Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +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=+avx512f | FileCheck %s --check-prefixes=AVX512,AVX512F |
| 3 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VL |
| 4 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 5 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi2 | FileCheck %s --check-prefixes=AVX512,AVX512VBMI2 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 6 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VLBW |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 7 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi2,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VLVBMI2 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 8 | |
| 9 | declare <8 x i64> @llvm.fshl.v8i64(<8 x i64>, <8 x i64>, <8 x i64>) |
| 10 | declare <16 x i32> @llvm.fshl.v16i32(<16 x i32>, <16 x i32>, <16 x i32>) |
| 11 | declare <32 x i16> @llvm.fshl.v32i16(<32 x i16>, <32 x i16>, <32 x i16>) |
| 12 | declare <64 x i8> @llvm.fshl.v64i8(<64 x i8>, <64 x i8>, <64 x i8>) |
| 13 | |
| 14 | ; |
| 15 | ; Variable Shifts |
| 16 | ; |
| 17 | |
| 18 | define <8 x i64> @var_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %amt) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 19 | ; AVX512F-LABEL: var_funnnel_v8i64: |
| 20 | ; AVX512F: # %bb.0: |
| 21 | ; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 22 | ; AVX512F-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 23 | ; AVX512F-NEXT: vpsllvq %zmm4, %zmm0, %zmm5 |
| 24 | ; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64] |
| 25 | ; AVX512F-NEXT: vpsubq %zmm4, %zmm6, %zmm4 |
| 26 | ; AVX512F-NEXT: vpsrlvq %zmm4, %zmm1, %zmm1 |
| 27 | ; AVX512F-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 28 | ; AVX512F-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 29 | ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 30 | ; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 31 | ; AVX512F-NEXT: retq |
| 32 | ; |
| 33 | ; AVX512VL-LABEL: var_funnnel_v8i64: |
| 34 | ; AVX512VL: # %bb.0: |
| 35 | ; AVX512VL-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 36 | ; AVX512VL-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 37 | ; AVX512VL-NEXT: vpsllvq %zmm4, %zmm0, %zmm5 |
| 38 | ; AVX512VL-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64] |
| 39 | ; AVX512VL-NEXT: vpsubq %zmm4, %zmm6, %zmm4 |
| 40 | ; AVX512VL-NEXT: vpsrlvq %zmm4, %zmm1, %zmm1 |
| 41 | ; AVX512VL-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 42 | ; AVX512VL-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 43 | ; AVX512VL-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 44 | ; AVX512VL-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 45 | ; AVX512VL-NEXT: retq |
| 46 | ; |
| 47 | ; AVX512BW-LABEL: var_funnnel_v8i64: |
| 48 | ; AVX512BW: # %bb.0: |
| 49 | ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 50 | ; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 51 | ; AVX512BW-NEXT: vpsllvq %zmm4, %zmm0, %zmm5 |
| 52 | ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64] |
| 53 | ; AVX512BW-NEXT: vpsubq %zmm4, %zmm6, %zmm4 |
| 54 | ; AVX512BW-NEXT: vpsrlvq %zmm4, %zmm1, %zmm1 |
| 55 | ; AVX512BW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 56 | ; AVX512BW-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 57 | ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 58 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 59 | ; AVX512BW-NEXT: retq |
| 60 | ; |
| 61 | ; AVX512VBMI2-LABEL: var_funnnel_v8i64: |
| 62 | ; AVX512VBMI2: # %bb.0: |
| 63 | ; AVX512VBMI2-NEXT: vpshldvq %zmm2, %zmm1, %zmm0 |
| 64 | ; AVX512VBMI2-NEXT: retq |
| 65 | ; |
| 66 | ; AVX512VLBW-LABEL: var_funnnel_v8i64: |
| 67 | ; AVX512VLBW: # %bb.0: |
| 68 | ; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 69 | ; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 70 | ; AVX512VLBW-NEXT: vpsllvq %zmm4, %zmm0, %zmm5 |
| 71 | ; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64] |
| 72 | ; AVX512VLBW-NEXT: vpsubq %zmm4, %zmm6, %zmm4 |
| 73 | ; AVX512VLBW-NEXT: vpsrlvq %zmm4, %zmm1, %zmm1 |
| 74 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 75 | ; AVX512VLBW-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 76 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 77 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 78 | ; AVX512VLBW-NEXT: retq |
| 79 | ; |
| 80 | ; AVX512VLVBMI2-LABEL: var_funnnel_v8i64: |
| 81 | ; AVX512VLVBMI2: # %bb.0: |
| 82 | ; AVX512VLVBMI2-NEXT: vpshldvq %zmm2, %zmm1, %zmm0 |
| 83 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 84 | %res = call <8 x i64> @llvm.fshl.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %amt) |
| 85 | ret <8 x i64> %res |
| 86 | } |
| 87 | |
| 88 | define <16 x i32> @var_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %amt) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 89 | ; AVX512F-LABEL: var_funnnel_v16i32: |
| 90 | ; AVX512F: # %bb.0: |
| 91 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 92 | ; AVX512F-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 93 | ; AVX512F-NEXT: vpsllvd %zmm4, %zmm0, %zmm5 |
| 94 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32] |
| 95 | ; AVX512F-NEXT: vpsubd %zmm4, %zmm6, %zmm4 |
| 96 | ; AVX512F-NEXT: vpsrlvd %zmm4, %zmm1, %zmm1 |
| 97 | ; AVX512F-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 98 | ; AVX512F-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 99 | ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 100 | ; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 101 | ; AVX512F-NEXT: retq |
| 102 | ; |
| 103 | ; AVX512VL-LABEL: var_funnnel_v16i32: |
| 104 | ; AVX512VL: # %bb.0: |
| 105 | ; AVX512VL-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 106 | ; AVX512VL-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 107 | ; AVX512VL-NEXT: vpsllvd %zmm4, %zmm0, %zmm5 |
| 108 | ; AVX512VL-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32] |
| 109 | ; AVX512VL-NEXT: vpsubd %zmm4, %zmm6, %zmm4 |
| 110 | ; AVX512VL-NEXT: vpsrlvd %zmm4, %zmm1, %zmm1 |
| 111 | ; AVX512VL-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 112 | ; AVX512VL-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 113 | ; AVX512VL-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 114 | ; AVX512VL-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 115 | ; AVX512VL-NEXT: retq |
| 116 | ; |
| 117 | ; AVX512BW-LABEL: var_funnnel_v16i32: |
| 118 | ; AVX512BW: # %bb.0: |
| 119 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 120 | ; AVX512BW-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 121 | ; AVX512BW-NEXT: vpsllvd %zmm4, %zmm0, %zmm5 |
| 122 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32] |
| 123 | ; AVX512BW-NEXT: vpsubd %zmm4, %zmm6, %zmm4 |
| 124 | ; AVX512BW-NEXT: vpsrlvd %zmm4, %zmm1, %zmm1 |
| 125 | ; AVX512BW-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 126 | ; AVX512BW-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 127 | ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 128 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 129 | ; AVX512BW-NEXT: retq |
| 130 | ; |
| 131 | ; AVX512VBMI2-LABEL: var_funnnel_v16i32: |
| 132 | ; AVX512VBMI2: # %bb.0: |
| 133 | ; AVX512VBMI2-NEXT: vpshldvd %zmm2, %zmm1, %zmm0 |
| 134 | ; AVX512VBMI2-NEXT: retq |
| 135 | ; |
| 136 | ; AVX512VLBW-LABEL: var_funnnel_v16i32: |
| 137 | ; AVX512VLBW: # %bb.0: |
| 138 | ; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 139 | ; AVX512VLBW-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 140 | ; AVX512VLBW-NEXT: vpsllvd %zmm4, %zmm0, %zmm5 |
| 141 | ; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32] |
| 142 | ; AVX512VLBW-NEXT: vpsubd %zmm4, %zmm6, %zmm4 |
| 143 | ; AVX512VLBW-NEXT: vpsrlvd %zmm4, %zmm1, %zmm1 |
| 144 | ; AVX512VLBW-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 145 | ; AVX512VLBW-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 146 | ; AVX512VLBW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 147 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 148 | ; AVX512VLBW-NEXT: retq |
| 149 | ; |
| 150 | ; AVX512VLVBMI2-LABEL: var_funnnel_v16i32: |
| 151 | ; AVX512VLVBMI2: # %bb.0: |
| 152 | ; AVX512VLVBMI2-NEXT: vpshldvd %zmm2, %zmm1, %zmm0 |
| 153 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 154 | %res = call <16 x i32> @llvm.fshl.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %amt) |
| 155 | ret <16 x i32> %res |
| 156 | } |
| 157 | |
| 158 | define <32 x i16> @var_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %amt) nounwind { |
| 159 | ; AVX512F-LABEL: var_funnnel_v32i16: |
| 160 | ; AVX512F: # %bb.0: |
| 161 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 162 | ; AVX512F-NEXT: vpand %ymm6, %ymm4, %ymm4 |
| 163 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm7 = ymm4[0],zero,ymm4[1],zero,ymm4[2],zero,ymm4[3],zero,ymm4[4],zero,ymm4[5],zero,ymm4[6],zero,ymm4[7],zero,ymm4[8],zero,ymm4[9],zero,ymm4[10],zero,ymm4[11],zero,ymm4[12],zero,ymm4[13],zero,ymm4[14],zero,ymm4[15],zero |
| 164 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm8 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero |
| 165 | ; AVX512F-NEXT: vpsllvd %zmm7, %zmm8, %zmm7 |
| 166 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm8 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16] |
| 167 | ; AVX512F-NEXT: vpsubw %ymm4, %ymm8, %ymm9 |
| 168 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm9 = ymm9[0],zero,ymm9[1],zero,ymm9[2],zero,ymm9[3],zero,ymm9[4],zero,ymm9[5],zero,ymm9[6],zero,ymm9[7],zero,ymm9[8],zero,ymm9[9],zero,ymm9[10],zero,ymm9[11],zero,ymm9[12],zero,ymm9[13],zero,ymm9[14],zero,ymm9[15],zero |
| 169 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm2 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero |
| 170 | ; AVX512F-NEXT: vpsrlvd %zmm9, %zmm2, %zmm2 |
| 171 | ; AVX512F-NEXT: vpord %zmm2, %zmm7, %zmm2 |
| 172 | ; AVX512F-NEXT: vpmovdw %zmm2, %ymm2 |
| 173 | ; AVX512F-NEXT: vpxor %xmm7, %xmm7, %xmm7 |
| 174 | ; AVX512F-NEXT: vpcmpeqw %ymm7, %ymm4, %ymm4 |
| 175 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 176 | ; AVX512F-NEXT: vpand %ymm6, %ymm5, %ymm2 |
| 177 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm4 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero |
| 178 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm1[0],zero,ymm1[1],zero,ymm1[2],zero,ymm1[3],zero,ymm1[4],zero,ymm1[5],zero,ymm1[6],zero,ymm1[7],zero,ymm1[8],zero,ymm1[9],zero,ymm1[10],zero,ymm1[11],zero,ymm1[12],zero,ymm1[13],zero,ymm1[14],zero,ymm1[15],zero |
| 179 | ; AVX512F-NEXT: vpsllvd %zmm4, %zmm5, %zmm4 |
| 180 | ; AVX512F-NEXT: vpsubw %ymm2, %ymm8, %ymm5 |
| 181 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm5[0],zero,ymm5[1],zero,ymm5[2],zero,ymm5[3],zero,ymm5[4],zero,ymm5[5],zero,ymm5[6],zero,ymm5[7],zero,ymm5[8],zero,ymm5[9],zero,ymm5[10],zero,ymm5[11],zero,ymm5[12],zero,ymm5[13],zero,ymm5[14],zero,ymm5[15],zero |
| 182 | ; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm3 = ymm3[0],zero,ymm3[1],zero,ymm3[2],zero,ymm3[3],zero,ymm3[4],zero,ymm3[5],zero,ymm3[6],zero,ymm3[7],zero,ymm3[8],zero,ymm3[9],zero,ymm3[10],zero,ymm3[11],zero,ymm3[12],zero,ymm3[13],zero,ymm3[14],zero,ymm3[15],zero |
| 183 | ; AVX512F-NEXT: vpsrlvd %zmm5, %zmm3, %zmm3 |
| 184 | ; AVX512F-NEXT: vpord %zmm3, %zmm4, %zmm3 |
| 185 | ; AVX512F-NEXT: vpmovdw %zmm3, %ymm3 |
| 186 | ; AVX512F-NEXT: vpcmpeqw %ymm7, %ymm2, %ymm2 |
| 187 | ; AVX512F-NEXT: vpblendvb %ymm2, %ymm1, %ymm3, %ymm1 |
| 188 | ; AVX512F-NEXT: retq |
| 189 | ; |
| 190 | ; AVX512VL-LABEL: var_funnnel_v32i16: |
| 191 | ; AVX512VL: # %bb.0: |
| 192 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 193 | ; AVX512VL-NEXT: vpand %ymm6, %ymm4, %ymm4 |
| 194 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm7 = ymm4[0],zero,ymm4[1],zero,ymm4[2],zero,ymm4[3],zero,ymm4[4],zero,ymm4[5],zero,ymm4[6],zero,ymm4[7],zero,ymm4[8],zero,ymm4[9],zero,ymm4[10],zero,ymm4[11],zero,ymm4[12],zero,ymm4[13],zero,ymm4[14],zero,ymm4[15],zero |
| 195 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm8 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero |
| 196 | ; AVX512VL-NEXT: vpsllvd %zmm7, %zmm8, %zmm7 |
| 197 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm8 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16] |
| 198 | ; AVX512VL-NEXT: vpsubw %ymm4, %ymm8, %ymm9 |
| 199 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm9 = ymm9[0],zero,ymm9[1],zero,ymm9[2],zero,ymm9[3],zero,ymm9[4],zero,ymm9[5],zero,ymm9[6],zero,ymm9[7],zero,ymm9[8],zero,ymm9[9],zero,ymm9[10],zero,ymm9[11],zero,ymm9[12],zero,ymm9[13],zero,ymm9[14],zero,ymm9[15],zero |
| 200 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm2 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero |
| 201 | ; AVX512VL-NEXT: vpsrlvd %zmm9, %zmm2, %zmm2 |
| 202 | ; AVX512VL-NEXT: vpord %zmm2, %zmm7, %zmm2 |
| 203 | ; AVX512VL-NEXT: vpmovdw %zmm2, %ymm2 |
| 204 | ; AVX512VL-NEXT: vpxor %xmm7, %xmm7, %xmm7 |
| 205 | ; AVX512VL-NEXT: vpcmpeqw %ymm7, %ymm4, %ymm4 |
| 206 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 207 | ; AVX512VL-NEXT: vpand %ymm6, %ymm5, %ymm2 |
| 208 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm4 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero |
| 209 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm1[0],zero,ymm1[1],zero,ymm1[2],zero,ymm1[3],zero,ymm1[4],zero,ymm1[5],zero,ymm1[6],zero,ymm1[7],zero,ymm1[8],zero,ymm1[9],zero,ymm1[10],zero,ymm1[11],zero,ymm1[12],zero,ymm1[13],zero,ymm1[14],zero,ymm1[15],zero |
| 210 | ; AVX512VL-NEXT: vpsllvd %zmm4, %zmm5, %zmm4 |
| 211 | ; AVX512VL-NEXT: vpsubw %ymm2, %ymm8, %ymm5 |
| 212 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm5[0],zero,ymm5[1],zero,ymm5[2],zero,ymm5[3],zero,ymm5[4],zero,ymm5[5],zero,ymm5[6],zero,ymm5[7],zero,ymm5[8],zero,ymm5[9],zero,ymm5[10],zero,ymm5[11],zero,ymm5[12],zero,ymm5[13],zero,ymm5[14],zero,ymm5[15],zero |
| 213 | ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm3 = ymm3[0],zero,ymm3[1],zero,ymm3[2],zero,ymm3[3],zero,ymm3[4],zero,ymm3[5],zero,ymm3[6],zero,ymm3[7],zero,ymm3[8],zero,ymm3[9],zero,ymm3[10],zero,ymm3[11],zero,ymm3[12],zero,ymm3[13],zero,ymm3[14],zero,ymm3[15],zero |
| 214 | ; AVX512VL-NEXT: vpsrlvd %zmm5, %zmm3, %zmm3 |
| 215 | ; AVX512VL-NEXT: vpord %zmm3, %zmm4, %zmm3 |
| 216 | ; AVX512VL-NEXT: vpmovdw %zmm3, %ymm3 |
| 217 | ; AVX512VL-NEXT: vpcmpeqw %ymm7, %ymm2, %ymm2 |
| 218 | ; AVX512VL-NEXT: vpblendvb %ymm2, %ymm1, %ymm3, %ymm1 |
| 219 | ; AVX512VL-NEXT: retq |
| 220 | ; |
| 221 | ; AVX512BW-LABEL: var_funnnel_v32i16: |
| 222 | ; AVX512BW: # %bb.0: |
| 223 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 224 | ; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 225 | ; AVX512BW-NEXT: vpsllvw %zmm4, %zmm0, %zmm5 |
| 226 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16] |
| 227 | ; AVX512BW-NEXT: vpsubw %zmm4, %zmm6, %zmm4 |
| 228 | ; AVX512BW-NEXT: vpsrlvw %zmm4, %zmm1, %zmm1 |
| 229 | ; AVX512BW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 230 | ; AVX512BW-NEXT: vptestnmw %zmm3, %zmm2, %k1 |
| 231 | ; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm1 {%k1} |
| 232 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 233 | ; AVX512BW-NEXT: retq |
| 234 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 235 | ; AVX512VBMI2-LABEL: var_funnnel_v32i16: |
| 236 | ; AVX512VBMI2: # %bb.0: |
| 237 | ; AVX512VBMI2-NEXT: vpshldvw %zmm2, %zmm1, %zmm0 |
| 238 | ; AVX512VBMI2-NEXT: retq |
| 239 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 240 | ; AVX512VLBW-LABEL: var_funnnel_v32i16: |
| 241 | ; AVX512VLBW: # %bb.0: |
| 242 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 243 | ; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 244 | ; AVX512VLBW-NEXT: vpsllvw %zmm4, %zmm0, %zmm5 |
| 245 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16] |
| 246 | ; AVX512VLBW-NEXT: vpsubw %zmm4, %zmm6, %zmm4 |
| 247 | ; AVX512VLBW-NEXT: vpsrlvw %zmm4, %zmm1, %zmm1 |
| 248 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 249 | ; AVX512VLBW-NEXT: vptestnmw %zmm3, %zmm2, %k1 |
| 250 | ; AVX512VLBW-NEXT: vmovdqu16 %zmm0, %zmm1 {%k1} |
| 251 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 252 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 253 | ; |
| 254 | ; AVX512VLVBMI2-LABEL: var_funnnel_v32i16: |
| 255 | ; AVX512VLVBMI2: # %bb.0: |
| 256 | ; AVX512VLVBMI2-NEXT: vpshldvw %zmm2, %zmm1, %zmm0 |
| 257 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 258 | %res = call <32 x i16> @llvm.fshl.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %amt) |
| 259 | ret <32 x i16> %res |
| 260 | } |
| 261 | |
| 262 | define <64 x i8> @var_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %amt) nounwind { |
| 263 | ; AVX512F-LABEL: var_funnnel_v64i8: |
| 264 | ; AVX512F: # %bb.0: |
| 265 | ; AVX512F-NEXT: vpsllw $4, %ymm0, %ymm7 |
| 266 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240] |
| 267 | ; AVX512F-NEXT: vpand %ymm6, %ymm7, %ymm8 |
| 268 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm7 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 269 | ; AVX512F-NEXT: vpand %ymm7, %ymm4, %ymm9 |
| 270 | ; AVX512F-NEXT: vpsllw $5, %ymm9, %ymm10 |
| 271 | ; AVX512F-NEXT: vpblendvb %ymm10, %ymm8, %ymm0, %ymm8 |
| 272 | ; AVX512F-NEXT: vpsllw $2, %ymm8, %ymm11 |
| 273 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252] |
| 274 | ; AVX512F-NEXT: vpand %ymm4, %ymm11, %ymm11 |
| 275 | ; AVX512F-NEXT: vpaddb %ymm10, %ymm10, %ymm10 |
| 276 | ; AVX512F-NEXT: vpblendvb %ymm10, %ymm11, %ymm8, %ymm8 |
| 277 | ; AVX512F-NEXT: vpaddb %ymm8, %ymm8, %ymm11 |
| 278 | ; AVX512F-NEXT: vpaddb %ymm10, %ymm10, %ymm10 |
| 279 | ; AVX512F-NEXT: vpblendvb %ymm10, %ymm11, %ymm8, %ymm10 |
| 280 | ; AVX512F-NEXT: vpsrlw $4, %ymm2, %ymm11 |
| 281 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm8 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 282 | ; AVX512F-NEXT: vpand %ymm8, %ymm11, %ymm11 |
| 283 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm12 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 284 | ; AVX512F-NEXT: vpsubb %ymm9, %ymm12, %ymm13 |
| 285 | ; AVX512F-NEXT: vpsllw $5, %ymm13, %ymm13 |
| 286 | ; AVX512F-NEXT: vpblendvb %ymm13, %ymm11, %ymm2, %ymm2 |
| 287 | ; AVX512F-NEXT: vpsrlw $2, %ymm2, %ymm11 |
| 288 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm14 = [63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63] |
| 289 | ; AVX512F-NEXT: vpand %ymm14, %ymm11, %ymm11 |
| 290 | ; AVX512F-NEXT: vpaddb %ymm13, %ymm13, %ymm13 |
| 291 | ; AVX512F-NEXT: vpblendvb %ymm13, %ymm11, %ymm2, %ymm2 |
| 292 | ; AVX512F-NEXT: vpsrlw $1, %ymm2, %ymm11 |
| 293 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm15 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127] |
| 294 | ; AVX512F-NEXT: vpand %ymm15, %ymm11, %ymm11 |
| 295 | ; AVX512F-NEXT: vpaddb %ymm13, %ymm13, %ymm13 |
| 296 | ; AVX512F-NEXT: vpblendvb %ymm13, %ymm11, %ymm2, %ymm2 |
| 297 | ; AVX512F-NEXT: vpor %ymm2, %ymm10, %ymm2 |
| 298 | ; AVX512F-NEXT: vpxor %xmm10, %xmm10, %xmm10 |
| 299 | ; AVX512F-NEXT: vpcmpeqb %ymm10, %ymm9, %ymm9 |
| 300 | ; AVX512F-NEXT: vpblendvb %ymm9, %ymm0, %ymm2, %ymm0 |
| 301 | ; AVX512F-NEXT: vpsllw $4, %ymm1, %ymm2 |
| 302 | ; AVX512F-NEXT: vpand %ymm6, %ymm2, %ymm2 |
| 303 | ; AVX512F-NEXT: vpand %ymm7, %ymm5, %ymm5 |
| 304 | ; AVX512F-NEXT: vpsllw $5, %ymm5, %ymm6 |
| 305 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm2 |
| 306 | ; AVX512F-NEXT: vpsllw $2, %ymm2, %ymm7 |
| 307 | ; AVX512F-NEXT: vpand %ymm4, %ymm7, %ymm4 |
| 308 | ; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6 |
| 309 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2 |
| 310 | ; AVX512F-NEXT: vpaddb %ymm2, %ymm2, %ymm4 |
| 311 | ; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6 |
| 312 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2 |
| 313 | ; AVX512F-NEXT: vpsrlw $4, %ymm3, %ymm4 |
| 314 | ; AVX512F-NEXT: vpand %ymm8, %ymm4, %ymm4 |
| 315 | ; AVX512F-NEXT: vpsubb %ymm5, %ymm12, %ymm6 |
| 316 | ; AVX512F-NEXT: vpsllw $5, %ymm6, %ymm6 |
| 317 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm3, %ymm3 |
| 318 | ; AVX512F-NEXT: vpsrlw $2, %ymm3, %ymm4 |
| 319 | ; AVX512F-NEXT: vpand %ymm14, %ymm4, %ymm4 |
| 320 | ; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6 |
| 321 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm3, %ymm3 |
| 322 | ; AVX512F-NEXT: vpsrlw $1, %ymm3, %ymm4 |
| 323 | ; AVX512F-NEXT: vpand %ymm15, %ymm4, %ymm4 |
| 324 | ; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6 |
| 325 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm3, %ymm3 |
| 326 | ; AVX512F-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 327 | ; AVX512F-NEXT: vpcmpeqb %ymm10, %ymm5, %ymm3 |
| 328 | ; AVX512F-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 |
| 329 | ; AVX512F-NEXT: retq |
| 330 | ; |
| 331 | ; AVX512VL-LABEL: var_funnnel_v64i8: |
| 332 | ; AVX512VL: # %bb.0: |
| 333 | ; AVX512VL-NEXT: vpsllw $4, %ymm0, %ymm6 |
| 334 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm7 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240] |
| 335 | ; AVX512VL-NEXT: vpand %ymm7, %ymm6, %ymm6 |
| 336 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm8 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 337 | ; AVX512VL-NEXT: vpand %ymm8, %ymm4, %ymm4 |
| 338 | ; AVX512VL-NEXT: vpsllw $5, %ymm4, %ymm9 |
| 339 | ; AVX512VL-NEXT: vpblendvb %ymm9, %ymm6, %ymm0, %ymm6 |
| 340 | ; AVX512VL-NEXT: vpsllw $2, %ymm6, %ymm10 |
| 341 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm11 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252] |
| 342 | ; AVX512VL-NEXT: vpand %ymm11, %ymm10, %ymm10 |
| 343 | ; AVX512VL-NEXT: vpaddb %ymm9, %ymm9, %ymm9 |
| 344 | ; AVX512VL-NEXT: vpblendvb %ymm9, %ymm10, %ymm6, %ymm6 |
| 345 | ; AVX512VL-NEXT: vpaddb %ymm6, %ymm6, %ymm10 |
| 346 | ; AVX512VL-NEXT: vpaddb %ymm9, %ymm9, %ymm9 |
| 347 | ; AVX512VL-NEXT: vpblendvb %ymm9, %ymm10, %ymm6, %ymm6 |
| 348 | ; AVX512VL-NEXT: vpsrlw $4, %ymm2, %ymm9 |
| 349 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm10 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 350 | ; AVX512VL-NEXT: vpand %ymm10, %ymm9, %ymm9 |
| 351 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm12 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 352 | ; AVX512VL-NEXT: vpsubb %ymm4, %ymm12, %ymm13 |
| 353 | ; AVX512VL-NEXT: vpsllw $5, %ymm13, %ymm13 |
| 354 | ; AVX512VL-NEXT: vpblendvb %ymm13, %ymm9, %ymm2, %ymm2 |
| 355 | ; AVX512VL-NEXT: vpsrlw $2, %ymm2, %ymm9 |
| 356 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm14 = [63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63] |
| 357 | ; AVX512VL-NEXT: vpand %ymm14, %ymm9, %ymm9 |
| 358 | ; AVX512VL-NEXT: vpaddb %ymm13, %ymm13, %ymm13 |
| 359 | ; AVX512VL-NEXT: vpblendvb %ymm13, %ymm9, %ymm2, %ymm2 |
| 360 | ; AVX512VL-NEXT: vpsrlw $1, %ymm2, %ymm9 |
| 361 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm15 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127] |
| 362 | ; AVX512VL-NEXT: vpand %ymm15, %ymm9, %ymm9 |
| 363 | ; AVX512VL-NEXT: vpaddb %ymm13, %ymm13, %ymm13 |
| 364 | ; AVX512VL-NEXT: vpblendvb %ymm13, %ymm9, %ymm2, %ymm2 |
| 365 | ; AVX512VL-NEXT: vpor %ymm2, %ymm6, %ymm2 |
| 366 | ; AVX512VL-NEXT: vpxor %xmm6, %xmm6, %xmm6 |
| 367 | ; AVX512VL-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm4 |
| 368 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 369 | ; AVX512VL-NEXT: vpsllw $4, %ymm1, %ymm2 |
| 370 | ; AVX512VL-NEXT: vpand %ymm7, %ymm2, %ymm2 |
| 371 | ; AVX512VL-NEXT: vpand %ymm8, %ymm5, %ymm4 |
| 372 | ; AVX512VL-NEXT: vpsllw $5, %ymm4, %ymm5 |
| 373 | ; AVX512VL-NEXT: vpblendvb %ymm5, %ymm2, %ymm1, %ymm2 |
| 374 | ; AVX512VL-NEXT: vpsllw $2, %ymm2, %ymm7 |
| 375 | ; AVX512VL-NEXT: vpand %ymm11, %ymm7, %ymm7 |
| 376 | ; AVX512VL-NEXT: vpaddb %ymm5, %ymm5, %ymm5 |
| 377 | ; AVX512VL-NEXT: vpblendvb %ymm5, %ymm7, %ymm2, %ymm2 |
| 378 | ; AVX512VL-NEXT: vpaddb %ymm2, %ymm2, %ymm7 |
| 379 | ; AVX512VL-NEXT: vpaddb %ymm5, %ymm5, %ymm5 |
| 380 | ; AVX512VL-NEXT: vpblendvb %ymm5, %ymm7, %ymm2, %ymm2 |
| 381 | ; AVX512VL-NEXT: vpsrlw $4, %ymm3, %ymm5 |
| 382 | ; AVX512VL-NEXT: vpand %ymm10, %ymm5, %ymm5 |
| 383 | ; AVX512VL-NEXT: vpsubb %ymm4, %ymm12, %ymm7 |
| 384 | ; AVX512VL-NEXT: vpsllw $5, %ymm7, %ymm7 |
| 385 | ; AVX512VL-NEXT: vpblendvb %ymm7, %ymm5, %ymm3, %ymm3 |
| 386 | ; AVX512VL-NEXT: vpsrlw $2, %ymm3, %ymm5 |
| 387 | ; AVX512VL-NEXT: vpand %ymm14, %ymm5, %ymm5 |
| 388 | ; AVX512VL-NEXT: vpaddb %ymm7, %ymm7, %ymm7 |
| 389 | ; AVX512VL-NEXT: vpblendvb %ymm7, %ymm5, %ymm3, %ymm3 |
| 390 | ; AVX512VL-NEXT: vpsrlw $1, %ymm3, %ymm5 |
| 391 | ; AVX512VL-NEXT: vpand %ymm15, %ymm5, %ymm5 |
| 392 | ; AVX512VL-NEXT: vpaddb %ymm7, %ymm7, %ymm7 |
| 393 | ; AVX512VL-NEXT: vpblendvb %ymm7, %ymm5, %ymm3, %ymm3 |
| 394 | ; AVX512VL-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 395 | ; AVX512VL-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm3 |
| 396 | ; AVX512VL-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 |
| 397 | ; AVX512VL-NEXT: retq |
| 398 | ; |
| 399 | ; AVX512BW-LABEL: var_funnnel_v64i8: |
| 400 | ; AVX512BW: # %bb.0: |
| 401 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 402 | ; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 403 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 404 | ; AVX512BW-NEXT: vpsubb %zmm4, %zmm5, %zmm5 |
| 405 | ; AVX512BW-NEXT: vpsllw $5, %zmm5, %zmm5 |
| 406 | ; AVX512BW-NEXT: vpaddb %zmm5, %zmm5, %zmm6 |
| 407 | ; AVX512BW-NEXT: vpmovb2m %zmm6, %k1 |
| 408 | ; AVX512BW-NEXT: vpmovb2m %zmm5, %k2 |
| 409 | ; AVX512BW-NEXT: vpsrlw $4, %zmm1, %zmm5 |
| 410 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 411 | ; AVX512BW-NEXT: vmovdqu8 %zmm5, %zmm1 {%k2} |
| 412 | ; AVX512BW-NEXT: vpsrlw $2, %zmm1, %zmm5 |
| 413 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 414 | ; AVX512BW-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 415 | ; AVX512BW-NEXT: vpsrlw $1, %zmm1, %zmm5 |
| 416 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 417 | ; AVX512BW-NEXT: vpaddb %zmm6, %zmm6, %zmm6 |
| 418 | ; AVX512BW-NEXT: vpmovb2m %zmm6, %k1 |
| 419 | ; AVX512BW-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 420 | ; AVX512BW-NEXT: vpsllw $5, %zmm4, %zmm4 |
| 421 | ; AVX512BW-NEXT: vpaddb %zmm4, %zmm4, %zmm5 |
| 422 | ; AVX512BW-NEXT: vpmovb2m %zmm5, %k1 |
| 423 | ; AVX512BW-NEXT: vpmovb2m %zmm4, %k2 |
| 424 | ; AVX512BW-NEXT: vpsllw $4, %zmm0, %zmm4 |
| 425 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 426 | ; AVX512BW-NEXT: vpblendmb %zmm4, %zmm0, %zmm4 {%k2} |
| 427 | ; AVX512BW-NEXT: vpsllw $2, %zmm4, %zmm6 |
| 428 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm6, %zmm6 |
| 429 | ; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm4 {%k1} |
| 430 | ; AVX512BW-NEXT: vpaddb %zmm5, %zmm5, %zmm5 |
| 431 | ; AVX512BW-NEXT: vpmovb2m %zmm5, %k1 |
| 432 | ; AVX512BW-NEXT: vpaddb %zmm4, %zmm4, %zmm4 {%k1} |
| 433 | ; AVX512BW-NEXT: vporq %zmm1, %zmm4, %zmm1 |
| 434 | ; AVX512BW-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 435 | ; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 436 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 437 | ; AVX512BW-NEXT: retq |
| 438 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 439 | ; AVX512VBMI2-LABEL: var_funnnel_v64i8: |
| 440 | ; AVX512VBMI2: # %bb.0: |
| 441 | ; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 442 | ; AVX512VBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 443 | ; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm5 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 444 | ; AVX512VBMI2-NEXT: vpsubb %zmm4, %zmm5, %zmm5 |
| 445 | ; AVX512VBMI2-NEXT: vpsllw $5, %zmm5, %zmm5 |
| 446 | ; AVX512VBMI2-NEXT: vpaddb %zmm5, %zmm5, %zmm6 |
| 447 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm6, %k1 |
| 448 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm5, %k2 |
| 449 | ; AVX512VBMI2-NEXT: vpsrlw $4, %zmm1, %zmm5 |
| 450 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 451 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm5, %zmm1 {%k2} |
| 452 | ; AVX512VBMI2-NEXT: vpsrlw $2, %zmm1, %zmm5 |
| 453 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 454 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 455 | ; AVX512VBMI2-NEXT: vpsrlw $1, %zmm1, %zmm5 |
| 456 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 457 | ; AVX512VBMI2-NEXT: vpaddb %zmm6, %zmm6, %zmm6 |
| 458 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm6, %k1 |
| 459 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 460 | ; AVX512VBMI2-NEXT: vpsllw $5, %zmm4, %zmm4 |
| 461 | ; AVX512VBMI2-NEXT: vpaddb %zmm4, %zmm4, %zmm5 |
| 462 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm5, %k1 |
| 463 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm4, %k2 |
| 464 | ; AVX512VBMI2-NEXT: vpsllw $4, %zmm0, %zmm4 |
| 465 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 466 | ; AVX512VBMI2-NEXT: vpblendmb %zmm4, %zmm0, %zmm4 {%k2} |
| 467 | ; AVX512VBMI2-NEXT: vpsllw $2, %zmm4, %zmm6 |
| 468 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm6, %zmm6 |
| 469 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm6, %zmm4 {%k1} |
| 470 | ; AVX512VBMI2-NEXT: vpaddb %zmm5, %zmm5, %zmm5 |
| 471 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm5, %k1 |
| 472 | ; AVX512VBMI2-NEXT: vpaddb %zmm4, %zmm4, %zmm4 {%k1} |
| 473 | ; AVX512VBMI2-NEXT: vporq %zmm1, %zmm4, %zmm1 |
| 474 | ; AVX512VBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 475 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 476 | ; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 477 | ; AVX512VBMI2-NEXT: retq |
| 478 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 479 | ; AVX512VLBW-LABEL: var_funnnel_v64i8: |
| 480 | ; AVX512VLBW: # %bb.0: |
| 481 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 482 | ; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 483 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 484 | ; AVX512VLBW-NEXT: vpsubb %zmm4, %zmm5, %zmm5 |
| 485 | ; AVX512VLBW-NEXT: vpsllw $5, %zmm5, %zmm5 |
| 486 | ; AVX512VLBW-NEXT: vpaddb %zmm5, %zmm5, %zmm6 |
| 487 | ; AVX512VLBW-NEXT: vpmovb2m %zmm6, %k1 |
| 488 | ; AVX512VLBW-NEXT: vpmovb2m %zmm5, %k2 |
| 489 | ; AVX512VLBW-NEXT: vpsrlw $4, %zmm1, %zmm5 |
| 490 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 491 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm5, %zmm1 {%k2} |
| 492 | ; AVX512VLBW-NEXT: vpsrlw $2, %zmm1, %zmm5 |
| 493 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 494 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 495 | ; AVX512VLBW-NEXT: vpsrlw $1, %zmm1, %zmm5 |
| 496 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 497 | ; AVX512VLBW-NEXT: vpaddb %zmm6, %zmm6, %zmm6 |
| 498 | ; AVX512VLBW-NEXT: vpmovb2m %zmm6, %k1 |
| 499 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 500 | ; AVX512VLBW-NEXT: vpsllw $5, %zmm4, %zmm4 |
| 501 | ; AVX512VLBW-NEXT: vpaddb %zmm4, %zmm4, %zmm5 |
| 502 | ; AVX512VLBW-NEXT: vpmovb2m %zmm5, %k1 |
| 503 | ; AVX512VLBW-NEXT: vpmovb2m %zmm4, %k2 |
| 504 | ; AVX512VLBW-NEXT: vpsllw $4, %zmm0, %zmm4 |
| 505 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 506 | ; AVX512VLBW-NEXT: vpblendmb %zmm4, %zmm0, %zmm4 {%k2} |
| 507 | ; AVX512VLBW-NEXT: vpsllw $2, %zmm4, %zmm6 |
| 508 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm6, %zmm6 |
| 509 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm6, %zmm4 {%k1} |
| 510 | ; AVX512VLBW-NEXT: vpaddb %zmm5, %zmm5, %zmm5 |
| 511 | ; AVX512VLBW-NEXT: vpmovb2m %zmm5, %k1 |
| 512 | ; AVX512VLBW-NEXT: vpaddb %zmm4, %zmm4, %zmm4 {%k1} |
| 513 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm4, %zmm1 |
| 514 | ; AVX512VLBW-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 515 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 516 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 517 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 518 | ; |
| 519 | ; AVX512VLVBMI2-LABEL: var_funnnel_v64i8: |
| 520 | ; AVX512VLVBMI2: # %bb.0: |
| 521 | ; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 522 | ; AVX512VLVBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 523 | ; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm5 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 524 | ; AVX512VLVBMI2-NEXT: vpsubb %zmm4, %zmm5, %zmm5 |
| 525 | ; AVX512VLVBMI2-NEXT: vpsllw $5, %zmm5, %zmm5 |
| 526 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm5, %zmm5, %zmm6 |
| 527 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm6, %k1 |
| 528 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm5, %k2 |
| 529 | ; AVX512VLVBMI2-NEXT: vpsrlw $4, %zmm1, %zmm5 |
| 530 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 531 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm5, %zmm1 {%k2} |
| 532 | ; AVX512VLVBMI2-NEXT: vpsrlw $2, %zmm1, %zmm5 |
| 533 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 534 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 535 | ; AVX512VLVBMI2-NEXT: vpsrlw $1, %zmm1, %zmm5 |
| 536 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5 |
| 537 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm6, %zmm6, %zmm6 |
| 538 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm6, %k1 |
| 539 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1} |
| 540 | ; AVX512VLVBMI2-NEXT: vpsllw $5, %zmm4, %zmm4 |
| 541 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm4, %zmm4, %zmm5 |
| 542 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm5, %k1 |
| 543 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm4, %k2 |
| 544 | ; AVX512VLVBMI2-NEXT: vpsllw $4, %zmm0, %zmm4 |
| 545 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 546 | ; AVX512VLVBMI2-NEXT: vpblendmb %zmm4, %zmm0, %zmm4 {%k2} |
| 547 | ; AVX512VLVBMI2-NEXT: vpsllw $2, %zmm4, %zmm6 |
| 548 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm6, %zmm6 |
| 549 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm6, %zmm4 {%k1} |
| 550 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm5, %zmm5, %zmm5 |
| 551 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm5, %k1 |
| 552 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm4, %zmm4, %zmm4 {%k1} |
| 553 | ; AVX512VLVBMI2-NEXT: vporq %zmm1, %zmm4, %zmm1 |
| 554 | ; AVX512VLVBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 555 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 556 | ; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 557 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 558 | %res = call <64 x i8> @llvm.fshl.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %amt) |
| 559 | ret <64 x i8> %res |
| 560 | } |
| 561 | |
| 562 | ; |
| 563 | ; Uniform Variable Shifts |
| 564 | ; |
| 565 | |
| 566 | define <8 x i64> @splatvar_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %amt) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 567 | ; AVX512F-LABEL: splatvar_funnnel_v8i64: |
| 568 | ; AVX512F: # %bb.0: |
| 569 | ; AVX512F-NEXT: vpbroadcastq %xmm2, %zmm2 |
| 570 | ; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 571 | ; AVX512F-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 572 | ; AVX512F-NEXT: vpsllq %xmm4, %zmm0, %zmm5 |
| 573 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64] |
| 574 | ; AVX512F-NEXT: vpsubq %xmm4, %xmm6, %xmm4 |
| 575 | ; AVX512F-NEXT: vpsrlq %xmm4, %zmm1, %zmm1 |
| 576 | ; AVX512F-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 577 | ; AVX512F-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 578 | ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 579 | ; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 580 | ; AVX512F-NEXT: retq |
| 581 | ; |
| 582 | ; AVX512VL-LABEL: splatvar_funnnel_v8i64: |
| 583 | ; AVX512VL: # %bb.0: |
| 584 | ; AVX512VL-NEXT: vpbroadcastq %xmm2, %zmm2 |
| 585 | ; AVX512VL-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 586 | ; AVX512VL-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 587 | ; AVX512VL-NEXT: vpsllq %xmm4, %zmm0, %zmm5 |
| 588 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64] |
| 589 | ; AVX512VL-NEXT: vpsubq %xmm4, %xmm6, %xmm4 |
| 590 | ; AVX512VL-NEXT: vpsrlq %xmm4, %zmm1, %zmm1 |
| 591 | ; AVX512VL-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 592 | ; AVX512VL-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 593 | ; AVX512VL-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 594 | ; AVX512VL-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 595 | ; AVX512VL-NEXT: retq |
| 596 | ; |
| 597 | ; AVX512BW-LABEL: splatvar_funnnel_v8i64: |
| 598 | ; AVX512BW: # %bb.0: |
| 599 | ; AVX512BW-NEXT: vpbroadcastq %xmm2, %zmm2 |
| 600 | ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 601 | ; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 602 | ; AVX512BW-NEXT: vpsllq %xmm4, %zmm0, %zmm5 |
| 603 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64] |
| 604 | ; AVX512BW-NEXT: vpsubq %xmm4, %xmm6, %xmm4 |
| 605 | ; AVX512BW-NEXT: vpsrlq %xmm4, %zmm1, %zmm1 |
| 606 | ; AVX512BW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 607 | ; AVX512BW-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 608 | ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 609 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 610 | ; AVX512BW-NEXT: retq |
| 611 | ; |
| 612 | ; AVX512VBMI2-LABEL: splatvar_funnnel_v8i64: |
| 613 | ; AVX512VBMI2: # %bb.0: |
| 614 | ; AVX512VBMI2-NEXT: vpbroadcastq %xmm2, %zmm2 |
| 615 | ; AVX512VBMI2-NEXT: vpshldvq %zmm2, %zmm1, %zmm0 |
| 616 | ; AVX512VBMI2-NEXT: retq |
| 617 | ; |
| 618 | ; AVX512VLBW-LABEL: splatvar_funnnel_v8i64: |
| 619 | ; AVX512VLBW: # %bb.0: |
| 620 | ; AVX512VLBW-NEXT: vpbroadcastq %xmm2, %zmm2 |
| 621 | ; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63] |
| 622 | ; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 623 | ; AVX512VLBW-NEXT: vpsllq %xmm4, %zmm0, %zmm5 |
| 624 | ; AVX512VLBW-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64] |
| 625 | ; AVX512VLBW-NEXT: vpsubq %xmm4, %xmm6, %xmm4 |
| 626 | ; AVX512VLBW-NEXT: vpsrlq %xmm4, %zmm1, %zmm1 |
| 627 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 628 | ; AVX512VLBW-NEXT: vptestnmq %zmm3, %zmm2, %k1 |
| 629 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1} |
| 630 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 631 | ; AVX512VLBW-NEXT: retq |
| 632 | ; |
| 633 | ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v8i64: |
| 634 | ; AVX512VLVBMI2: # %bb.0: |
| 635 | ; AVX512VLVBMI2-NEXT: vpbroadcastq %xmm2, %zmm2 |
| 636 | ; AVX512VLVBMI2-NEXT: vpshldvq %zmm2, %zmm1, %zmm0 |
| 637 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 638 | %splat = shufflevector <8 x i64> %amt, <8 x i64> undef, <8 x i32> zeroinitializer |
| 639 | %res = call <8 x i64> @llvm.fshl.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %splat) |
| 640 | ret <8 x i64> %res |
| 641 | } |
| 642 | |
| 643 | define <16 x i32> @splatvar_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %amt) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 644 | ; AVX512F-LABEL: splatvar_funnnel_v16i32: |
| 645 | ; AVX512F: # %bb.0: |
| 646 | ; AVX512F-NEXT: vpbroadcastd %xmm2, %zmm2 |
| 647 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 648 | ; AVX512F-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 649 | ; AVX512F-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero |
| 650 | ; AVX512F-NEXT: vpslld %xmm5, %zmm0, %zmm5 |
| 651 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32] |
| 652 | ; AVX512F-NEXT: vpsubd %xmm4, %xmm6, %xmm4 |
| 653 | ; AVX512F-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero |
| 654 | ; AVX512F-NEXT: vpsrld %xmm4, %zmm1, %zmm1 |
| 655 | ; AVX512F-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 656 | ; AVX512F-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 657 | ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 658 | ; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 659 | ; AVX512F-NEXT: retq |
| 660 | ; |
| 661 | ; AVX512VL-LABEL: splatvar_funnnel_v16i32: |
| 662 | ; AVX512VL: # %bb.0: |
| 663 | ; AVX512VL-NEXT: vpbroadcastd %xmm2, %zmm2 |
| 664 | ; AVX512VL-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 665 | ; AVX512VL-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 666 | ; AVX512VL-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero |
| 667 | ; AVX512VL-NEXT: vpslld %xmm5, %zmm0, %zmm5 |
| 668 | ; AVX512VL-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32] |
| 669 | ; AVX512VL-NEXT: vpsubd %xmm4, %xmm6, %xmm4 |
| 670 | ; AVX512VL-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero |
| 671 | ; AVX512VL-NEXT: vpsrld %xmm4, %zmm1, %zmm1 |
| 672 | ; AVX512VL-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 673 | ; AVX512VL-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 674 | ; AVX512VL-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 675 | ; AVX512VL-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 676 | ; AVX512VL-NEXT: retq |
| 677 | ; |
| 678 | ; AVX512BW-LABEL: splatvar_funnnel_v16i32: |
| 679 | ; AVX512BW: # %bb.0: |
| 680 | ; AVX512BW-NEXT: vpbroadcastd %xmm2, %zmm2 |
| 681 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 682 | ; AVX512BW-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 683 | ; AVX512BW-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero |
| 684 | ; AVX512BW-NEXT: vpslld %xmm5, %zmm0, %zmm5 |
| 685 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32] |
| 686 | ; AVX512BW-NEXT: vpsubd %xmm4, %xmm6, %xmm4 |
| 687 | ; AVX512BW-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero |
| 688 | ; AVX512BW-NEXT: vpsrld %xmm4, %zmm1, %zmm1 |
| 689 | ; AVX512BW-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 690 | ; AVX512BW-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 691 | ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 692 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 693 | ; AVX512BW-NEXT: retq |
| 694 | ; |
| 695 | ; AVX512VBMI2-LABEL: splatvar_funnnel_v16i32: |
| 696 | ; AVX512VBMI2: # %bb.0: |
| 697 | ; AVX512VBMI2-NEXT: vpbroadcastd %xmm2, %zmm2 |
| 698 | ; AVX512VBMI2-NEXT: vpshldvd %zmm2, %zmm1, %zmm0 |
| 699 | ; AVX512VBMI2-NEXT: retq |
| 700 | ; |
| 701 | ; AVX512VLBW-LABEL: splatvar_funnnel_v16i32: |
| 702 | ; AVX512VLBW: # %bb.0: |
| 703 | ; AVX512VLBW-NEXT: vpbroadcastd %xmm2, %zmm2 |
| 704 | ; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31] |
| 705 | ; AVX512VLBW-NEXT: vpandd %zmm3, %zmm2, %zmm4 |
| 706 | ; AVX512VLBW-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero |
| 707 | ; AVX512VLBW-NEXT: vpslld %xmm5, %zmm0, %zmm5 |
| 708 | ; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32] |
| 709 | ; AVX512VLBW-NEXT: vpsubd %xmm4, %xmm6, %xmm4 |
| 710 | ; AVX512VLBW-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero |
| 711 | ; AVX512VLBW-NEXT: vpsrld %xmm4, %zmm1, %zmm1 |
| 712 | ; AVX512VLBW-NEXT: vpord %zmm1, %zmm5, %zmm1 |
| 713 | ; AVX512VLBW-NEXT: vptestnmd %zmm3, %zmm2, %k1 |
| 714 | ; AVX512VLBW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| 715 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 716 | ; AVX512VLBW-NEXT: retq |
| 717 | ; |
| 718 | ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v16i32: |
| 719 | ; AVX512VLVBMI2: # %bb.0: |
| 720 | ; AVX512VLVBMI2-NEXT: vpbroadcastd %xmm2, %zmm2 |
| 721 | ; AVX512VLVBMI2-NEXT: vpshldvd %zmm2, %zmm1, %zmm0 |
| 722 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 723 | %splat = shufflevector <16 x i32> %amt, <16 x i32> undef, <16 x i32> zeroinitializer |
| 724 | %res = call <16 x i32> @llvm.fshl.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %splat) |
| 725 | ret <16 x i32> %res |
| 726 | } |
| 727 | |
| 728 | define <32 x i16> @splatvar_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %amt) nounwind { |
| 729 | ; AVX512F-LABEL: splatvar_funnnel_v32i16: |
| 730 | ; AVX512F: # %bb.0: |
| 731 | ; AVX512F-NEXT: vpbroadcastw %xmm4, %ymm4 |
| 732 | ; AVX512F-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4 |
| 733 | ; AVX512F-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero |
| 734 | ; AVX512F-NEXT: vpsllw %xmm5, %ymm0, %ymm6 |
| 735 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm7 = [16,16,16,16,16,16,16,16] |
| 736 | ; AVX512F-NEXT: vpsubw %xmm4, %xmm7, %xmm7 |
| 737 | ; AVX512F-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero |
| 738 | ; AVX512F-NEXT: vpsrlw %xmm7, %ymm2, %ymm2 |
| 739 | ; AVX512F-NEXT: vpor %ymm2, %ymm6, %ymm2 |
| 740 | ; AVX512F-NEXT: vpxor %xmm6, %xmm6, %xmm6 |
| 741 | ; AVX512F-NEXT: vpcmpeqw %ymm6, %ymm4, %ymm4 |
| 742 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 743 | ; AVX512F-NEXT: vpsllw %xmm5, %ymm1, %ymm2 |
| 744 | ; AVX512F-NEXT: vpsrlw %xmm7, %ymm3, %ymm3 |
| 745 | ; AVX512F-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 746 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm1, %ymm2, %ymm1 |
| 747 | ; AVX512F-NEXT: retq |
| 748 | ; |
| 749 | ; AVX512VL-LABEL: splatvar_funnnel_v32i16: |
| 750 | ; AVX512VL: # %bb.0: |
| 751 | ; AVX512VL-NEXT: vpbroadcastw %xmm4, %ymm4 |
| 752 | ; AVX512VL-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4 |
| 753 | ; AVX512VL-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero |
| 754 | ; AVX512VL-NEXT: vpsllw %xmm5, %ymm0, %ymm6 |
| 755 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm7 = [16,16,16,16,16,16,16,16] |
| 756 | ; AVX512VL-NEXT: vpsubw %xmm4, %xmm7, %xmm7 |
| 757 | ; AVX512VL-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero |
| 758 | ; AVX512VL-NEXT: vpsrlw %xmm7, %ymm2, %ymm2 |
| 759 | ; AVX512VL-NEXT: vpor %ymm2, %ymm6, %ymm2 |
| 760 | ; AVX512VL-NEXT: vpxor %xmm6, %xmm6, %xmm6 |
| 761 | ; AVX512VL-NEXT: vpcmpeqw %ymm6, %ymm4, %ymm4 |
| 762 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 763 | ; AVX512VL-NEXT: vpsllw %xmm5, %ymm1, %ymm2 |
| 764 | ; AVX512VL-NEXT: vpsrlw %xmm7, %ymm3, %ymm3 |
| 765 | ; AVX512VL-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 766 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm1, %ymm2, %ymm1 |
| 767 | ; AVX512VL-NEXT: retq |
| 768 | ; |
| 769 | ; AVX512BW-LABEL: splatvar_funnnel_v32i16: |
| 770 | ; AVX512BW: # %bb.0: |
| 771 | ; AVX512BW-NEXT: vpbroadcastw %xmm2, %zmm2 |
| 772 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 773 | ; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 774 | ; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero |
| 775 | ; AVX512BW-NEXT: vpsllw %xmm5, %zmm0, %zmm5 |
| 776 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [16,16,16,16,16,16,16,16] |
| 777 | ; AVX512BW-NEXT: vpsubw %xmm4, %xmm6, %xmm4 |
| 778 | ; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero |
| 779 | ; AVX512BW-NEXT: vpsrlw %xmm4, %zmm1, %zmm1 |
| 780 | ; AVX512BW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 781 | ; AVX512BW-NEXT: vptestnmw %zmm3, %zmm2, %k1 |
| 782 | ; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm1 {%k1} |
| 783 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 784 | ; AVX512BW-NEXT: retq |
| 785 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 786 | ; AVX512VBMI2-LABEL: splatvar_funnnel_v32i16: |
| 787 | ; AVX512VBMI2: # %bb.0: |
| 788 | ; AVX512VBMI2-NEXT: vpbroadcastw %xmm2, %zmm2 |
| 789 | ; AVX512VBMI2-NEXT: vpshldvw %zmm2, %zmm1, %zmm0 |
| 790 | ; AVX512VBMI2-NEXT: retq |
| 791 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 792 | ; AVX512VLBW-LABEL: splatvar_funnnel_v32i16: |
| 793 | ; AVX512VLBW: # %bb.0: |
| 794 | ; AVX512VLBW-NEXT: vpbroadcastw %xmm2, %zmm2 |
| 795 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] |
| 796 | ; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 797 | ; AVX512VLBW-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero |
| 798 | ; AVX512VLBW-NEXT: vpsllw %xmm5, %zmm0, %zmm5 |
| 799 | ; AVX512VLBW-NEXT: vmovdqa {{.*#+}} xmm6 = [16,16,16,16,16,16,16,16] |
| 800 | ; AVX512VLBW-NEXT: vpsubw %xmm4, %xmm6, %xmm4 |
| 801 | ; AVX512VLBW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero |
| 802 | ; AVX512VLBW-NEXT: vpsrlw %xmm4, %zmm1, %zmm1 |
| 803 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 804 | ; AVX512VLBW-NEXT: vptestnmw %zmm3, %zmm2, %k1 |
| 805 | ; AVX512VLBW-NEXT: vmovdqu16 %zmm0, %zmm1 {%k1} |
| 806 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 807 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 808 | ; |
| 809 | ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v32i16: |
| 810 | ; AVX512VLVBMI2: # %bb.0: |
| 811 | ; AVX512VLVBMI2-NEXT: vpbroadcastw %xmm2, %zmm2 |
| 812 | ; AVX512VLVBMI2-NEXT: vpshldvw %zmm2, %zmm1, %zmm0 |
| 813 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 814 | %splat = shufflevector <32 x i16> %amt, <32 x i16> undef, <32 x i32> zeroinitializer |
| 815 | %res = call <32 x i16> @llvm.fshl.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %splat) |
| 816 | ret <32 x i16> %res |
| 817 | } |
| 818 | |
| 819 | define <64 x i8> @splatvar_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %amt) nounwind { |
| 820 | ; AVX512F-LABEL: splatvar_funnnel_v64i8: |
| 821 | ; AVX512F: # %bb.0: |
| 822 | ; AVX512F-NEXT: vpbroadcastb %xmm4, %ymm4 |
| 823 | ; AVX512F-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4 |
| 824 | ; AVX512F-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 825 | ; AVX512F-NEXT: vpsllw %xmm5, %ymm0, %ymm6 |
| 826 | ; AVX512F-NEXT: vpcmpeqd %ymm9, %ymm9, %ymm9 |
| 827 | ; AVX512F-NEXT: vpsllw %xmm5, %ymm9, %ymm8 |
| 828 | ; AVX512F-NEXT: vpbroadcastb %xmm8, %ymm8 |
| 829 | ; AVX512F-NEXT: vpand %ymm8, %ymm6, %ymm6 |
| 830 | ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm7 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 831 | ; AVX512F-NEXT: vpsubb %xmm4, %xmm7, %xmm7 |
| 832 | ; AVX512F-NEXT: vpmovzxbq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,zero,zero,zero,zero,xmm7[1],zero,zero,zero,zero,zero,zero,zero |
| 833 | ; AVX512F-NEXT: vpsrlw %xmm7, %ymm2, %ymm2 |
| 834 | ; AVX512F-NEXT: vpsrlw %xmm7, %ymm9, %ymm9 |
| 835 | ; AVX512F-NEXT: vpsrlw $8, %ymm9, %ymm9 |
| 836 | ; AVX512F-NEXT: vpbroadcastb %xmm9, %ymm9 |
| 837 | ; AVX512F-NEXT: vpand %ymm9, %ymm2, %ymm2 |
| 838 | ; AVX512F-NEXT: vpor %ymm2, %ymm6, %ymm2 |
| 839 | ; AVX512F-NEXT: vpxor %xmm6, %xmm6, %xmm6 |
| 840 | ; AVX512F-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm4 |
| 841 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 842 | ; AVX512F-NEXT: vpsllw %xmm5, %ymm1, %ymm2 |
| 843 | ; AVX512F-NEXT: vpand %ymm8, %ymm2, %ymm2 |
| 844 | ; AVX512F-NEXT: vpsrlw %xmm7, %ymm3, %ymm3 |
| 845 | ; AVX512F-NEXT: vpand %ymm9, %ymm3, %ymm3 |
| 846 | ; AVX512F-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 847 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm1, %ymm2, %ymm1 |
| 848 | ; AVX512F-NEXT: retq |
| 849 | ; |
| 850 | ; AVX512VL-LABEL: splatvar_funnnel_v64i8: |
| 851 | ; AVX512VL: # %bb.0: |
| 852 | ; AVX512VL-NEXT: vpbroadcastb %xmm4, %ymm4 |
| 853 | ; AVX512VL-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4 |
| 854 | ; AVX512VL-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 855 | ; AVX512VL-NEXT: vpsllw %xmm5, %ymm0, %ymm6 |
| 856 | ; AVX512VL-NEXT: vpcmpeqd %ymm9, %ymm9, %ymm9 |
| 857 | ; AVX512VL-NEXT: vpsllw %xmm5, %ymm9, %ymm8 |
| 858 | ; AVX512VL-NEXT: vpbroadcastb %xmm8, %ymm8 |
| 859 | ; AVX512VL-NEXT: vpand %ymm8, %ymm6, %ymm6 |
| 860 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm7 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 861 | ; AVX512VL-NEXT: vpsubb %xmm4, %xmm7, %xmm7 |
| 862 | ; AVX512VL-NEXT: vpmovzxbq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,zero,zero,zero,zero,xmm7[1],zero,zero,zero,zero,zero,zero,zero |
| 863 | ; AVX512VL-NEXT: vpsrlw %xmm7, %ymm2, %ymm2 |
| 864 | ; AVX512VL-NEXT: vpsrlw %xmm7, %ymm9, %ymm9 |
| 865 | ; AVX512VL-NEXT: vpsrlw $8, %ymm9, %ymm9 |
| 866 | ; AVX512VL-NEXT: vpbroadcastb %xmm9, %ymm9 |
| 867 | ; AVX512VL-NEXT: vpand %ymm9, %ymm2, %ymm2 |
| 868 | ; AVX512VL-NEXT: vpor %ymm2, %ymm6, %ymm2 |
| 869 | ; AVX512VL-NEXT: vpxor %xmm6, %xmm6, %xmm6 |
| 870 | ; AVX512VL-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm4 |
| 871 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0 |
| 872 | ; AVX512VL-NEXT: vpsllw %xmm5, %ymm1, %ymm2 |
| 873 | ; AVX512VL-NEXT: vpand %ymm8, %ymm2, %ymm2 |
| 874 | ; AVX512VL-NEXT: vpsrlw %xmm7, %ymm3, %ymm3 |
| 875 | ; AVX512VL-NEXT: vpand %ymm9, %ymm3, %ymm3 |
| 876 | ; AVX512VL-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 877 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm1, %ymm2, %ymm1 |
| 878 | ; AVX512VL-NEXT: retq |
| 879 | ; |
| 880 | ; AVX512BW-LABEL: splatvar_funnnel_v64i8: |
| 881 | ; AVX512BW: # %bb.0: |
| 882 | ; AVX512BW-NEXT: vpbroadcastb %xmm2, %zmm2 |
| 883 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 884 | ; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 885 | ; AVX512BW-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 886 | ; AVX512BW-NEXT: vpsllw %xmm5, %zmm0, %zmm6 |
| 887 | ; AVX512BW-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7 |
| 888 | ; AVX512BW-NEXT: vpsllw %xmm5, %zmm7, %zmm5 |
| 889 | ; AVX512BW-NEXT: vpbroadcastb %xmm5, %zmm5 |
| 890 | ; AVX512BW-NEXT: vpandq %zmm5, %zmm6, %zmm5 |
| 891 | ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 892 | ; AVX512BW-NEXT: vpsubb %xmm4, %xmm6, %xmm4 |
| 893 | ; AVX512BW-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 894 | ; AVX512BW-NEXT: vpsrlw %xmm4, %zmm1, %zmm1 |
| 895 | ; AVX512BW-NEXT: vpsrlw %xmm4, %zmm7, %zmm4 |
| 896 | ; AVX512BW-NEXT: vpsrlw $8, %zmm4, %zmm4 |
| 897 | ; AVX512BW-NEXT: vpbroadcastb %xmm4, %zmm4 |
| 898 | ; AVX512BW-NEXT: vpandq %zmm4, %zmm1, %zmm1 |
| 899 | ; AVX512BW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 900 | ; AVX512BW-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 901 | ; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 902 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 903 | ; AVX512BW-NEXT: retq |
| 904 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 905 | ; AVX512VBMI2-LABEL: splatvar_funnnel_v64i8: |
| 906 | ; AVX512VBMI2: # %bb.0: |
| 907 | ; AVX512VBMI2-NEXT: vpbroadcastb %xmm2, %zmm2 |
| 908 | ; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 909 | ; AVX512VBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 910 | ; AVX512VBMI2-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 911 | ; AVX512VBMI2-NEXT: vpsllw %xmm5, %zmm0, %zmm6 |
| 912 | ; AVX512VBMI2-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7 |
| 913 | ; AVX512VBMI2-NEXT: vpsllw %xmm5, %zmm7, %zmm5 |
| 914 | ; AVX512VBMI2-NEXT: vpbroadcastb %xmm5, %zmm5 |
| 915 | ; AVX512VBMI2-NEXT: vpandq %zmm5, %zmm6, %zmm5 |
| 916 | ; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 917 | ; AVX512VBMI2-NEXT: vpsubb %xmm4, %xmm6, %xmm4 |
| 918 | ; AVX512VBMI2-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 919 | ; AVX512VBMI2-NEXT: vpsrlw %xmm4, %zmm1, %zmm1 |
| 920 | ; AVX512VBMI2-NEXT: vpsrlw %xmm4, %zmm7, %zmm4 |
| 921 | ; AVX512VBMI2-NEXT: vpsrlw $8, %zmm4, %zmm4 |
| 922 | ; AVX512VBMI2-NEXT: vpbroadcastb %xmm4, %zmm4 |
| 923 | ; AVX512VBMI2-NEXT: vpandq %zmm4, %zmm1, %zmm1 |
| 924 | ; AVX512VBMI2-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 925 | ; AVX512VBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 926 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 927 | ; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 928 | ; AVX512VBMI2-NEXT: retq |
| 929 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 930 | ; AVX512VLBW-LABEL: splatvar_funnnel_v64i8: |
| 931 | ; AVX512VLBW: # %bb.0: |
| 932 | ; AVX512VLBW-NEXT: vpbroadcastb %xmm2, %zmm2 |
| 933 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 934 | ; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 935 | ; AVX512VLBW-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 936 | ; AVX512VLBW-NEXT: vpsllw %xmm5, %zmm0, %zmm6 |
| 937 | ; AVX512VLBW-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7 |
| 938 | ; AVX512VLBW-NEXT: vpsllw %xmm5, %zmm7, %zmm5 |
| 939 | ; AVX512VLBW-NEXT: vpbroadcastb %xmm5, %zmm5 |
| 940 | ; AVX512VLBW-NEXT: vpandq %zmm5, %zmm6, %zmm5 |
| 941 | ; AVX512VLBW-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 942 | ; AVX512VLBW-NEXT: vpsubb %xmm4, %xmm6, %xmm4 |
| 943 | ; AVX512VLBW-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 944 | ; AVX512VLBW-NEXT: vpsrlw %xmm4, %zmm1, %zmm1 |
| 945 | ; AVX512VLBW-NEXT: vpsrlw %xmm4, %zmm7, %zmm4 |
| 946 | ; AVX512VLBW-NEXT: vpsrlw $8, %zmm4, %zmm4 |
| 947 | ; AVX512VLBW-NEXT: vpbroadcastb %xmm4, %zmm4 |
| 948 | ; AVX512VLBW-NEXT: vpandq %zmm4, %zmm1, %zmm1 |
| 949 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 950 | ; AVX512VLBW-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 951 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 952 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 953 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 954 | ; |
| 955 | ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v64i8: |
| 956 | ; AVX512VLVBMI2: # %bb.0: |
| 957 | ; AVX512VLVBMI2-NEXT: vpbroadcastb %xmm2, %zmm2 |
| 958 | ; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7] |
| 959 | ; AVX512VLVBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4 |
| 960 | ; AVX512VLVBMI2-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 961 | ; AVX512VLVBMI2-NEXT: vpsllw %xmm5, %zmm0, %zmm6 |
| 962 | ; AVX512VLVBMI2-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7 |
| 963 | ; AVX512VLVBMI2-NEXT: vpsllw %xmm5, %zmm7, %zmm5 |
| 964 | ; AVX512VLVBMI2-NEXT: vpbroadcastb %xmm5, %zmm5 |
| 965 | ; AVX512VLVBMI2-NEXT: vpandq %zmm5, %zmm6, %zmm5 |
| 966 | ; AVX512VLVBMI2-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8] |
| 967 | ; AVX512VLVBMI2-NEXT: vpsubb %xmm4, %xmm6, %xmm4 |
| 968 | ; AVX512VLVBMI2-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero |
| 969 | ; AVX512VLVBMI2-NEXT: vpsrlw %xmm4, %zmm1, %zmm1 |
| 970 | ; AVX512VLVBMI2-NEXT: vpsrlw %xmm4, %zmm7, %zmm4 |
| 971 | ; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm4, %zmm4 |
| 972 | ; AVX512VLVBMI2-NEXT: vpbroadcastb %xmm4, %zmm4 |
| 973 | ; AVX512VLVBMI2-NEXT: vpandq %zmm4, %zmm1, %zmm1 |
| 974 | ; AVX512VLVBMI2-NEXT: vporq %zmm1, %zmm5, %zmm1 |
| 975 | ; AVX512VLVBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1 |
| 976 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 977 | ; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 978 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 979 | %splat = shufflevector <64 x i8> %amt, <64 x i8> undef, <64 x i32> zeroinitializer |
| 980 | %res = call <64 x i8> @llvm.fshl.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %splat) |
| 981 | ret <64 x i8> %res |
| 982 | } |
| 983 | |
| 984 | ; |
| 985 | ; Constant Shifts |
| 986 | ; |
| 987 | |
| 988 | define <8 x i64> @constant_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 989 | ; AVX512F-LABEL: constant_funnnel_v8i64: |
| 990 | ; AVX512F: # %bb.0: |
| 991 | ; AVX512F-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1 |
| 992 | ; AVX512F-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0 |
| 993 | ; AVX512F-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 994 | ; AVX512F-NEXT: retq |
| 995 | ; |
| 996 | ; AVX512VL-LABEL: constant_funnnel_v8i64: |
| 997 | ; AVX512VL: # %bb.0: |
| 998 | ; AVX512VL-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1 |
| 999 | ; AVX512VL-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0 |
| 1000 | ; AVX512VL-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1001 | ; AVX512VL-NEXT: retq |
| 1002 | ; |
| 1003 | ; AVX512BW-LABEL: constant_funnnel_v8i64: |
| 1004 | ; AVX512BW: # %bb.0: |
| 1005 | ; AVX512BW-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1 |
| 1006 | ; AVX512BW-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0 |
| 1007 | ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1008 | ; AVX512BW-NEXT: retq |
| 1009 | ; |
| 1010 | ; AVX512VBMI2-LABEL: constant_funnnel_v8i64: |
| 1011 | ; AVX512VBMI2: # %bb.0: |
| 1012 | ; AVX512VBMI2-NEXT: vpshldvq {{.*}}(%rip), %zmm1, %zmm0 |
| 1013 | ; AVX512VBMI2-NEXT: retq |
| 1014 | ; |
| 1015 | ; AVX512VLBW-LABEL: constant_funnnel_v8i64: |
| 1016 | ; AVX512VLBW: # %bb.0: |
| 1017 | ; AVX512VLBW-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1 |
| 1018 | ; AVX512VLBW-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0 |
| 1019 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1020 | ; AVX512VLBW-NEXT: retq |
| 1021 | ; |
| 1022 | ; AVX512VLVBMI2-LABEL: constant_funnnel_v8i64: |
| 1023 | ; AVX512VLVBMI2: # %bb.0: |
| 1024 | ; AVX512VLVBMI2-NEXT: vpshldvq {{.*}}(%rip), %zmm1, %zmm0 |
| 1025 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1026 | %res = call <8 x i64> @llvm.fshl.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> <i64 4, i64 14, i64 50, i64 60, i64 4, i64 14, i64 50, i64 60>) |
| 1027 | ret <8 x i64> %res |
| 1028 | } |
| 1029 | |
| 1030 | define <16 x i32> @constant_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1031 | ; AVX512F-LABEL: constant_funnnel_v16i32: |
| 1032 | ; AVX512F: # %bb.0: |
| 1033 | ; AVX512F-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1 |
| 1034 | ; AVX512F-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0 |
| 1035 | ; AVX512F-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1036 | ; AVX512F-NEXT: retq |
| 1037 | ; |
| 1038 | ; AVX512VL-LABEL: constant_funnnel_v16i32: |
| 1039 | ; AVX512VL: # %bb.0: |
| 1040 | ; AVX512VL-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1 |
| 1041 | ; AVX512VL-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0 |
| 1042 | ; AVX512VL-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1043 | ; AVX512VL-NEXT: retq |
| 1044 | ; |
| 1045 | ; AVX512BW-LABEL: constant_funnnel_v16i32: |
| 1046 | ; AVX512BW: # %bb.0: |
| 1047 | ; AVX512BW-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1 |
| 1048 | ; AVX512BW-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0 |
| 1049 | ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1050 | ; AVX512BW-NEXT: retq |
| 1051 | ; |
| 1052 | ; AVX512VBMI2-LABEL: constant_funnnel_v16i32: |
| 1053 | ; AVX512VBMI2: # %bb.0: |
| 1054 | ; AVX512VBMI2-NEXT: vpshldvd {{.*}}(%rip), %zmm1, %zmm0 |
| 1055 | ; AVX512VBMI2-NEXT: retq |
| 1056 | ; |
| 1057 | ; AVX512VLBW-LABEL: constant_funnnel_v16i32: |
| 1058 | ; AVX512VLBW: # %bb.0: |
| 1059 | ; AVX512VLBW-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1 |
| 1060 | ; AVX512VLBW-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0 |
| 1061 | ; AVX512VLBW-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1062 | ; AVX512VLBW-NEXT: retq |
| 1063 | ; |
| 1064 | ; AVX512VLVBMI2-LABEL: constant_funnnel_v16i32: |
| 1065 | ; AVX512VLVBMI2: # %bb.0: |
| 1066 | ; AVX512VLVBMI2-NEXT: vpshldvd {{.*}}(%rip), %zmm1, %zmm0 |
| 1067 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1068 | %res = call <16 x i32> @llvm.fshl.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>) |
| 1069 | ret <16 x i32> %res |
| 1070 | } |
| 1071 | |
| 1072 | define <32 x i16> @constant_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y) nounwind { |
| 1073 | ; AVX512F-LABEL: constant_funnnel_v32i16: |
| 1074 | ; AVX512F: # %bb.0: |
| 1075 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = [1,2,4,8,16,32,64,128,256,512,1024,2048,4096,8192,16384,32768] |
| 1076 | ; AVX512F-NEXT: vpmulhuw %ymm4, %ymm2, %ymm2 |
| 1077 | ; AVX512F-NEXT: vpmullw %ymm4, %ymm0, %ymm5 |
| 1078 | ; AVX512F-NEXT: vpor %ymm2, %ymm5, %ymm2 |
| 1079 | ; AVX512F-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3,4,5,6,7],ymm0[8],ymm2[9,10,11,12,13,14,15] |
| 1080 | ; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7] |
| 1081 | ; AVX512F-NEXT: vpmulhuw %ymm4, %ymm3, %ymm2 |
| 1082 | ; AVX512F-NEXT: vpmullw %ymm4, %ymm1, %ymm3 |
| 1083 | ; AVX512F-NEXT: vpor %ymm2, %ymm3, %ymm2 |
| 1084 | ; AVX512F-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3,4,5,6,7],ymm1[8],ymm2[9,10,11,12,13,14,15] |
| 1085 | ; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| 1086 | ; AVX512F-NEXT: retq |
| 1087 | ; |
| 1088 | ; AVX512VL-LABEL: constant_funnnel_v32i16: |
| 1089 | ; AVX512VL: # %bb.0: |
| 1090 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm4 = [1,2,4,8,16,32,64,128,256,512,1024,2048,4096,8192,16384,32768] |
| 1091 | ; AVX512VL-NEXT: vpmulhuw %ymm4, %ymm2, %ymm2 |
| 1092 | ; AVX512VL-NEXT: vpmullw %ymm4, %ymm0, %ymm5 |
| 1093 | ; AVX512VL-NEXT: vpor %ymm2, %ymm5, %ymm2 |
| 1094 | ; AVX512VL-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3,4,5,6,7],ymm0[8],ymm2[9,10,11,12,13,14,15] |
| 1095 | ; AVX512VL-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7] |
| 1096 | ; AVX512VL-NEXT: vpmulhuw %ymm4, %ymm3, %ymm2 |
| 1097 | ; AVX512VL-NEXT: vpmullw %ymm4, %ymm1, %ymm3 |
| 1098 | ; AVX512VL-NEXT: vpor %ymm2, %ymm3, %ymm2 |
| 1099 | ; AVX512VL-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3,4,5,6,7],ymm1[8],ymm2[9,10,11,12,13,14,15] |
| 1100 | ; AVX512VL-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| 1101 | ; AVX512VL-NEXT: retq |
| 1102 | ; |
| 1103 | ; AVX512BW-LABEL: constant_funnnel_v32i16: |
| 1104 | ; AVX512BW: # %bb.0: |
| 1105 | ; AVX512BW-NEXT: vpsrlvw {{.*}}(%rip), %zmm1, %zmm1 |
| 1106 | ; AVX512BW-NEXT: vpsllvw {{.*}}(%rip), %zmm0, %zmm2 |
| 1107 | ; AVX512BW-NEXT: vporq %zmm1, %zmm2, %zmm1 |
| 1108 | ; AVX512BW-NEXT: movl $65537, %eax # imm = 0x10001 |
| 1109 | ; AVX512BW-NEXT: kmovd %eax, %k1 |
| 1110 | ; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm1 {%k1} |
| 1111 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 1112 | ; AVX512BW-NEXT: retq |
| 1113 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1114 | ; AVX512VBMI2-LABEL: constant_funnnel_v32i16: |
| 1115 | ; AVX512VBMI2: # %bb.0: |
| 1116 | ; AVX512VBMI2-NEXT: vpshldvw {{.*}}(%rip), %zmm1, %zmm0 |
| 1117 | ; AVX512VBMI2-NEXT: retq |
| 1118 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1119 | ; AVX512VLBW-LABEL: constant_funnnel_v32i16: |
| 1120 | ; AVX512VLBW: # %bb.0: |
| 1121 | ; AVX512VLBW-NEXT: vpsrlvw {{.*}}(%rip), %zmm1, %zmm1 |
| 1122 | ; AVX512VLBW-NEXT: vpsllvw {{.*}}(%rip), %zmm0, %zmm2 |
| 1123 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm2, %zmm1 |
| 1124 | ; AVX512VLBW-NEXT: movl $65537, %eax # imm = 0x10001 |
| 1125 | ; AVX512VLBW-NEXT: kmovd %eax, %k1 |
| 1126 | ; AVX512VLBW-NEXT: vmovdqu16 %zmm0, %zmm1 {%k1} |
| 1127 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 1128 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1129 | ; |
| 1130 | ; AVX512VLVBMI2-LABEL: constant_funnnel_v32i16: |
| 1131 | ; AVX512VLVBMI2: # %bb.0: |
| 1132 | ; AVX512VLVBMI2-NEXT: vpshldvw {{.*}}(%rip), %zmm1, %zmm0 |
| 1133 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1134 | %res = call <32 x i16> @llvm.fshl.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>) |
| 1135 | ret <32 x i16> %res |
| 1136 | } |
| 1137 | |
| 1138 | define <64 x i8> @constant_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y) nounwind { |
| 1139 | ; AVX512F-LABEL: constant_funnnel_v64i8: |
| 1140 | ; AVX512F: # %bb.0: |
| 1141 | ; AVX512F-NEXT: vpsllw $4, %ymm0, %ymm4 |
| 1142 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm5 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240] |
| 1143 | ; AVX512F-NEXT: vpand %ymm5, %ymm4, %ymm4 |
| 1144 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256] |
| 1145 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm0, %ymm4 |
| 1146 | ; AVX512F-NEXT: vpsllw $2, %ymm4, %ymm7 |
| 1147 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm8 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252] |
| 1148 | ; AVX512F-NEXT: vpand %ymm8, %ymm7, %ymm7 |
| 1149 | ; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm9 |
| 1150 | ; AVX512F-NEXT: vpblendvb %ymm9, %ymm7, %ymm4, %ymm4 |
| 1151 | ; AVX512F-NEXT: vpaddb %ymm4, %ymm4, %ymm7 |
| 1152 | ; AVX512F-NEXT: vpaddb %ymm9, %ymm9, %ymm10 |
| 1153 | ; AVX512F-NEXT: vpblendvb %ymm10, %ymm7, %ymm4, %ymm4 |
| 1154 | ; AVX512F-NEXT: vpxor %xmm7, %xmm7, %xmm7 |
| 1155 | ; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm2[8],ymm7[8],ymm2[9],ymm7[9],ymm2[10],ymm7[10],ymm2[11],ymm7[11],ymm2[12],ymm7[12],ymm2[13],ymm7[13],ymm2[14],ymm7[14],ymm2[15],ymm7[15],ymm2[24],ymm7[24],ymm2[25],ymm7[25],ymm2[26],ymm7[26],ymm2[27],ymm7[27],ymm2[28],ymm7[28],ymm2[29],ymm7[29],ymm2[30],ymm7[30],ymm2[31],ymm7[31] |
| 1156 | ; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [1,128,64,32,16,8,4,2,1,128,64,32,16,8,4,2] |
| 1157 | ; AVX512F-NEXT: # ymm12 = mem[0,1,0,1] |
| 1158 | ; AVX512F-NEXT: vpmullw %ymm12, %ymm11, %ymm11 |
| 1159 | ; AVX512F-NEXT: vpsrlw $8, %ymm11, %ymm11 |
| 1160 | ; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm7[0],ymm2[1],ymm7[1],ymm2[2],ymm7[2],ymm2[3],ymm7[3],ymm2[4],ymm7[4],ymm2[5],ymm7[5],ymm2[6],ymm7[6],ymm2[7],ymm7[7],ymm2[16],ymm7[16],ymm2[17],ymm7[17],ymm2[18],ymm7[18],ymm2[19],ymm7[19],ymm2[20],ymm7[20],ymm2[21],ymm7[21],ymm2[22],ymm7[22],ymm2[23],ymm7[23] |
| 1161 | ; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] |
| 1162 | ; AVX512F-NEXT: # ymm13 = mem[0,1,0,1] |
| 1163 | ; AVX512F-NEXT: vpmullw %ymm13, %ymm2, %ymm2 |
| 1164 | ; AVX512F-NEXT: vpsrlw $8, %ymm2, %ymm2 |
| 1165 | ; AVX512F-NEXT: vpackuswb %ymm11, %ymm2, %ymm2 |
| 1166 | ; AVX512F-NEXT: vpor %ymm2, %ymm4, %ymm2 |
| 1167 | ; AVX512F-NEXT: vpbroadcastq {{.*#+}} ymm4 = [18446744073709551360,18446744073709551360,18446744073709551360,18446744073709551360] |
| 1168 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0 |
| 1169 | ; AVX512F-NEXT: vpsllw $4, %ymm1, %ymm2 |
| 1170 | ; AVX512F-NEXT: vpand %ymm5, %ymm2, %ymm2 |
| 1171 | ; AVX512F-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm2 |
| 1172 | ; AVX512F-NEXT: vpsllw $2, %ymm2, %ymm5 |
| 1173 | ; AVX512F-NEXT: vpand %ymm8, %ymm5, %ymm5 |
| 1174 | ; AVX512F-NEXT: vpblendvb %ymm9, %ymm5, %ymm2, %ymm2 |
| 1175 | ; AVX512F-NEXT: vpaddb %ymm2, %ymm2, %ymm5 |
| 1176 | ; AVX512F-NEXT: vpblendvb %ymm10, %ymm5, %ymm2, %ymm2 |
| 1177 | ; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm3[8],ymm7[8],ymm3[9],ymm7[9],ymm3[10],ymm7[10],ymm3[11],ymm7[11],ymm3[12],ymm7[12],ymm3[13],ymm7[13],ymm3[14],ymm7[14],ymm3[15],ymm7[15],ymm3[24],ymm7[24],ymm3[25],ymm7[25],ymm3[26],ymm7[26],ymm3[27],ymm7[27],ymm3[28],ymm7[28],ymm3[29],ymm7[29],ymm3[30],ymm7[30],ymm3[31],ymm7[31] |
| 1178 | ; AVX512F-NEXT: vpmullw %ymm12, %ymm5, %ymm5 |
| 1179 | ; AVX512F-NEXT: vpsrlw $8, %ymm5, %ymm5 |
| 1180 | ; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm7[0],ymm3[1],ymm7[1],ymm3[2],ymm7[2],ymm3[3],ymm7[3],ymm3[4],ymm7[4],ymm3[5],ymm7[5],ymm3[6],ymm7[6],ymm3[7],ymm7[7],ymm3[16],ymm7[16],ymm3[17],ymm7[17],ymm3[18],ymm7[18],ymm3[19],ymm7[19],ymm3[20],ymm7[20],ymm3[21],ymm7[21],ymm3[22],ymm7[22],ymm3[23],ymm7[23] |
| 1181 | ; AVX512F-NEXT: vpmullw %ymm13, %ymm3, %ymm3 |
| 1182 | ; AVX512F-NEXT: vpsrlw $8, %ymm3, %ymm3 |
| 1183 | ; AVX512F-NEXT: vpackuswb %ymm5, %ymm3, %ymm3 |
| 1184 | ; AVX512F-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 1185 | ; AVX512F-NEXT: vpblendvb %ymm4, %ymm2, %ymm1, %ymm1 |
| 1186 | ; AVX512F-NEXT: retq |
| 1187 | ; |
| 1188 | ; AVX512VL-LABEL: constant_funnnel_v64i8: |
| 1189 | ; AVX512VL: # %bb.0: |
| 1190 | ; AVX512VL-NEXT: vpsllw $4, %ymm0, %ymm4 |
| 1191 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm5 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240] |
| 1192 | ; AVX512VL-NEXT: vpand %ymm5, %ymm4, %ymm4 |
| 1193 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm6 = [8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256] |
| 1194 | ; AVX512VL-NEXT: vpblendvb %ymm6, %ymm4, %ymm0, %ymm4 |
| 1195 | ; AVX512VL-NEXT: vpsllw $2, %ymm4, %ymm7 |
| 1196 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm8 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252] |
| 1197 | ; AVX512VL-NEXT: vpand %ymm8, %ymm7, %ymm7 |
| 1198 | ; AVX512VL-NEXT: vpaddb %ymm6, %ymm6, %ymm9 |
| 1199 | ; AVX512VL-NEXT: vpblendvb %ymm9, %ymm7, %ymm4, %ymm4 |
| 1200 | ; AVX512VL-NEXT: vpaddb %ymm4, %ymm4, %ymm7 |
| 1201 | ; AVX512VL-NEXT: vpaddb %ymm9, %ymm9, %ymm10 |
| 1202 | ; AVX512VL-NEXT: vpblendvb %ymm10, %ymm7, %ymm4, %ymm4 |
| 1203 | ; AVX512VL-NEXT: vpunpckhbw {{.*#+}} ymm7 = ymm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31] |
| 1204 | ; AVX512VL-NEXT: vpsrlw $8, %ymm7, %ymm7 |
| 1205 | ; AVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [1,128,64,32,16,8,4,2,1,128,64,32,16,8,4,2] |
| 1206 | ; AVX512VL-NEXT: # ymm11 = mem[0,1,0,1] |
| 1207 | ; AVX512VL-NEXT: vpmullw %ymm11, %ymm7, %ymm7 |
| 1208 | ; AVX512VL-NEXT: vpsrlw $8, %ymm7, %ymm7 |
| 1209 | ; AVX512VL-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23] |
| 1210 | ; AVX512VL-NEXT: vpsrlw $8, %ymm2, %ymm2 |
| 1211 | ; AVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] |
| 1212 | ; AVX512VL-NEXT: # ymm12 = mem[0,1,0,1] |
| 1213 | ; AVX512VL-NEXT: vpmullw %ymm12, %ymm2, %ymm2 |
| 1214 | ; AVX512VL-NEXT: vpsrlw $8, %ymm2, %ymm2 |
| 1215 | ; AVX512VL-NEXT: vpackuswb %ymm7, %ymm2, %ymm2 |
| 1216 | ; AVX512VL-NEXT: vpor %ymm2, %ymm4, %ymm2 |
| 1217 | ; AVX512VL-NEXT: vpbroadcastq {{.*#+}} ymm4 = [18446744073709551360,18446744073709551360,18446744073709551360,18446744073709551360] |
| 1218 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0 |
| 1219 | ; AVX512VL-NEXT: vpsllw $4, %ymm1, %ymm2 |
| 1220 | ; AVX512VL-NEXT: vpand %ymm5, %ymm2, %ymm2 |
| 1221 | ; AVX512VL-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm2 |
| 1222 | ; AVX512VL-NEXT: vpsllw $2, %ymm2, %ymm5 |
| 1223 | ; AVX512VL-NEXT: vpand %ymm8, %ymm5, %ymm5 |
| 1224 | ; AVX512VL-NEXT: vpblendvb %ymm9, %ymm5, %ymm2, %ymm2 |
| 1225 | ; AVX512VL-NEXT: vpaddb %ymm2, %ymm2, %ymm5 |
| 1226 | ; AVX512VL-NEXT: vpblendvb %ymm10, %ymm5, %ymm2, %ymm2 |
| 1227 | ; AVX512VL-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31] |
| 1228 | ; AVX512VL-NEXT: vpsrlw $8, %ymm5, %ymm5 |
| 1229 | ; AVX512VL-NEXT: vpmullw %ymm11, %ymm5, %ymm5 |
| 1230 | ; AVX512VL-NEXT: vpsrlw $8, %ymm5, %ymm5 |
| 1231 | ; AVX512VL-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23] |
| 1232 | ; AVX512VL-NEXT: vpsrlw $8, %ymm3, %ymm3 |
| 1233 | ; AVX512VL-NEXT: vpmullw %ymm12, %ymm3, %ymm3 |
| 1234 | ; AVX512VL-NEXT: vpsrlw $8, %ymm3, %ymm3 |
| 1235 | ; AVX512VL-NEXT: vpackuswb %ymm5, %ymm3, %ymm3 |
| 1236 | ; AVX512VL-NEXT: vpor %ymm3, %ymm2, %ymm2 |
| 1237 | ; AVX512VL-NEXT: vpblendvb %ymm4, %ymm2, %ymm1, %ymm1 |
| 1238 | ; AVX512VL-NEXT: retq |
| 1239 | ; |
| 1240 | ; AVX512BW-LABEL: constant_funnnel_v64i8: |
| 1241 | ; AVX512BW: # %bb.0: |
| 1242 | ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256] |
| 1243 | ; AVX512BW-NEXT: vpmovb2m %zmm2, %k1 |
| 1244 | ; AVX512BW-NEXT: vpsllw $4, %zmm0, %zmm3 |
| 1245 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3 |
| 1246 | ; AVX512BW-NEXT: vpblendmb %zmm3, %zmm0, %zmm3 {%k1} |
| 1247 | ; AVX512BW-NEXT: vpsllw $2, %zmm3, %zmm4 |
| 1248 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 1249 | ; AVX512BW-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1250 | ; AVX512BW-NEXT: vpmovb2m %zmm2, %k1 |
| 1251 | ; AVX512BW-NEXT: vmovdqu8 %zmm4, %zmm3 {%k1} |
| 1252 | ; AVX512BW-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1253 | ; AVX512BW-NEXT: vpmovb2m %zmm2, %k1 |
| 1254 | ; AVX512BW-NEXT: vpaddb %zmm3, %zmm3, %zmm3 {%k1} |
| 1255 | ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63] |
| 1256 | ; AVX512BW-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1257 | ; AVX512BW-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2 |
| 1258 | ; AVX512BW-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1259 | ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} zmm1 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55] |
| 1260 | ; AVX512BW-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1261 | ; AVX512BW-NEXT: vpsllvw {{.*}}(%rip), %zmm1, %zmm1 |
| 1262 | ; AVX512BW-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1263 | ; AVX512BW-NEXT: vpackuswb %zmm2, %zmm1, %zmm1 |
| 1264 | ; AVX512BW-NEXT: vporq %zmm1, %zmm3, %zmm1 |
| 1265 | ; AVX512BW-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101 |
| 1266 | ; AVX512BW-NEXT: kmovq %rax, %k1 |
| 1267 | ; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 1268 | ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 1269 | ; AVX512BW-NEXT: retq |
| 1270 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1271 | ; AVX512VBMI2-LABEL: constant_funnnel_v64i8: |
| 1272 | ; AVX512VBMI2: # %bb.0: |
| 1273 | ; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm2 = [8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256] |
| 1274 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm2, %k1 |
| 1275 | ; AVX512VBMI2-NEXT: vpsllw $4, %zmm0, %zmm3 |
| 1276 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3 |
| 1277 | ; AVX512VBMI2-NEXT: vpblendmb %zmm3, %zmm0, %zmm3 {%k1} |
| 1278 | ; AVX512VBMI2-NEXT: vpsllw $2, %zmm3, %zmm4 |
| 1279 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 1280 | ; AVX512VBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1281 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm2, %k1 |
| 1282 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm4, %zmm3 {%k1} |
| 1283 | ; AVX512VBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1284 | ; AVX512VBMI2-NEXT: vpmovb2m %zmm2, %k1 |
| 1285 | ; AVX512VBMI2-NEXT: vpaddb %zmm3, %zmm3, %zmm3 {%k1} |
| 1286 | ; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63] |
| 1287 | ; AVX512VBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1288 | ; AVX512VBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2 |
| 1289 | ; AVX512VBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1290 | ; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} zmm1 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55] |
| 1291 | ; AVX512VBMI2-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1292 | ; AVX512VBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm1, %zmm1 |
| 1293 | ; AVX512VBMI2-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1294 | ; AVX512VBMI2-NEXT: vpackuswb %zmm2, %zmm1, %zmm1 |
| 1295 | ; AVX512VBMI2-NEXT: vporq %zmm1, %zmm3, %zmm1 |
| 1296 | ; AVX512VBMI2-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101 |
| 1297 | ; AVX512VBMI2-NEXT: kmovq %rax, %k1 |
| 1298 | ; AVX512VBMI2-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 1299 | ; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 1300 | ; AVX512VBMI2-NEXT: retq |
| 1301 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1302 | ; AVX512VLBW-LABEL: constant_funnnel_v64i8: |
| 1303 | ; AVX512VLBW: # %bb.0: |
| 1304 | ; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256] |
| 1305 | ; AVX512VLBW-NEXT: vpmovb2m %zmm2, %k1 |
| 1306 | ; AVX512VLBW-NEXT: vpsllw $4, %zmm0, %zmm3 |
| 1307 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3 |
| 1308 | ; AVX512VLBW-NEXT: vpblendmb %zmm3, %zmm0, %zmm3 {%k1} |
| 1309 | ; AVX512VLBW-NEXT: vpsllw $2, %zmm3, %zmm4 |
| 1310 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 1311 | ; AVX512VLBW-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1312 | ; AVX512VLBW-NEXT: vpmovb2m %zmm2, %k1 |
| 1313 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm4, %zmm3 {%k1} |
| 1314 | ; AVX512VLBW-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1315 | ; AVX512VLBW-NEXT: vpmovb2m %zmm2, %k1 |
| 1316 | ; AVX512VLBW-NEXT: vpaddb %zmm3, %zmm3, %zmm3 {%k1} |
| 1317 | ; AVX512VLBW-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63] |
| 1318 | ; AVX512VLBW-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1319 | ; AVX512VLBW-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2 |
| 1320 | ; AVX512VLBW-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1321 | ; AVX512VLBW-NEXT: vpunpcklbw {{.*#+}} zmm1 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55] |
| 1322 | ; AVX512VLBW-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1323 | ; AVX512VLBW-NEXT: vpsllvw {{.*}}(%rip), %zmm1, %zmm1 |
| 1324 | ; AVX512VLBW-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1325 | ; AVX512VLBW-NEXT: vpackuswb %zmm2, %zmm1, %zmm1 |
| 1326 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm3, %zmm1 |
| 1327 | ; AVX512VLBW-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101 |
| 1328 | ; AVX512VLBW-NEXT: kmovq %rax, %k1 |
| 1329 | ; AVX512VLBW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 1330 | ; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 1331 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1332 | ; |
| 1333 | ; AVX512VLVBMI2-LABEL: constant_funnnel_v64i8: |
| 1334 | ; AVX512VLVBMI2: # %bb.0: |
| 1335 | ; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm2 = [8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256,8192,24640,41088,57536,57344,41152,24704,8256] |
| 1336 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm2, %k1 |
| 1337 | ; AVX512VLVBMI2-NEXT: vpsllw $4, %zmm0, %zmm3 |
| 1338 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3 |
| 1339 | ; AVX512VLVBMI2-NEXT: vpblendmb %zmm3, %zmm0, %zmm3 {%k1} |
| 1340 | ; AVX512VLVBMI2-NEXT: vpsllw $2, %zmm3, %zmm4 |
| 1341 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4 |
| 1342 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1343 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm2, %k1 |
| 1344 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm4, %zmm3 {%k1} |
| 1345 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2 |
| 1346 | ; AVX512VLVBMI2-NEXT: vpmovb2m %zmm2, %k1 |
| 1347 | ; AVX512VLVBMI2-NEXT: vpaddb %zmm3, %zmm3, %zmm3 {%k1} |
| 1348 | ; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63] |
| 1349 | ; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1350 | ; AVX512VLVBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2 |
| 1351 | ; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2 |
| 1352 | ; AVX512VLVBMI2-NEXT: vpunpcklbw {{.*#+}} zmm1 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55] |
| 1353 | ; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1354 | ; AVX512VLVBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm1, %zmm1 |
| 1355 | ; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm1, %zmm1 |
| 1356 | ; AVX512VLVBMI2-NEXT: vpackuswb %zmm2, %zmm1, %zmm1 |
| 1357 | ; AVX512VLVBMI2-NEXT: vporq %zmm1, %zmm3, %zmm1 |
| 1358 | ; AVX512VLVBMI2-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101 |
| 1359 | ; AVX512VLVBMI2-NEXT: kmovq %rax, %k1 |
| 1360 | ; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1} |
| 1361 | ; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0 |
| 1362 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1363 | %res = call <64 x i8> @llvm.fshl.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1>) |
| 1364 | ret <64 x i8> %res |
| 1365 | } |
| 1366 | |
| 1367 | ; |
| 1368 | ; Uniform Constant Shifts |
| 1369 | ; |
| 1370 | |
| 1371 | define <8 x i64> @splatconstant_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1372 | ; AVX512F-LABEL: splatconstant_funnnel_v8i64: |
| 1373 | ; AVX512F: # %bb.0: |
| 1374 | ; AVX512F-NEXT: vpsrlq $50, %zmm1, %zmm1 |
| 1375 | ; AVX512F-NEXT: vpsllq $14, %zmm0, %zmm0 |
| 1376 | ; AVX512F-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1377 | ; AVX512F-NEXT: retq |
| 1378 | ; |
| 1379 | ; AVX512VL-LABEL: splatconstant_funnnel_v8i64: |
| 1380 | ; AVX512VL: # %bb.0: |
| 1381 | ; AVX512VL-NEXT: vpsrlq $50, %zmm1, %zmm1 |
| 1382 | ; AVX512VL-NEXT: vpsllq $14, %zmm0, %zmm0 |
| 1383 | ; AVX512VL-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1384 | ; AVX512VL-NEXT: retq |
| 1385 | ; |
| 1386 | ; AVX512BW-LABEL: splatconstant_funnnel_v8i64: |
| 1387 | ; AVX512BW: # %bb.0: |
| 1388 | ; AVX512BW-NEXT: vpsrlq $50, %zmm1, %zmm1 |
| 1389 | ; AVX512BW-NEXT: vpsllq $14, %zmm0, %zmm0 |
| 1390 | ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1391 | ; AVX512BW-NEXT: retq |
| 1392 | ; |
| 1393 | ; AVX512VBMI2-LABEL: splatconstant_funnnel_v8i64: |
| 1394 | ; AVX512VBMI2: # %bb.0: |
| 1395 | ; AVX512VBMI2-NEXT: vpshldq $14, %zmm1, %zmm0, %zmm0 |
| 1396 | ; AVX512VBMI2-NEXT: retq |
| 1397 | ; |
| 1398 | ; AVX512VLBW-LABEL: splatconstant_funnnel_v8i64: |
| 1399 | ; AVX512VLBW: # %bb.0: |
| 1400 | ; AVX512VLBW-NEXT: vpsrlq $50, %zmm1, %zmm1 |
| 1401 | ; AVX512VLBW-NEXT: vpsllq $14, %zmm0, %zmm0 |
| 1402 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1403 | ; AVX512VLBW-NEXT: retq |
| 1404 | ; |
| 1405 | ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v8i64: |
| 1406 | ; AVX512VLVBMI2: # %bb.0: |
| 1407 | ; AVX512VLVBMI2-NEXT: vpshldq $14, %zmm1, %zmm0, %zmm0 |
| 1408 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1409 | %res = call <8 x i64> @llvm.fshl.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> <i64 14, i64 14, i64 14, i64 14, i64 14, i64 14, i64 14, i64 14>) |
| 1410 | ret <8 x i64> %res |
| 1411 | } |
| 1412 | |
| 1413 | define <16 x i32> @splatconstant_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y) nounwind { |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1414 | ; AVX512F-LABEL: splatconstant_funnnel_v16i32: |
| 1415 | ; AVX512F: # %bb.0: |
| 1416 | ; AVX512F-NEXT: vpsrld $28, %zmm1, %zmm1 |
| 1417 | ; AVX512F-NEXT: vpslld $4, %zmm0, %zmm0 |
| 1418 | ; AVX512F-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1419 | ; AVX512F-NEXT: retq |
| 1420 | ; |
| 1421 | ; AVX512VL-LABEL: splatconstant_funnnel_v16i32: |
| 1422 | ; AVX512VL: # %bb.0: |
| 1423 | ; AVX512VL-NEXT: vpsrld $28, %zmm1, %zmm1 |
| 1424 | ; AVX512VL-NEXT: vpslld $4, %zmm0, %zmm0 |
| 1425 | ; AVX512VL-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1426 | ; AVX512VL-NEXT: retq |
| 1427 | ; |
| 1428 | ; AVX512BW-LABEL: splatconstant_funnnel_v16i32: |
| 1429 | ; AVX512BW: # %bb.0: |
| 1430 | ; AVX512BW-NEXT: vpsrld $28, %zmm1, %zmm1 |
| 1431 | ; AVX512BW-NEXT: vpslld $4, %zmm0, %zmm0 |
| 1432 | ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1433 | ; AVX512BW-NEXT: retq |
| 1434 | ; |
| 1435 | ; AVX512VBMI2-LABEL: splatconstant_funnnel_v16i32: |
| 1436 | ; AVX512VBMI2: # %bb.0: |
| 1437 | ; AVX512VBMI2-NEXT: vpshldd $4, %zmm1, %zmm0, %zmm0 |
| 1438 | ; AVX512VBMI2-NEXT: retq |
| 1439 | ; |
| 1440 | ; AVX512VLBW-LABEL: splatconstant_funnnel_v16i32: |
| 1441 | ; AVX512VLBW: # %bb.0: |
| 1442 | ; AVX512VLBW-NEXT: vpsrld $28, %zmm1, %zmm1 |
| 1443 | ; AVX512VLBW-NEXT: vpslld $4, %zmm0, %zmm0 |
| 1444 | ; AVX512VLBW-NEXT: vpord %zmm1, %zmm0, %zmm0 |
| 1445 | ; AVX512VLBW-NEXT: retq |
| 1446 | ; |
| 1447 | ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v16i32: |
| 1448 | ; AVX512VLVBMI2: # %bb.0: |
| 1449 | ; AVX512VLVBMI2-NEXT: vpshldd $4, %zmm1, %zmm0, %zmm0 |
| 1450 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1451 | %res = call <16 x i32> @llvm.fshl.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> <i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4>) |
| 1452 | ret <16 x i32> %res |
| 1453 | } |
| 1454 | |
| 1455 | define <32 x i16> @splatconstant_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y) nounwind { |
| 1456 | ; AVX512F-LABEL: splatconstant_funnnel_v32i16: |
| 1457 | ; AVX512F: # %bb.0: |
| 1458 | ; AVX512F-NEXT: vpsrlw $9, %ymm2, %ymm2 |
| 1459 | ; AVX512F-NEXT: vpsllw $7, %ymm0, %ymm0 |
| 1460 | ; AVX512F-NEXT: vpor %ymm2, %ymm0, %ymm0 |
| 1461 | ; AVX512F-NEXT: vpsrlw $9, %ymm3, %ymm2 |
| 1462 | ; AVX512F-NEXT: vpsllw $7, %ymm1, %ymm1 |
| 1463 | ; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1 |
| 1464 | ; AVX512F-NEXT: retq |
| 1465 | ; |
| 1466 | ; AVX512VL-LABEL: splatconstant_funnnel_v32i16: |
| 1467 | ; AVX512VL: # %bb.0: |
| 1468 | ; AVX512VL-NEXT: vpsrlw $9, %ymm2, %ymm2 |
| 1469 | ; AVX512VL-NEXT: vpsllw $7, %ymm0, %ymm0 |
| 1470 | ; AVX512VL-NEXT: vpor %ymm2, %ymm0, %ymm0 |
| 1471 | ; AVX512VL-NEXT: vpsrlw $9, %ymm3, %ymm2 |
| 1472 | ; AVX512VL-NEXT: vpsllw $7, %ymm1, %ymm1 |
| 1473 | ; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1 |
| 1474 | ; AVX512VL-NEXT: retq |
| 1475 | ; |
| 1476 | ; AVX512BW-LABEL: splatconstant_funnnel_v32i16: |
| 1477 | ; AVX512BW: # %bb.0: |
| 1478 | ; AVX512BW-NEXT: vpsrlw $9, %zmm1, %zmm1 |
| 1479 | ; AVX512BW-NEXT: vpsllw $7, %zmm0, %zmm0 |
| 1480 | ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1481 | ; AVX512BW-NEXT: retq |
| 1482 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1483 | ; AVX512VBMI2-LABEL: splatconstant_funnnel_v32i16: |
| 1484 | ; AVX512VBMI2: # %bb.0: |
| 1485 | ; AVX512VBMI2-NEXT: vpshldw $7, %zmm1, %zmm0, %zmm0 |
| 1486 | ; AVX512VBMI2-NEXT: retq |
| 1487 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1488 | ; AVX512VLBW-LABEL: splatconstant_funnnel_v32i16: |
| 1489 | ; AVX512VLBW: # %bb.0: |
| 1490 | ; AVX512VLBW-NEXT: vpsrlw $9, %zmm1, %zmm1 |
| 1491 | ; AVX512VLBW-NEXT: vpsllw $7, %zmm0, %zmm0 |
| 1492 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1493 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1494 | ; |
| 1495 | ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v32i16: |
| 1496 | ; AVX512VLVBMI2: # %bb.0: |
| 1497 | ; AVX512VLVBMI2-NEXT: vpshldw $7, %zmm1, %zmm0, %zmm0 |
| 1498 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1499 | %res = call <32 x i16> @llvm.fshl.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>) |
| 1500 | ret <32 x i16> %res |
| 1501 | } |
| 1502 | |
| 1503 | define <64 x i8> @splatconstant_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y) nounwind { |
| 1504 | ; AVX512F-LABEL: splatconstant_funnnel_v64i8: |
| 1505 | ; AVX512F: # %bb.0: |
| 1506 | ; AVX512F-NEXT: vpsrlw $4, %ymm2, %ymm2 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1507 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240] |
| 1508 | ; AVX512F-NEXT: vpandn %ymm2, %ymm4, %ymm2 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1509 | ; AVX512F-NEXT: vpsllw $4, %ymm0, %ymm0 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1510 | ; AVX512F-NEXT: vpand %ymm4, %ymm0, %ymm0 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1511 | ; AVX512F-NEXT: vpor %ymm2, %ymm0, %ymm0 |
| 1512 | ; AVX512F-NEXT: vpsrlw $4, %ymm3, %ymm2 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1513 | ; AVX512F-NEXT: vpandn %ymm2, %ymm4, %ymm2 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1514 | ; AVX512F-NEXT: vpsllw $4, %ymm1, %ymm1 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1515 | ; AVX512F-NEXT: vpand %ymm4, %ymm1, %ymm1 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1516 | ; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1 |
| 1517 | ; AVX512F-NEXT: retq |
| 1518 | ; |
| 1519 | ; AVX512VL-LABEL: splatconstant_funnnel_v64i8: |
| 1520 | ; AVX512VL: # %bb.0: |
| 1521 | ; AVX512VL-NEXT: vpsrlw $4, %ymm2, %ymm2 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1522 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm4 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240] |
| 1523 | ; AVX512VL-NEXT: vpandn %ymm2, %ymm4, %ymm2 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1524 | ; AVX512VL-NEXT: vpsllw $4, %ymm0, %ymm0 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1525 | ; AVX512VL-NEXT: vpand %ymm4, %ymm0, %ymm0 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1526 | ; AVX512VL-NEXT: vpor %ymm2, %ymm0, %ymm0 |
| 1527 | ; AVX512VL-NEXT: vpsrlw $4, %ymm3, %ymm2 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1528 | ; AVX512VL-NEXT: vpandn %ymm2, %ymm4, %ymm2 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1529 | ; AVX512VL-NEXT: vpsllw $4, %ymm1, %ymm1 |
Simon Pilgrim | 933673d | 2019-01-22 13:44:49 +0000 | [diff] [blame] | 1530 | ; AVX512VL-NEXT: vpand %ymm4, %ymm1, %ymm1 |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1531 | ; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1 |
| 1532 | ; AVX512VL-NEXT: retq |
| 1533 | ; |
| 1534 | ; AVX512BW-LABEL: splatconstant_funnnel_v64i8: |
| 1535 | ; AVX512BW: # %bb.0: |
| 1536 | ; AVX512BW-NEXT: vpsrlw $4, %zmm1, %zmm1 |
| 1537 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1 |
| 1538 | ; AVX512BW-NEXT: vpsllw $4, %zmm0, %zmm0 |
| 1539 | ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0 |
| 1540 | ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1541 | ; AVX512BW-NEXT: retq |
| 1542 | ; |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1543 | ; AVX512VBMI2-LABEL: splatconstant_funnnel_v64i8: |
| 1544 | ; AVX512VBMI2: # %bb.0: |
| 1545 | ; AVX512VBMI2-NEXT: vpsrlw $4, %zmm1, %zmm1 |
| 1546 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1 |
| 1547 | ; AVX512VBMI2-NEXT: vpsllw $4, %zmm0, %zmm0 |
| 1548 | ; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0 |
| 1549 | ; AVX512VBMI2-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1550 | ; AVX512VBMI2-NEXT: retq |
| 1551 | ; |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1552 | ; AVX512VLBW-LABEL: splatconstant_funnnel_v64i8: |
| 1553 | ; AVX512VLBW: # %bb.0: |
| 1554 | ; AVX512VLBW-NEXT: vpsrlw $4, %zmm1, %zmm1 |
| 1555 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1 |
| 1556 | ; AVX512VLBW-NEXT: vpsllw $4, %zmm0, %zmm0 |
| 1557 | ; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0 |
| 1558 | ; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1559 | ; AVX512VLBW-NEXT: retq |
Craig Topper | 6ffeeb7 | 2019-01-06 18:10:18 +0000 | [diff] [blame] | 1560 | ; |
| 1561 | ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v64i8: |
| 1562 | ; AVX512VLVBMI2: # %bb.0: |
| 1563 | ; AVX512VLVBMI2-NEXT: vpsrlw $4, %zmm1, %zmm1 |
| 1564 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1 |
| 1565 | ; AVX512VLVBMI2-NEXT: vpsllw $4, %zmm0, %zmm0 |
| 1566 | ; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0 |
| 1567 | ; AVX512VLVBMI2-NEXT: vporq %zmm1, %zmm0, %zmm0 |
| 1568 | ; AVX512VLVBMI2-NEXT: retq |
Simon Pilgrim | ba8e84b | 2018-12-18 10:32:54 +0000 | [diff] [blame] | 1569 | %res = call <64 x i8> @llvm.fshl.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> <i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4>) |
| 1570 | ret <64 x i8> %res |
| 1571 | } |