 -1) |
| 131 | ret void |
| 132 | } |
| 133 | |
| 134 | define void @memset_256_nonzero_bytes(i8* %x) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 135 | ; SSE-LABEL: memset_256_nonzero_bytes: |
| 136 | ; SSE: pushq %rax |
| 137 | ; SSE-NEXT: .Ltmp0: |
| 138 | ; SSE-NEXT: .cfi_def_cfa_offset 16 |
| 139 | ; SSE-NEXT: movl $42, %esi |
| 140 | ; SSE-NEXT: movl $256, %edx # imm = 0x100 |
| 141 | ; SSE-NEXT: callq memset |
| 142 | ; SSE-NEXT: popq %rax |
| 143 | ; SSE-NEXT: retq |
Sanjay Patel | 92d5ea5e | 2016-03-31 17:30:06 +0000 | [diff] [blame] | 144 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 145 | ; SSE2FAST-LABEL: memset_256_nonzero_bytes: |
| 146 | ; SSE2FAST: movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42] |
| 147 | ; SSE2FAST-NEXT: movups %xmm0, 240(%rdi) |
| 148 | ; SSE2FAST-NEXT: movups %xmm0, 224(%rdi) |
| 149 | ; SSE2FAST-NEXT: movups %xmm0, 208(%rdi) |
| 150 | ; SSE2FAST-NEXT: movups %xmm0, 192(%rdi) |
| 151 | ; SSE2FAST-NEXT: movups %xmm0, 176(%rdi) |
| 152 | ; SSE2FAST-NEXT: movups %xmm0, 160(%rdi) |
| 153 | ; SSE2FAST-NEXT: movups %xmm0, 144(%rdi) |
| 154 | ; SSE2FAST-NEXT: movups %xmm0, 128(%rdi) |
| 155 | ; SSE2FAST-NEXT: movups %xmm0, 112(%rdi) |
| 156 | ; SSE2FAST-NEXT: movups %xmm0, 96(%rdi) |
| 157 | ; SSE2FAST-NEXT: movups %xmm0, 80(%rdi) |
| 158 | ; SSE2FAST-NEXT: movups %xmm0, 64(%rdi) |
| 159 | ; SSE2FAST-NEXT: movups %xmm0, 48(%rdi) |
| 160 | ; SSE2FAST-NEXT: movups %xmm0, 32(%rdi) |
| 161 | ; SSE2FAST-NEXT: movups %xmm0, 16(%rdi) |
| 162 | ; SSE2FAST-NEXT: movups %xmm0, (%rdi) |
| 163 | ; SSE2FAST-NEXT: retq |
| 164 | ; |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 165 | ; AVX-LABEL: memset_256_nonzero_bytes: |
| 166 | ; 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] |
| 167 | ; AVX-NEXT: vmovups %ymm0, 224(%rdi) |
| 168 | ; AVX-NEXT: vmovups %ymm0, 192(%rdi) |
| 169 | ; AVX-NEXT: vmovups %ymm0, 160(%rdi) |
| 170 | ; AVX-NEXT: vmovups %ymm0, 128(%rdi) |
| 171 | ; AVX-NEXT: vmovups %ymm0, 96(%rdi) |
| 172 | ; AVX-NEXT: vmovups %ymm0, 64(%rdi) |
| 173 | ; AVX-NEXT: vmovups %ymm0, 32(%rdi) |
| 174 | ; AVX-NEXT: vmovups %ymm0, (%rdi) |
| 175 | ; AVX-NEXT: vzeroupper |
| 176 | ; AVX-NEXT: retq |
Sanjay Patel | 7c61507 | 2016-03-29 17:09:27 +0000 | [diff] [blame] | 177 | ; |
| 178 | %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 256, i64 -1) |
| 179 | ret void |
| 180 | } |
| 181 | |
| 182 | declare i8* @__memset_chk(i8*, i32, i64, i64) |
| 183 | |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 184 | ; Repeat with a non-constant value for the stores. |
| 185 | |
| 186 | define void @memset_16_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 187 | ; SSE-LABEL: memset_16_nonconst_bytes: |
| 188 | ; SSE: movzbl %sil, %eax |
| 189 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 190 | ; SSE-NEXT: imulq %rax, %rcx |
| 191 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 192 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 193 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 194 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 195 | ; SSE2FAST-LABEL: memset_16_nonconst_bytes: |
| 196 | ; SSE2FAST: movd %esi, %xmm0 |
| 197 | ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] |
| 198 | ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7] |
| 199 | ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1] |
| 200 | ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi) |
| 201 | ; SSE2FAST-NEXT: retq |
| 202 | ; |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 203 | ; AVX1-LABEL: memset_16_nonconst_bytes: |
Sanjay Patel | 9f41336 | 2016-04-01 17:36:45 +0000 | [diff] [blame] | 204 | ; AVX1: vmovd %esi, %xmm0 |
| 205 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 206 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 207 | ; AVX1-NEXT: vmovdqu %xmm0, (%rdi) |
| 208 | ; AVX1-NEXT: retq |
| 209 | ; |
| 210 | ; AVX2-LABEL: memset_16_nonconst_bytes: |
Sanjay Patel | 9f41336 | 2016-04-01 17:36:45 +0000 | [diff] [blame] | 211 | ; AVX2: vmovd %esi, %xmm0 |
| 212 | ; AVX2-NEXT: vpbroadcastb %xmm0, %xmm0 |
| 213 | ; AVX2-NEXT: vmovdqu %xmm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 214 | ; AVX2-NEXT: retq |
| 215 | ; |
| 216 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 16, i32 1, i1 false) |
| 217 | ret void |
| 218 | } |
| 219 | |
| 220 | define void @memset_32_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 221 | ; SSE-LABEL: memset_32_nonconst_bytes: |
| 222 | ; SSE: movzbl %sil, %eax |
| 223 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 224 | ; SSE-NEXT: imulq %rax, %rcx |
| 225 | ; SSE-NEXT: movq %rcx, 24(%rdi) |
| 226 | ; SSE-NEXT: movq %rcx, 16(%rdi) |
| 227 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 228 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 229 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 230 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 231 | ; SSE2FAST-LABEL: memset_32_nonconst_bytes: |
| 232 | ; SSE2FAST: movd %esi, %xmm0 |
| 233 | ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] |
| 234 | ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7] |
| 235 | ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1] |
| 236 | ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi) |
| 237 | ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi) |
| 238 | ; SSE2FAST-NEXT: retq |
| 239 | ; |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 240 | ; AVX1-LABEL: memset_32_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 241 | ; AVX1: vmovd %esi, %xmm0 |
| 242 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 243 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 244 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 245 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 246 | ; AVX1-NEXT: vzeroupper |
| 247 | ; AVX1-NEXT: retq |
| 248 | ; |
| 249 | ; AVX2-LABEL: memset_32_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 250 | ; AVX2: vmovd %esi, %xmm0 |
| 251 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 252 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 253 | ; AVX2-NEXT: vzeroupper |
| 254 | ; AVX2-NEXT: retq |
| 255 | ; |
| 256 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 32, i32 1, i1 false) |
| 257 | ret void |
| 258 | } |
| 259 | |
| 260 | define void @memset_64_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 261 | ; SSE-LABEL: memset_64_nonconst_bytes: |
| 262 | ; SSE: movzbl %sil, %eax |
| 263 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 264 | ; SSE-NEXT: imulq %rax, %rcx |
| 265 | ; SSE-NEXT: movq %rcx, 56(%rdi) |
| 266 | ; SSE-NEXT: movq %rcx, 48(%rdi) |
| 267 | ; SSE-NEXT: movq %rcx, 40(%rdi) |
| 268 | ; SSE-NEXT: movq %rcx, 32(%rdi) |
| 269 | ; SSE-NEXT: movq %rcx, 24(%rdi) |
| 270 | ; SSE-NEXT: movq %rcx, 16(%rdi) |
| 271 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 272 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 273 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 274 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 275 | ; SSE2FAST-LABEL: memset_64_nonconst_bytes: |
| 276 | ; SSE2FAST: movd %esi, %xmm0 |
| 277 | ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] |
| 278 | ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7] |
| 279 | ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1] |
| 280 | ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi) |
| 281 | ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi) |
| 282 | ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi) |
| 283 | ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi) |
| 284 | ; SSE2FAST-NEXT: retq |
| 285 | ; |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 286 | ; AVX1-LABEL: memset_64_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 287 | ; AVX1: vmovd %esi, %xmm0 |
| 288 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 289 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 290 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 291 | ; AVX1-NEXT: vmovups %ymm0, 32(%rdi) |
| 292 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 293 | ; AVX1-NEXT: vzeroupper |
| 294 | ; AVX1-NEXT: retq |
| 295 | ; |
| 296 | ; AVX2-LABEL: memset_64_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 297 | ; AVX2: vmovd %esi, %xmm0 |
| 298 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 299 | ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi) |
| 300 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 301 | ; AVX2-NEXT: vzeroupper |
| 302 | ; AVX2-NEXT: retq |
| 303 | ; |
| 304 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 64, i32 1, i1 false) |
| 305 | ret void |
| 306 | } |
| 307 | |
| 308 | define void @memset_128_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 309 | ; SSE-LABEL: memset_128_nonconst_bytes: |
| 310 | ; SSE: movzbl %sil, %eax |
| 311 | ; SSE-NEXT: movabsq $72340172838076673, %rcx # imm = 0x101010101010101 |
| 312 | ; SSE-NEXT: imulq %rax, %rcx |
| 313 | ; SSE-NEXT: movq %rcx, 120(%rdi) |
| 314 | ; SSE-NEXT: movq %rcx, 112(%rdi) |
| 315 | ; SSE-NEXT: movq %rcx, 104(%rdi) |
| 316 | ; SSE-NEXT: movq %rcx, 96(%rdi) |
| 317 | ; SSE-NEXT: movq %rcx, 88(%rdi) |
| 318 | ; SSE-NEXT: movq %rcx, 80(%rdi) |
| 319 | ; SSE-NEXT: movq %rcx, 72(%rdi) |
| 320 | ; SSE-NEXT: movq %rcx, 64(%rdi) |
| 321 | ; SSE-NEXT: movq %rcx, 56(%rdi) |
| 322 | ; SSE-NEXT: movq %rcx, 48(%rdi) |
| 323 | ; SSE-NEXT: movq %rcx, 40(%rdi) |
| 324 | ; SSE-NEXT: movq %rcx, 32(%rdi) |
| 325 | ; SSE-NEXT: movq %rcx, 24(%rdi) |
| 326 | ; SSE-NEXT: movq %rcx, 16(%rdi) |
| 327 | ; SSE-NEXT: movq %rcx, 8(%rdi) |
| 328 | ; SSE-NEXT: movq %rcx, (%rdi) |
| 329 | ; SSE-NEXT: retq |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 330 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 331 | ; SSE2FAST-LABEL: memset_128_nonconst_bytes: |
| 332 | ; SSE2FAST: movd %esi, %xmm0 |
| 333 | ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] |
| 334 | ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7] |
| 335 | ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1] |
| 336 | ; SSE2FAST-NEXT: movdqu %xmm0, 112(%rdi) |
| 337 | ; SSE2FAST-NEXT: movdqu %xmm0, 96(%rdi) |
| 338 | ; SSE2FAST-NEXT: movdqu %xmm0, 80(%rdi) |
| 339 | ; SSE2FAST-NEXT: movdqu %xmm0, 64(%rdi) |
| 340 | ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi) |
| 341 | ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi) |
| 342 | ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi) |
| 343 | ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi) |
| 344 | ; SSE2FAST-NEXT: retq |
| 345 | ; |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 346 | ; AVX1-LABEL: memset_128_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 347 | ; AVX1: vmovd %esi, %xmm0 |
| 348 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 349 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 350 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 351 | ; AVX1-NEXT: vmovups %ymm0, 96(%rdi) |
| 352 | ; AVX1-NEXT: vmovups %ymm0, 64(%rdi) |
| 353 | ; AVX1-NEXT: vmovups %ymm0, 32(%rdi) |
| 354 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 355 | ; AVX1-NEXT: vzeroupper |
| 356 | ; AVX1-NEXT: retq |
| 357 | ; |
| 358 | ; AVX2-LABEL: memset_128_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 359 | ; AVX2: vmovd %esi, %xmm0 |
| 360 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 361 | ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi) |
| 362 | ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi) |
| 363 | ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi) |
| 364 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 365 | ; AVX2-NEXT: vzeroupper |
| 366 | ; AVX2-NEXT: retq |
| 367 | ; |
| 368 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 128, i32 1, i1 false) |
| 369 | ret void |
| 370 | } |
| 371 | |
| 372 | define void @memset_256_nonconst_bytes(i8* %x, i8 %c) { |
Sanjay Patel | d3e3d48 | 2016-04-01 18:11:30 +0000 | [diff] [blame] | 373 | ; SSE-LABEL: memset_256_nonconst_bytes: |
| 374 | ; SSE: movl $256, %edx # imm = 0x100 |
| 375 | ; SSE-NEXT: jmp memset # TAILCALL |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 376 | ; |
Sanjay Patel | 9b5b5c8 | 2016-04-01 18:29:25 +0000 | [diff] [blame^] | 377 | ; SSE2FAST-LABEL: memset_256_nonconst_bytes: |
| 378 | ; SSE2FAST: movd %esi, %xmm0 |
| 379 | ; SSE2FAST-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] |
| 380 | ; SSE2FAST-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7] |
| 381 | ; SSE2FAST-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1] |
| 382 | ; SSE2FAST-NEXT: movdqu %xmm0, 240(%rdi) |
| 383 | ; SSE2FAST-NEXT: movdqu %xmm0, 224(%rdi) |
| 384 | ; SSE2FAST-NEXT: movdqu %xmm0, 208(%rdi) |
| 385 | ; SSE2FAST-NEXT: movdqu %xmm0, 192(%rdi) |
| 386 | ; SSE2FAST-NEXT: movdqu %xmm0, 176(%rdi) |
| 387 | ; SSE2FAST-NEXT: movdqu %xmm0, 160(%rdi) |
| 388 | ; SSE2FAST-NEXT: movdqu %xmm0, 144(%rdi) |
| 389 | ; SSE2FAST-NEXT: movdqu %xmm0, 128(%rdi) |
| 390 | ; SSE2FAST-NEXT: movdqu %xmm0, 112(%rdi) |
| 391 | ; SSE2FAST-NEXT: movdqu %xmm0, 96(%rdi) |
| 392 | ; SSE2FAST-NEXT: movdqu %xmm0, 80(%rdi) |
| 393 | ; SSE2FAST-NEXT: movdqu %xmm0, 64(%rdi) |
| 394 | ; SSE2FAST-NEXT: movdqu %xmm0, 48(%rdi) |
| 395 | ; SSE2FAST-NEXT: movdqu %xmm0, 32(%rdi) |
| 396 | ; SSE2FAST-NEXT: movdqu %xmm0, 16(%rdi) |
| 397 | ; SSE2FAST-NEXT: movdqu %xmm0, (%rdi) |
| 398 | ; SSE2FAST-NEXT: retq |
| 399 | ; |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 400 | ; AVX1-LABEL: memset_256_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 401 | ; AVX1: vmovd %esi, %xmm0 |
| 402 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 403 | ; AVX1-NEXT: vpshufb %xmm1, %xmm0, %xmm0 |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 404 | ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| 405 | ; AVX1-NEXT: vmovups %ymm0, 224(%rdi) |
| 406 | ; AVX1-NEXT: vmovups %ymm0, 192(%rdi) |
| 407 | ; AVX1-NEXT: vmovups %ymm0, 160(%rdi) |
| 408 | ; AVX1-NEXT: vmovups %ymm0, 128(%rdi) |
| 409 | ; AVX1-NEXT: vmovups %ymm0, 96(%rdi) |
| 410 | ; AVX1-NEXT: vmovups %ymm0, 64(%rdi) |
| 411 | ; AVX1-NEXT: vmovups %ymm0, 32(%rdi) |
| 412 | ; AVX1-NEXT: vmovups %ymm0, (%rdi) |
| 413 | ; AVX1-NEXT: vzeroupper |
| 414 | ; AVX1-NEXT: retq |
| 415 | ; |
| 416 | ; AVX2-LABEL: memset_256_nonconst_bytes: |
Sanjay Patel | a05e0ff | 2016-04-01 16:27:14 +0000 | [diff] [blame] | 417 | ; AVX2: vmovd %esi, %xmm0 |
| 418 | ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0 |
| 419 | ; AVX2-NEXT: vmovdqu %ymm0, 224(%rdi) |
| 420 | ; AVX2-NEXT: vmovdqu %ymm0, 192(%rdi) |
| 421 | ; AVX2-NEXT: vmovdqu %ymm0, 160(%rdi) |
| 422 | ; AVX2-NEXT: vmovdqu %ymm0, 128(%rdi) |
| 423 | ; AVX2-NEXT: vmovdqu %ymm0, 96(%rdi) |
| 424 | ; AVX2-NEXT: vmovdqu %ymm0, 64(%rdi) |
| 425 | ; AVX2-NEXT: vmovdqu %ymm0, 32(%rdi) |
| 426 | ; AVX2-NEXT: vmovdqu %ymm0, (%rdi) |
Sanjay Patel | 61e1324 | 2016-03-31 20:40:32 +0000 | [diff] [blame] | 427 | ; AVX2-NEXT: vzeroupper |
| 428 | ; AVX2-NEXT: retq |
| 429 | ; |
| 430 | tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 256, i32 1, i1 false) |
| 431 | ret void |
| 432 | } |
| 433 | |
| 434 | declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) #1 |
| 435 | |