| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1 | ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py |
| 2 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1 |
| 3 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2 |
| 4 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512F |
| 5 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512VL |
| 6 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512BW |
| 7 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512BWVL |
| 8 | |
| 9 | define void @shuffle_v32i8_to_v16i8_1(<32 x i8>* %L, <16 x i8>* %S) nounwind { |
| 10 | ; AVX1-LABEL: shuffle_v32i8_to_v16i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 11 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 12 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 13 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 14 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u> |
| 15 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 16 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 17 | ; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 18 | ; AVX1-NEXT: vmovdqa %xmm0, (%rsi) |
| 19 | ; AVX1-NEXT: vzeroupper |
| 20 | ; AVX1-NEXT: retq |
| 21 | ; |
| 22 | ; AVX2-LABEL: shuffle_v32i8_to_v16i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 23 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 24 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| 25 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 26 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u> |
| 27 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 28 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 29 | ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 30 | ; AVX2-NEXT: vmovdqa %xmm0, (%rsi) |
| 31 | ; AVX2-NEXT: vzeroupper |
| 32 | ; AVX2-NEXT: retq |
| 33 | ; |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 34 | ; AVX512-LABEL: shuffle_v32i8_to_v16i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 35 | ; AVX512: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 36 | ; AVX512-NEXT: vmovdqa (%rdi), %ymm0 |
| 37 | ; AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 38 | ; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = <1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u> |
| 39 | ; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 40 | ; AVX512-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 41 | ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 42 | ; AVX512-NEXT: vmovdqa %xmm0, (%rsi) |
| 43 | ; AVX512-NEXT: vzeroupper |
| 44 | ; AVX512-NEXT: retq |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 45 | %vec = load <32 x i8>, <32 x i8>* %L |
| 46 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <16 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31> |
| 47 | store <16 x i8> %strided.vec, <16 x i8>* %S |
| 48 | ret void |
| 49 | } |
| 50 | |
| 51 | define void @shuffle_v16i16_to_v8i16_1(<16 x i16>* %L, <8 x i16>* %S) nounwind { |
| 52 | ; AVX1-LABEL: shuffle_v16i16_to_v8i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 53 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 54 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 55 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 56 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15] |
| 57 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 58 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 59 | ; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 60 | ; AVX1-NEXT: vmovdqa %xmm0, (%rsi) |
| 61 | ; AVX1-NEXT: vzeroupper |
| 62 | ; AVX1-NEXT: retq |
| 63 | ; |
| 64 | ; AVX2-LABEL: shuffle_v16i16_to_v8i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 65 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 66 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 67 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 68 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15] |
| 69 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 70 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 71 | ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 72 | ; AVX2-NEXT: vmovdqa %xmm0, (%rsi) |
| 73 | ; AVX2-NEXT: vzeroupper |
| 74 | ; AVX2-NEXT: retq |
| 75 | ; |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 76 | ; AVX512-LABEL: shuffle_v16i16_to_v8i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 77 | ; AVX512: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 78 | ; AVX512-NEXT: vmovdqa (%rdi), %ymm0 |
| 79 | ; AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 80 | ; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15] |
| 81 | ; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 82 | ; AVX512-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 83 | ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 84 | ; AVX512-NEXT: vmovdqa %xmm0, (%rsi) |
| 85 | ; AVX512-NEXT: vzeroupper |
| 86 | ; AVX512-NEXT: retq |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 87 | %vec = load <16 x i16>, <16 x i16>* %L |
| 88 | %strided.vec = shufflevector <16 x i16> %vec, <16 x i16> undef, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15> |
| 89 | store <8 x i16> %strided.vec, <8 x i16>* %S |
| 90 | ret void |
| 91 | } |
| 92 | |
| 93 | define void @shuffle_v8i32_to_v4i32_1(<8 x i32>* %L, <4 x i32>* %S) nounwind { |
| 94 | ; AVX-LABEL: shuffle_v8i32_to_v4i32_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 95 | ; AVX: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 96 | ; AVX-NEXT: vmovaps (%rdi), %ymm0 |
| 97 | ; AVX-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 98 | ; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] |
| 99 | ; AVX-NEXT: vmovaps %xmm0, (%rsi) |
| 100 | ; AVX-NEXT: vzeroupper |
| 101 | ; AVX-NEXT: retq |
| 102 | ; |
| Craig Topper | 410d252 | 2017-07-31 22:07:29 +0000 | [diff] [blame] | 103 | ; AVX512-LABEL: shuffle_v8i32_to_v4i32_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 104 | ; AVX512: # %bb.0: |
| Craig Topper | 410d252 | 2017-07-31 22:07:29 +0000 | [diff] [blame] | 105 | ; AVX512-NEXT: vmovaps (%rdi), %ymm0 |
| 106 | ; AVX512-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 107 | ; AVX512-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] |
| 108 | ; AVX512-NEXT: vmovaps %xmm0, (%rsi) |
| 109 | ; AVX512-NEXT: vzeroupper |
| 110 | ; AVX512-NEXT: retq |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 111 | %vec = load <8 x i32>, <8 x i32>* %L |
| 112 | %strided.vec = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 1, i32 3, i32 5, i32 7> |
| 113 | store <4 x i32> %strided.vec, <4 x i32>* %S |
| 114 | ret void |
| 115 | } |
| 116 | |
| 117 | define void @shuffle_v32i8_to_v8i8_1(<32 x i8>* %L, <8 x i8>* %S) nounwind { |
| 118 | ; AVX1-LABEL: shuffle_v32i8_to_v8i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 119 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 120 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 121 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 122 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> |
| 123 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 124 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 125 | ; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 126 | ; AVX1-NEXT: vmovq %xmm0, (%rsi) |
| 127 | ; AVX1-NEXT: vzeroupper |
| 128 | ; AVX1-NEXT: retq |
| 129 | ; |
| 130 | ; AVX2-LABEL: shuffle_v32i8_to_v8i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 131 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 132 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 133 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 134 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> |
| 135 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 136 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 137 | ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 138 | ; AVX2-NEXT: vmovq %xmm0, (%rsi) |
| 139 | ; AVX2-NEXT: vzeroupper |
| 140 | ; AVX2-NEXT: retq |
| 141 | ; |
| 142 | ; AVX512F-LABEL: shuffle_v32i8_to_v8i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 143 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 144 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 145 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 146 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> |
| 147 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 148 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 149 | ; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 150 | ; AVX512F-NEXT: vmovq %xmm0, (%rsi) |
| 151 | ; AVX512F-NEXT: vzeroupper |
| 152 | ; AVX512F-NEXT: retq |
| 153 | ; |
| 154 | ; AVX512VL-LABEL: shuffle_v32i8_to_v8i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 155 | ; AVX512VL: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 156 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 157 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 158 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> |
| 159 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 160 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 161 | ; AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 162 | ; AVX512VL-NEXT: vmovq %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 163 | ; AVX512VL-NEXT: vzeroupper |
| 164 | ; AVX512VL-NEXT: retq |
| 165 | ; |
| 166 | ; AVX512BW-LABEL: shuffle_v32i8_to_v8i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 167 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 168 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 169 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 170 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> |
| 171 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 172 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 173 | ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 174 | ; AVX512BW-NEXT: vmovq %xmm0, (%rsi) |
| 175 | ; AVX512BW-NEXT: vzeroupper |
| 176 | ; AVX512BW-NEXT: retq |
| 177 | ; |
| 178 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v8i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 179 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 180 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 181 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 182 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [1,1,5,5,9,9,13,13,13,13,5,5,12,12,13,13] |
| 183 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 184 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 185 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 186 | ; AVX512BWVL-NEXT: vpmovwb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 187 | ; AVX512BWVL-NEXT: vzeroupper |
| 188 | ; AVX512BWVL-NEXT: retq |
| 189 | %vec = load <32 x i8>, <32 x i8>* %L |
| 190 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <8 x i32> <i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 29> |
| 191 | store <8 x i8> %strided.vec, <8 x i8>* %S |
| 192 | ret void |
| 193 | } |
| 194 | |
| 195 | define void @shuffle_v32i8_to_v8i8_2(<32 x i8>* %L, <8 x i8>* %S) nounwind { |
| 196 | ; AVX1-LABEL: shuffle_v32i8_to_v8i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 197 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 198 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 199 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 200 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> |
| 201 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 202 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 203 | ; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 204 | ; AVX1-NEXT: vmovq %xmm0, (%rsi) |
| 205 | ; AVX1-NEXT: vzeroupper |
| 206 | ; AVX1-NEXT: retq |
| 207 | ; |
| 208 | ; AVX2-LABEL: shuffle_v32i8_to_v8i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 209 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 210 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 211 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 212 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> |
| 213 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 214 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 215 | ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 216 | ; AVX2-NEXT: vmovq %xmm0, (%rsi) |
| 217 | ; AVX2-NEXT: vzeroupper |
| 218 | ; AVX2-NEXT: retq |
| 219 | ; |
| 220 | ; AVX512F-LABEL: shuffle_v32i8_to_v8i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 221 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 222 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 223 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 224 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> |
| 225 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 226 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 227 | ; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 228 | ; AVX512F-NEXT: vmovq %xmm0, (%rsi) |
| 229 | ; AVX512F-NEXT: vzeroupper |
| 230 | ; AVX512F-NEXT: retq |
| 231 | ; |
| 232 | ; AVX512VL-LABEL: shuffle_v32i8_to_v8i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 233 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 234 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 235 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 236 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> |
| 237 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 238 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 239 | ; AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 240 | ; AVX512VL-NEXT: vmovq %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 241 | ; AVX512VL-NEXT: vzeroupper |
| 242 | ; AVX512VL-NEXT: retq |
| 243 | ; |
| 244 | ; AVX512BW-LABEL: shuffle_v32i8_to_v8i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 245 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 246 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 247 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 248 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> |
| 249 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 250 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 251 | ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 252 | ; AVX512BW-NEXT: vmovq %xmm0, (%rsi) |
| 253 | ; AVX512BW-NEXT: vzeroupper |
| 254 | ; AVX512BW-NEXT: retq |
| 255 | ; |
| 256 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v8i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 257 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 258 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 259 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 260 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [2,3,6,7,10,11,14,15,14,15,10,11,12,13,14,15] |
| 261 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 262 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 263 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 264 | ; AVX512BWVL-NEXT: vpmovwb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 265 | ; AVX512BWVL-NEXT: vzeroupper |
| 266 | ; AVX512BWVL-NEXT: retq |
| 267 | %vec = load <32 x i8>, <32 x i8>* %L |
| 268 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <8 x i32> <i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 30> |
| 269 | store <8 x i8> %strided.vec, <8 x i8>* %S |
| 270 | ret void |
| 271 | } |
| 272 | |
| 273 | define void @shuffle_v32i8_to_v8i8_3(<32 x i8>* %L, <8 x i8>* %S) nounwind { |
| 274 | ; AVX1-LABEL: shuffle_v32i8_to_v8i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 275 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 276 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 277 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 278 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> |
| 279 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 280 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 281 | ; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 282 | ; AVX1-NEXT: vmovq %xmm0, (%rsi) |
| 283 | ; AVX1-NEXT: vzeroupper |
| 284 | ; AVX1-NEXT: retq |
| 285 | ; |
| 286 | ; AVX2-LABEL: shuffle_v32i8_to_v8i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 287 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 288 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 289 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 290 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> |
| 291 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 292 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 293 | ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 294 | ; AVX2-NEXT: vmovq %xmm0, (%rsi) |
| 295 | ; AVX2-NEXT: vzeroupper |
| 296 | ; AVX2-NEXT: retq |
| 297 | ; |
| 298 | ; AVX512F-LABEL: shuffle_v32i8_to_v8i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 299 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 300 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 301 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 302 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> |
| 303 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 304 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 305 | ; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 306 | ; AVX512F-NEXT: vmovq %xmm0, (%rsi) |
| 307 | ; AVX512F-NEXT: vzeroupper |
| 308 | ; AVX512F-NEXT: retq |
| 309 | ; |
| 310 | ; AVX512VL-LABEL: shuffle_v32i8_to_v8i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 311 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 312 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 313 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 314 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> |
| 315 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 316 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 317 | ; AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 318 | ; AVX512VL-NEXT: vmovq %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 319 | ; AVX512VL-NEXT: vzeroupper |
| 320 | ; AVX512VL-NEXT: retq |
| 321 | ; |
| 322 | ; AVX512BW-LABEL: shuffle_v32i8_to_v8i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 323 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 324 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 325 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 326 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> |
| 327 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 328 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 329 | ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 330 | ; AVX512BW-NEXT: vmovq %xmm0, (%rsi) |
| 331 | ; AVX512BW-NEXT: vzeroupper |
| 332 | ; AVX512BW-NEXT: retq |
| 333 | ; |
| 334 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v8i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 335 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 336 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 337 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 338 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [3,3,7,7,11,11,15,15,7,7,15,15,6,6,7,7] |
| 339 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 340 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 341 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 342 | ; AVX512BWVL-NEXT: vpmovwb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 343 | ; AVX512BWVL-NEXT: vzeroupper |
| 344 | ; AVX512BWVL-NEXT: retq |
| 345 | %vec = load <32 x i8>, <32 x i8>* %L |
| 346 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <8 x i32> <i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27, i32 31> |
| 347 | store <8 x i8> %strided.vec, <8 x i8>* %S |
| 348 | ret void |
| 349 | } |
| 350 | |
| 351 | define void @shuffle_v16i16_to_v4i16_1(<16 x i16>* %L, <4 x i16>* %S) nounwind { |
| 352 | ; AVX1-LABEL: shuffle_v16i16_to_v4i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 353 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 354 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 355 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 356 | ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 357 | ; AVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7] |
| 358 | ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 359 | ; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7] |
| 360 | ; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 361 | ; AVX1-NEXT: vmovq %xmm0, (%rsi) |
| 362 | ; AVX1-NEXT: vzeroupper |
| 363 | ; AVX1-NEXT: retq |
| 364 | ; |
| 365 | ; AVX2-LABEL: shuffle_v16i16_to_v4i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 366 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 367 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 368 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 369 | ; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 370 | ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7] |
| 371 | ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 372 | ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7] |
| 373 | ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 374 | ; AVX2-NEXT: vmovq %xmm0, (%rsi) |
| 375 | ; AVX2-NEXT: vzeroupper |
| 376 | ; AVX2-NEXT: retq |
| 377 | ; |
| 378 | ; AVX512F-LABEL: shuffle_v16i16_to_v4i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 379 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 380 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 381 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 382 | ; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 383 | ; AVX512F-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7] |
| 384 | ; AVX512F-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 385 | ; AVX512F-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7] |
| 386 | ; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 387 | ; AVX512F-NEXT: vmovq %xmm0, (%rsi) |
| 388 | ; AVX512F-NEXT: vzeroupper |
| 389 | ; AVX512F-NEXT: retq |
| 390 | ; |
| 391 | ; AVX512VL-LABEL: shuffle_v16i16_to_v4i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 392 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 393 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 394 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 395 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 396 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,1,3,3,4,5,6,7] |
| 397 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 398 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,1,3,3,4,5,6,7] |
| 399 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 400 | ; AVX512VL-NEXT: vpmovdw %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 401 | ; AVX512VL-NEXT: vzeroupper |
| 402 | ; AVX512VL-NEXT: retq |
| 403 | ; |
| 404 | ; AVX512BW-LABEL: shuffle_v16i16_to_v4i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 405 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 406 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 407 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 408 | ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 409 | ; AVX512BW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7] |
| 410 | ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 411 | ; AVX512BW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7] |
| 412 | ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 413 | ; AVX512BW-NEXT: vmovq %xmm0, (%rsi) |
| 414 | ; AVX512BW-NEXT: vzeroupper |
| 415 | ; AVX512BW-NEXT: retq |
| 416 | ; |
| 417 | ; AVX512BWVL-LABEL: shuffle_v16i16_to_v4i16_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 418 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 419 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 420 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 421 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 422 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,1,3,3,4,5,6,7] |
| 423 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 424 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,1,3,3,4,5,6,7] |
| 425 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 426 | ; AVX512BWVL-NEXT: vpmovdw %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 427 | ; AVX512BWVL-NEXT: vzeroupper |
| 428 | ; AVX512BWVL-NEXT: retq |
| 429 | %vec = load <16 x i16>, <16 x i16>* %L |
| 430 | %strided.vec = shufflevector <16 x i16> %vec, <16 x i16> undef, <4 x i32> <i32 1, i32 5, i32 9, i32 13> |
| 431 | store <4 x i16> %strided.vec, <4 x i16>* %S |
| 432 | ret void |
| 433 | } |
| 434 | |
| 435 | define void @shuffle_v16i16_to_v4i16_2(<16 x i16>* %L, <4 x i16>* %S) nounwind { |
| 436 | ; AVX1-LABEL: shuffle_v16i16_to_v4i16_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 437 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 438 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 439 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 440 | ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 441 | ; AVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7] |
| 442 | ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 443 | ; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7] |
| 444 | ; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 445 | ; AVX1-NEXT: vmovq %xmm0, (%rsi) |
| 446 | ; AVX1-NEXT: vzeroupper |
| 447 | ; AVX1-NEXT: retq |
| 448 | ; |
| 449 | ; AVX2-LABEL: shuffle_v16i16_to_v4i16_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 450 | ; AVX2: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 451 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| 452 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 453 | ; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 454 | ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7] |
| 455 | ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 456 | ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7] |
| 457 | ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 458 | ; AVX2-NEXT: vmovq %xmm0, (%rsi) |
| 459 | ; AVX2-NEXT: vzeroupper |
| 460 | ; AVX2-NEXT: retq |
| 461 | ; |
| 462 | ; AVX512F-LABEL: shuffle_v16i16_to_v4i16_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 463 | ; AVX512F: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 464 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| 465 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 466 | ; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 467 | ; AVX512F-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7] |
| 468 | ; AVX512F-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 469 | ; AVX512F-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7] |
| 470 | ; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 471 | ; AVX512F-NEXT: vmovq %xmm0, (%rsi) |
| 472 | ; AVX512F-NEXT: vzeroupper |
| 473 | ; AVX512F-NEXT: retq |
| 474 | ; |
| 475 | ; AVX512VL-LABEL: shuffle_v16i16_to_v4i16_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 476 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 477 | ; AVX512VL-NEXT: vmovaps (%rdi), %ymm0 |
| 478 | ; AVX512VL-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 479 | ; AVX512VL-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] |
| 480 | ; AVX512VL-NEXT: vpmovdw %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 481 | ; AVX512VL-NEXT: vzeroupper |
| 482 | ; AVX512VL-NEXT: retq |
| 483 | ; |
| 484 | ; AVX512BW-LABEL: shuffle_v16i16_to_v4i16_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 485 | ; AVX512BW: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 486 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| 487 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 488 | ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 489 | ; AVX512BW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7] |
| 490 | ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 491 | ; AVX512BW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7] |
| 492 | ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 493 | ; AVX512BW-NEXT: vmovq %xmm0, (%rsi) |
| 494 | ; AVX512BW-NEXT: vzeroupper |
| 495 | ; AVX512BW-NEXT: retq |
| 496 | ; |
| 497 | ; AVX512BWVL-LABEL: shuffle_v16i16_to_v4i16_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 498 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 499 | ; AVX512BWVL-NEXT: vmovaps (%rdi), %ymm0 |
| 500 | ; AVX512BWVL-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 501 | ; AVX512BWVL-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] |
| 502 | ; AVX512BWVL-NEXT: vpmovdw %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 503 | ; AVX512BWVL-NEXT: vzeroupper |
| 504 | ; AVX512BWVL-NEXT: retq |
| 505 | %vec = load <16 x i16>, <16 x i16>* %L |
| 506 | %strided.vec = shufflevector <16 x i16> %vec, <16 x i16> undef, <4 x i32> <i32 2, i32 6, i32 10, i32 14> |
| 507 | store <4 x i16> %strided.vec, <4 x i16>* %S |
| 508 | ret void |
| 509 | } |
| 510 | |
| 511 | define void @shuffle_v16i16_to_v4i16_3(<16 x i16>* %L, <4 x i16>* %S) nounwind { |
| 512 | ; AVX1-LABEL: shuffle_v16i16_to_v4i16_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 513 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 514 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 515 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 516 | ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 517 | ; AVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7] |
| 518 | ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 519 | ; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7] |
| 520 | ; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 521 | ; AVX1-NEXT: vmovq %xmm0, (%rsi) |
| 522 | ; AVX1-NEXT: vzeroupper |
| 523 | ; AVX1-NEXT: retq |
| 524 | ; |
| 525 | ; AVX2-LABEL: shuffle_v16i16_to_v4i16_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 526 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 527 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 528 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 529 | ; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 530 | ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7] |
| 531 | ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 532 | ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7] |
| 533 | ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 534 | ; AVX2-NEXT: vmovq %xmm0, (%rsi) |
| 535 | ; AVX2-NEXT: vzeroupper |
| 536 | ; AVX2-NEXT: retq |
| 537 | ; |
| 538 | ; AVX512F-LABEL: shuffle_v16i16_to_v4i16_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 539 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 540 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 541 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 542 | ; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 543 | ; AVX512F-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7] |
| 544 | ; AVX512F-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 545 | ; AVX512F-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7] |
| 546 | ; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 547 | ; AVX512F-NEXT: vmovq %xmm0, (%rsi) |
| 548 | ; AVX512F-NEXT: vzeroupper |
| 549 | ; AVX512F-NEXT: retq |
| 550 | ; |
| 551 | ; AVX512VL-LABEL: shuffle_v16i16_to_v4i16_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 552 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 553 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 554 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 555 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 556 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,1,3,4,5,6,7] |
| 557 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 558 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,1,3,4,5,6,7] |
| 559 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 560 | ; AVX512VL-NEXT: vpmovdw %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 561 | ; AVX512VL-NEXT: vzeroupper |
| 562 | ; AVX512VL-NEXT: retq |
| 563 | ; |
| 564 | ; AVX512BW-LABEL: shuffle_v16i16_to_v4i16_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 565 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 566 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 567 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 568 | ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 569 | ; AVX512BW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7] |
| 570 | ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 571 | ; AVX512BW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7] |
| 572 | ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 573 | ; AVX512BW-NEXT: vmovq %xmm0, (%rsi) |
| 574 | ; AVX512BW-NEXT: vzeroupper |
| 575 | ; AVX512BW-NEXT: retq |
| 576 | ; |
| 577 | ; AVX512BWVL-LABEL: shuffle_v16i16_to_v4i16_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 578 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 579 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 580 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 581 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 582 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,1,3,4,5,6,7] |
| 583 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 584 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,1,3,4,5,6,7] |
| 585 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 586 | ; AVX512BWVL-NEXT: vpmovdw %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 587 | ; AVX512BWVL-NEXT: vzeroupper |
| 588 | ; AVX512BWVL-NEXT: retq |
| 589 | %vec = load <16 x i16>, <16 x i16>* %L |
| 590 | %strided.vec = shufflevector <16 x i16> %vec, <16 x i16> undef, <4 x i32> <i32 3, i32 7, i32 11, i32 15> |
| 591 | store <4 x i16> %strided.vec, <4 x i16>* %S |
| 592 | ret void |
| 593 | } |
| 594 | |
| 595 | define void @shuffle_v32i8_to_v4i8_1(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 596 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 597 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 598 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 599 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 600 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 601 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 602 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 603 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 604 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 605 | ; AVX1-NEXT: vzeroupper |
| 606 | ; AVX1-NEXT: retq |
| 607 | ; |
| 608 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 609 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 610 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 611 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 612 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 613 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 614 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 615 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 616 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 617 | ; AVX2-NEXT: vzeroupper |
| 618 | ; AVX2-NEXT: retq |
| 619 | ; |
| 620 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 621 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 622 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 623 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 624 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 625 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 626 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 627 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 628 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 629 | ; AVX512F-NEXT: vzeroupper |
| 630 | ; AVX512F-NEXT: retq |
| 631 | ; |
| 632 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 633 | ; AVX512VL: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 634 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 635 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | d873b6f | 2017-12-19 16:54:07 +0000 | [diff] [blame] | 636 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = [1,1,255,255,9,9,255,255,9,9,255,255,11,11,255,255] |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 637 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 638 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 639 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 640 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 641 | ; AVX512VL-NEXT: vzeroupper |
| 642 | ; AVX512VL-NEXT: retq |
| 643 | ; |
| 644 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 645 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 646 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 647 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 648 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 649 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 650 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 651 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 652 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 653 | ; AVX512BW-NEXT: vzeroupper |
| 654 | ; AVX512BW-NEXT: retq |
| 655 | ; |
| 656 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_1: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 657 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 658 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 659 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | d873b6f | 2017-12-19 16:54:07 +0000 | [diff] [blame] | 660 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [1,1,255,255,9,9,255,255,9,9,255,255,11,11,255,255] |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 661 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 662 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 663 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 664 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 665 | ; AVX512BWVL-NEXT: vzeroupper |
| 666 | ; AVX512BWVL-NEXT: retq |
| 667 | %vec = load <32 x i8>, <32 x i8>* %L |
| 668 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 1, i32 9, i32 17, i32 25> |
| 669 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 670 | ret void |
| 671 | } |
| 672 | |
| 673 | define void @shuffle_v32i8_to_v4i8_2(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 674 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 675 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 676 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 677 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 678 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 679 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 680 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 681 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 682 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 683 | ; AVX1-NEXT: vzeroupper |
| 684 | ; AVX1-NEXT: retq |
| 685 | ; |
| 686 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 687 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 688 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 689 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 690 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 691 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 692 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 693 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 694 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 695 | ; AVX2-NEXT: vzeroupper |
| 696 | ; AVX2-NEXT: retq |
| 697 | ; |
| 698 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 699 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 700 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 701 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 702 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 703 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 704 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 705 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 706 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 707 | ; AVX512F-NEXT: vzeroupper |
| 708 | ; AVX512F-NEXT: retq |
| 709 | ; |
| 710 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 711 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 712 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 713 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 714 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 715 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,1,3,3,4,5,6,7] |
| 716 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 717 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,1,3,3,4,5,6,7] |
| 718 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 719 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 720 | ; AVX512VL-NEXT: vzeroupper |
| 721 | ; AVX512VL-NEXT: retq |
| 722 | ; |
| 723 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 724 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 725 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 726 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 727 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 728 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 729 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 730 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 731 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 732 | ; AVX512BW-NEXT: vzeroupper |
| 733 | ; AVX512BW-NEXT: retq |
| 734 | ; |
| 735 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_2: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 736 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 737 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 738 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 739 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] |
| 740 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,1,3,3,4,5,6,7] |
| 741 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| 742 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,1,3,3,4,5,6,7] |
| 743 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 744 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 745 | ; AVX512BWVL-NEXT: vzeroupper |
| 746 | ; AVX512BWVL-NEXT: retq |
| 747 | %vec = load <32 x i8>, <32 x i8>* %L |
| 748 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 2, i32 10, i32 18, i32 26> |
| 749 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 750 | ret void |
| 751 | } |
| 752 | |
| 753 | define void @shuffle_v32i8_to_v4i8_3(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 754 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 755 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 756 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 757 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 758 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 759 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 760 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 761 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 762 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 763 | ; AVX1-NEXT: vzeroupper |
| 764 | ; AVX1-NEXT: retq |
| 765 | ; |
| 766 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 767 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 768 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 769 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 770 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 771 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 772 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 773 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 774 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 775 | ; AVX2-NEXT: vzeroupper |
| 776 | ; AVX2-NEXT: retq |
| 777 | ; |
| 778 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 779 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 780 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 781 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 782 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 783 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 784 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 785 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 786 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 787 | ; AVX512F-NEXT: vzeroupper |
| 788 | ; AVX512F-NEXT: retq |
| 789 | ; |
| 790 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 791 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 792 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 793 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | d873b6f | 2017-12-19 16:54:07 +0000 | [diff] [blame] | 794 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = [3,3,11,11,11,11,3,3,8,8,9,9,10,10,11,11] |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 795 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 796 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 797 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 798 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 799 | ; AVX512VL-NEXT: vzeroupper |
| 800 | ; AVX512VL-NEXT: retq |
| 801 | ; |
| 802 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 803 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 804 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 805 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 806 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 807 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 808 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 809 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 810 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 811 | ; AVX512BW-NEXT: vzeroupper |
| 812 | ; AVX512BW-NEXT: retq |
| 813 | ; |
| 814 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_3: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 815 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 816 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 817 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | d873b6f | 2017-12-19 16:54:07 +0000 | [diff] [blame] | 818 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [3,3,11,11,11,11,3,3,8,8,9,9,10,10,11,11] |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 819 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 820 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 821 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 822 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 823 | ; AVX512BWVL-NEXT: vzeroupper |
| 824 | ; AVX512BWVL-NEXT: retq |
| 825 | %vec = load <32 x i8>, <32 x i8>* %L |
| 826 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 3, i32 11, i32 19, i32 27> |
| 827 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 828 | ret void |
| 829 | } |
| 830 | |
| 831 | define void @shuffle_v32i8_to_v4i8_4(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 832 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_4: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 833 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 834 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 835 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 836 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 837 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 838 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 839 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 840 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 841 | ; AVX1-NEXT: vzeroupper |
| 842 | ; AVX1-NEXT: retq |
| 843 | ; |
| 844 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_4: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 845 | ; AVX2: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 846 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| 847 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 848 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 849 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 850 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 851 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 852 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 853 | ; AVX2-NEXT: vzeroupper |
| 854 | ; AVX2-NEXT: retq |
| 855 | ; |
| 856 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_4: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 857 | ; AVX512F: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 858 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| 859 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 860 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 861 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 862 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 863 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 864 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 865 | ; AVX512F-NEXT: vzeroupper |
| 866 | ; AVX512F-NEXT: retq |
| 867 | ; |
| 868 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_4: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 869 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 870 | ; AVX512VL-NEXT: vmovaps (%rdi), %ymm0 |
| 871 | ; AVX512VL-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 872 | ; AVX512VL-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] |
| 873 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 874 | ; AVX512VL-NEXT: vzeroupper |
| 875 | ; AVX512VL-NEXT: retq |
| 876 | ; |
| 877 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_4: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 878 | ; AVX512BW: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 879 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| 880 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 881 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 882 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 883 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 884 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 885 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 886 | ; AVX512BW-NEXT: vzeroupper |
| 887 | ; AVX512BW-NEXT: retq |
| 888 | ; |
| 889 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_4: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 890 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 891 | ; AVX512BWVL-NEXT: vmovaps (%rdi), %ymm0 |
| 892 | ; AVX512BWVL-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 893 | ; AVX512BWVL-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] |
| 894 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 895 | ; AVX512BWVL-NEXT: vzeroupper |
| 896 | ; AVX512BWVL-NEXT: retq |
| 897 | %vec = load <32 x i8>, <32 x i8>* %L |
| 898 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 4, i32 12, i32 20, i32 28> |
| 899 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 900 | ret void |
| 901 | } |
| 902 | |
| 903 | define void @shuffle_v32i8_to_v4i8_5(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 904 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_5: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 905 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 906 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 907 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 908 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 909 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 910 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 911 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 912 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 913 | ; AVX1-NEXT: vzeroupper |
| 914 | ; AVX1-NEXT: retq |
| 915 | ; |
| 916 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_5: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 917 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 918 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 919 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 920 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 921 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 922 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 923 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 924 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 925 | ; AVX2-NEXT: vzeroupper |
| 926 | ; AVX2-NEXT: retq |
| 927 | ; |
| 928 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_5: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 929 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 930 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 931 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 932 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 933 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 934 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 935 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 936 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 937 | ; AVX512F-NEXT: vzeroupper |
| 938 | ; AVX512F-NEXT: retq |
| 939 | ; |
| 940 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_5: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 941 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 942 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 943 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | d873b6f | 2017-12-19 16:54:07 +0000 | [diff] [blame] | 944 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = [5,5,13,13,13,13,5,5,4,4,5,5,6,6,7,7] |
| 945 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 946 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 947 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 948 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 949 | ; AVX512VL-NEXT: vzeroupper |
| 950 | ; AVX512VL-NEXT: retq |
| 951 | ; |
| 952 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_5: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 953 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 954 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 955 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 956 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 957 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 958 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 959 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 960 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 961 | ; AVX512BW-NEXT: vzeroupper |
| 962 | ; AVX512BW-NEXT: retq |
| 963 | ; |
| 964 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_5: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 965 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 966 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 967 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | d873b6f | 2017-12-19 16:54:07 +0000 | [diff] [blame] | 968 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [5,5,13,13,13,13,5,5,4,4,5,5,6,6,7,7] |
| 969 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 970 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 971 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 972 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 973 | ; AVX512BWVL-NEXT: vzeroupper |
| 974 | ; AVX512BWVL-NEXT: retq |
| 975 | %vec = load <32 x i8>, <32 x i8>* %L |
| 976 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 5, i32 13, i32 21, i32 29> |
| 977 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 978 | ret void |
| 979 | } |
| 980 | |
| 981 | define void @shuffle_v32i8_to_v4i8_6(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 982 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_6: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 983 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 984 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 985 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 986 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 987 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 988 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 989 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 990 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 991 | ; AVX1-NEXT: vzeroupper |
| 992 | ; AVX1-NEXT: retq |
| 993 | ; |
| 994 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_6: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 995 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 996 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 997 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 998 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 999 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1000 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1001 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1002 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 1003 | ; AVX2-NEXT: vzeroupper |
| 1004 | ; AVX2-NEXT: retq |
| 1005 | ; |
| 1006 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_6: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1007 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1008 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1009 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1010 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 1011 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1012 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1013 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1014 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 1015 | ; AVX512F-NEXT: vzeroupper |
| 1016 | ; AVX512F-NEXT: retq |
| 1017 | ; |
| 1018 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_6: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1019 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1020 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 1021 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1022 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 1023 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,1,3,4,5,6,7] |
| 1024 | ; AVX512VL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 1025 | ; AVX512VL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,1,3,4,5,6,7] |
| 1026 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 1027 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1028 | ; AVX512VL-NEXT: vzeroupper |
| 1029 | ; AVX512VL-NEXT: retq |
| 1030 | ; |
| 1031 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_6: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1032 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1033 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1034 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1035 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 1036 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1037 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1038 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1039 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 1040 | ; AVX512BW-NEXT: vzeroupper |
| 1041 | ; AVX512BW-NEXT: retq |
| 1042 | ; |
| 1043 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_6: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1044 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1045 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 1046 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1047 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] |
| 1048 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[3,1,1,3,4,5,6,7] |
| 1049 | ; AVX512BWVL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] |
| 1050 | ; AVX512BWVL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,1,1,3,4,5,6,7] |
| 1051 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 1052 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1053 | ; AVX512BWVL-NEXT: vzeroupper |
| 1054 | ; AVX512BWVL-NEXT: retq |
| 1055 | %vec = load <32 x i8>, <32 x i8>* %L |
| 1056 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 6, i32 14, i32 22, i32 30> |
| 1057 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 1058 | ret void |
| 1059 | } |
| 1060 | |
| 1061 | define void @shuffle_v32i8_to_v4i8_7(<32 x i8>* %L, <4 x i8>* %S) nounwind { |
| 1062 | ; AVX1-LABEL: shuffle_v32i8_to_v4i8_7: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1063 | ; AVX1: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1064 | ; AVX1-NEXT: vmovdqa (%rdi), %ymm0 |
| 1065 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 1066 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 1067 | ; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1068 | ; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1069 | ; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| 1070 | ; AVX1-NEXT: vmovd %xmm0, (%rsi) |
| 1071 | ; AVX1-NEXT: vzeroupper |
| 1072 | ; AVX1-NEXT: retq |
| 1073 | ; |
| 1074 | ; AVX2-LABEL: shuffle_v32i8_to_v4i8_7: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1075 | ; AVX2: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1076 | ; AVX2-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1077 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1078 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 1079 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1080 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1081 | ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1082 | ; AVX2-NEXT: vmovd %xmm0, (%rsi) |
| 1083 | ; AVX2-NEXT: vzeroupper |
| 1084 | ; AVX2-NEXT: retq |
| 1085 | ; |
| 1086 | ; AVX512F-LABEL: shuffle_v32i8_to_v4i8_7: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1087 | ; AVX512F: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1088 | ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1089 | ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1090 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm2 = <7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 1091 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1092 | ; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1093 | ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1094 | ; AVX512F-NEXT: vmovd %xmm0, (%rsi) |
| 1095 | ; AVX512F-NEXT: vzeroupper |
| 1096 | ; AVX512F-NEXT: retq |
| 1097 | ; |
| 1098 | ; AVX512VL-LABEL: shuffle_v32i8_to_v4i8_7: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1099 | ; AVX512VL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1100 | ; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0 |
| 1101 | ; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1102 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm2 = [7,7,14,14,15,15,14,14,15,15,4,4,5,5,6,6] |
| 1103 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1104 | ; AVX512VL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1105 | ; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 1106 | ; AVX512VL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1107 | ; AVX512VL-NEXT: vzeroupper |
| 1108 | ; AVX512VL-NEXT: retq |
| 1109 | ; |
| 1110 | ; AVX512BW-LABEL: shuffle_v32i8_to_v4i8_7: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1111 | ; AVX512BW: # %bb.0: |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1112 | ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1113 | ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1114 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = <7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u> |
| 1115 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1116 | ; AVX512BW-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1117 | ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1118 | ; AVX512BW-NEXT: vmovd %xmm0, (%rsi) |
| 1119 | ; AVX512BW-NEXT: vzeroupper |
| 1120 | ; AVX512BW-NEXT: retq |
| 1121 | ; |
| 1122 | ; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8_7: |
| Francis Visoiu Mistrih | 25528d6 | 2017-12-04 17:18:51 +0000 | [diff] [blame] | 1123 | ; AVX512BWVL: # %bb.0: |
| Wei Mi | 1736efd | 2017-10-12 00:24:52 +0000 | [diff] [blame] | 1124 | ; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0 |
| 1125 | ; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1126 | ; AVX512BWVL-NEXT: vmovdqa {{.*#+}} xmm2 = [7,7,14,14,15,15,14,14,15,15,4,4,5,5,6,6] |
| 1127 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 1128 | ; AVX512BWVL-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 1129 | ; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 1130 | ; AVX512BWVL-NEXT: vpmovdb %xmm0, (%rsi) |
| Zvi Rackover | eac8e7c | 2017-07-20 21:03:36 +0000 | [diff] [blame] | 1131 | ; AVX512BWVL-NEXT: vzeroupper |
| 1132 | ; AVX512BWVL-NEXT: retq |
| 1133 | %vec = load <32 x i8>, <32 x i8>* %L |
| 1134 | %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 7, i32 15, i32 23, i32 31> |
| 1135 | store <4 x i8> %strided.vec, <4 x i8>* %S |
| 1136 | ret void |
| 1137 | } |
| 1138 | |