| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +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=+sse2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE2 |
| 3 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+ssse3 | FileCheck %s --check-prefix=SSE --check-prefix=SSSE3 |
| 4 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE --check-prefix=SSE41 |
| 5 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1 |
| 6 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-SLOW |
| 7 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-FAST |
| 8 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512F |
| 9 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512VL |
| 10 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512BW |
| 11 | ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512BWVL |
| 12 | |
| 13 | ; |
| 14 | ; PACKUS saturation truncation to vXi32 |
| 15 | ; |
| 16 | |
| 17 | define <4 x i32> @trunc_packus_v4i64_v4i32(<4 x i64> %a0) { |
| 18 | ; SSE2-LABEL: trunc_packus_v4i64_v4i32: |
| 19 | ; SSE2: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 20 | ; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 21 | ; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0] |
| 22 | ; SSE2-NEXT: movdqa %xmm0, %xmm3 |
| 23 | ; SSE2-NEXT: pxor %xmm2, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 24 | ; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [2147483647,2147483647] |
| 25 | ; SSE2-NEXT: movdqa %xmm5, %xmm6 |
| 26 | ; SSE2-NEXT: pcmpgtd %xmm3, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 27 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 28 | ; SSE2-NEXT: pcmpeqd %xmm5, %xmm3 |
| 29 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3] |
| 30 | ; SSE2-NEXT: pand %xmm7, %xmm4 |
| 31 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3] |
| 32 | ; SSE2-NEXT: por %xmm4, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 33 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 34 | ; SSE2-NEXT: pandn %xmm8, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 35 | ; SSE2-NEXT: por %xmm0, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 36 | ; SSE2-NEXT: movdqa %xmm1, %xmm0 |
| 37 | ; SSE2-NEXT: pxor %xmm2, %xmm0 |
| 38 | ; SSE2-NEXT: movdqa %xmm5, %xmm4 |
| 39 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm4 |
| 40 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 41 | ; SSE2-NEXT: pcmpeqd %xmm5, %xmm0 |
| 42 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 43 | ; SSE2-NEXT: pand %xmm6, %xmm0 |
| 44 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 45 | ; SSE2-NEXT: por %xmm0, %xmm4 |
| 46 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| 47 | ; SSE2-NEXT: pandn %xmm8, %xmm4 |
| 48 | ; SSE2-NEXT: por %xmm1, %xmm4 |
| 49 | ; SSE2-NEXT: movdqa %xmm4, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 50 | ; SSE2-NEXT: pxor %xmm2, %xmm0 |
| 51 | ; SSE2-NEXT: movdqa %xmm0, %xmm1 |
| 52 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm1 |
| 53 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2] |
| 54 | ; SSE2-NEXT: pcmpeqd %xmm2, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 55 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 56 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 57 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 58 | ; SSE2-NEXT: por %xmm0, %xmm1 |
| 59 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| 60 | ; SSE2-NEXT: movdqa %xmm3, %xmm0 |
| 61 | ; SSE2-NEXT: pxor %xmm2, %xmm0 |
| 62 | ; SSE2-NEXT: movdqa %xmm0, %xmm4 |
| 63 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm4 |
| 64 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 65 | ; SSE2-NEXT: pcmpeqd %xmm2, %xmm0 |
| 66 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3] |
| 67 | ; SSE2-NEXT: pand %xmm5, %xmm2 |
| 68 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 69 | ; SSE2-NEXT: por %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 70 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 71 | ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 72 | ; SSE2-NEXT: retq |
| 73 | ; |
| 74 | ; SSSE3-LABEL: trunc_packus_v4i64_v4i32: |
| 75 | ; SSSE3: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 76 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 77 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0] |
| 78 | ; SSSE3-NEXT: movdqa %xmm0, %xmm3 |
| 79 | ; SSSE3-NEXT: pxor %xmm2, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 80 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm5 = [2147483647,2147483647] |
| 81 | ; SSSE3-NEXT: movdqa %xmm5, %xmm6 |
| 82 | ; SSSE3-NEXT: pcmpgtd %xmm3, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 83 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 84 | ; SSSE3-NEXT: pcmpeqd %xmm5, %xmm3 |
| 85 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3] |
| 86 | ; SSSE3-NEXT: pand %xmm7, %xmm4 |
| 87 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3] |
| 88 | ; SSSE3-NEXT: por %xmm4, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 89 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 90 | ; SSSE3-NEXT: pandn %xmm8, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 91 | ; SSSE3-NEXT: por %xmm0, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 92 | ; SSSE3-NEXT: movdqa %xmm1, %xmm0 |
| 93 | ; SSSE3-NEXT: pxor %xmm2, %xmm0 |
| 94 | ; SSSE3-NEXT: movdqa %xmm5, %xmm4 |
| 95 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4 |
| 96 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 97 | ; SSSE3-NEXT: pcmpeqd %xmm5, %xmm0 |
| 98 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 99 | ; SSSE3-NEXT: pand %xmm6, %xmm0 |
| 100 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 101 | ; SSSE3-NEXT: por %xmm0, %xmm4 |
| 102 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| 103 | ; SSSE3-NEXT: pandn %xmm8, %xmm4 |
| 104 | ; SSSE3-NEXT: por %xmm1, %xmm4 |
| 105 | ; SSSE3-NEXT: movdqa %xmm4, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 106 | ; SSSE3-NEXT: pxor %xmm2, %xmm0 |
| 107 | ; SSSE3-NEXT: movdqa %xmm0, %xmm1 |
| 108 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1 |
| 109 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2] |
| 110 | ; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 111 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 112 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 113 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 114 | ; SSSE3-NEXT: por %xmm0, %xmm1 |
| 115 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| 116 | ; SSSE3-NEXT: movdqa %xmm3, %xmm0 |
| 117 | ; SSSE3-NEXT: pxor %xmm2, %xmm0 |
| 118 | ; SSSE3-NEXT: movdqa %xmm0, %xmm4 |
| 119 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4 |
| 120 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 121 | ; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0 |
| 122 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3] |
| 123 | ; SSSE3-NEXT: pand %xmm5, %xmm2 |
| 124 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 125 | ; SSSE3-NEXT: por %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 126 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 127 | ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 128 | ; SSSE3-NEXT: retq |
| 129 | ; |
| 130 | ; SSE41-LABEL: trunc_packus_v4i64_v4i32: |
| 131 | ; SSE41: # %bb.0: |
| 132 | ; SSE41-NEXT: movdqa %xmm0, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 133 | ; SSE41-NEXT: movapd {{.*#+}} xmm4 = [4294967295,4294967295] |
| 134 | ; SSE41-NEXT: movdqa {{.*#+}} xmm8 = [2147483648,0,2147483648,0] |
| 135 | ; SSE41-NEXT: pxor %xmm8, %xmm0 |
| 136 | ; SSE41-NEXT: movdqa {{.*#+}} xmm6 = [2147483647,2147483647] |
| 137 | ; SSE41-NEXT: movdqa %xmm6, %xmm5 |
| 138 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm5 |
| 139 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] |
| 140 | ; SSE41-NEXT: pcmpeqd %xmm6, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 141 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 142 | ; SSE41-NEXT: pand %xmm7, %xmm3 |
| 143 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 144 | ; SSE41-NEXT: por %xmm3, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 145 | ; SSE41-NEXT: movapd %xmm4, %xmm5 |
| 146 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm5 |
| 147 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| 148 | ; SSE41-NEXT: pxor %xmm8, %xmm0 |
| 149 | ; SSE41-NEXT: movdqa %xmm6, %xmm2 |
| 150 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 151 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 152 | ; SSE41-NEXT: pcmpeqd %xmm6, %xmm0 |
| 153 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 154 | ; SSE41-NEXT: pand %xmm3, %xmm6 |
| 155 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 156 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 157 | ; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm4 |
| 158 | ; SSE41-NEXT: xorpd %xmm1, %xmm1 |
| 159 | ; SSE41-NEXT: movapd %xmm4, %xmm0 |
| 160 | ; SSE41-NEXT: xorpd %xmm8, %xmm0 |
| 161 | ; SSE41-NEXT: movapd %xmm0, %xmm2 |
| 162 | ; SSE41-NEXT: pcmpgtd %xmm8, %xmm2 |
| 163 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2] |
| 164 | ; SSE41-NEXT: pcmpeqd %xmm8, %xmm0 |
| 165 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 166 | ; SSE41-NEXT: pand %xmm3, %xmm6 |
| 167 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 168 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 169 | ; SSE41-NEXT: pxor %xmm2, %xmm2 |
| 170 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm2 |
| 171 | ; SSE41-NEXT: movapd %xmm5, %xmm0 |
| 172 | ; SSE41-NEXT: xorpd %xmm8, %xmm0 |
| 173 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 174 | ; SSE41-NEXT: pcmpgtd %xmm8, %xmm3 |
| 175 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 176 | ; SSE41-NEXT: pcmpeqd %xmm8, %xmm0 |
| 177 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 178 | ; SSE41-NEXT: pand %xmm4, %xmm6 |
| 179 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 180 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 181 | ; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm1 |
| 182 | ; SSE41-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm2[0,2] |
| 183 | ; SSE41-NEXT: movaps %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 184 | ; SSE41-NEXT: retq |
| 185 | ; |
| 186 | ; AVX1-LABEL: trunc_packus_v4i64_v4i32: |
| 187 | ; AVX1: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 188 | ; AVX1-NEXT: vmovapd {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295] |
| 189 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2 |
| 190 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [4294967295,4294967295] |
| 191 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm2 |
| 192 | ; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm3 |
| 193 | ; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2 |
| 194 | ; AVX1-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 195 | ; AVX1-NEXT: vxorpd %xmm1, %xmm1, %xmm1 |
| 196 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2 |
| 197 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3 |
| 198 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm1 |
| 199 | ; AVX1-NEXT: vpand %xmm3, %xmm1, %xmm1 |
| 200 | ; AVX1-NEXT: vpand %xmm0, %xmm2, %xmm0 |
| 201 | ; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] |
| 202 | ; AVX1-NEXT: vzeroupper |
| 203 | ; AVX1-NEXT: retq |
| 204 | ; |
| 205 | ; AVX2-SLOW-LABEL: trunc_packus_v4i64_v4i32: |
| 206 | ; AVX2-SLOW: # %bb.0: |
| 207 | ; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295] |
| 208 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2 |
| 209 | ; AVX2-SLOW-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0 |
| 210 | ; AVX2-SLOW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 211 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1 |
| 212 | ; AVX2-SLOW-NEXT: vpand %ymm0, %ymm1, %ymm0 |
| 213 | ; AVX2-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] |
| 214 | ; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 215 | ; AVX2-SLOW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 216 | ; AVX2-SLOW-NEXT: vzeroupper |
| 217 | ; AVX2-SLOW-NEXT: retq |
| 218 | ; |
| 219 | ; AVX2-FAST-LABEL: trunc_packus_v4i64_v4i32: |
| 220 | ; AVX2-FAST: # %bb.0: |
| 221 | ; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295] |
| 222 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2 |
| 223 | ; AVX2-FAST-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0 |
| 224 | ; AVX2-FAST-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 225 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1 |
| 226 | ; AVX2-FAST-NEXT: vpand %ymm0, %ymm1, %ymm0 |
| 227 | ; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [0,2,4,6,4,6,6,7] |
| 228 | ; AVX2-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 229 | ; AVX2-FAST-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 230 | ; AVX2-FAST-NEXT: vzeroupper |
| 231 | ; AVX2-FAST-NEXT: retq |
| 232 | ; |
| 233 | ; AVX512F-LABEL: trunc_packus_v4i64_v4i32: |
| 234 | ; AVX512F: # %bb.0: |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 235 | ; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 236 | ; AVX512F-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295] |
| 237 | ; AVX512F-NEXT: vpminsq %zmm1, %zmm0, %zmm0 |
| 238 | ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 239 | ; AVX512F-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 |
| 240 | ; AVX512F-NEXT: vpmovqd %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 241 | ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 242 | ; AVX512F-NEXT: vzeroupper |
| 243 | ; AVX512F-NEXT: retq |
| 244 | ; |
| 245 | ; AVX512VL-LABEL: trunc_packus_v4i64_v4i32: |
| 246 | ; AVX512VL: # %bb.0: |
| 247 | ; AVX512VL-NEXT: vpminsq {{.*}}(%rip){1to4}, %ymm0, %ymm0 |
| 248 | ; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 249 | ; AVX512VL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0 |
| 250 | ; AVX512VL-NEXT: vpmovqd %ymm0, %xmm0 |
| 251 | ; AVX512VL-NEXT: vzeroupper |
| 252 | ; AVX512VL-NEXT: retq |
| 253 | ; |
| 254 | ; AVX512BW-LABEL: trunc_packus_v4i64_v4i32: |
| 255 | ; AVX512BW: # %bb.0: |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 256 | ; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 257 | ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295] |
| 258 | ; AVX512BW-NEXT: vpminsq %zmm1, %zmm0, %zmm0 |
| 259 | ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 260 | ; AVX512BW-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 |
| 261 | ; AVX512BW-NEXT: vpmovqd %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 262 | ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 263 | ; AVX512BW-NEXT: vzeroupper |
| 264 | ; AVX512BW-NEXT: retq |
| 265 | ; |
| 266 | ; AVX512BWVL-LABEL: trunc_packus_v4i64_v4i32: |
| 267 | ; AVX512BWVL: # %bb.0: |
| 268 | ; AVX512BWVL-NEXT: vpminsq {{.*}}(%rip){1to4}, %ymm0, %ymm0 |
| 269 | ; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 270 | ; AVX512BWVL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0 |
| 271 | ; AVX512BWVL-NEXT: vpmovqd %ymm0, %xmm0 |
| 272 | ; AVX512BWVL-NEXT: vzeroupper |
| 273 | ; AVX512BWVL-NEXT: retq |
| 274 | %1 = icmp slt <4 x i64> %a0, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> |
| 275 | %2 = select <4 x i1> %1, <4 x i64> %a0, <4 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> |
| 276 | %3 = icmp sgt <4 x i64> %2, zeroinitializer |
| 277 | %4 = select <4 x i1> %3, <4 x i64> %2, <4 x i64> zeroinitializer |
| 278 | %5 = trunc <4 x i64> %4 to <4 x i32> |
| 279 | ret <4 x i32> %5 |
| 280 | } |
| 281 | |
| 282 | |
| 283 | define <8 x i32> @trunc_packus_v8i64_v8i32(<8 x i64> %a0) { |
| 284 | ; SSE2-LABEL: trunc_packus_v8i64_v8i32: |
| 285 | ; SSE2: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 286 | ; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295] |
| 287 | ; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 288 | ; SSE2-NEXT: movdqa %xmm0, %xmm5 |
| 289 | ; SSE2-NEXT: pxor %xmm10, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 290 | ; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483647,2147483647] |
| 291 | ; SSE2-NEXT: movdqa %xmm9, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 292 | ; SSE2-NEXT: pcmpgtd %xmm5, %xmm6 |
| 293 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 294 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 295 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| 296 | ; SSE2-NEXT: pand %xmm7, %xmm4 |
| 297 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 298 | ; SSE2-NEXT: por %xmm4, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 299 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| 300 | ; SSE2-NEXT: pandn %xmm8, %xmm5 |
| 301 | ; SSE2-NEXT: por %xmm0, %xmm5 |
| 302 | ; SSE2-NEXT: movdqa %xmm1, %xmm0 |
| 303 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 304 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 305 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm4 |
| 306 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 307 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 308 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 309 | ; SSE2-NEXT: pand %xmm6, %xmm7 |
| 310 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 311 | ; SSE2-NEXT: por %xmm7, %xmm0 |
| 312 | ; SSE2-NEXT: pand %xmm0, %xmm1 |
| 313 | ; SSE2-NEXT: pandn %xmm8, %xmm0 |
| 314 | ; SSE2-NEXT: por %xmm1, %xmm0 |
| 315 | ; SSE2-NEXT: movdqa %xmm2, %xmm1 |
| 316 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 317 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 318 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm4 |
| 319 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 320 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 321 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 322 | ; SSE2-NEXT: pand %xmm6, %xmm1 |
| 323 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 324 | ; SSE2-NEXT: por %xmm1, %xmm6 |
| 325 | ; SSE2-NEXT: pand %xmm6, %xmm2 |
| 326 | ; SSE2-NEXT: pandn %xmm8, %xmm6 |
| 327 | ; SSE2-NEXT: por %xmm2, %xmm6 |
| 328 | ; SSE2-NEXT: movdqa %xmm3, %xmm1 |
| 329 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 330 | ; SSE2-NEXT: movdqa %xmm9, %xmm2 |
| 331 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm2 |
| 332 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 333 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm1 |
| 334 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 335 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 336 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| 337 | ; SSE2-NEXT: por %xmm1, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 338 | ; SSE2-NEXT: pand %xmm2, %xmm3 |
| 339 | ; SSE2-NEXT: pandn %xmm8, %xmm2 |
| 340 | ; SSE2-NEXT: por %xmm3, %xmm2 |
| 341 | ; SSE2-NEXT: movdqa %xmm2, %xmm1 |
| 342 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 343 | ; SSE2-NEXT: movdqa %xmm1, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 344 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm3 |
| 345 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 346 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm1 |
| 347 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 348 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 349 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 350 | ; SSE2-NEXT: por %xmm1, %xmm3 |
| 351 | ; SSE2-NEXT: pand %xmm2, %xmm3 |
| 352 | ; SSE2-NEXT: movdqa %xmm6, %xmm1 |
| 353 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 354 | ; SSE2-NEXT: movdqa %xmm1, %xmm2 |
| 355 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm2 |
| 356 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 357 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm1 |
| 358 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] |
| 359 | ; SSE2-NEXT: pand %xmm4, %xmm7 |
| 360 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3] |
| 361 | ; SSE2-NEXT: por %xmm7, %xmm1 |
| 362 | ; SSE2-NEXT: pand %xmm6, %xmm1 |
| 363 | ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm3[0,2] |
| 364 | ; SSE2-NEXT: movdqa %xmm0, %xmm2 |
| 365 | ; SSE2-NEXT: pxor %xmm10, %xmm2 |
| 366 | ; SSE2-NEXT: movdqa %xmm2, %xmm3 |
| 367 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm3 |
| 368 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 369 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm2 |
| 370 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| 371 | ; SSE2-NEXT: pand %xmm4, %xmm2 |
| 372 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 373 | ; SSE2-NEXT: por %xmm2, %xmm3 |
| 374 | ; SSE2-NEXT: pand %xmm0, %xmm3 |
| 375 | ; SSE2-NEXT: movdqa %xmm5, %xmm0 |
| 376 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 377 | ; SSE2-NEXT: movdqa %xmm0, %xmm2 |
| 378 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm2 |
| 379 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 380 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm0 |
| 381 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 382 | ; SSE2-NEXT: pand %xmm4, %xmm6 |
| 383 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 384 | ; SSE2-NEXT: por %xmm6, %xmm0 |
| 385 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| 386 | ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 387 | ; SSE2-NEXT: retq |
| 388 | ; |
| 389 | ; SSSE3-LABEL: trunc_packus_v8i64_v8i32: |
| 390 | ; SSSE3: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 391 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295] |
| 392 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 393 | ; SSSE3-NEXT: movdqa %xmm0, %xmm5 |
| 394 | ; SSSE3-NEXT: pxor %xmm10, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 395 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483647,2147483647] |
| 396 | ; SSSE3-NEXT: movdqa %xmm9, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 397 | ; SSSE3-NEXT: pcmpgtd %xmm5, %xmm6 |
| 398 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 399 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 400 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| 401 | ; SSSE3-NEXT: pand %xmm7, %xmm4 |
| 402 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 403 | ; SSSE3-NEXT: por %xmm4, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 404 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| 405 | ; SSSE3-NEXT: pandn %xmm8, %xmm5 |
| 406 | ; SSSE3-NEXT: por %xmm0, %xmm5 |
| 407 | ; SSSE3-NEXT: movdqa %xmm1, %xmm0 |
| 408 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 409 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 410 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4 |
| 411 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 412 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 413 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 414 | ; SSSE3-NEXT: pand %xmm6, %xmm7 |
| 415 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 416 | ; SSSE3-NEXT: por %xmm7, %xmm0 |
| 417 | ; SSSE3-NEXT: pand %xmm0, %xmm1 |
| 418 | ; SSSE3-NEXT: pandn %xmm8, %xmm0 |
| 419 | ; SSSE3-NEXT: por %xmm1, %xmm0 |
| 420 | ; SSSE3-NEXT: movdqa %xmm2, %xmm1 |
| 421 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 422 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 423 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4 |
| 424 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 425 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 426 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 427 | ; SSSE3-NEXT: pand %xmm6, %xmm1 |
| 428 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 429 | ; SSSE3-NEXT: por %xmm1, %xmm6 |
| 430 | ; SSSE3-NEXT: pand %xmm6, %xmm2 |
| 431 | ; SSSE3-NEXT: pandn %xmm8, %xmm6 |
| 432 | ; SSSE3-NEXT: por %xmm2, %xmm6 |
| 433 | ; SSSE3-NEXT: movdqa %xmm3, %xmm1 |
| 434 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 435 | ; SSSE3-NEXT: movdqa %xmm9, %xmm2 |
| 436 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm2 |
| 437 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 438 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1 |
| 439 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 440 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 441 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| 442 | ; SSSE3-NEXT: por %xmm1, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 443 | ; SSSE3-NEXT: pand %xmm2, %xmm3 |
| 444 | ; SSSE3-NEXT: pandn %xmm8, %xmm2 |
| 445 | ; SSSE3-NEXT: por %xmm3, %xmm2 |
| 446 | ; SSSE3-NEXT: movdqa %xmm2, %xmm1 |
| 447 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 448 | ; SSSE3-NEXT: movdqa %xmm1, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 449 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3 |
| 450 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 451 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1 |
| 452 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 453 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 454 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 455 | ; SSSE3-NEXT: por %xmm1, %xmm3 |
| 456 | ; SSSE3-NEXT: pand %xmm2, %xmm3 |
| 457 | ; SSSE3-NEXT: movdqa %xmm6, %xmm1 |
| 458 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 459 | ; SSSE3-NEXT: movdqa %xmm1, %xmm2 |
| 460 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2 |
| 461 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 462 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1 |
| 463 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] |
| 464 | ; SSSE3-NEXT: pand %xmm4, %xmm7 |
| 465 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3] |
| 466 | ; SSSE3-NEXT: por %xmm7, %xmm1 |
| 467 | ; SSSE3-NEXT: pand %xmm6, %xmm1 |
| 468 | ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm3[0,2] |
| 469 | ; SSSE3-NEXT: movdqa %xmm0, %xmm2 |
| 470 | ; SSSE3-NEXT: pxor %xmm10, %xmm2 |
| 471 | ; SSSE3-NEXT: movdqa %xmm2, %xmm3 |
| 472 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3 |
| 473 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 474 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm2 |
| 475 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| 476 | ; SSSE3-NEXT: pand %xmm4, %xmm2 |
| 477 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 478 | ; SSSE3-NEXT: por %xmm2, %xmm3 |
| 479 | ; SSSE3-NEXT: pand %xmm0, %xmm3 |
| 480 | ; SSSE3-NEXT: movdqa %xmm5, %xmm0 |
| 481 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 482 | ; SSSE3-NEXT: movdqa %xmm0, %xmm2 |
| 483 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2 |
| 484 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 485 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0 |
| 486 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 487 | ; SSSE3-NEXT: pand %xmm4, %xmm6 |
| 488 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 489 | ; SSSE3-NEXT: por %xmm6, %xmm0 |
| 490 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| 491 | ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 492 | ; SSSE3-NEXT: retq |
| 493 | ; |
| 494 | ; SSE41-LABEL: trunc_packus_v8i64_v8i32: |
| 495 | ; SSE41: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 496 | ; SSE41-NEXT: movdqa %xmm0, %xmm4 |
| 497 | ; SSE41-NEXT: movapd {{.*#+}} xmm7 = [4294967295,4294967295] |
| 498 | ; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 499 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 500 | ; SSE41-NEXT: movdqa {{.*#+}} xmm11 = [2147483647,2147483647] |
| 501 | ; SSE41-NEXT: movdqa %xmm11, %xmm6 |
| 502 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm6 |
| 503 | ; SSE41-NEXT: pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2] |
| 504 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| 505 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 506 | ; SSE41-NEXT: pand %xmm8, %xmm5 |
| 507 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3] |
| 508 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 509 | ; SSE41-NEXT: movapd %xmm7, %xmm8 |
| 510 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm8 |
| 511 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| 512 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 513 | ; SSE41-NEXT: movdqa %xmm11, %xmm4 |
| 514 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm4 |
| 515 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 516 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| 517 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 518 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 519 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 520 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 521 | ; SSE41-NEXT: movapd %xmm7, %xmm9 |
| 522 | ; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm9 |
| 523 | ; SSE41-NEXT: movdqa %xmm2, %xmm0 |
| 524 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 525 | ; SSE41-NEXT: movdqa %xmm11, %xmm1 |
| 526 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm1 |
| 527 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] |
| 528 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| 529 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 530 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 531 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 532 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 533 | ; SSE41-NEXT: movapd %xmm7, %xmm4 |
| 534 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4 |
| 535 | ; SSE41-NEXT: movdqa %xmm3, %xmm0 |
| 536 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 537 | ; SSE41-NEXT: movdqa %xmm11, %xmm1 |
| 538 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm1 |
| 539 | ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] |
| 540 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| 541 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 542 | ; SSE41-NEXT: pand %xmm2, %xmm5 |
| 543 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 544 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 545 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm7 |
| 546 | ; SSE41-NEXT: pxor %xmm2, %xmm2 |
| 547 | ; SSE41-NEXT: movapd %xmm7, %xmm0 |
| 548 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 549 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 550 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm1 |
| 551 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2] |
| 552 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 553 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 554 | ; SSE41-NEXT: pand %xmm3, %xmm5 |
| 555 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 556 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 557 | ; SSE41-NEXT: pxor %xmm3, %xmm3 |
| 558 | ; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm3 |
| 559 | ; SSE41-NEXT: movapd %xmm4, %xmm0 |
| 560 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 561 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 562 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm1 |
| 563 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2] |
| 564 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 565 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 566 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 567 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 568 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 569 | ; SSE41-NEXT: pxor %xmm1, %xmm1 |
| 570 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm1 |
| 571 | ; SSE41-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm3[0,2] |
| 572 | ; SSE41-NEXT: movapd %xmm9, %xmm0 |
| 573 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 574 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 575 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 576 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 577 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 578 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 579 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 580 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 581 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 582 | ; SSE41-NEXT: pxor %xmm3, %xmm3 |
| 583 | ; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm3 |
| 584 | ; SSE41-NEXT: movapd %xmm8, %xmm0 |
| 585 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 586 | ; SSE41-NEXT: movapd %xmm0, %xmm4 |
| 587 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm4 |
| 588 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 589 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 590 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 591 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 592 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 593 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 594 | ; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm2 |
| 595 | ; SSE41-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2] |
| 596 | ; SSE41-NEXT: movaps %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 597 | ; SSE41-NEXT: retq |
| 598 | ; |
| 599 | ; AVX1-LABEL: trunc_packus_v8i64_v8i32: |
| 600 | ; AVX1: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 601 | ; AVX1-NEXT: vmovapd {{.*#+}} ymm2 = [4294967295,4294967295,4294967295,4294967295] |
| 602 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3 |
| 603 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [4294967295,4294967295] |
| 604 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 605 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm4, %xmm5 |
| 606 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3 |
| 607 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| 608 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3 |
| 609 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 610 | ; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm4 |
| 611 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3 |
| 612 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 613 | ; AVX1-NEXT: vxorpd %xmm2, %xmm2, %xmm2 |
| 614 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm3 |
| 615 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4 |
| 616 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm4, %xmm5 |
| 617 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm6 |
| 618 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm7 |
| 619 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm7, %xmm2 |
| 620 | ; AVX1-NEXT: vpand %xmm7, %xmm2, %xmm2 |
| 621 | ; AVX1-NEXT: vpand %xmm1, %xmm6, %xmm1 |
| 622 | ; AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,2],xmm2[0,2] |
| 623 | ; AVX1-NEXT: vpand %xmm4, %xmm5, %xmm2 |
| 624 | ; AVX1-NEXT: vpand %xmm0, %xmm3, %xmm0 |
| 625 | ; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2] |
| 626 | ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 |
| 627 | ; AVX1-NEXT: retq |
| 628 | ; |
| 629 | ; AVX2-SLOW-LABEL: trunc_packus_v8i64_v8i32: |
| 630 | ; AVX2-SLOW: # %bb.0: |
| 631 | ; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm2 = [4294967295,4294967295,4294967295,4294967295] |
| 632 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 633 | ; AVX2-SLOW-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 634 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3 |
| 635 | ; AVX2-SLOW-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 636 | ; AVX2-SLOW-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 637 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3 |
| 638 | ; AVX2-SLOW-NEXT: vpand %ymm1, %ymm3, %ymm1 |
| 639 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2 |
| 640 | ; AVX2-SLOW-NEXT: vpand %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 641 | ; AVX2-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] |
| 642 | ; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] |
| 643 | ; AVX2-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] |
| 644 | ; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3] |
| 645 | ; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| 646 | ; AVX2-SLOW-NEXT: retq |
| 647 | ; |
| 648 | ; AVX2-FAST-LABEL: trunc_packus_v8i64_v8i32: |
| 649 | ; AVX2-FAST: # %bb.0: |
| 650 | ; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm2 = [4294967295,4294967295,4294967295,4294967295] |
| 651 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 652 | ; AVX2-FAST-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 653 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3 |
| 654 | ; AVX2-FAST-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 655 | ; AVX2-FAST-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 656 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3 |
| 657 | ; AVX2-FAST-NEXT: vpand %ymm1, %ymm3, %ymm1 |
| 658 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2 |
| 659 | ; AVX2-FAST-NEXT: vpand %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 660 | ; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [0,2,4,6,4,6,6,7] |
| 661 | ; AVX2-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0 |
| 662 | ; AVX2-FAST-NEXT: vpermd %ymm1, %ymm2, %ymm1 |
| 663 | ; AVX2-FAST-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| 664 | ; AVX2-FAST-NEXT: retq |
| 665 | ; |
| 666 | ; AVX512-LABEL: trunc_packus_v8i64_v8i32: |
| 667 | ; AVX512: # %bb.0: |
| 668 | ; AVX512-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0 |
| 669 | ; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 670 | ; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 |
| 671 | ; AVX512-NEXT: vpmovqd %zmm0, %ymm0 |
| 672 | ; AVX512-NEXT: retq |
| 673 | %1 = icmp slt <8 x i64> %a0, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> |
| 674 | %2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> |
| 675 | %3 = icmp sgt <8 x i64> %2, zeroinitializer |
| 676 | %4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer |
| 677 | %5 = trunc <8 x i64> %4 to <8 x i32> |
| 678 | ret <8 x i32> %5 |
| 679 | } |
| 680 | |
| 681 | ; |
| 682 | ; PACKUS saturation truncation to vXi16 |
| 683 | ; |
| 684 | |
| 685 | define <8 x i16> @trunc_packus_v8i64_v8i16(<8 x i64> %a0) { |
| 686 | ; SSE2-LABEL: trunc_packus_v8i64_v8i16: |
| 687 | ; SSE2: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 688 | ; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535] |
| 689 | ; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 690 | ; SSE2-NEXT: movdqa %xmm1, %xmm5 |
| 691 | ; SSE2-NEXT: pxor %xmm10, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 692 | ; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147549183,2147549183] |
| 693 | ; SSE2-NEXT: movdqa %xmm9, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 694 | ; SSE2-NEXT: pcmpgtd %xmm5, %xmm6 |
| 695 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 696 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 697 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| 698 | ; SSE2-NEXT: pand %xmm7, %xmm4 |
| 699 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3] |
| 700 | ; SSE2-NEXT: por %xmm4, %xmm5 |
| 701 | ; SSE2-NEXT: pand %xmm5, %xmm1 |
| 702 | ; SSE2-NEXT: pandn %xmm8, %xmm5 |
| 703 | ; SSE2-NEXT: por %xmm1, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 704 | ; SSE2-NEXT: movdqa %xmm0, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 705 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 706 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 707 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 708 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 709 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm1 |
| 710 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] |
| 711 | ; SSE2-NEXT: pand %xmm6, %xmm7 |
| 712 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] |
| 713 | ; SSE2-NEXT: por %xmm7, %xmm1 |
| 714 | ; SSE2-NEXT: pand %xmm1, %xmm0 |
| 715 | ; SSE2-NEXT: pandn %xmm8, %xmm1 |
| 716 | ; SSE2-NEXT: por %xmm0, %xmm1 |
| 717 | ; SSE2-NEXT: movdqa %xmm3, %xmm0 |
| 718 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 719 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 720 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm4 |
| 721 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 722 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 723 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 724 | ; SSE2-NEXT: pand %xmm6, %xmm0 |
| 725 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 726 | ; SSE2-NEXT: por %xmm0, %xmm6 |
| 727 | ; SSE2-NEXT: pand %xmm6, %xmm3 |
| 728 | ; SSE2-NEXT: pandn %xmm8, %xmm6 |
| 729 | ; SSE2-NEXT: por %xmm3, %xmm6 |
| 730 | ; SSE2-NEXT: movdqa %xmm2, %xmm0 |
| 731 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 732 | ; SSE2-NEXT: movdqa %xmm9, %xmm3 |
| 733 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm3 |
| 734 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 735 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 736 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 737 | ; SSE2-NEXT: pand %xmm4, %xmm0 |
| 738 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 739 | ; SSE2-NEXT: por %xmm0, %xmm3 |
| 740 | ; SSE2-NEXT: pand %xmm3, %xmm2 |
| 741 | ; SSE2-NEXT: pandn %xmm8, %xmm3 |
| 742 | ; SSE2-NEXT: por %xmm2, %xmm3 |
| 743 | ; SSE2-NEXT: movdqa %xmm3, %xmm0 |
| 744 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 745 | ; SSE2-NEXT: movdqa %xmm0, %xmm2 |
| 746 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm2 |
| 747 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 748 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm0 |
| 749 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 750 | ; SSE2-NEXT: pand %xmm4, %xmm7 |
| 751 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 752 | ; SSE2-NEXT: por %xmm7, %xmm0 |
| 753 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| 754 | ; SSE2-NEXT: movdqa %xmm6, %xmm2 |
| 755 | ; SSE2-NEXT: pxor %xmm10, %xmm2 |
| 756 | ; SSE2-NEXT: movdqa %xmm2, %xmm3 |
| 757 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm3 |
| 758 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 759 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm2 |
| 760 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] |
| 761 | ; SSE2-NEXT: pand %xmm4, %xmm7 |
| 762 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3] |
| 763 | ; SSE2-NEXT: por %xmm7, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 764 | ; SSE2-NEXT: pand %xmm6, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 765 | ; SSE2-NEXT: movdqa %xmm1, %xmm3 |
| 766 | ; SSE2-NEXT: pxor %xmm10, %xmm3 |
| 767 | ; SSE2-NEXT: movdqa %xmm3, %xmm4 |
| 768 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm4 |
| 769 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 770 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm3 |
| 771 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 772 | ; SSE2-NEXT: pand %xmm6, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 773 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 774 | ; SSE2-NEXT: por %xmm3, %xmm4 |
| 775 | ; SSE2-NEXT: pand %xmm1, %xmm4 |
| 776 | ; SSE2-NEXT: movdqa %xmm5, %xmm1 |
| 777 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 778 | ; SSE2-NEXT: movdqa %xmm1, %xmm3 |
| 779 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm3 |
| 780 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] |
| 781 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm1 |
| 782 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 783 | ; SSE2-NEXT: pand %xmm6, %xmm1 |
| 784 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 785 | ; SSE2-NEXT: por %xmm1, %xmm3 |
| 786 | ; SSE2-NEXT: pand %xmm5, %xmm3 |
| 787 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 788 | ; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 789 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3] |
| 790 | ; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7] |
| 791 | ; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] |
| 792 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3] |
| 793 | ; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7] |
| 794 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 795 | ; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 796 | ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 797 | ; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 798 | ; SSE2-NEXT: retq |
| 799 | ; |
| 800 | ; SSSE3-LABEL: trunc_packus_v8i64_v8i16: |
| 801 | ; SSSE3: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 802 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535] |
| 803 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 804 | ; SSSE3-NEXT: movdqa %xmm1, %xmm5 |
| 805 | ; SSSE3-NEXT: pxor %xmm10, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 806 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147549183,2147549183] |
| 807 | ; SSSE3-NEXT: movdqa %xmm9, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 808 | ; SSSE3-NEXT: pcmpgtd %xmm5, %xmm6 |
| 809 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 810 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 811 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| 812 | ; SSSE3-NEXT: pand %xmm7, %xmm4 |
| 813 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3] |
| 814 | ; SSSE3-NEXT: por %xmm4, %xmm5 |
| 815 | ; SSSE3-NEXT: pand %xmm5, %xmm1 |
| 816 | ; SSSE3-NEXT: pandn %xmm8, %xmm5 |
| 817 | ; SSSE3-NEXT: por %xmm1, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 818 | ; SSSE3-NEXT: movdqa %xmm0, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 819 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 820 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 821 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 822 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 823 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1 |
| 824 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] |
| 825 | ; SSSE3-NEXT: pand %xmm6, %xmm7 |
| 826 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] |
| 827 | ; SSSE3-NEXT: por %xmm7, %xmm1 |
| 828 | ; SSSE3-NEXT: pand %xmm1, %xmm0 |
| 829 | ; SSSE3-NEXT: pandn %xmm8, %xmm1 |
| 830 | ; SSSE3-NEXT: por %xmm0, %xmm1 |
| 831 | ; SSSE3-NEXT: movdqa %xmm3, %xmm0 |
| 832 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 833 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 834 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4 |
| 835 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 836 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 837 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 838 | ; SSSE3-NEXT: pand %xmm6, %xmm0 |
| 839 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 840 | ; SSSE3-NEXT: por %xmm0, %xmm6 |
| 841 | ; SSSE3-NEXT: pand %xmm6, %xmm3 |
| 842 | ; SSSE3-NEXT: pandn %xmm8, %xmm6 |
| 843 | ; SSSE3-NEXT: por %xmm3, %xmm6 |
| 844 | ; SSSE3-NEXT: movdqa %xmm2, %xmm0 |
| 845 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 846 | ; SSSE3-NEXT: movdqa %xmm9, %xmm3 |
| 847 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm3 |
| 848 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 849 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 850 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 851 | ; SSSE3-NEXT: pand %xmm4, %xmm0 |
| 852 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 853 | ; SSSE3-NEXT: por %xmm0, %xmm3 |
| 854 | ; SSSE3-NEXT: pand %xmm3, %xmm2 |
| 855 | ; SSSE3-NEXT: pandn %xmm8, %xmm3 |
| 856 | ; SSSE3-NEXT: por %xmm2, %xmm3 |
| 857 | ; SSSE3-NEXT: movdqa %xmm3, %xmm0 |
| 858 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 859 | ; SSSE3-NEXT: movdqa %xmm0, %xmm2 |
| 860 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2 |
| 861 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 862 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0 |
| 863 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 864 | ; SSSE3-NEXT: pand %xmm4, %xmm7 |
| 865 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 866 | ; SSSE3-NEXT: por %xmm7, %xmm0 |
| 867 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| 868 | ; SSSE3-NEXT: movdqa %xmm6, %xmm2 |
| 869 | ; SSSE3-NEXT: pxor %xmm10, %xmm2 |
| 870 | ; SSSE3-NEXT: movdqa %xmm2, %xmm3 |
| 871 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3 |
| 872 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 873 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm2 |
| 874 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] |
| 875 | ; SSSE3-NEXT: pand %xmm4, %xmm7 |
| 876 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3] |
| 877 | ; SSSE3-NEXT: por %xmm7, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 878 | ; SSSE3-NEXT: pand %xmm6, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 879 | ; SSSE3-NEXT: movdqa %xmm1, %xmm3 |
| 880 | ; SSSE3-NEXT: pxor %xmm10, %xmm3 |
| 881 | ; SSSE3-NEXT: movdqa %xmm3, %xmm4 |
| 882 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm4 |
| 883 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 884 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm3 |
| 885 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 886 | ; SSSE3-NEXT: pand %xmm6, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 887 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 888 | ; SSSE3-NEXT: por %xmm3, %xmm4 |
| 889 | ; SSSE3-NEXT: pand %xmm1, %xmm4 |
| 890 | ; SSSE3-NEXT: movdqa %xmm5, %xmm1 |
| 891 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 892 | ; SSSE3-NEXT: movdqa %xmm1, %xmm3 |
| 893 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3 |
| 894 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] |
| 895 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1 |
| 896 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 897 | ; SSSE3-NEXT: pand %xmm6, %xmm1 |
| 898 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 899 | ; SSSE3-NEXT: por %xmm1, %xmm3 |
| 900 | ; SSSE3-NEXT: pand %xmm5, %xmm3 |
| 901 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 902 | ; SSSE3-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 903 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3] |
| 904 | ; SSSE3-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7] |
| 905 | ; SSSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] |
| 906 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3] |
| 907 | ; SSSE3-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7] |
| 908 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 909 | ; SSSE3-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 910 | ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 911 | ; SSSE3-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 912 | ; SSSE3-NEXT: retq |
| 913 | ; |
| 914 | ; SSE41-LABEL: trunc_packus_v8i64_v8i16: |
| 915 | ; SSE41: # %bb.0: |
| 916 | ; SSE41-NEXT: movdqa %xmm0, %xmm8 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 917 | ; SSE41-NEXT: movapd {{.*#+}} xmm7 = [65535,65535] |
| 918 | ; SSE41-NEXT: movdqa {{.*#+}} xmm9 = [2147483648,0,2147483648,0] |
| 919 | ; SSE41-NEXT: movdqa %xmm3, %xmm0 |
| 920 | ; SSE41-NEXT: pxor %xmm9, %xmm0 |
| 921 | ; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147549183,2147549183] |
| 922 | ; SSE41-NEXT: movdqa %xmm10, %xmm6 |
| 923 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm6 |
| 924 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm6[0,0,2,2] |
| 925 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 926 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3] |
| 927 | ; SSE41-NEXT: pand %xmm5, %xmm4 |
| 928 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3] |
| 929 | ; SSE41-NEXT: por %xmm4, %xmm0 |
| 930 | ; SSE41-NEXT: movapd %xmm7, %xmm6 |
| 931 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm6 |
| 932 | ; SSE41-NEXT: movdqa %xmm2, %xmm0 |
| 933 | ; SSE41-NEXT: pxor %xmm9, %xmm0 |
| 934 | ; SSE41-NEXT: movdqa %xmm10, %xmm3 |
| 935 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm3 |
| 936 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 937 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 938 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 939 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 940 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 941 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 942 | ; SSE41-NEXT: movapd %xmm7, %xmm3 |
| 943 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm3 |
| 944 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| 945 | ; SSE41-NEXT: pxor %xmm9, %xmm0 |
| 946 | ; SSE41-NEXT: movdqa %xmm10, %xmm2 |
| 947 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm2 |
| 948 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 949 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 950 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 951 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 952 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 953 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 954 | ; SSE41-NEXT: movapd %xmm7, %xmm2 |
| 955 | ; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm2 |
| 956 | ; SSE41-NEXT: movdqa %xmm8, %xmm0 |
| 957 | ; SSE41-NEXT: pxor %xmm9, %xmm0 |
| 958 | ; SSE41-NEXT: movdqa %xmm10, %xmm1 |
| 959 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm1 |
| 960 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] |
| 961 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 962 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 963 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 964 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 965 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 966 | ; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm7 |
| 967 | ; SSE41-NEXT: movapd %xmm7, %xmm0 |
| 968 | ; SSE41-NEXT: xorpd %xmm9, %xmm0 |
| 969 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 970 | ; SSE41-NEXT: pcmpgtd %xmm9, %xmm1 |
| 971 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] |
| 972 | ; SSE41-NEXT: pcmpeqd %xmm9, %xmm0 |
| 973 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 974 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 975 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 976 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 977 | ; SSE41-NEXT: pxor %xmm8, %xmm8 |
| 978 | ; SSE41-NEXT: pxor %xmm1, %xmm1 |
| 979 | ; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm1 |
| 980 | ; SSE41-NEXT: movapd %xmm2, %xmm0 |
| 981 | ; SSE41-NEXT: xorpd %xmm9, %xmm0 |
| 982 | ; SSE41-NEXT: movapd %xmm0, %xmm5 |
| 983 | ; SSE41-NEXT: pcmpgtd %xmm9, %xmm5 |
| 984 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] |
| 985 | ; SSE41-NEXT: pcmpeqd %xmm9, %xmm0 |
| 986 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3] |
| 987 | ; SSE41-NEXT: pand %xmm7, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 988 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] |
| 989 | ; SSE41-NEXT: por %xmm4, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 990 | ; SSE41-NEXT: pxor %xmm7, %xmm7 |
| 991 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm7 |
| 992 | ; SSE41-NEXT: movapd %xmm3, %xmm0 |
| 993 | ; SSE41-NEXT: xorpd %xmm9, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 994 | ; SSE41-NEXT: movapd %xmm0, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 995 | ; SSE41-NEXT: pcmpgtd %xmm9, %xmm2 |
| 996 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 997 | ; SSE41-NEXT: pcmpeqd %xmm9, %xmm0 |
| 998 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 999 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 1000 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 1001 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 1002 | ; SSE41-NEXT: pxor %xmm2, %xmm2 |
| 1003 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm2 |
| 1004 | ; SSE41-NEXT: movapd %xmm6, %xmm0 |
| 1005 | ; SSE41-NEXT: xorpd %xmm9, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1006 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1007 | ; SSE41-NEXT: pcmpgtd %xmm9, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1008 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1009 | ; SSE41-NEXT: pcmpeqd %xmm9, %xmm0 |
| 1010 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 1011 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 1012 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 1013 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 1014 | ; SSE41-NEXT: pxor %xmm3, %xmm3 |
| 1015 | ; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm3 |
| 1016 | ; SSE41-NEXT: pblendw {{.*#+}} xmm3 = xmm3[0],xmm8[1,2,3],xmm3[4],xmm8[5,6,7] |
| 1017 | ; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0],xmm8[1,2,3],xmm2[4],xmm8[5,6,7] |
| 1018 | ; SSE41-NEXT: packusdw %xmm3, %xmm2 |
| 1019 | ; SSE41-NEXT: pblendw {{.*#+}} xmm7 = xmm7[0],xmm8[1,2,3],xmm7[4],xmm8[5,6,7] |
| 1020 | ; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0],xmm8[1,2,3],xmm1[4],xmm8[5,6,7] |
| 1021 | ; SSE41-NEXT: packusdw %xmm7, %xmm1 |
| 1022 | ; SSE41-NEXT: packusdw %xmm2, %xmm1 |
| 1023 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1024 | ; SSE41-NEXT: retq |
| 1025 | ; |
| 1026 | ; AVX1-LABEL: trunc_packus_v8i64_v8i16: |
| 1027 | ; AVX1: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1028 | ; AVX1-NEXT: vmovapd {{.*#+}} ymm2 = [65535,65535,65535,65535] |
| 1029 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3 |
| 1030 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [65535,65535] |
| 1031 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 1032 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm4, %xmm5 |
| 1033 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3 |
| 1034 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| 1035 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3 |
| 1036 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 1037 | ; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm4 |
| 1038 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3 |
| 1039 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1040 | ; AVX1-NEXT: vxorpd %xmm2, %xmm2, %xmm2 |
| 1041 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm8 |
| 1042 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4 |
| 1043 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm4, %xmm5 |
| 1044 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm6 |
| 1045 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm7 |
| 1046 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm7, %xmm3 |
| 1047 | ; AVX1-NEXT: vpand %xmm7, %xmm3, %xmm3 |
| 1048 | ; AVX1-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1,2,3],xmm3[4],xmm2[5,6,7] |
| 1049 | ; AVX1-NEXT: vpand %xmm1, %xmm6, %xmm1 |
| 1050 | ; AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3],xmm1[4],xmm2[5,6,7] |
| 1051 | ; AVX1-NEXT: vpackusdw %xmm3, %xmm1, %xmm1 |
| 1052 | ; AVX1-NEXT: vpand %xmm4, %xmm5, %xmm3 |
| 1053 | ; AVX1-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1,2,3],xmm3[4],xmm2[5,6,7] |
| 1054 | ; AVX1-NEXT: vpand %xmm0, %xmm8, %xmm0 |
| 1055 | ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3],xmm0[4],xmm2[5,6,7] |
| 1056 | ; AVX1-NEXT: vpackusdw %xmm3, %xmm0, %xmm0 |
| 1057 | ; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0 |
| 1058 | ; AVX1-NEXT: vzeroupper |
| 1059 | ; AVX1-NEXT: retq |
| 1060 | ; |
| 1061 | ; AVX2-SLOW-LABEL: trunc_packus_v8i64_v8i16: |
| 1062 | ; AVX2-SLOW: # %bb.0: |
| 1063 | ; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm2 = [65535,65535,65535,65535] |
| 1064 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1065 | ; AVX2-SLOW-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1066 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3 |
| 1067 | ; AVX2-SLOW-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1068 | ; AVX2-SLOW-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1069 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3 |
| 1070 | ; AVX2-SLOW-NEXT: vpand %ymm1, %ymm3, %ymm1 |
| 1071 | ; AVX2-SLOW-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2 |
| 1072 | ; AVX2-SLOW-NEXT: vpand %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1073 | ; AVX2-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] |
| 1074 | ; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] |
| 1075 | ; AVX2-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] |
| 1076 | ; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3] |
| 1077 | ; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| 1078 | ; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,24,25,28,29,28,29,30,31] |
| 1079 | ; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 1080 | ; AVX2-SLOW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1081 | ; AVX2-SLOW-NEXT: vzeroupper |
| 1082 | ; AVX2-SLOW-NEXT: retq |
| 1083 | ; |
| 1084 | ; AVX2-FAST-LABEL: trunc_packus_v8i64_v8i16: |
| 1085 | ; AVX2-FAST: # %bb.0: |
| 1086 | ; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm2 = [65535,65535,65535,65535] |
| 1087 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1088 | ; AVX2-FAST-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1089 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3 |
| 1090 | ; AVX2-FAST-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1091 | ; AVX2-FAST-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1092 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3 |
| 1093 | ; AVX2-FAST-NEXT: vpand %ymm1, %ymm3, %ymm1 |
| 1094 | ; AVX2-FAST-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2 |
| 1095 | ; AVX2-FAST-NEXT: vpand %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1096 | ; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [0,2,4,6,4,6,6,7] |
| 1097 | ; AVX2-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0 |
| 1098 | ; AVX2-FAST-NEXT: vpermd %ymm1, %ymm2, %ymm1 |
| 1099 | ; AVX2-FAST-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| 1100 | ; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,24,25,28,29,28,29,30,31] |
| 1101 | ; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 1102 | ; AVX2-FAST-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1103 | ; AVX2-FAST-NEXT: vzeroupper |
| 1104 | ; AVX2-FAST-NEXT: retq |
| 1105 | ; |
| 1106 | ; AVX512-LABEL: trunc_packus_v8i64_v8i16: |
| 1107 | ; AVX512: # %bb.0: |
| 1108 | ; AVX512-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0 |
| 1109 | ; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1110 | ; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 |
| 1111 | ; AVX512-NEXT: vpmovqw %zmm0, %xmm0 |
| 1112 | ; AVX512-NEXT: vzeroupper |
| 1113 | ; AVX512-NEXT: retq |
| 1114 | %1 = icmp slt <8 x i64> %a0, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> |
| 1115 | %2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> |
| 1116 | %3 = icmp sgt <8 x i64> %2, zeroinitializer |
| 1117 | %4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer |
| 1118 | %5 = trunc <8 x i64> %4 to <8 x i16> |
| 1119 | ret <8 x i16> %5 |
| 1120 | } |
| 1121 | |
| 1122 | define <8 x i16> @trunc_packus_v8i32_v8i16(<8 x i32> %a0) { |
| 1123 | ; SSE2-LABEL: trunc_packus_v8i32_v8i16: |
| 1124 | ; SSE2: # %bb.0: |
| 1125 | ; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535] |
| 1126 | ; SSE2-NEXT: movdqa %xmm2, %xmm3 |
| 1127 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1128 | ; SSE2-NEXT: pand %xmm3, %xmm1 |
| 1129 | ; SSE2-NEXT: pandn %xmm2, %xmm3 |
| 1130 | ; SSE2-NEXT: por %xmm1, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1131 | ; SSE2-NEXT: movdqa %xmm2, %xmm1 |
| 1132 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm1 |
| 1133 | ; SSE2-NEXT: pand %xmm1, %xmm0 |
| 1134 | ; SSE2-NEXT: pandn %xmm2, %xmm1 |
| 1135 | ; SSE2-NEXT: por %xmm0, %xmm1 |
| 1136 | ; SSE2-NEXT: pxor %xmm2, %xmm2 |
| 1137 | ; SSE2-NEXT: movdqa %xmm1, %xmm0 |
| 1138 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm0 |
| 1139 | ; SSE2-NEXT: pand %xmm1, %xmm0 |
| 1140 | ; SSE2-NEXT: movdqa %xmm3, %xmm1 |
| 1141 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm1 |
| 1142 | ; SSE2-NEXT: pand %xmm3, %xmm1 |
| 1143 | ; SSE2-NEXT: pslld $16, %xmm1 |
| 1144 | ; SSE2-NEXT: psrad $16, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1145 | ; SSE2-NEXT: pslld $16, %xmm0 |
| 1146 | ; SSE2-NEXT: psrad $16, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1147 | ; SSE2-NEXT: packssdw %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1148 | ; SSE2-NEXT: retq |
| 1149 | ; |
| 1150 | ; SSSE3-LABEL: trunc_packus_v8i32_v8i16: |
| 1151 | ; SSSE3: # %bb.0: |
| 1152 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535] |
| 1153 | ; SSSE3-NEXT: movdqa %xmm2, %xmm3 |
| 1154 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1155 | ; SSSE3-NEXT: pand %xmm3, %xmm1 |
| 1156 | ; SSSE3-NEXT: pandn %xmm2, %xmm3 |
| 1157 | ; SSSE3-NEXT: por %xmm1, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1158 | ; SSSE3-NEXT: movdqa %xmm2, %xmm1 |
| 1159 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm1 |
| 1160 | ; SSSE3-NEXT: pand %xmm1, %xmm0 |
| 1161 | ; SSSE3-NEXT: pandn %xmm2, %xmm1 |
| 1162 | ; SSSE3-NEXT: por %xmm0, %xmm1 |
| 1163 | ; SSSE3-NEXT: pxor %xmm2, %xmm2 |
| 1164 | ; SSSE3-NEXT: movdqa %xmm1, %xmm0 |
| 1165 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm0 |
| 1166 | ; SSSE3-NEXT: pand %xmm1, %xmm0 |
| 1167 | ; SSSE3-NEXT: movdqa %xmm3, %xmm1 |
| 1168 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1 |
| 1169 | ; SSSE3-NEXT: pand %xmm3, %xmm1 |
| 1170 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] |
| 1171 | ; SSSE3-NEXT: pshufb %xmm2, %xmm1 |
| 1172 | ; SSSE3-NEXT: pshufb %xmm2, %xmm0 |
| 1173 | ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1174 | ; SSSE3-NEXT: retq |
| 1175 | ; |
| 1176 | ; SSE41-LABEL: trunc_packus_v8i32_v8i16: |
| 1177 | ; SSE41: # %bb.0: |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 1178 | ; SSE41-NEXT: packusdw %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1179 | ; SSE41-NEXT: retq |
| 1180 | ; |
| 1181 | ; AVX1-LABEL: trunc_packus_v8i32_v8i16: |
| 1182 | ; AVX1: # %bb.0: |
| 1183 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | b4e789e | 2018-02-07 15:48:44 +0000 | [diff] [blame] | 1184 | ; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1185 | ; AVX1-NEXT: vzeroupper |
| 1186 | ; AVX1-NEXT: retq |
| 1187 | ; |
| 1188 | ; AVX2-LABEL: trunc_packus_v8i32_v8i16: |
| 1189 | ; AVX2: # %bb.0: |
| Simon Pilgrim | b4e789e | 2018-02-07 15:48:44 +0000 | [diff] [blame] | 1190 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1191 | ; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1192 | ; AVX2-NEXT: vzeroupper |
| 1193 | ; AVX2-NEXT: retq |
| 1194 | ; |
| 1195 | ; AVX512F-LABEL: trunc_packus_v8i32_v8i16: |
| 1196 | ; AVX512F: # %bb.0: |
| 1197 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} ymm1 = [65535,65535,65535,65535,65535,65535,65535,65535] |
| 1198 | ; AVX512F-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 1199 | ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1200 | ; AVX512F-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 1201 | ; AVX512F-NEXT: vpmovdw %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 1202 | ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1203 | ; AVX512F-NEXT: vzeroupper |
| 1204 | ; AVX512F-NEXT: retq |
| 1205 | ; |
| 1206 | ; AVX512VL-LABEL: trunc_packus_v8i32_v8i16: |
| 1207 | ; AVX512VL: # %bb.0: |
| 1208 | ; AVX512VL-NEXT: vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 |
| 1209 | ; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1210 | ; AVX512VL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 1211 | ; AVX512VL-NEXT: vpmovdw %ymm0, %xmm0 |
| 1212 | ; AVX512VL-NEXT: vzeroupper |
| 1213 | ; AVX512VL-NEXT: retq |
| 1214 | ; |
| 1215 | ; AVX512BW-LABEL: trunc_packus_v8i32_v8i16: |
| 1216 | ; AVX512BW: # %bb.0: |
| 1217 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm1 = [65535,65535,65535,65535,65535,65535,65535,65535] |
| 1218 | ; AVX512BW-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 1219 | ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1220 | ; AVX512BW-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 1221 | ; AVX512BW-NEXT: vpmovdw %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 1222 | ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1223 | ; AVX512BW-NEXT: vzeroupper |
| 1224 | ; AVX512BW-NEXT: retq |
| 1225 | ; |
| 1226 | ; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i16: |
| 1227 | ; AVX512BWVL: # %bb.0: |
| 1228 | ; AVX512BWVL-NEXT: vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 |
| 1229 | ; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1230 | ; AVX512BWVL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 1231 | ; AVX512BWVL-NEXT: vpmovdw %ymm0, %xmm0 |
| 1232 | ; AVX512BWVL-NEXT: vzeroupper |
| 1233 | ; AVX512BWVL-NEXT: retq |
| 1234 | %1 = icmp slt <8 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> |
| 1235 | %2 = select <8 x i1> %1, <8 x i32> %a0, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> |
| 1236 | %3 = icmp sgt <8 x i32> %2, zeroinitializer |
| 1237 | %4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer |
| 1238 | %5 = trunc <8 x i32> %4 to <8 x i16> |
| 1239 | ret <8 x i16> %5 |
| 1240 | } |
| 1241 | |
| 1242 | define <16 x i16> @trunc_packus_v16i32_v16i16(<16 x i32> %a0) { |
| 1243 | ; SSE2-LABEL: trunc_packus_v16i32_v16i16: |
| 1244 | ; SSE2: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1245 | ; SSE2-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535] |
| 1246 | ; SSE2-NEXT: movdqa %xmm6, %xmm4 |
| 1247 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm4 |
| 1248 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| 1249 | ; SSE2-NEXT: pandn %xmm6, %xmm4 |
| 1250 | ; SSE2-NEXT: por %xmm1, %xmm4 |
| 1251 | ; SSE2-NEXT: movdqa %xmm6, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1252 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1253 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1254 | ; SSE2-NEXT: pandn %xmm6, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1255 | ; SSE2-NEXT: por %xmm0, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1256 | ; SSE2-NEXT: movdqa %xmm6, %xmm0 |
| 1257 | ; SSE2-NEXT: pcmpgtd %xmm3, %xmm0 |
| 1258 | ; SSE2-NEXT: pand %xmm0, %xmm3 |
| 1259 | ; SSE2-NEXT: pandn %xmm6, %xmm0 |
| 1260 | ; SSE2-NEXT: por %xmm3, %xmm0 |
| 1261 | ; SSE2-NEXT: movdqa %xmm6, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1262 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1263 | ; SSE2-NEXT: pand %xmm3, %xmm2 |
| 1264 | ; SSE2-NEXT: pandn %xmm6, %xmm3 |
| 1265 | ; SSE2-NEXT: por %xmm2, %xmm3 |
| 1266 | ; SSE2-NEXT: pxor %xmm2, %xmm2 |
| 1267 | ; SSE2-NEXT: movdqa %xmm3, %xmm1 |
| 1268 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm1 |
| 1269 | ; SSE2-NEXT: pand %xmm3, %xmm1 |
| 1270 | ; SSE2-NEXT: movdqa %xmm0, %xmm3 |
| 1271 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm3 |
| 1272 | ; SSE2-NEXT: pand %xmm0, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1273 | ; SSE2-NEXT: movdqa %xmm5, %xmm0 |
| 1274 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1275 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1276 | ; SSE2-NEXT: movdqa %xmm4, %xmm5 |
| 1277 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm5 |
| 1278 | ; SSE2-NEXT: pand %xmm4, %xmm5 |
| 1279 | ; SSE2-NEXT: pslld $16, %xmm5 |
| 1280 | ; SSE2-NEXT: psrad $16, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1281 | ; SSE2-NEXT: pslld $16, %xmm0 |
| 1282 | ; SSE2-NEXT: psrad $16, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1283 | ; SSE2-NEXT: packssdw %xmm5, %xmm0 |
| 1284 | ; SSE2-NEXT: pslld $16, %xmm3 |
| 1285 | ; SSE2-NEXT: psrad $16, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1286 | ; SSE2-NEXT: pslld $16, %xmm1 |
| 1287 | ; SSE2-NEXT: psrad $16, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1288 | ; SSE2-NEXT: packssdw %xmm3, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1289 | ; SSE2-NEXT: retq |
| 1290 | ; |
| 1291 | ; SSSE3-LABEL: trunc_packus_v16i32_v16i16: |
| 1292 | ; SSSE3: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1293 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535] |
| 1294 | ; SSSE3-NEXT: movdqa %xmm6, %xmm4 |
| 1295 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4 |
| 1296 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| 1297 | ; SSSE3-NEXT: pandn %xmm6, %xmm4 |
| 1298 | ; SSSE3-NEXT: por %xmm1, %xmm4 |
| 1299 | ; SSSE3-NEXT: movdqa %xmm6, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1300 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1301 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1302 | ; SSSE3-NEXT: pandn %xmm6, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1303 | ; SSSE3-NEXT: por %xmm0, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1304 | ; SSSE3-NEXT: movdqa %xmm6, %xmm0 |
| 1305 | ; SSSE3-NEXT: pcmpgtd %xmm3, %xmm0 |
| 1306 | ; SSSE3-NEXT: pand %xmm0, %xmm3 |
| 1307 | ; SSSE3-NEXT: pandn %xmm6, %xmm0 |
| 1308 | ; SSSE3-NEXT: por %xmm3, %xmm0 |
| 1309 | ; SSSE3-NEXT: movdqa %xmm6, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1310 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1311 | ; SSSE3-NEXT: pand %xmm3, %xmm2 |
| 1312 | ; SSSE3-NEXT: pandn %xmm6, %xmm3 |
| 1313 | ; SSSE3-NEXT: por %xmm2, %xmm3 |
| 1314 | ; SSSE3-NEXT: pxor %xmm2, %xmm2 |
| 1315 | ; SSSE3-NEXT: movdqa %xmm3, %xmm1 |
| 1316 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1 |
| 1317 | ; SSSE3-NEXT: pand %xmm3, %xmm1 |
| 1318 | ; SSSE3-NEXT: movdqa %xmm0, %xmm3 |
| 1319 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm3 |
| 1320 | ; SSSE3-NEXT: pand %xmm0, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1321 | ; SSSE3-NEXT: movdqa %xmm5, %xmm0 |
| 1322 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1323 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1324 | ; SSSE3-NEXT: movdqa %xmm4, %xmm5 |
| 1325 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm5 |
| 1326 | ; SSSE3-NEXT: pand %xmm4, %xmm5 |
| 1327 | ; SSSE3-NEXT: pslld $16, %xmm5 |
| 1328 | ; SSSE3-NEXT: psrad $16, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1329 | ; SSSE3-NEXT: pslld $16, %xmm0 |
| 1330 | ; SSSE3-NEXT: psrad $16, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1331 | ; SSSE3-NEXT: packssdw %xmm5, %xmm0 |
| 1332 | ; SSSE3-NEXT: pslld $16, %xmm3 |
| 1333 | ; SSSE3-NEXT: psrad $16, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1334 | ; SSSE3-NEXT: pslld $16, %xmm1 |
| 1335 | ; SSSE3-NEXT: psrad $16, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1336 | ; SSSE3-NEXT: packssdw %xmm3, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1337 | ; SSSE3-NEXT: retq |
| 1338 | ; |
| 1339 | ; SSE41-LABEL: trunc_packus_v16i32_v16i16: |
| 1340 | ; SSE41: # %bb.0: |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1341 | ; SSE41-NEXT: packusdw %xmm1, %xmm0 |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 1342 | ; SSE41-NEXT: packusdw %xmm3, %xmm2 |
| 1343 | ; SSE41-NEXT: movdqa %xmm2, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1344 | ; SSE41-NEXT: retq |
| 1345 | ; |
| 1346 | ; AVX1-LABEL: trunc_packus_v16i32_v16i16: |
| 1347 | ; AVX1: # %bb.0: |
| 1348 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1349 | ; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1 |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 1350 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1351 | ; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0 |
| 1352 | ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 |
| 1353 | ; AVX1-NEXT: retq |
| 1354 | ; |
| 1355 | ; AVX2-LABEL: trunc_packus_v16i32_v16i16: |
| 1356 | ; AVX2: # %bb.0: |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 1357 | ; AVX2-NEXT: vpackusdw %ymm1, %ymm0, %ymm0 |
| 1358 | ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1359 | ; AVX2-NEXT: retq |
| 1360 | ; |
| 1361 | ; AVX512-LABEL: trunc_packus_v16i32_v16i16: |
| 1362 | ; AVX512: # %bb.0: |
| 1363 | ; AVX512-NEXT: vpminsd {{.*}}(%rip){1to16}, %zmm0, %zmm0 |
| 1364 | ; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1365 | ; AVX512-NEXT: vpmaxsd %zmm1, %zmm0, %zmm0 |
| 1366 | ; AVX512-NEXT: vpmovdw %zmm0, %ymm0 |
| 1367 | ; AVX512-NEXT: retq |
| 1368 | %1 = icmp slt <16 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> |
| 1369 | %2 = select <16 x i1> %1, <16 x i32> %a0, <16 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> |
| 1370 | %3 = icmp sgt <16 x i32> %2, zeroinitializer |
| 1371 | %4 = select <16 x i1> %3, <16 x i32> %2, <16 x i32> zeroinitializer |
| 1372 | %5 = trunc <16 x i32> %4 to <16 x i16> |
| 1373 | ret <16 x i16> %5 |
| 1374 | } |
| 1375 | |
| 1376 | ; |
| 1377 | ; PACKUS saturation truncation to v16i8 |
| 1378 | ; |
| 1379 | |
| 1380 | define <8 x i8> @trunc_packus_v8i64_v8i8(<8 x i64> %a0) { |
| 1381 | ; SSE2-LABEL: trunc_packus_v8i64_v8i8: |
| 1382 | ; SSE2: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1383 | ; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,255] |
| 1384 | ; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1385 | ; SSE2-NEXT: movdqa %xmm2, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1386 | ; SSE2-NEXT: pxor %xmm10, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1387 | ; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1388 | ; SSE2-NEXT: movdqa %xmm9, %xmm7 |
| 1389 | ; SSE2-NEXT: pcmpgtd %xmm5, %xmm7 |
| 1390 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1391 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1392 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1393 | ; SSE2-NEXT: pand %xmm6, %xmm4 |
| 1394 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1395 | ; SSE2-NEXT: por %xmm4, %xmm5 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1396 | ; SSE2-NEXT: pand %xmm5, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1397 | ; SSE2-NEXT: pandn %xmm8, %xmm5 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1398 | ; SSE2-NEXT: por %xmm2, %xmm5 |
| 1399 | ; SSE2-NEXT: movdqa %xmm3, %xmm2 |
| 1400 | ; SSE2-NEXT: pxor %xmm10, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1401 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1402 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1403 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1404 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm2 |
| 1405 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1406 | ; SSE2-NEXT: pand %xmm6, %xmm7 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1407 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3] |
| 1408 | ; SSE2-NEXT: por %xmm7, %xmm2 |
| 1409 | ; SSE2-NEXT: pand %xmm2, %xmm3 |
| 1410 | ; SSE2-NEXT: pandn %xmm8, %xmm2 |
| 1411 | ; SSE2-NEXT: por %xmm3, %xmm2 |
| 1412 | ; SSE2-NEXT: movdqa %xmm0, %xmm3 |
| 1413 | ; SSE2-NEXT: pxor %xmm10, %xmm3 |
| 1414 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 1415 | ; SSE2-NEXT: pcmpgtd %xmm3, %xmm4 |
| 1416 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 1417 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm3 |
| 1418 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3] |
| 1419 | ; SSE2-NEXT: pand %xmm6, %xmm7 |
| 1420 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3] |
| 1421 | ; SSE2-NEXT: por %xmm7, %xmm3 |
| 1422 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| 1423 | ; SSE2-NEXT: pandn %xmm8, %xmm3 |
| 1424 | ; SSE2-NEXT: por %xmm0, %xmm3 |
| 1425 | ; SSE2-NEXT: movdqa %xmm1, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1426 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 1427 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 1428 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm4 |
| 1429 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 1430 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 1431 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 1432 | ; SSE2-NEXT: pand %xmm6, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1433 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 1434 | ; SSE2-NEXT: por %xmm0, %xmm4 |
| 1435 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| 1436 | ; SSE2-NEXT: pandn %xmm8, %xmm4 |
| 1437 | ; SSE2-NEXT: por %xmm1, %xmm4 |
| 1438 | ; SSE2-NEXT: movdqa %xmm4, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1439 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1440 | ; SSE2-NEXT: movdqa %xmm0, %xmm1 |
| 1441 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm1 |
| 1442 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2] |
| 1443 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1444 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1445 | ; SSE2-NEXT: pand %xmm6, %xmm0 |
| 1446 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 1447 | ; SSE2-NEXT: por %xmm0, %xmm1 |
| 1448 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1449 | ; SSE2-NEXT: movdqa %xmm3, %xmm0 |
| 1450 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1451 | ; SSE2-NEXT: movdqa %xmm0, %xmm4 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1452 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm4 |
| 1453 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1454 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1455 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 1456 | ; SSE2-NEXT: pand %xmm6, %xmm7 |
| 1457 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 1458 | ; SSE2-NEXT: por %xmm7, %xmm0 |
| 1459 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| 1460 | ; SSE2-NEXT: packssdw %xmm1, %xmm0 |
| 1461 | ; SSE2-NEXT: movdqa %xmm2, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1462 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 1463 | ; SSE2-NEXT: movdqa %xmm1, %xmm3 |
| 1464 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1465 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1466 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm1 |
| 1467 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1468 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1469 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 1470 | ; SSE2-NEXT: por %xmm1, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1471 | ; SSE2-NEXT: pand %xmm2, %xmm3 |
| 1472 | ; SSE2-NEXT: movdqa %xmm5, %xmm1 |
| 1473 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 1474 | ; SSE2-NEXT: movdqa %xmm1, %xmm2 |
| 1475 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm2 |
| 1476 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 1477 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm1 |
| 1478 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 1479 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| 1480 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| 1481 | ; SSE2-NEXT: por %xmm1, %xmm2 |
| 1482 | ; SSE2-NEXT: pand %xmm5, %xmm2 |
| 1483 | ; SSE2-NEXT: packssdw %xmm3, %xmm2 |
| 1484 | ; SSE2-NEXT: packssdw %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1485 | ; SSE2-NEXT: retq |
| 1486 | ; |
| 1487 | ; SSSE3-LABEL: trunc_packus_v8i64_v8i8: |
| 1488 | ; SSSE3: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1489 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [255,255] |
| 1490 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1491 | ; SSSE3-NEXT: movdqa %xmm2, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1492 | ; SSSE3-NEXT: pxor %xmm10, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1493 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1494 | ; SSSE3-NEXT: movdqa %xmm9, %xmm7 |
| 1495 | ; SSSE3-NEXT: pcmpgtd %xmm5, %xmm7 |
| 1496 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1497 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1498 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1499 | ; SSSE3-NEXT: pand %xmm6, %xmm4 |
| 1500 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1501 | ; SSSE3-NEXT: por %xmm4, %xmm5 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1502 | ; SSSE3-NEXT: pand %xmm5, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1503 | ; SSSE3-NEXT: pandn %xmm8, %xmm5 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1504 | ; SSSE3-NEXT: por %xmm2, %xmm5 |
| 1505 | ; SSSE3-NEXT: movdqa %xmm3, %xmm2 |
| 1506 | ; SSSE3-NEXT: pxor %xmm10, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1507 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1508 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1509 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1510 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm2 |
| 1511 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1512 | ; SSSE3-NEXT: pand %xmm6, %xmm7 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1513 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3] |
| 1514 | ; SSSE3-NEXT: por %xmm7, %xmm2 |
| 1515 | ; SSSE3-NEXT: pand %xmm2, %xmm3 |
| 1516 | ; SSSE3-NEXT: pandn %xmm8, %xmm2 |
| 1517 | ; SSSE3-NEXT: por %xmm3, %xmm2 |
| 1518 | ; SSSE3-NEXT: movdqa %xmm0, %xmm3 |
| 1519 | ; SSSE3-NEXT: pxor %xmm10, %xmm3 |
| 1520 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 1521 | ; SSSE3-NEXT: pcmpgtd %xmm3, %xmm4 |
| 1522 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 1523 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm3 |
| 1524 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3] |
| 1525 | ; SSSE3-NEXT: pand %xmm6, %xmm7 |
| 1526 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3] |
| 1527 | ; SSSE3-NEXT: por %xmm7, %xmm3 |
| 1528 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| 1529 | ; SSSE3-NEXT: pandn %xmm8, %xmm3 |
| 1530 | ; SSSE3-NEXT: por %xmm0, %xmm3 |
| 1531 | ; SSSE3-NEXT: movdqa %xmm1, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1532 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 1533 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 1534 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4 |
| 1535 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 1536 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 1537 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 1538 | ; SSSE3-NEXT: pand %xmm6, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1539 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 1540 | ; SSSE3-NEXT: por %xmm0, %xmm4 |
| 1541 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| 1542 | ; SSSE3-NEXT: pandn %xmm8, %xmm4 |
| 1543 | ; SSSE3-NEXT: por %xmm1, %xmm4 |
| 1544 | ; SSSE3-NEXT: movdqa %xmm4, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1545 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1546 | ; SSSE3-NEXT: movdqa %xmm0, %xmm1 |
| 1547 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm1 |
| 1548 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2] |
| 1549 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1550 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1551 | ; SSSE3-NEXT: pand %xmm6, %xmm0 |
| 1552 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 1553 | ; SSSE3-NEXT: por %xmm0, %xmm1 |
| 1554 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1555 | ; SSSE3-NEXT: movdqa %xmm3, %xmm0 |
| 1556 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1557 | ; SSSE3-NEXT: movdqa %xmm0, %xmm4 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1558 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm4 |
| 1559 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1560 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1561 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 1562 | ; SSSE3-NEXT: pand %xmm6, %xmm7 |
| 1563 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 1564 | ; SSSE3-NEXT: por %xmm7, %xmm0 |
| 1565 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| 1566 | ; SSSE3-NEXT: packssdw %xmm1, %xmm0 |
| 1567 | ; SSSE3-NEXT: movdqa %xmm2, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1568 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 1569 | ; SSSE3-NEXT: movdqa %xmm1, %xmm3 |
| 1570 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1571 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1572 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1 |
| 1573 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1574 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1575 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 1576 | ; SSSE3-NEXT: por %xmm1, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1577 | ; SSSE3-NEXT: pand %xmm2, %xmm3 |
| 1578 | ; SSSE3-NEXT: movdqa %xmm5, %xmm1 |
| 1579 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 1580 | ; SSSE3-NEXT: movdqa %xmm1, %xmm2 |
| 1581 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2 |
| 1582 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 1583 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1 |
| 1584 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 1585 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| 1586 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| 1587 | ; SSSE3-NEXT: por %xmm1, %xmm2 |
| 1588 | ; SSSE3-NEXT: pand %xmm5, %xmm2 |
| 1589 | ; SSSE3-NEXT: packssdw %xmm3, %xmm2 |
| 1590 | ; SSSE3-NEXT: packssdw %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1591 | ; SSSE3-NEXT: retq |
| 1592 | ; |
| 1593 | ; SSE41-LABEL: trunc_packus_v8i64_v8i8: |
| 1594 | ; SSE41: # %bb.0: |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1595 | ; SSE41-NEXT: movdqa %xmm0, %xmm9 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1596 | ; SSE41-NEXT: movapd {{.*#+}} xmm7 = [255,255] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1597 | ; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1598 | ; SSE41-NEXT: movdqa %xmm2, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1599 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 1600 | ; SSE41-NEXT: movdqa {{.*#+}} xmm11 = [2147483903,2147483903] |
| 1601 | ; SSE41-NEXT: movdqa %xmm11, %xmm4 |
| 1602 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm4 |
| 1603 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1604 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| 1605 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 1606 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 1607 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 1608 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 1609 | ; SSE41-NEXT: movapd %xmm7, %xmm8 |
| 1610 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm8 |
| 1611 | ; SSE41-NEXT: movdqa %xmm3, %xmm0 |
| 1612 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 1613 | ; SSE41-NEXT: movdqa %xmm11, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1614 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm2 |
| 1615 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1616 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1617 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 1618 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 1619 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 1620 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 1621 | ; SSE41-NEXT: movapd %xmm7, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1622 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm2 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1623 | ; SSE41-NEXT: movdqa %xmm9, %xmm0 |
| 1624 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 1625 | ; SSE41-NEXT: movdqa %xmm11, %xmm3 |
| 1626 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1627 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1628 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1629 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 1630 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 1631 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 1632 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1633 | ; SSE41-NEXT: movapd %xmm7, %xmm4 |
| 1634 | ; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm4 |
| 1635 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| 1636 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 1637 | ; SSE41-NEXT: movdqa %xmm11, %xmm3 |
| 1638 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm3 |
| 1639 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] |
| 1640 | ; SSE41-NEXT: pcmpeqd %xmm11, %xmm0 |
| 1641 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 1642 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 1643 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 1644 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 1645 | ; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm7 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1646 | ; SSE41-NEXT: pxor %xmm3, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1647 | ; SSE41-NEXT: movapd %xmm7, %xmm0 |
| 1648 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 1649 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 1650 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm1 |
| 1651 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2] |
| 1652 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1653 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 1654 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 1655 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 1656 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 1657 | ; SSE41-NEXT: pxor %xmm5, %xmm5 |
| 1658 | ; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm5 |
| 1659 | ; SSE41-NEXT: movapd %xmm4, %xmm0 |
| 1660 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 1661 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 1662 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm1 |
| 1663 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2] |
| 1664 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1665 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 1666 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 1667 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 1668 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 1669 | ; SSE41-NEXT: pxor %xmm1, %xmm1 |
| 1670 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm1 |
| 1671 | ; SSE41-NEXT: packssdw %xmm5, %xmm1 |
| 1672 | ; SSE41-NEXT: movapd %xmm2, %xmm0 |
| 1673 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 1674 | ; SSE41-NEXT: movapd %xmm0, %xmm4 |
| 1675 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm4 |
| 1676 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1677 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1678 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 1679 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 1680 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 1681 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 1682 | ; SSE41-NEXT: pxor %xmm4, %xmm4 |
| 1683 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4 |
| 1684 | ; SSE41-NEXT: movapd %xmm8, %xmm0 |
| 1685 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 1686 | ; SSE41-NEXT: movapd %xmm0, %xmm2 |
| 1687 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm2 |
| 1688 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2] |
| 1689 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1690 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 1691 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 1692 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 1693 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 1694 | ; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm3 |
| 1695 | ; SSE41-NEXT: packssdw %xmm4, %xmm3 |
| 1696 | ; SSE41-NEXT: packssdw %xmm3, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1697 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1698 | ; SSE41-NEXT: retq |
| 1699 | ; |
| 1700 | ; AVX1-LABEL: trunc_packus_v8i64_v8i8: |
| 1701 | ; AVX1: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1702 | ; AVX1-NEXT: vmovapd {{.*#+}} ymm2 = [255,255,255,255] |
| 1703 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3 |
| 1704 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [255,255] |
| 1705 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 1706 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm4, %xmm5 |
| 1707 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3 |
| 1708 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| 1709 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3 |
| 1710 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 1711 | ; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm4 |
| 1712 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3 |
| 1713 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1714 | ; AVX1-NEXT: vxorpd %xmm2, %xmm2, %xmm2 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1715 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1716 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4 |
| 1717 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm4, %xmm5 |
| 1718 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm6 |
| 1719 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm7 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1720 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm7, %xmm2 |
| 1721 | ; AVX1-NEXT: vpand %xmm7, %xmm2, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1722 | ; AVX1-NEXT: vpand %xmm1, %xmm6, %xmm1 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1723 | ; AVX1-NEXT: vpackssdw %xmm2, %xmm1, %xmm1 |
| 1724 | ; AVX1-NEXT: vpand %xmm4, %xmm5, %xmm2 |
| 1725 | ; AVX1-NEXT: vpand %xmm0, %xmm3, %xmm0 |
| 1726 | ; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0 |
| 1727 | ; AVX1-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1728 | ; AVX1-NEXT: vzeroupper |
| 1729 | ; AVX1-NEXT: retq |
| 1730 | ; |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 1731 | ; AVX2-LABEL: trunc_packus_v8i64_v8i8: |
| 1732 | ; AVX2: # %bb.0: |
| 1733 | ; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255] |
| 1734 | ; AVX2-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3 |
| 1735 | ; AVX2-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| 1736 | ; AVX2-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3 |
| 1737 | ; AVX2-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| 1738 | ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| 1739 | ; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3 |
| 1740 | ; AVX2-NEXT: vpand %ymm1, %ymm3, %ymm1 |
| 1741 | ; AVX2-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2 |
| 1742 | ; AVX2-NEXT: vpand %ymm0, %ymm2, %ymm0 |
| 1743 | ; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0 |
| 1744 | ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| 1745 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 1746 | ; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 |
| 1747 | ; AVX2-NEXT: vzeroupper |
| 1748 | ; AVX2-NEXT: retq |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1749 | ; |
| 1750 | ; AVX512-LABEL: trunc_packus_v8i64_v8i8: |
| 1751 | ; AVX512: # %bb.0: |
| 1752 | ; AVX512-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0 |
| 1753 | ; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 1754 | ; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 |
| 1755 | ; AVX512-NEXT: vpmovqw %zmm0, %xmm0 |
| 1756 | ; AVX512-NEXT: vzeroupper |
| 1757 | ; AVX512-NEXT: retq |
| 1758 | %1 = icmp slt <8 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> |
| 1759 | %2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> |
| 1760 | %3 = icmp sgt <8 x i64> %2, zeroinitializer |
| 1761 | %4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer |
| 1762 | %5 = trunc <8 x i64> %4 to <8 x i8> |
| 1763 | ret <8 x i8> %5 |
| 1764 | } |
| 1765 | |
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 1766 | define void @trunc_packus_v8i64_v8i8_store(<8 x i64> %a0, <8 x i8> *%p1) { |
| 1767 | ; SSE2-LABEL: trunc_packus_v8i64_v8i8_store: |
| 1768 | ; SSE2: # %bb.0: |
| 1769 | ; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,255] |
| 1770 | ; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 1771 | ; SSE2-NEXT: movdqa %xmm3, %xmm4 |
| 1772 | ; SSE2-NEXT: pxor %xmm10, %xmm4 |
| 1773 | ; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903] |
| 1774 | ; SSE2-NEXT: movdqa %xmm9, %xmm6 |
| 1775 | ; SSE2-NEXT: pcmpgtd %xmm4, %xmm6 |
| 1776 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 1777 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm4 |
| 1778 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 1779 | ; SSE2-NEXT: pand %xmm7, %xmm4 |
| 1780 | ; SSE2-NEXT: pshufd {{.*#+}} xmm11 = xmm6[1,1,3,3] |
| 1781 | ; SSE2-NEXT: por %xmm4, %xmm11 |
| 1782 | ; SSE2-NEXT: pand %xmm11, %xmm3 |
| 1783 | ; SSE2-NEXT: pandn %xmm8, %xmm11 |
| 1784 | ; SSE2-NEXT: por %xmm3, %xmm11 |
| 1785 | ; SSE2-NEXT: movdqa %xmm2, %xmm3 |
| 1786 | ; SSE2-NEXT: pxor %xmm10, %xmm3 |
| 1787 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 1788 | ; SSE2-NEXT: pcmpgtd %xmm3, %xmm4 |
| 1789 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2] |
| 1790 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm3 |
| 1791 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm3[1,1,3,3] |
| 1792 | ; SSE2-NEXT: pand %xmm7, %xmm5 |
| 1793 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3] |
| 1794 | ; SSE2-NEXT: por %xmm5, %xmm3 |
| 1795 | ; SSE2-NEXT: pand %xmm3, %xmm2 |
| 1796 | ; SSE2-NEXT: pandn %xmm8, %xmm3 |
| 1797 | ; SSE2-NEXT: por %xmm2, %xmm3 |
| 1798 | ; SSE2-NEXT: movdqa %xmm1, %xmm2 |
| 1799 | ; SSE2-NEXT: pxor %xmm10, %xmm2 |
| 1800 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 1801 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm4 |
| 1802 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1803 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm2 |
| 1804 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] |
| 1805 | ; SSE2-NEXT: pand %xmm5, %xmm7 |
| 1806 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3] |
| 1807 | ; SSE2-NEXT: por %xmm7, %xmm2 |
| 1808 | ; SSE2-NEXT: pand %xmm2, %xmm1 |
| 1809 | ; SSE2-NEXT: pandn %xmm8, %xmm2 |
| 1810 | ; SSE2-NEXT: por %xmm1, %xmm2 |
| 1811 | ; SSE2-NEXT: movdqa %xmm0, %xmm1 |
| 1812 | ; SSE2-NEXT: pxor %xmm10, %xmm1 |
| 1813 | ; SSE2-NEXT: movdqa %xmm9, %xmm4 |
| 1814 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm4 |
| 1815 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1816 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm1 |
| 1817 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] |
| 1818 | ; SSE2-NEXT: pand %xmm5, %xmm7 |
| 1819 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] |
| 1820 | ; SSE2-NEXT: por %xmm7, %xmm1 |
| 1821 | ; SSE2-NEXT: pand %xmm1, %xmm0 |
| 1822 | ; SSE2-NEXT: pandn %xmm8, %xmm1 |
| 1823 | ; SSE2-NEXT: por %xmm0, %xmm1 |
| 1824 | ; SSE2-NEXT: movdqa %xmm1, %xmm0 |
| 1825 | ; SSE2-NEXT: pxor %xmm10, %xmm0 |
| 1826 | ; SSE2-NEXT: movdqa %xmm0, %xmm4 |
| 1827 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm4 |
| 1828 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1829 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1830 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 1831 | ; SSE2-NEXT: pand %xmm5, %xmm7 |
| 1832 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 1833 | ; SSE2-NEXT: por %xmm7, %xmm0 |
| 1834 | ; SSE2-NEXT: movdqa %xmm2, %xmm4 |
| 1835 | ; SSE2-NEXT: pxor %xmm10, %xmm4 |
| 1836 | ; SSE2-NEXT: movdqa %xmm4, %xmm5 |
| 1837 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm5 |
| 1838 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] |
| 1839 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm4 |
| 1840 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 1841 | ; SSE2-NEXT: pand %xmm7, %xmm4 |
| 1842 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm5[1,1,3,3] |
| 1843 | ; SSE2-NEXT: por %xmm4, %xmm7 |
| 1844 | ; SSE2-NEXT: movdqa %xmm3, %xmm4 |
| 1845 | ; SSE2-NEXT: pxor %xmm10, %xmm4 |
| 1846 | ; SSE2-NEXT: movdqa %xmm4, %xmm5 |
| 1847 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm5 |
| 1848 | ; SSE2-NEXT: pshufd {{.*#+}} xmm9 = xmm5[0,0,2,2] |
| 1849 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm4 |
| 1850 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 1851 | ; SSE2-NEXT: pand %xmm9, %xmm6 |
| 1852 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| 1853 | ; SSE2-NEXT: por %xmm6, %xmm4 |
| 1854 | ; SSE2-NEXT: movdqa %xmm11, %xmm5 |
| 1855 | ; SSE2-NEXT: pxor %xmm10, %xmm5 |
| 1856 | ; SSE2-NEXT: movdqa %xmm5, %xmm6 |
| 1857 | ; SSE2-NEXT: pcmpgtd %xmm10, %xmm6 |
| 1858 | ; SSE2-NEXT: pshufd {{.*#+}} xmm9 = xmm6[0,0,2,2] |
| 1859 | ; SSE2-NEXT: pcmpeqd %xmm10, %xmm5 |
| 1860 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 1861 | ; SSE2-NEXT: pand %xmm9, %xmm5 |
| 1862 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] |
| 1863 | ; SSE2-NEXT: por %xmm5, %xmm6 |
| 1864 | ; SSE2-NEXT: pand %xmm8, %xmm6 |
| 1865 | ; SSE2-NEXT: pand %xmm11, %xmm6 |
| 1866 | ; SSE2-NEXT: pand %xmm8, %xmm4 |
| 1867 | ; SSE2-NEXT: pand %xmm3, %xmm4 |
| 1868 | ; SSE2-NEXT: packuswb %xmm6, %xmm4 |
| 1869 | ; SSE2-NEXT: pand %xmm8, %xmm7 |
| 1870 | ; SSE2-NEXT: pand %xmm2, %xmm7 |
| 1871 | ; SSE2-NEXT: pand %xmm8, %xmm0 |
| 1872 | ; SSE2-NEXT: pand %xmm1, %xmm0 |
| 1873 | ; SSE2-NEXT: packuswb %xmm7, %xmm0 |
| 1874 | ; SSE2-NEXT: packuswb %xmm4, %xmm0 |
| 1875 | ; SSE2-NEXT: packuswb %xmm0, %xmm0 |
| 1876 | ; SSE2-NEXT: movq %xmm0, (%rdi) |
| 1877 | ; SSE2-NEXT: retq |
| 1878 | ; |
| 1879 | ; SSSE3-LABEL: trunc_packus_v8i64_v8i8_store: |
| 1880 | ; SSSE3: # %bb.0: |
| 1881 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [255,255] |
| 1882 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 1883 | ; SSSE3-NEXT: movdqa %xmm3, %xmm4 |
| 1884 | ; SSSE3-NEXT: pxor %xmm10, %xmm4 |
| 1885 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903] |
| 1886 | ; SSSE3-NEXT: movdqa %xmm9, %xmm6 |
| 1887 | ; SSSE3-NEXT: pcmpgtd %xmm4, %xmm6 |
| 1888 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 1889 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm4 |
| 1890 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 1891 | ; SSSE3-NEXT: pand %xmm7, %xmm4 |
| 1892 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm11 = xmm6[1,1,3,3] |
| 1893 | ; SSSE3-NEXT: por %xmm4, %xmm11 |
| 1894 | ; SSSE3-NEXT: pand %xmm11, %xmm3 |
| 1895 | ; SSSE3-NEXT: pandn %xmm8, %xmm11 |
| 1896 | ; SSSE3-NEXT: por %xmm3, %xmm11 |
| 1897 | ; SSSE3-NEXT: movdqa %xmm2, %xmm3 |
| 1898 | ; SSSE3-NEXT: pxor %xmm10, %xmm3 |
| 1899 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 1900 | ; SSSE3-NEXT: pcmpgtd %xmm3, %xmm4 |
| 1901 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2] |
| 1902 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm3 |
| 1903 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm3[1,1,3,3] |
| 1904 | ; SSSE3-NEXT: pand %xmm7, %xmm5 |
| 1905 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3] |
| 1906 | ; SSSE3-NEXT: por %xmm5, %xmm3 |
| 1907 | ; SSSE3-NEXT: pand %xmm3, %xmm2 |
| 1908 | ; SSSE3-NEXT: pandn %xmm8, %xmm3 |
| 1909 | ; SSSE3-NEXT: por %xmm2, %xmm3 |
| 1910 | ; SSSE3-NEXT: movdqa %xmm1, %xmm2 |
| 1911 | ; SSSE3-NEXT: pxor %xmm10, %xmm2 |
| 1912 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 1913 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4 |
| 1914 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1915 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm2 |
| 1916 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] |
| 1917 | ; SSSE3-NEXT: pand %xmm5, %xmm7 |
| 1918 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3] |
| 1919 | ; SSSE3-NEXT: por %xmm7, %xmm2 |
| 1920 | ; SSSE3-NEXT: pand %xmm2, %xmm1 |
| 1921 | ; SSSE3-NEXT: pandn %xmm8, %xmm2 |
| 1922 | ; SSSE3-NEXT: por %xmm1, %xmm2 |
| 1923 | ; SSSE3-NEXT: movdqa %xmm0, %xmm1 |
| 1924 | ; SSSE3-NEXT: pxor %xmm10, %xmm1 |
| 1925 | ; SSSE3-NEXT: movdqa %xmm9, %xmm4 |
| 1926 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4 |
| 1927 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1928 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1 |
| 1929 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] |
| 1930 | ; SSSE3-NEXT: pand %xmm5, %xmm7 |
| 1931 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] |
| 1932 | ; SSSE3-NEXT: por %xmm7, %xmm1 |
| 1933 | ; SSSE3-NEXT: pand %xmm1, %xmm0 |
| 1934 | ; SSSE3-NEXT: pandn %xmm8, %xmm1 |
| 1935 | ; SSSE3-NEXT: por %xmm0, %xmm1 |
| 1936 | ; SSSE3-NEXT: movdqa %xmm1, %xmm0 |
| 1937 | ; SSSE3-NEXT: pxor %xmm10, %xmm0 |
| 1938 | ; SSSE3-NEXT: movdqa %xmm0, %xmm4 |
| 1939 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm4 |
| 1940 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 1941 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0 |
| 1942 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 1943 | ; SSSE3-NEXT: pand %xmm5, %xmm7 |
| 1944 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 1945 | ; SSSE3-NEXT: por %xmm7, %xmm0 |
| 1946 | ; SSSE3-NEXT: movdqa %xmm2, %xmm4 |
| 1947 | ; SSSE3-NEXT: pxor %xmm10, %xmm4 |
| 1948 | ; SSSE3-NEXT: movdqa %xmm4, %xmm5 |
| 1949 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm5 |
| 1950 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] |
| 1951 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm4 |
| 1952 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| 1953 | ; SSSE3-NEXT: pand %xmm7, %xmm4 |
| 1954 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm5[1,1,3,3] |
| 1955 | ; SSSE3-NEXT: por %xmm4, %xmm7 |
| 1956 | ; SSSE3-NEXT: movdqa %xmm3, %xmm4 |
| 1957 | ; SSSE3-NEXT: pxor %xmm10, %xmm4 |
| 1958 | ; SSSE3-NEXT: movdqa %xmm4, %xmm5 |
| 1959 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm5 |
| 1960 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm9 = xmm5[0,0,2,2] |
| 1961 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm4 |
| 1962 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 1963 | ; SSSE3-NEXT: pand %xmm9, %xmm6 |
| 1964 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] |
| 1965 | ; SSSE3-NEXT: por %xmm6, %xmm4 |
| 1966 | ; SSSE3-NEXT: movdqa %xmm11, %xmm5 |
| 1967 | ; SSSE3-NEXT: pxor %xmm10, %xmm5 |
| 1968 | ; SSSE3-NEXT: movdqa %xmm5, %xmm6 |
| 1969 | ; SSSE3-NEXT: pcmpgtd %xmm10, %xmm6 |
| 1970 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm9 = xmm6[0,0,2,2] |
| 1971 | ; SSSE3-NEXT: pcmpeqd %xmm10, %xmm5 |
| 1972 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 1973 | ; SSSE3-NEXT: pand %xmm9, %xmm5 |
| 1974 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] |
| 1975 | ; SSSE3-NEXT: por %xmm5, %xmm6 |
| 1976 | ; SSSE3-NEXT: pand %xmm8, %xmm6 |
| 1977 | ; SSSE3-NEXT: pand %xmm11, %xmm6 |
| 1978 | ; SSSE3-NEXT: pand %xmm8, %xmm4 |
| 1979 | ; SSSE3-NEXT: pand %xmm3, %xmm4 |
| 1980 | ; SSSE3-NEXT: packuswb %xmm6, %xmm4 |
| 1981 | ; SSSE3-NEXT: pand %xmm8, %xmm7 |
| 1982 | ; SSSE3-NEXT: pand %xmm2, %xmm7 |
| 1983 | ; SSSE3-NEXT: pand %xmm8, %xmm0 |
| 1984 | ; SSSE3-NEXT: pand %xmm1, %xmm0 |
| 1985 | ; SSSE3-NEXT: packuswb %xmm7, %xmm0 |
| 1986 | ; SSSE3-NEXT: packuswb %xmm4, %xmm0 |
| 1987 | ; SSSE3-NEXT: packuswb %xmm0, %xmm0 |
| 1988 | ; SSSE3-NEXT: movq %xmm0, (%rdi) |
| 1989 | ; SSSE3-NEXT: retq |
| 1990 | ; |
| 1991 | ; SSE41-LABEL: trunc_packus_v8i64_v8i8_store: |
| 1992 | ; SSE41: # %bb.0: |
| 1993 | ; SSE41-NEXT: movdqa %xmm0, %xmm9 |
| 1994 | ; SSE41-NEXT: movapd {{.*#+}} xmm8 = [255,255] |
| 1995 | ; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| 1996 | ; SSE41-NEXT: movdqa %xmm3, %xmm0 |
| 1997 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 1998 | ; SSE41-NEXT: movdqa {{.*#+}} xmm5 = [2147483903,2147483903] |
| 1999 | ; SSE41-NEXT: movdqa %xmm5, %xmm4 |
| 2000 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm4 |
| 2001 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2] |
| 2002 | ; SSE41-NEXT: pcmpeqd %xmm5, %xmm0 |
| 2003 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 2004 | ; SSE41-NEXT: pand %xmm7, %xmm6 |
| 2005 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 2006 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 2007 | ; SSE41-NEXT: movapd %xmm8, %xmm11 |
| 2008 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm11 |
| 2009 | ; SSE41-NEXT: movdqa %xmm2, %xmm0 |
| 2010 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2011 | ; SSE41-NEXT: movdqa %xmm5, %xmm3 |
| 2012 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm3 |
| 2013 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 2014 | ; SSE41-NEXT: pcmpeqd %xmm5, %xmm0 |
| 2015 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 2016 | ; SSE41-NEXT: pand %xmm4, %xmm6 |
| 2017 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2018 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 2019 | ; SSE41-NEXT: movapd %xmm8, %xmm3 |
| 2020 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm3 |
| 2021 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| 2022 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2023 | ; SSE41-NEXT: movdqa %xmm5, %xmm2 |
| 2024 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm2 |
| 2025 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 2026 | ; SSE41-NEXT: pcmpeqd %xmm5, %xmm0 |
| 2027 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 2028 | ; SSE41-NEXT: pand %xmm4, %xmm6 |
| 2029 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 2030 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 2031 | ; SSE41-NEXT: movapd %xmm8, %xmm4 |
| 2032 | ; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm4 |
| 2033 | ; SSE41-NEXT: movdqa %xmm9, %xmm0 |
| 2034 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2035 | ; SSE41-NEXT: movdqa %xmm5, %xmm1 |
| 2036 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm1 |
| 2037 | ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] |
| 2038 | ; SSE41-NEXT: pcmpeqd %xmm5, %xmm0 |
| 2039 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 2040 | ; SSE41-NEXT: pand %xmm2, %xmm5 |
| 2041 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 2042 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 2043 | ; SSE41-NEXT: movapd %xmm8, %xmm5 |
| 2044 | ; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm5 |
| 2045 | ; SSE41-NEXT: movapd %xmm5, %xmm0 |
| 2046 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2047 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 2048 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm1 |
| 2049 | ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] |
| 2050 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2051 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 2052 | ; SSE41-NEXT: pand %xmm2, %xmm6 |
| 2053 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 2054 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 2055 | ; SSE41-NEXT: pxor %xmm2, %xmm2 |
| 2056 | ; SSE41-NEXT: pxor %xmm1, %xmm1 |
| 2057 | ; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm1 |
| 2058 | ; SSE41-NEXT: movapd %xmm4, %xmm0 |
| 2059 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2060 | ; SSE41-NEXT: movapd %xmm0, %xmm5 |
| 2061 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm5 |
| 2062 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] |
| 2063 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2064 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2065 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 2066 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] |
| 2067 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2068 | ; SSE41-NEXT: pxor %xmm5, %xmm5 |
| 2069 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm5 |
| 2070 | ; SSE41-NEXT: movapd %xmm3, %xmm0 |
| 2071 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2072 | ; SSE41-NEXT: movapd %xmm0, %xmm4 |
| 2073 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm4 |
| 2074 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 2075 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2076 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2077 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 2078 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 2079 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2080 | ; SSE41-NEXT: pxor %xmm4, %xmm4 |
| 2081 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm4 |
| 2082 | ; SSE41-NEXT: movapd %xmm11, %xmm0 |
| 2083 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2084 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 2085 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 2086 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] |
| 2087 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2088 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2089 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 2090 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2091 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2092 | ; SSE41-NEXT: blendvpd %xmm0, %xmm11, %xmm2 |
| 2093 | ; SSE41-NEXT: andpd %xmm8, %xmm2 |
| 2094 | ; SSE41-NEXT: andpd %xmm8, %xmm4 |
| 2095 | ; SSE41-NEXT: packuswb %xmm2, %xmm4 |
| 2096 | ; SSE41-NEXT: andpd %xmm8, %xmm5 |
| 2097 | ; SSE41-NEXT: andpd %xmm8, %xmm1 |
| 2098 | ; SSE41-NEXT: packuswb %xmm5, %xmm1 |
| 2099 | ; SSE41-NEXT: packuswb %xmm4, %xmm1 |
| 2100 | ; SSE41-NEXT: packuswb %xmm1, %xmm1 |
| 2101 | ; SSE41-NEXT: movq %xmm1, (%rdi) |
| 2102 | ; SSE41-NEXT: retq |
| 2103 | ; |
| 2104 | ; AVX1-LABEL: trunc_packus_v8i64_v8i8_store: |
| 2105 | ; AVX1: # %bb.0: |
| 2106 | ; AVX1-NEXT: vmovapd {{.*#+}} ymm2 = [255,255,255,255] |
| 2107 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3 |
| 2108 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [255,255] |
| 2109 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 2110 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm4, %xmm5 |
| 2111 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3 |
| 2112 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| 2113 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3 |
| 2114 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3 |
| 2115 | ; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm5 |
| 2116 | ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3 |
| 2117 | ; AVX1-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| 2118 | ; AVX1-NEXT: vxorpd %xmm2, %xmm2, %xmm2 |
| 2119 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm8 |
| 2120 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm5 |
| 2121 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm5, %xmm6 |
| 2122 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm7 |
| 2123 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3 |
| 2124 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm2 |
| 2125 | ; AVX1-NEXT: vpand %xmm4, %xmm3, %xmm3 |
| 2126 | ; AVX1-NEXT: vpand %xmm3, %xmm2, %xmm2 |
| 2127 | ; AVX1-NEXT: vpand %xmm4, %xmm1, %xmm1 |
| 2128 | ; AVX1-NEXT: vpand %xmm1, %xmm7, %xmm1 |
| 2129 | ; AVX1-NEXT: vpackuswb %xmm2, %xmm1, %xmm1 |
| 2130 | ; AVX1-NEXT: vpand %xmm4, %xmm5, %xmm2 |
| 2131 | ; AVX1-NEXT: vpand %xmm2, %xmm6, %xmm2 |
| 2132 | ; AVX1-NEXT: vpand %xmm4, %xmm0, %xmm0 |
| 2133 | ; AVX1-NEXT: vpand %xmm0, %xmm8, %xmm0 |
| 2134 | ; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0 |
| 2135 | ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 |
| 2136 | ; AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 |
| 2137 | ; AVX1-NEXT: vmovq %xmm0, (%rdi) |
| 2138 | ; AVX1-NEXT: vzeroupper |
| 2139 | ; AVX1-NEXT: retq |
| 2140 | ; |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 2141 | ; AVX2-LABEL: trunc_packus_v8i64_v8i8_store: |
| 2142 | ; AVX2: # %bb.0: |
| 2143 | ; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255] |
| 2144 | ; AVX2-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3 |
| 2145 | ; AVX2-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 |
| 2146 | ; AVX2-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3 |
| 2147 | ; AVX2-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 |
| 2148 | ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| 2149 | ; AVX2-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm3 |
| 2150 | ; AVX2-NEXT: vpand %ymm0, %ymm3, %ymm0 |
| 2151 | ; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2 |
| 2152 | ; AVX2-NEXT: vpand %ymm1, %ymm2, %ymm1 |
| 2153 | ; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2 |
| 2154 | ; AVX2-NEXT: vpackssdw %xmm2, %xmm1, %xmm1 |
| 2155 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] |
| 2156 | ; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1 |
| 2157 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm3 |
| 2158 | ; AVX2-NEXT: vpackssdw %xmm3, %xmm0, %xmm0 |
| 2159 | ; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0 |
| 2160 | ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| 2161 | ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] |
| 2162 | ; AVX2-NEXT: vmovq %xmm0, (%rdi) |
| 2163 | ; AVX2-NEXT: vzeroupper |
| 2164 | ; AVX2-NEXT: retq |
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 2165 | ; |
| 2166 | ; AVX512-LABEL: trunc_packus_v8i64_v8i8_store: |
| 2167 | ; AVX512: # %bb.0: |
| 2168 | ; AVX512-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0 |
| 2169 | ; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 2170 | ; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 |
| 2171 | ; AVX512-NEXT: vpmovqb %zmm0, (%rdi) |
| 2172 | ; AVX512-NEXT: vzeroupper |
| 2173 | ; AVX512-NEXT: retq |
| 2174 | %1 = icmp slt <8 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> |
| 2175 | %2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> |
| 2176 | %3 = icmp sgt <8 x i64> %2, zeroinitializer |
| 2177 | %4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer |
| 2178 | %5 = trunc <8 x i64> %4 to <8 x i8> |
| 2179 | store <8 x i8> %5, <8 x i8> *%p1 |
| 2180 | ret void |
| 2181 | } |
| 2182 | |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2183 | define <16 x i8> @trunc_packus_v16i64_v16i8(<16 x i64> %a0) { |
| 2184 | ; SSE2-LABEL: trunc_packus_v16i64_v16i8: |
| 2185 | ; SSE2: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2186 | ; SSE2-NEXT: movdqa {{.*#+}} xmm13 = [255,255] |
| 2187 | ; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483648,0,2147483648,0] |
| 2188 | ; SSE2-NEXT: movdqa %xmm7, %xmm8 |
| 2189 | ; SSE2-NEXT: pxor %xmm9, %xmm8 |
| 2190 | ; SSE2-NEXT: movdqa {{.*#+}} xmm11 = [2147483903,2147483903] |
| 2191 | ; SSE2-NEXT: movdqa %xmm11, %xmm10 |
| 2192 | ; SSE2-NEXT: pcmpgtd %xmm8, %xmm10 |
| 2193 | ; SSE2-NEXT: pshufd {{.*#+}} xmm12 = xmm10[0,0,2,2] |
| 2194 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm8 |
| 2195 | ; SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3] |
| 2196 | ; SSE2-NEXT: pand %xmm12, %xmm8 |
| 2197 | ; SSE2-NEXT: pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3] |
| 2198 | ; SSE2-NEXT: por %xmm8, %xmm10 |
| 2199 | ; SSE2-NEXT: pand %xmm10, %xmm7 |
| 2200 | ; SSE2-NEXT: pandn %xmm13, %xmm10 |
| 2201 | ; SSE2-NEXT: por %xmm7, %xmm10 |
| 2202 | ; SSE2-NEXT: movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| 2203 | ; SSE2-NEXT: movdqa %xmm6, %xmm7 |
| 2204 | ; SSE2-NEXT: pxor %xmm9, %xmm7 |
| 2205 | ; SSE2-NEXT: movdqa %xmm11, %xmm8 |
| 2206 | ; SSE2-NEXT: pcmpgtd %xmm7, %xmm8 |
| 2207 | ; SSE2-NEXT: pshufd {{.*#+}} xmm12 = xmm8[0,0,2,2] |
| 2208 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm7 |
| 2209 | ; SSE2-NEXT: pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] |
| 2210 | ; SSE2-NEXT: pand %xmm12, %xmm10 |
| 2211 | ; SSE2-NEXT: pshufd {{.*#+}} xmm14 = xmm8[1,1,3,3] |
| 2212 | ; SSE2-NEXT: por %xmm10, %xmm14 |
| 2213 | ; SSE2-NEXT: pand %xmm14, %xmm6 |
| 2214 | ; SSE2-NEXT: pandn %xmm13, %xmm14 |
| 2215 | ; SSE2-NEXT: por %xmm6, %xmm14 |
| 2216 | ; SSE2-NEXT: movdqa %xmm5, %xmm6 |
| 2217 | ; SSE2-NEXT: pxor %xmm9, %xmm6 |
| 2218 | ; SSE2-NEXT: movdqa %xmm11, %xmm7 |
| 2219 | ; SSE2-NEXT: pcmpgtd %xmm6, %xmm7 |
| 2220 | ; SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2] |
| 2221 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm6 |
| 2222 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] |
| 2223 | ; SSE2-NEXT: pand %xmm8, %xmm6 |
| 2224 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3] |
| 2225 | ; SSE2-NEXT: por %xmm6, %xmm7 |
| 2226 | ; SSE2-NEXT: pand %xmm7, %xmm5 |
| 2227 | ; SSE2-NEXT: pandn %xmm13, %xmm7 |
| 2228 | ; SSE2-NEXT: por %xmm5, %xmm7 |
| 2229 | ; SSE2-NEXT: movdqa %xmm7, %xmm10 |
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2230 | ; SSE2-NEXT: movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2231 | ; SSE2-NEXT: movdqa %xmm4, %xmm5 |
| 2232 | ; SSE2-NEXT: pxor %xmm9, %xmm5 |
| 2233 | ; SSE2-NEXT: movdqa %xmm11, %xmm6 |
| 2234 | ; SSE2-NEXT: pcmpgtd %xmm5, %xmm6 |
| 2235 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 2236 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm5 |
| 2237 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 2238 | ; SSE2-NEXT: pand %xmm7, %xmm5 |
| 2239 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] |
| 2240 | ; SSE2-NEXT: por %xmm5, %xmm6 |
| 2241 | ; SSE2-NEXT: pand %xmm6, %xmm4 |
| 2242 | ; SSE2-NEXT: pandn %xmm13, %xmm6 |
| 2243 | ; SSE2-NEXT: por %xmm4, %xmm6 |
| 2244 | ; SSE2-NEXT: movdqa %xmm6, %xmm7 |
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2245 | ; SSE2-NEXT: movdqa %xmm7, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2246 | ; SSE2-NEXT: movdqa %xmm3, %xmm4 |
| 2247 | ; SSE2-NEXT: pxor %xmm9, %xmm4 |
| 2248 | ; SSE2-NEXT: movdqa %xmm11, %xmm5 |
| 2249 | ; SSE2-NEXT: pcmpgtd %xmm4, %xmm5 |
| 2250 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] |
| 2251 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2252 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2253 | ; SSE2-NEXT: pand %xmm6, %xmm4 |
| 2254 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 2255 | ; SSE2-NEXT: por %xmm4, %xmm5 |
| 2256 | ; SSE2-NEXT: pand %xmm5, %xmm3 |
| 2257 | ; SSE2-NEXT: pandn %xmm13, %xmm5 |
| 2258 | ; SSE2-NEXT: por %xmm3, %xmm5 |
| 2259 | ; SSE2-NEXT: movdqa %xmm5, %xmm8 |
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2260 | ; SSE2-NEXT: movdqa %xmm8, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2261 | ; SSE2-NEXT: movdqa %xmm2, %xmm3 |
| 2262 | ; SSE2-NEXT: pxor %xmm9, %xmm3 |
| 2263 | ; SSE2-NEXT: movdqa %xmm11, %xmm4 |
| 2264 | ; SSE2-NEXT: pcmpgtd %xmm3, %xmm4 |
| 2265 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2266 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm3 |
| 2267 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 2268 | ; SSE2-NEXT: pand %xmm5, %xmm3 |
| 2269 | ; SSE2-NEXT: pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3] |
| 2270 | ; SSE2-NEXT: por %xmm3, %xmm15 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2271 | ; SSE2-NEXT: pand %xmm15, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2272 | ; SSE2-NEXT: pandn %xmm13, %xmm15 |
| 2273 | ; SSE2-NEXT: movdqa %xmm13, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2274 | ; SSE2-NEXT: por %xmm2, %xmm15 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2275 | ; SSE2-NEXT: movdqa %xmm1, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2276 | ; SSE2-NEXT: pxor %xmm9, %xmm2 |
| 2277 | ; SSE2-NEXT: movdqa %xmm11, %xmm3 |
| 2278 | ; SSE2-NEXT: pcmpgtd %xmm2, %xmm3 |
| 2279 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 2280 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2281 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2282 | ; SSE2-NEXT: pand %xmm4, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2283 | ; SSE2-NEXT: pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3] |
| 2284 | ; SSE2-NEXT: por %xmm2, %xmm13 |
| 2285 | ; SSE2-NEXT: pand %xmm13, %xmm1 |
| 2286 | ; SSE2-NEXT: pandn %xmm5, %xmm13 |
| 2287 | ; SSE2-NEXT: por %xmm1, %xmm13 |
| 2288 | ; SSE2-NEXT: movdqa %xmm0, %xmm1 |
| 2289 | ; SSE2-NEXT: pxor %xmm9, %xmm1 |
| 2290 | ; SSE2-NEXT: movdqa %xmm11, %xmm2 |
| 2291 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm2 |
| 2292 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 2293 | ; SSE2-NEXT: pcmpeqd %xmm11, %xmm1 |
| 2294 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 2295 | ; SSE2-NEXT: pand %xmm4, %xmm1 |
| 2296 | ; SSE2-NEXT: pshufd {{.*#+}} xmm11 = xmm2[1,1,3,3] |
| 2297 | ; SSE2-NEXT: por %xmm1, %xmm11 |
| 2298 | ; SSE2-NEXT: pand %xmm11, %xmm0 |
| 2299 | ; SSE2-NEXT: pandn %xmm5, %xmm11 |
| 2300 | ; SSE2-NEXT: por %xmm0, %xmm11 |
| 2301 | ; SSE2-NEXT: movdqa %xmm11, %xmm0 |
| 2302 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2303 | ; SSE2-NEXT: movdqa %xmm0, %xmm1 |
| 2304 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm1 |
| 2305 | ; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] |
| 2306 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2307 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2308 | ; SSE2-NEXT: pand %xmm4, %xmm0 |
| 2309 | ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3] |
| 2310 | ; SSE2-NEXT: por %xmm0, %xmm2 |
| 2311 | ; SSE2-NEXT: movdqa %xmm13, %xmm0 |
| 2312 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2313 | ; SSE2-NEXT: movdqa %xmm0, %xmm4 |
| 2314 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm4 |
| 2315 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2316 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2317 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2318 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| 2319 | ; SSE2-NEXT: pshufd {{.*#+}} xmm12 = xmm4[1,1,3,3] |
| 2320 | ; SSE2-NEXT: por %xmm0, %xmm12 |
| 2321 | ; SSE2-NEXT: movdqa %xmm15, %xmm0 |
| 2322 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2323 | ; SSE2-NEXT: movdqa %xmm0, %xmm4 |
| 2324 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm4 |
| 2325 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2326 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2327 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2328 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| 2329 | ; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 2330 | ; SSE2-NEXT: por %xmm0, %xmm6 |
| 2331 | ; SSE2-NEXT: movdqa %xmm8, %xmm0 |
| 2332 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2333 | ; SSE2-NEXT: movdqa %xmm0, %xmm4 |
| 2334 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm4 |
| 2335 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2336 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2337 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2338 | ; SSE2-NEXT: pand %xmm5, %xmm0 |
| 2339 | ; SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm4[1,1,3,3] |
| 2340 | ; SSE2-NEXT: por %xmm0, %xmm8 |
| 2341 | ; SSE2-NEXT: movdqa %xmm7, %xmm0 |
| 2342 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2343 | ; SSE2-NEXT: movdqa %xmm0, %xmm5 |
| 2344 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm5 |
| 2345 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2346 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2347 | ; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] |
| 2348 | ; SSE2-NEXT: pand %xmm7, %xmm0 |
| 2349 | ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 2350 | ; SSE2-NEXT: por %xmm0, %xmm5 |
| 2351 | ; SSE2-NEXT: movdqa %xmm10, %xmm0 |
| 2352 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2353 | ; SSE2-NEXT: movdqa %xmm0, %xmm7 |
| 2354 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm7 |
| 2355 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2356 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2357 | ; SSE2-NEXT: pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2] |
| 2358 | ; SSE2-NEXT: pand %xmm10, %xmm0 |
| 2359 | ; SSE2-NEXT: pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] |
| 2360 | ; SSE2-NEXT: por %xmm0, %xmm10 |
| 2361 | ; SSE2-NEXT: movdqa %xmm14, %xmm0 |
| 2362 | ; SSE2-NEXT: pxor %xmm9, %xmm0 |
| 2363 | ; SSE2-NEXT: movdqa %xmm0, %xmm7 |
| 2364 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm7 |
| 2365 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2366 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3] |
| 2367 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2] |
| 2368 | ; SSE2-NEXT: pand %xmm0, %xmm1 |
| 2369 | ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3] |
| 2370 | ; SSE2-NEXT: por %xmm1, %xmm0 |
| 2371 | ; SSE2-NEXT: movdqa -{{[0-9]+}}(%rsp), %xmm4 # 16-byte Reload |
| 2372 | ; SSE2-NEXT: movdqa %xmm4, %xmm1 |
| 2373 | ; SSE2-NEXT: pxor %xmm9, %xmm1 |
| 2374 | ; SSE2-NEXT: movdqa %xmm1, %xmm7 |
| 2375 | ; SSE2-NEXT: pcmpgtd %xmm9, %xmm7 |
| 2376 | ; SSE2-NEXT: pcmpeqd %xmm9, %xmm1 |
| 2377 | ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 2378 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2379 | ; SSE2-NEXT: pand %xmm3, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2380 | ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm7[1,1,3,3] |
| 2381 | ; SSE2-NEXT: por %xmm1, %xmm3 |
| 2382 | ; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [255,255] |
| 2383 | ; SSE2-NEXT: pand %xmm1, %xmm3 |
| 2384 | ; SSE2-NEXT: pand %xmm4, %xmm3 |
| 2385 | ; SSE2-NEXT: pand %xmm1, %xmm0 |
| 2386 | ; SSE2-NEXT: pand %xmm14, %xmm0 |
| 2387 | ; SSE2-NEXT: packuswb %xmm3, %xmm0 |
| 2388 | ; SSE2-NEXT: pand %xmm1, %xmm10 |
| 2389 | ; SSE2-NEXT: pand -{{[0-9]+}}(%rsp), %xmm10 # 16-byte Folded Reload |
| 2390 | ; SSE2-NEXT: pand %xmm1, %xmm5 |
| 2391 | ; SSE2-NEXT: pand -{{[0-9]+}}(%rsp), %xmm5 # 16-byte Folded Reload |
| 2392 | ; SSE2-NEXT: packuswb %xmm10, %xmm5 |
| 2393 | ; SSE2-NEXT: packuswb %xmm0, %xmm5 |
| 2394 | ; SSE2-NEXT: pand %xmm1, %xmm8 |
| 2395 | ; SSE2-NEXT: pand -{{[0-9]+}}(%rsp), %xmm8 # 16-byte Folded Reload |
| 2396 | ; SSE2-NEXT: pand %xmm1, %xmm6 |
| 2397 | ; SSE2-NEXT: pand %xmm15, %xmm6 |
| 2398 | ; SSE2-NEXT: packuswb %xmm8, %xmm6 |
| 2399 | ; SSE2-NEXT: pand %xmm1, %xmm12 |
| 2400 | ; SSE2-NEXT: pand %xmm13, %xmm12 |
| 2401 | ; SSE2-NEXT: pand %xmm1, %xmm2 |
| 2402 | ; SSE2-NEXT: pand %xmm11, %xmm2 |
| 2403 | ; SSE2-NEXT: packuswb %xmm12, %xmm2 |
| 2404 | ; SSE2-NEXT: packuswb %xmm6, %xmm2 |
| 2405 | ; SSE2-NEXT: packuswb %xmm5, %xmm2 |
| 2406 | ; SSE2-NEXT: movdqa %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2407 | ; SSE2-NEXT: retq |
| 2408 | ; |
| 2409 | ; SSSE3-LABEL: trunc_packus_v16i64_v16i8: |
| 2410 | ; SSSE3: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2411 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm13 = [255,255] |
| 2412 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483648,0,2147483648,0] |
| 2413 | ; SSSE3-NEXT: movdqa %xmm7, %xmm8 |
| 2414 | ; SSSE3-NEXT: pxor %xmm9, %xmm8 |
| 2415 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm11 = [2147483903,2147483903] |
| 2416 | ; SSSE3-NEXT: movdqa %xmm11, %xmm10 |
| 2417 | ; SSSE3-NEXT: pcmpgtd %xmm8, %xmm10 |
| 2418 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm12 = xmm10[0,0,2,2] |
| 2419 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm8 |
| 2420 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3] |
| 2421 | ; SSSE3-NEXT: pand %xmm12, %xmm8 |
| 2422 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3] |
| 2423 | ; SSSE3-NEXT: por %xmm8, %xmm10 |
| 2424 | ; SSSE3-NEXT: pand %xmm10, %xmm7 |
| 2425 | ; SSSE3-NEXT: pandn %xmm13, %xmm10 |
| 2426 | ; SSSE3-NEXT: por %xmm7, %xmm10 |
| 2427 | ; SSSE3-NEXT: movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| 2428 | ; SSSE3-NEXT: movdqa %xmm6, %xmm7 |
| 2429 | ; SSSE3-NEXT: pxor %xmm9, %xmm7 |
| 2430 | ; SSSE3-NEXT: movdqa %xmm11, %xmm8 |
| 2431 | ; SSSE3-NEXT: pcmpgtd %xmm7, %xmm8 |
| 2432 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm12 = xmm8[0,0,2,2] |
| 2433 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm7 |
| 2434 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] |
| 2435 | ; SSSE3-NEXT: pand %xmm12, %xmm10 |
| 2436 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm14 = xmm8[1,1,3,3] |
| 2437 | ; SSSE3-NEXT: por %xmm10, %xmm14 |
| 2438 | ; SSSE3-NEXT: pand %xmm14, %xmm6 |
| 2439 | ; SSSE3-NEXT: pandn %xmm13, %xmm14 |
| 2440 | ; SSSE3-NEXT: por %xmm6, %xmm14 |
| 2441 | ; SSSE3-NEXT: movdqa %xmm5, %xmm6 |
| 2442 | ; SSSE3-NEXT: pxor %xmm9, %xmm6 |
| 2443 | ; SSSE3-NEXT: movdqa %xmm11, %xmm7 |
| 2444 | ; SSSE3-NEXT: pcmpgtd %xmm6, %xmm7 |
| 2445 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2] |
| 2446 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm6 |
| 2447 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] |
| 2448 | ; SSSE3-NEXT: pand %xmm8, %xmm6 |
| 2449 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3] |
| 2450 | ; SSSE3-NEXT: por %xmm6, %xmm7 |
| 2451 | ; SSSE3-NEXT: pand %xmm7, %xmm5 |
| 2452 | ; SSSE3-NEXT: pandn %xmm13, %xmm7 |
| 2453 | ; SSSE3-NEXT: por %xmm5, %xmm7 |
| 2454 | ; SSSE3-NEXT: movdqa %xmm7, %xmm10 |
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2455 | ; SSSE3-NEXT: movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2456 | ; SSSE3-NEXT: movdqa %xmm4, %xmm5 |
| 2457 | ; SSSE3-NEXT: pxor %xmm9, %xmm5 |
| 2458 | ; SSSE3-NEXT: movdqa %xmm11, %xmm6 |
| 2459 | ; SSSE3-NEXT: pcmpgtd %xmm5, %xmm6 |
| 2460 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] |
| 2461 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm5 |
| 2462 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 2463 | ; SSSE3-NEXT: pand %xmm7, %xmm5 |
| 2464 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] |
| 2465 | ; SSSE3-NEXT: por %xmm5, %xmm6 |
| 2466 | ; SSSE3-NEXT: pand %xmm6, %xmm4 |
| 2467 | ; SSSE3-NEXT: pandn %xmm13, %xmm6 |
| 2468 | ; SSSE3-NEXT: por %xmm4, %xmm6 |
| 2469 | ; SSSE3-NEXT: movdqa %xmm6, %xmm7 |
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2470 | ; SSSE3-NEXT: movdqa %xmm7, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2471 | ; SSSE3-NEXT: movdqa %xmm3, %xmm4 |
| 2472 | ; SSSE3-NEXT: pxor %xmm9, %xmm4 |
| 2473 | ; SSSE3-NEXT: movdqa %xmm11, %xmm5 |
| 2474 | ; SSSE3-NEXT: pcmpgtd %xmm4, %xmm5 |
| 2475 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] |
| 2476 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm4 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2477 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2478 | ; SSSE3-NEXT: pand %xmm6, %xmm4 |
| 2479 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 2480 | ; SSSE3-NEXT: por %xmm4, %xmm5 |
| 2481 | ; SSSE3-NEXT: pand %xmm5, %xmm3 |
| 2482 | ; SSSE3-NEXT: pandn %xmm13, %xmm5 |
| 2483 | ; SSSE3-NEXT: por %xmm3, %xmm5 |
| 2484 | ; SSSE3-NEXT: movdqa %xmm5, %xmm8 |
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2485 | ; SSSE3-NEXT: movdqa %xmm8, -{{[0-9]+}}(%rsp) # 16-byte Spill |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2486 | ; SSSE3-NEXT: movdqa %xmm2, %xmm3 |
| 2487 | ; SSSE3-NEXT: pxor %xmm9, %xmm3 |
| 2488 | ; SSSE3-NEXT: movdqa %xmm11, %xmm4 |
| 2489 | ; SSSE3-NEXT: pcmpgtd %xmm3, %xmm4 |
| 2490 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2491 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm3 |
| 2492 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] |
| 2493 | ; SSSE3-NEXT: pand %xmm5, %xmm3 |
| 2494 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3] |
| 2495 | ; SSSE3-NEXT: por %xmm3, %xmm15 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2496 | ; SSSE3-NEXT: pand %xmm15, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2497 | ; SSSE3-NEXT: pandn %xmm13, %xmm15 |
| 2498 | ; SSSE3-NEXT: movdqa %xmm13, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2499 | ; SSSE3-NEXT: por %xmm2, %xmm15 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2500 | ; SSSE3-NEXT: movdqa %xmm1, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2501 | ; SSSE3-NEXT: pxor %xmm9, %xmm2 |
| 2502 | ; SSSE3-NEXT: movdqa %xmm11, %xmm3 |
| 2503 | ; SSSE3-NEXT: pcmpgtd %xmm2, %xmm3 |
| 2504 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 2505 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2506 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2507 | ; SSSE3-NEXT: pand %xmm4, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2508 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3] |
| 2509 | ; SSSE3-NEXT: por %xmm2, %xmm13 |
| 2510 | ; SSSE3-NEXT: pand %xmm13, %xmm1 |
| 2511 | ; SSSE3-NEXT: pandn %xmm5, %xmm13 |
| 2512 | ; SSSE3-NEXT: por %xmm1, %xmm13 |
| 2513 | ; SSSE3-NEXT: movdqa %xmm0, %xmm1 |
| 2514 | ; SSSE3-NEXT: pxor %xmm9, %xmm1 |
| 2515 | ; SSSE3-NEXT: movdqa %xmm11, %xmm2 |
| 2516 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm2 |
| 2517 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] |
| 2518 | ; SSSE3-NEXT: pcmpeqd %xmm11, %xmm1 |
| 2519 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 2520 | ; SSSE3-NEXT: pand %xmm4, %xmm1 |
| 2521 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm11 = xmm2[1,1,3,3] |
| 2522 | ; SSSE3-NEXT: por %xmm1, %xmm11 |
| 2523 | ; SSSE3-NEXT: pand %xmm11, %xmm0 |
| 2524 | ; SSSE3-NEXT: pandn %xmm5, %xmm11 |
| 2525 | ; SSSE3-NEXT: por %xmm0, %xmm11 |
| 2526 | ; SSSE3-NEXT: movdqa %xmm11, %xmm0 |
| 2527 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2528 | ; SSSE3-NEXT: movdqa %xmm0, %xmm1 |
| 2529 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm1 |
| 2530 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] |
| 2531 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2532 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2533 | ; SSSE3-NEXT: pand %xmm4, %xmm0 |
| 2534 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3] |
| 2535 | ; SSSE3-NEXT: por %xmm0, %xmm2 |
| 2536 | ; SSSE3-NEXT: movdqa %xmm13, %xmm0 |
| 2537 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2538 | ; SSSE3-NEXT: movdqa %xmm0, %xmm4 |
| 2539 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm4 |
| 2540 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2541 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2542 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2543 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| 2544 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm12 = xmm4[1,1,3,3] |
| 2545 | ; SSSE3-NEXT: por %xmm0, %xmm12 |
| 2546 | ; SSSE3-NEXT: movdqa %xmm15, %xmm0 |
| 2547 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2548 | ; SSSE3-NEXT: movdqa %xmm0, %xmm4 |
| 2549 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm4 |
| 2550 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2551 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2552 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2553 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| 2554 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] |
| 2555 | ; SSSE3-NEXT: por %xmm0, %xmm6 |
| 2556 | ; SSSE3-NEXT: movdqa %xmm8, %xmm0 |
| 2557 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2558 | ; SSSE3-NEXT: movdqa %xmm0, %xmm4 |
| 2559 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm4 |
| 2560 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] |
| 2561 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2562 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2563 | ; SSSE3-NEXT: pand %xmm5, %xmm0 |
| 2564 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm8 = xmm4[1,1,3,3] |
| 2565 | ; SSSE3-NEXT: por %xmm0, %xmm8 |
| 2566 | ; SSSE3-NEXT: movdqa %xmm7, %xmm0 |
| 2567 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2568 | ; SSSE3-NEXT: movdqa %xmm0, %xmm5 |
| 2569 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm5 |
| 2570 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2571 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2572 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] |
| 2573 | ; SSSE3-NEXT: pand %xmm7, %xmm0 |
| 2574 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] |
| 2575 | ; SSSE3-NEXT: por %xmm0, %xmm5 |
| 2576 | ; SSSE3-NEXT: movdqa %xmm10, %xmm0 |
| 2577 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2578 | ; SSSE3-NEXT: movdqa %xmm0, %xmm7 |
| 2579 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm7 |
| 2580 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2581 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] |
| 2582 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2] |
| 2583 | ; SSSE3-NEXT: pand %xmm10, %xmm0 |
| 2584 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] |
| 2585 | ; SSSE3-NEXT: por %xmm0, %xmm10 |
| 2586 | ; SSSE3-NEXT: movdqa %xmm14, %xmm0 |
| 2587 | ; SSSE3-NEXT: pxor %xmm9, %xmm0 |
| 2588 | ; SSSE3-NEXT: movdqa %xmm0, %xmm7 |
| 2589 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm7 |
| 2590 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0 |
| 2591 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3] |
| 2592 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2] |
| 2593 | ; SSSE3-NEXT: pand %xmm0, %xmm1 |
| 2594 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3] |
| 2595 | ; SSSE3-NEXT: por %xmm1, %xmm0 |
| 2596 | ; SSSE3-NEXT: movdqa -{{[0-9]+}}(%rsp), %xmm4 # 16-byte Reload |
| 2597 | ; SSSE3-NEXT: movdqa %xmm4, %xmm1 |
| 2598 | ; SSSE3-NEXT: pxor %xmm9, %xmm1 |
| 2599 | ; SSSE3-NEXT: movdqa %xmm1, %xmm7 |
| 2600 | ; SSSE3-NEXT: pcmpgtd %xmm9, %xmm7 |
| 2601 | ; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1 |
| 2602 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] |
| 2603 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2604 | ; SSSE3-NEXT: pand %xmm3, %xmm1 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2605 | ; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm7[1,1,3,3] |
| 2606 | ; SSSE3-NEXT: por %xmm1, %xmm3 |
| 2607 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [255,255] |
| 2608 | ; SSSE3-NEXT: pand %xmm1, %xmm3 |
| 2609 | ; SSSE3-NEXT: pand %xmm4, %xmm3 |
| 2610 | ; SSSE3-NEXT: pand %xmm1, %xmm0 |
| 2611 | ; SSSE3-NEXT: pand %xmm14, %xmm0 |
| 2612 | ; SSSE3-NEXT: packuswb %xmm3, %xmm0 |
| 2613 | ; SSSE3-NEXT: pand %xmm1, %xmm10 |
| 2614 | ; SSSE3-NEXT: pand -{{[0-9]+}}(%rsp), %xmm10 # 16-byte Folded Reload |
| 2615 | ; SSSE3-NEXT: pand %xmm1, %xmm5 |
| 2616 | ; SSSE3-NEXT: pand -{{[0-9]+}}(%rsp), %xmm5 # 16-byte Folded Reload |
| 2617 | ; SSSE3-NEXT: packuswb %xmm10, %xmm5 |
| 2618 | ; SSSE3-NEXT: packuswb %xmm0, %xmm5 |
| 2619 | ; SSSE3-NEXT: pand %xmm1, %xmm8 |
| 2620 | ; SSSE3-NEXT: pand -{{[0-9]+}}(%rsp), %xmm8 # 16-byte Folded Reload |
| 2621 | ; SSSE3-NEXT: pand %xmm1, %xmm6 |
| 2622 | ; SSSE3-NEXT: pand %xmm15, %xmm6 |
| 2623 | ; SSSE3-NEXT: packuswb %xmm8, %xmm6 |
| 2624 | ; SSSE3-NEXT: pand %xmm1, %xmm12 |
| 2625 | ; SSSE3-NEXT: pand %xmm13, %xmm12 |
| 2626 | ; SSSE3-NEXT: pand %xmm1, %xmm2 |
| 2627 | ; SSSE3-NEXT: pand %xmm11, %xmm2 |
| 2628 | ; SSSE3-NEXT: packuswb %xmm12, %xmm2 |
| 2629 | ; SSSE3-NEXT: packuswb %xmm6, %xmm2 |
| 2630 | ; SSSE3-NEXT: packuswb %xmm5, %xmm2 |
| 2631 | ; SSSE3-NEXT: movdqa %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2632 | ; SSSE3-NEXT: retq |
| 2633 | ; |
| 2634 | ; SSE41-LABEL: trunc_packus_v16i64_v16i8: |
| 2635 | ; SSE41: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2636 | ; SSE41-NEXT: movdqa %xmm0, %xmm8 |
| 2637 | ; SSE41-NEXT: movapd {{.*#+}} xmm9 = [255,255] |
| 2638 | ; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2639 | ; SSE41-NEXT: movdqa %xmm7, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2640 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2641 | ; SSE41-NEXT: movdqa {{.*#+}} xmm12 = [2147483903,2147483903] |
| 2642 | ; SSE41-NEXT: movdqa %xmm12, %xmm11 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2643 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm11 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2644 | ; SSE41-NEXT: pshufd {{.*#+}} xmm13 = xmm11[0,0,2,2] |
| 2645 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2646 | ; SSE41-NEXT: pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2647 | ; SSE41-NEXT: pand %xmm13, %xmm14 |
| 2648 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm11[1,1,3,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2649 | ; SSE41-NEXT: por %xmm14, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2650 | ; SSE41-NEXT: movapd %xmm9, %xmm11 |
| 2651 | ; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm11 |
| 2652 | ; SSE41-NEXT: movdqa %xmm6, %xmm0 |
| 2653 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2654 | ; SSE41-NEXT: movdqa %xmm12, %xmm7 |
| 2655 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm7 |
| 2656 | ; SSE41-NEXT: pshufd {{.*#+}} xmm13 = xmm7[0,0,2,2] |
| 2657 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2658 | ; SSE41-NEXT: pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3] |
| 2659 | ; SSE41-NEXT: pand %xmm13, %xmm14 |
| 2660 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3] |
| 2661 | ; SSE41-NEXT: por %xmm14, %xmm0 |
| 2662 | ; SSE41-NEXT: movapd %xmm9, %xmm13 |
| 2663 | ; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm13 |
| 2664 | ; SSE41-NEXT: movdqa %xmm5, %xmm0 |
| 2665 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2666 | ; SSE41-NEXT: movdqa %xmm12, %xmm6 |
| 2667 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm6 |
| 2668 | ; SSE41-NEXT: pshufd {{.*#+}} xmm14 = xmm6[0,0,2,2] |
| 2669 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2670 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2671 | ; SSE41-NEXT: pand %xmm14, %xmm7 |
| 2672 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3] |
| 2673 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2674 | ; SSE41-NEXT: movapd %xmm9, %xmm14 |
| 2675 | ; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm14 |
| 2676 | ; SSE41-NEXT: movdqa %xmm4, %xmm0 |
| 2677 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2678 | ; SSE41-NEXT: movdqa %xmm12, %xmm5 |
| 2679 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm5 |
| 2680 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] |
| 2681 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2682 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2683 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 2684 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] |
| 2685 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2686 | ; SSE41-NEXT: movapd %xmm9, %xmm15 |
| 2687 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm15 |
| 2688 | ; SSE41-NEXT: movdqa %xmm3, %xmm0 |
| 2689 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2690 | ; SSE41-NEXT: movdqa %xmm12, %xmm4 |
| 2691 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm4 |
| 2692 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] |
| 2693 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2694 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2695 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 2696 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] |
| 2697 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2698 | ; SSE41-NEXT: movapd %xmm9, %xmm4 |
| 2699 | ; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm4 |
| 2700 | ; SSE41-NEXT: movdqa %xmm2, %xmm0 |
| 2701 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2702 | ; SSE41-NEXT: movdqa %xmm12, %xmm3 |
| 2703 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm3 |
| 2704 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] |
| 2705 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2706 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2707 | ; SSE41-NEXT: pand %xmm6, %xmm7 |
| 2708 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2709 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2710 | ; SSE41-NEXT: movapd %xmm9, %xmm6 |
| 2711 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2712 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2713 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2714 | ; SSE41-NEXT: movdqa %xmm12, %xmm2 |
| 2715 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm2 |
| 2716 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2] |
| 2717 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2718 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2719 | ; SSE41-NEXT: pand %xmm3, %xmm7 |
| 2720 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 2721 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2722 | ; SSE41-NEXT: movapd %xmm9, %xmm7 |
| 2723 | ; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm7 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2724 | ; SSE41-NEXT: movdqa %xmm8, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2725 | ; SSE41-NEXT: pxor %xmm10, %xmm0 |
| 2726 | ; SSE41-NEXT: movdqa %xmm12, %xmm1 |
| 2727 | ; SSE41-NEXT: pcmpgtd %xmm0, %xmm1 |
| 2728 | ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] |
| 2729 | ; SSE41-NEXT: pcmpeqd %xmm12, %xmm0 |
| 2730 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] |
| 2731 | ; SSE41-NEXT: pand %xmm2, %xmm3 |
| 2732 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 2733 | ; SSE41-NEXT: por %xmm3, %xmm0 |
| 2734 | ; SSE41-NEXT: movapd %xmm9, %xmm2 |
| 2735 | ; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2736 | ; SSE41-NEXT: movapd %xmm2, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2737 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2738 | ; SSE41-NEXT: movapd %xmm0, %xmm1 |
| 2739 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm1 |
| 2740 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2] |
| 2741 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2742 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 2743 | ; SSE41-NEXT: pand %xmm3, %xmm5 |
| 2744 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] |
| 2745 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 2746 | ; SSE41-NEXT: xorpd %xmm8, %xmm8 |
| 2747 | ; SSE41-NEXT: pxor %xmm1, %xmm1 |
| 2748 | ; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm1 |
| 2749 | ; SSE41-NEXT: movapd %xmm7, %xmm0 |
| 2750 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2751 | ; SSE41-NEXT: movapd %xmm0, %xmm2 |
| 2752 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm2 |
| 2753 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2] |
| 2754 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2755 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] |
| 2756 | ; SSE41-NEXT: pand %xmm5, %xmm3 |
| 2757 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 2758 | ; SSE41-NEXT: por %xmm3, %xmm0 |
| 2759 | ; SSE41-NEXT: pxor %xmm12, %xmm12 |
| 2760 | ; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm12 |
| 2761 | ; SSE41-NEXT: movapd %xmm6, %xmm0 |
| 2762 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2763 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 2764 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 2765 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] |
| 2766 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2767 | ; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] |
| 2768 | ; SSE41-NEXT: pand %xmm5, %xmm7 |
| 2769 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2770 | ; SSE41-NEXT: por %xmm7, %xmm0 |
| 2771 | ; SSE41-NEXT: pxor %xmm7, %xmm7 |
| 2772 | ; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm7 |
| 2773 | ; SSE41-NEXT: movapd %xmm4, %xmm0 |
| 2774 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2775 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 2776 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 2777 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] |
| 2778 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2779 | ; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] |
| 2780 | ; SSE41-NEXT: pand %xmm5, %xmm6 |
| 2781 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2782 | ; SSE41-NEXT: por %xmm6, %xmm0 |
| 2783 | ; SSE41-NEXT: pxor %xmm6, %xmm6 |
| 2784 | ; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2785 | ; SSE41-NEXT: movapd %xmm15, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2786 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2787 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 2788 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 2789 | ; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] |
| 2790 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2791 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 2792 | ; SSE41-NEXT: pand %xmm4, %xmm5 |
| 2793 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2794 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 2795 | ; SSE41-NEXT: pxor %xmm4, %xmm4 |
| 2796 | ; SSE41-NEXT: blendvpd %xmm0, %xmm15, %xmm4 |
| 2797 | ; SSE41-NEXT: movapd %xmm14, %xmm0 |
| 2798 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2799 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 2800 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 2801 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] |
| 2802 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2803 | ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3] |
| 2804 | ; SSE41-NEXT: pand %xmm5, %xmm2 |
| 2805 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2806 | ; SSE41-NEXT: por %xmm2, %xmm0 |
| 2807 | ; SSE41-NEXT: xorpd %xmm15, %xmm15 |
| 2808 | ; SSE41-NEXT: blendvpd %xmm0, %xmm14, %xmm15 |
| 2809 | ; SSE41-NEXT: movapd %xmm13, %xmm0 |
| 2810 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2811 | ; SSE41-NEXT: movapd %xmm0, %xmm2 |
| 2812 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm2 |
| 2813 | ; SSE41-NEXT: pshufd {{.*#+}} xmm14 = xmm2[0,0,2,2] |
| 2814 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2815 | ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] |
| 2816 | ; SSE41-NEXT: pand %xmm14, %xmm3 |
| 2817 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] |
| 2818 | ; SSE41-NEXT: por %xmm3, %xmm0 |
| 2819 | ; SSE41-NEXT: pxor %xmm2, %xmm2 |
| 2820 | ; SSE41-NEXT: blendvpd %xmm0, %xmm13, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2821 | ; SSE41-NEXT: movapd %xmm11, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2822 | ; SSE41-NEXT: xorpd %xmm10, %xmm0 |
| 2823 | ; SSE41-NEXT: movapd %xmm0, %xmm3 |
| 2824 | ; SSE41-NEXT: pcmpgtd %xmm10, %xmm3 |
| 2825 | ; SSE41-NEXT: pshufd {{.*#+}} xmm13 = xmm3[0,0,2,2] |
| 2826 | ; SSE41-NEXT: pcmpeqd %xmm10, %xmm0 |
| 2827 | ; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] |
| 2828 | ; SSE41-NEXT: pand %xmm13, %xmm5 |
| 2829 | ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] |
| 2830 | ; SSE41-NEXT: por %xmm5, %xmm0 |
| 2831 | ; SSE41-NEXT: blendvpd %xmm0, %xmm11, %xmm8 |
| 2832 | ; SSE41-NEXT: andpd %xmm9, %xmm8 |
| 2833 | ; SSE41-NEXT: andpd %xmm9, %xmm2 |
| 2834 | ; SSE41-NEXT: packuswb %xmm8, %xmm2 |
| 2835 | ; SSE41-NEXT: andpd %xmm9, %xmm15 |
| 2836 | ; SSE41-NEXT: andpd %xmm9, %xmm4 |
| 2837 | ; SSE41-NEXT: packuswb %xmm15, %xmm4 |
| 2838 | ; SSE41-NEXT: packuswb %xmm2, %xmm4 |
| 2839 | ; SSE41-NEXT: andpd %xmm9, %xmm6 |
| 2840 | ; SSE41-NEXT: andpd %xmm9, %xmm7 |
| 2841 | ; SSE41-NEXT: packuswb %xmm6, %xmm7 |
| 2842 | ; SSE41-NEXT: andpd %xmm9, %xmm12 |
| 2843 | ; SSE41-NEXT: andpd %xmm9, %xmm1 |
| 2844 | ; SSE41-NEXT: packuswb %xmm12, %xmm1 |
| 2845 | ; SSE41-NEXT: packuswb %xmm7, %xmm1 |
| 2846 | ; SSE41-NEXT: packuswb %xmm4, %xmm1 |
| 2847 | ; SSE41-NEXT: movdqa %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2848 | ; SSE41-NEXT: retq |
| 2849 | ; |
| 2850 | ; AVX1-LABEL: trunc_packus_v16i64_v16i8: |
| 2851 | ; AVX1: # %bb.0: |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2852 | ; AVX1-NEXT: vmovapd {{.*#+}} ymm5 = [255,255,255,255] |
| 2853 | ; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2854 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [255,255] |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2855 | ; AVX1-NEXT: vpcmpgtq %xmm6, %xmm4, %xmm6 |
| 2856 | ; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm7 |
| 2857 | ; AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm7, %ymm6 |
| 2858 | ; AVX1-NEXT: vblendvpd %ymm6, %ymm3, %ymm5, %ymm3 |
| 2859 | ; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm6 |
| 2860 | ; AVX1-NEXT: vpcmpgtq %xmm6, %xmm4, %xmm6 |
| 2861 | ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm4, %xmm7 |
| 2862 | ; AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm7, %ymm6 |
| 2863 | ; AVX1-NEXT: vblendvpd %ymm6, %ymm2, %ymm5, %ymm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2864 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm6 |
| 2865 | ; AVX1-NEXT: vpcmpgtq %xmm6, %xmm4, %xmm6 |
| 2866 | ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm4, %xmm7 |
| 2867 | ; AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm7, %ymm6 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2868 | ; AVX1-NEXT: vblendvpd %ymm6, %ymm1, %ymm5, %ymm1 |
| 2869 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6 |
| 2870 | ; AVX1-NEXT: vpcmpgtq %xmm6, %xmm4, %xmm6 |
| 2871 | ; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm7 |
| 2872 | ; AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm7, %ymm6 |
| 2873 | ; AVX1-NEXT: vblendvpd %ymm6, %ymm0, %ymm5, %ymm0 |
| 2874 | ; AVX1-NEXT: vxorpd %xmm5, %xmm5, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2875 | ; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm6 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2876 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm6, %xmm7 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2877 | ; AVX1-NEXT: vpand %xmm4, %xmm6, %xmm6 |
| 2878 | ; AVX1-NEXT: vpand %xmm6, %xmm7, %xmm6 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2879 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm3, %xmm7 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2880 | ; AVX1-NEXT: vpand %xmm4, %xmm3, %xmm3 |
| 2881 | ; AVX1-NEXT: vpand %xmm3, %xmm7, %xmm3 |
| 2882 | ; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm7 |
| 2883 | ; AVX1-NEXT: vpackuswb %xmm6, %xmm3, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2884 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm7, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2885 | ; AVX1-NEXT: vpand %xmm4, %xmm7, %xmm7 |
| 2886 | ; AVX1-NEXT: vpand %xmm7, %xmm6, %xmm6 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2887 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm2, %xmm7 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2888 | ; AVX1-NEXT: vpand %xmm4, %xmm2, %xmm2 |
| 2889 | ; AVX1-NEXT: vpand %xmm2, %xmm7, %xmm2 |
| 2890 | ; AVX1-NEXT: vpackuswb %xmm6, %xmm2, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2891 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm6 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2892 | ; AVX1-NEXT: vpackuswb %xmm3, %xmm2, %xmm2 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2893 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm6, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2894 | ; AVX1-NEXT: vpand %xmm4, %xmm6, %xmm6 |
| 2895 | ; AVX1-NEXT: vpand %xmm6, %xmm3, %xmm3 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2896 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm1, %xmm6 |
| 2897 | ; AVX1-NEXT: vpand %xmm4, %xmm1, %xmm1 |
| 2898 | ; AVX1-NEXT: vpand %xmm1, %xmm6, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2899 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2900 | ; AVX1-NEXT: vpackuswb %xmm3, %xmm1, %xmm1 |
| 2901 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm6, %xmm3 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2902 | ; AVX1-NEXT: vpand %xmm4, %xmm6, %xmm6 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2903 | ; AVX1-NEXT: vpand %xmm6, %xmm3, %xmm3 |
| 2904 | ; AVX1-NEXT: vpcmpgtq %xmm5, %xmm0, %xmm5 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2905 | ; AVX1-NEXT: vpand %xmm4, %xmm0, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2906 | ; AVX1-NEXT: vpand %xmm0, %xmm5, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2907 | ; AVX1-NEXT: vpackuswb %xmm3, %xmm0, %xmm0 |
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2908 | ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2909 | ; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0 |
| 2910 | ; AVX1-NEXT: vzeroupper |
| 2911 | ; AVX1-NEXT: retq |
| 2912 | ; |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 2913 | ; AVX2-LABEL: trunc_packus_v16i64_v16i8: |
| 2914 | ; AVX2: # %bb.0: |
| 2915 | ; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [255,255,255,255] |
| 2916 | ; AVX2-NEXT: vpcmpgtq %ymm3, %ymm4, %ymm5 |
| 2917 | ; AVX2-NEXT: vblendvpd %ymm5, %ymm3, %ymm4, %ymm3 |
| 2918 | ; AVX2-NEXT: vpcmpgtq %ymm2, %ymm4, %ymm5 |
| 2919 | ; AVX2-NEXT: vblendvpd %ymm5, %ymm2, %ymm4, %ymm2 |
| 2920 | ; AVX2-NEXT: vpcmpgtq %ymm1, %ymm4, %ymm5 |
| 2921 | ; AVX2-NEXT: vblendvpd %ymm5, %ymm1, %ymm4, %ymm1 |
| 2922 | ; AVX2-NEXT: vpcmpgtq %ymm0, %ymm4, %ymm5 |
| 2923 | ; AVX2-NEXT: vblendvpd %ymm5, %ymm0, %ymm4, %ymm0 |
| 2924 | ; AVX2-NEXT: vpxor %xmm4, %xmm4, %xmm4 |
| 2925 | ; AVX2-NEXT: vpcmpgtq %ymm4, %ymm0, %ymm5 |
| 2926 | ; AVX2-NEXT: vpand %ymm0, %ymm5, %ymm0 |
| 2927 | ; AVX2-NEXT: vpcmpgtq %ymm4, %ymm1, %ymm5 |
| 2928 | ; AVX2-NEXT: vpand %ymm1, %ymm5, %ymm1 |
| 2929 | ; AVX2-NEXT: vpcmpgtq %ymm4, %ymm2, %ymm5 |
| 2930 | ; AVX2-NEXT: vpand %ymm2, %ymm5, %ymm2 |
| 2931 | ; AVX2-NEXT: vpcmpgtq %ymm4, %ymm3, %ymm4 |
| 2932 | ; AVX2-NEXT: vpand %ymm3, %ymm4, %ymm3 |
| 2933 | ; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4 |
| 2934 | ; AVX2-NEXT: vpackssdw %xmm4, %xmm3, %xmm3 |
| 2935 | ; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm4 |
| 2936 | ; AVX2-NEXT: vpackssdw %xmm4, %xmm2, %xmm2 |
| 2937 | ; AVX2-NEXT: vpackssdw %xmm3, %xmm2, %xmm2 |
| 2938 | ; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u> |
| 2939 | ; AVX2-NEXT: vpshufb %xmm3, %xmm2, %xmm2 |
| 2940 | ; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm4 |
| 2941 | ; AVX2-NEXT: vpackssdw %xmm4, %xmm1, %xmm1 |
| 2942 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm4 |
| 2943 | ; AVX2-NEXT: vpackssdw %xmm4, %xmm0, %xmm0 |
| 2944 | ; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 |
| 2945 | ; AVX2-NEXT: vpshufb %xmm3, %xmm0, %xmm0 |
| 2946 | ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0] |
| 2947 | ; AVX2-NEXT: vzeroupper |
| 2948 | ; AVX2-NEXT: retq |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2949 | ; |
| 2950 | ; AVX512-LABEL: trunc_packus_v16i64_v16i8: |
| 2951 | ; AVX512: # %bb.0: |
| 2952 | ; AVX512-NEXT: vpbroadcastq {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255] |
| 2953 | ; AVX512-NEXT: vpminsq %zmm2, %zmm0, %zmm0 |
| 2954 | ; AVX512-NEXT: vpminsq %zmm2, %zmm1, %zmm1 |
| 2955 | ; AVX512-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| 2956 | ; AVX512-NEXT: vpmaxsq %zmm2, %zmm1, %zmm1 |
| 2957 | ; AVX512-NEXT: vpmaxsq %zmm2, %zmm0, %zmm0 |
| 2958 | ; AVX512-NEXT: vpmovqd %zmm0, %ymm0 |
| 2959 | ; AVX512-NEXT: vpmovqd %zmm1, %ymm1 |
| 2960 | ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 |
| 2961 | ; AVX512-NEXT: vpmovdb %zmm0, %xmm0 |
| 2962 | ; AVX512-NEXT: vzeroupper |
| 2963 | ; AVX512-NEXT: retq |
| 2964 | %1 = icmp slt <16 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> |
| 2965 | %2 = select <16 x i1> %1, <16 x i64> %a0, <16 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> |
| 2966 | %3 = icmp sgt <16 x i64> %2, zeroinitializer |
| 2967 | %4 = select <16 x i1> %3, <16 x i64> %2, <16 x i64> zeroinitializer |
| 2968 | %5 = trunc <16 x i64> %4 to <16 x i8> |
| 2969 | ret <16 x i8> %5 |
| 2970 | } |
| 2971 | |
| 2972 | define <8 x i8> @trunc_packus_v8i32_v8i8(<8 x i32> %a0) { |
| 2973 | ; SSE2-LABEL: trunc_packus_v8i32_v8i8: |
| 2974 | ; SSE2: # %bb.0: |
| 2975 | ; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255] |
| 2976 | ; SSE2-NEXT: movdqa %xmm2, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 2977 | ; SSE2-NEXT: pcmpgtd %xmm0, %xmm3 |
| 2978 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2979 | ; SSE2-NEXT: pandn %xmm2, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 2980 | ; SSE2-NEXT: por %xmm0, %xmm3 |
| 2981 | ; SSE2-NEXT: movdqa %xmm2, %xmm0 |
| 2982 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm0 |
| 2983 | ; SSE2-NEXT: pand %xmm0, %xmm1 |
| 2984 | ; SSE2-NEXT: pandn %xmm2, %xmm0 |
| 2985 | ; SSE2-NEXT: por %xmm1, %xmm0 |
| 2986 | ; SSE2-NEXT: pxor %xmm1, %xmm1 |
| 2987 | ; SSE2-NEXT: movdqa %xmm0, %xmm2 |
| 2988 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm2 |
| 2989 | ; SSE2-NEXT: pand %xmm0, %xmm2 |
| 2990 | ; SSE2-NEXT: movdqa %xmm3, %xmm0 |
| 2991 | ; SSE2-NEXT: pcmpgtd %xmm1, %xmm0 |
| 2992 | ; SSE2-NEXT: pand %xmm3, %xmm0 |
| 2993 | ; SSE2-NEXT: packssdw %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2994 | ; SSE2-NEXT: retq |
| 2995 | ; |
| 2996 | ; SSSE3-LABEL: trunc_packus_v8i32_v8i8: |
| 2997 | ; SSSE3: # %bb.0: |
| 2998 | ; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255] |
| 2999 | ; SSSE3-NEXT: movdqa %xmm2, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3000 | ; SSSE3-NEXT: pcmpgtd %xmm0, %xmm3 |
| 3001 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3002 | ; SSSE3-NEXT: pandn %xmm2, %xmm3 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3003 | ; SSSE3-NEXT: por %xmm0, %xmm3 |
| 3004 | ; SSSE3-NEXT: movdqa %xmm2, %xmm0 |
| 3005 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm0 |
| 3006 | ; SSSE3-NEXT: pand %xmm0, %xmm1 |
| 3007 | ; SSSE3-NEXT: pandn %xmm2, %xmm0 |
| 3008 | ; SSSE3-NEXT: por %xmm1, %xmm0 |
| 3009 | ; SSSE3-NEXT: pxor %xmm1, %xmm1 |
| 3010 | ; SSSE3-NEXT: movdqa %xmm0, %xmm2 |
| 3011 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm2 |
| 3012 | ; SSSE3-NEXT: pand %xmm0, %xmm2 |
| 3013 | ; SSSE3-NEXT: movdqa %xmm3, %xmm0 |
| 3014 | ; SSSE3-NEXT: pcmpgtd %xmm1, %xmm0 |
| 3015 | ; SSSE3-NEXT: pand %xmm3, %xmm0 |
| 3016 | ; SSSE3-NEXT: packssdw %xmm2, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3017 | ; SSSE3-NEXT: retq |
| 3018 | ; |
| 3019 | ; SSE41-LABEL: trunc_packus_v8i32_v8i8: |
| 3020 | ; SSE41: # %bb.0: |
| 3021 | ; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3022 | ; SSE41-NEXT: pminsd %xmm2, %xmm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3023 | ; SSE41-NEXT: pminsd %xmm2, %xmm1 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3024 | ; SSE41-NEXT: pxor %xmm2, %xmm2 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3025 | ; SSE41-NEXT: pmaxsd %xmm2, %xmm1 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3026 | ; SSE41-NEXT: pmaxsd %xmm2, %xmm0 |
| 3027 | ; SSE41-NEXT: packssdw %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3028 | ; SSE41-NEXT: retq |
| 3029 | ; |
| 3030 | ; AVX1-LABEL: trunc_packus_v8i32_v8i8: |
| 3031 | ; AVX1: # %bb.0: |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3032 | ; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [255,255,255,255] |
| 3033 | ; AVX1-NEXT: vpminsd %xmm1, %xmm0, %xmm2 |
| 3034 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0 |
| 3035 | ; AVX1-NEXT: vpminsd %xmm1, %xmm0, %xmm0 |
| 3036 | ; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3037 | ; AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0 |
| 3038 | ; AVX1-NEXT: vpmaxsd %xmm1, %xmm2, %xmm1 |
| 3039 | ; AVX1-NEXT: vpackssdw %xmm0, %xmm1, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3040 | ; AVX1-NEXT: vzeroupper |
| 3041 | ; AVX1-NEXT: retq |
| 3042 | ; |
| 3043 | ; AVX2-LABEL: trunc_packus_v8i32_v8i8: |
| 3044 | ; AVX2: # %bb.0: |
| 3045 | ; AVX2-NEXT: vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] |
| 3046 | ; AVX2-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 3047 | ; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3048 | ; AVX2-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3049 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 3050 | ; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3051 | ; AVX2-NEXT: vzeroupper |
| 3052 | ; AVX2-NEXT: retq |
| 3053 | ; |
| 3054 | ; AVX512F-LABEL: trunc_packus_v8i32_v8i8: |
| 3055 | ; AVX512F: # %bb.0: |
| 3056 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] |
| 3057 | ; AVX512F-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 3058 | ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3059 | ; AVX512F-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3060 | ; AVX512F-NEXT: vpmovdw %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3061 | ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3062 | ; AVX512F-NEXT: vzeroupper |
| 3063 | ; AVX512F-NEXT: retq |
| 3064 | ; |
| 3065 | ; AVX512VL-LABEL: trunc_packus_v8i32_v8i8: |
| 3066 | ; AVX512VL: # %bb.0: |
| 3067 | ; AVX512VL-NEXT: vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 |
| 3068 | ; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3069 | ; AVX512VL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3070 | ; AVX512VL-NEXT: vpmovdw %ymm0, %xmm0 |
| 3071 | ; AVX512VL-NEXT: vzeroupper |
| 3072 | ; AVX512VL-NEXT: retq |
| 3073 | ; |
| 3074 | ; AVX512BW-LABEL: trunc_packus_v8i32_v8i8: |
| 3075 | ; AVX512BW: # %bb.0: |
| 3076 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] |
| 3077 | ; AVX512BW-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 3078 | ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3079 | ; AVX512BW-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3080 | ; AVX512BW-NEXT: vpmovdw %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3081 | ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3082 | ; AVX512BW-NEXT: vzeroupper |
| 3083 | ; AVX512BW-NEXT: retq |
| 3084 | ; |
| 3085 | ; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i8: |
| 3086 | ; AVX512BWVL: # %bb.0: |
| 3087 | ; AVX512BWVL-NEXT: vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 |
| 3088 | ; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3089 | ; AVX512BWVL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3090 | ; AVX512BWVL-NEXT: vpmovdw %ymm0, %xmm0 |
| 3091 | ; AVX512BWVL-NEXT: vzeroupper |
| 3092 | ; AVX512BWVL-NEXT: retq |
| 3093 | %1 = icmp slt <8 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> |
| 3094 | %2 = select <8 x i1> %1, <8 x i32> %a0, <8 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> |
| 3095 | %3 = icmp sgt <8 x i32> %2, zeroinitializer |
| 3096 | %4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer |
| 3097 | %5 = trunc <8 x i32> %4 to <8 x i8> |
| 3098 | ret <8 x i8> %5 |
| 3099 | } |
| 3100 | |
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 3101 | define void @trunc_packus_v8i32_v8i8_store(<8 x i32> %a0, <8 x i8> *%p1) { |
| Simon Pilgrim | c302a58 | 2018-02-19 13:29:20 +0000 | [diff] [blame^] | 3102 | ; SSE-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3103 | ; SSE: # %bb.0: |
| 3104 | ; SSE-NEXT: packssdw %xmm1, %xmm0 |
| 3105 | ; SSE-NEXT: packuswb %xmm0, %xmm0 |
| 3106 | ; SSE-NEXT: movq %xmm0, (%rdi) |
| 3107 | ; SSE-NEXT: retq |
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 3108 | ; |
| 3109 | ; AVX1-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3110 | ; AVX1: # %bb.0: |
| Simon Pilgrim | c302a58 | 2018-02-19 13:29:20 +0000 | [diff] [blame^] | 3111 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| 3112 | ; AVX1-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 |
| 3113 | ; AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 |
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 3114 | ; AVX1-NEXT: vmovq %xmm0, (%rdi) |
| 3115 | ; AVX1-NEXT: vzeroupper |
| 3116 | ; AVX1-NEXT: retq |
| 3117 | ; |
| 3118 | ; AVX2-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3119 | ; AVX2: # %bb.0: |
| Simon Pilgrim | 7fae42e | 2018-02-17 22:19:50 +0000 | [diff] [blame] | 3120 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 3121 | ; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | c302a58 | 2018-02-19 13:29:20 +0000 | [diff] [blame^] | 3122 | ; AVX2-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 |
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 3123 | ; AVX2-NEXT: vmovq %xmm0, (%rdi) |
| 3124 | ; AVX2-NEXT: vzeroupper |
| 3125 | ; AVX2-NEXT: retq |
| 3126 | ; |
| 3127 | ; AVX512F-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3128 | ; AVX512F: # %bb.0: |
| 3129 | ; AVX512F-NEXT: vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] |
| 3130 | ; AVX512F-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 3131 | ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3132 | ; AVX512F-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3133 | ; AVX512F-NEXT: vpmovdw %zmm0, %ymm0 |
| 3134 | ; AVX512F-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] |
| 3135 | ; AVX512F-NEXT: vmovq %xmm0, (%rdi) |
| 3136 | ; AVX512F-NEXT: vzeroupper |
| 3137 | ; AVX512F-NEXT: retq |
| 3138 | ; |
| 3139 | ; AVX512VL-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3140 | ; AVX512VL: # %bb.0: |
| 3141 | ; AVX512VL-NEXT: vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 |
| 3142 | ; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3143 | ; AVX512VL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3144 | ; AVX512VL-NEXT: vpmovdb %ymm0, (%rdi) |
| 3145 | ; AVX512VL-NEXT: vzeroupper |
| 3146 | ; AVX512VL-NEXT: retq |
| 3147 | ; |
| 3148 | ; AVX512BW-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3149 | ; AVX512BW: # %bb.0: |
| 3150 | ; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] |
| 3151 | ; AVX512BW-NEXT: vpminsd %ymm1, %ymm0, %ymm0 |
| 3152 | ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3153 | ; AVX512BW-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3154 | ; AVX512BW-NEXT: vpmovdw %zmm0, %ymm0 |
| 3155 | ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] |
| 3156 | ; AVX512BW-NEXT: vmovq %xmm0, (%rdi) |
| 3157 | ; AVX512BW-NEXT: vzeroupper |
| 3158 | ; AVX512BW-NEXT: retq |
| 3159 | ; |
| 3160 | ; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i8_store: |
| 3161 | ; AVX512BWVL: # %bb.0: |
| 3162 | ; AVX512BWVL-NEXT: vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 |
| 3163 | ; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3164 | ; AVX512BWVL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0 |
| 3165 | ; AVX512BWVL-NEXT: vpmovdb %ymm0, (%rdi) |
| 3166 | ; AVX512BWVL-NEXT: vzeroupper |
| 3167 | ; AVX512BWVL-NEXT: retq |
| 3168 | %1 = icmp slt <8 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> |
| 3169 | %2 = select <8 x i1> %1, <8 x i32> %a0, <8 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> |
| 3170 | %3 = icmp sgt <8 x i32> %2, zeroinitializer |
| 3171 | %4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer |
| 3172 | %5 = trunc <8 x i32> %4 to <8 x i8> |
| 3173 | store <8 x i8> %5, <8 x i8> *%p1 |
| 3174 | ret void |
| 3175 | } |
| 3176 | |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3177 | define <16 x i8> @trunc_packus_v16i32_v16i8(<16 x i32> %a0) { |
| Simon Pilgrim | 17bb6f0 | 2018-02-15 14:37:59 +0000 | [diff] [blame] | 3178 | ; SSE-LABEL: trunc_packus_v16i32_v16i8: |
| 3179 | ; SSE: # %bb.0: |
| 3180 | ; SSE-NEXT: packssdw %xmm3, %xmm2 |
| 3181 | ; SSE-NEXT: packssdw %xmm1, %xmm0 |
| 3182 | ; SSE-NEXT: packuswb %xmm2, %xmm0 |
| 3183 | ; SSE-NEXT: retq |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3184 | ; |
| 3185 | ; AVX1-LABEL: trunc_packus_v16i32_v16i8: |
| 3186 | ; AVX1: # %bb.0: |
| 3187 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2 |
| Simon Pilgrim | 17bb6f0 | 2018-02-15 14:37:59 +0000 | [diff] [blame] | 3188 | ; AVX1-NEXT: vpackssdw %xmm2, %xmm1, %xmm1 |
| 3189 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2 |
| 3190 | ; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3191 | ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 |
| 3192 | ; AVX1-NEXT: vzeroupper |
| 3193 | ; AVX1-NEXT: retq |
| 3194 | ; |
| 3195 | ; AVX2-LABEL: trunc_packus_v16i32_v16i8: |
| 3196 | ; AVX2: # %bb.0: |
| Simon Pilgrim | 17bb6f0 | 2018-02-15 14:37:59 +0000 | [diff] [blame] | 3197 | ; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0 |
| 3198 | ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| 3199 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| 3200 | ; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3201 | ; AVX2-NEXT: vzeroupper |
| 3202 | ; AVX2-NEXT: retq |
| 3203 | ; |
| 3204 | ; AVX512-LABEL: trunc_packus_v16i32_v16i8: |
| 3205 | ; AVX512: # %bb.0: |
| 3206 | ; AVX512-NEXT: vpminsd {{.*}}(%rip){1to16}, %zmm0, %zmm0 |
| 3207 | ; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3208 | ; AVX512-NEXT: vpmaxsd %zmm1, %zmm0, %zmm0 |
| 3209 | ; AVX512-NEXT: vpmovdb %zmm0, %xmm0 |
| 3210 | ; AVX512-NEXT: vzeroupper |
| 3211 | ; AVX512-NEXT: retq |
| 3212 | %1 = icmp slt <16 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> |
| 3213 | %2 = select <16 x i1> %1, <16 x i32> %a0, <16 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> |
| 3214 | %3 = icmp sgt <16 x i32> %2, zeroinitializer |
| 3215 | %4 = select <16 x i1> %3, <16 x i32> %2, <16 x i32> zeroinitializer |
| 3216 | %5 = trunc <16 x i32> %4 to <16 x i8> |
| 3217 | ret <16 x i8> %5 |
| 3218 | } |
| 3219 | |
| 3220 | define <16 x i8> @trunc_packus_v16i16_v16i8(<16 x i16> %a0) { |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 3221 | ; SSE-LABEL: trunc_packus_v16i16_v16i8: |
| 3222 | ; SSE: # %bb.0: |
| 3223 | ; SSE-NEXT: packuswb %xmm1, %xmm0 |
| 3224 | ; SSE-NEXT: retq |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3225 | ; |
| 3226 | ; AVX1-LABEL: trunc_packus_v16i16_v16i8: |
| 3227 | ; AVX1: # %bb.0: |
| 3228 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | ae00a71 | 2018-02-06 14:07:46 +0000 | [diff] [blame] | 3229 | ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3230 | ; AVX1-NEXT: vzeroupper |
| 3231 | ; AVX1-NEXT: retq |
| 3232 | ; |
| 3233 | ; AVX2-LABEL: trunc_packus_v16i16_v16i8: |
| 3234 | ; AVX2: # %bb.0: |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3235 | ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 |
| Simon Pilgrim | ae00a71 | 2018-02-06 14:07:46 +0000 | [diff] [blame] | 3236 | ; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3237 | ; AVX2-NEXT: vzeroupper |
| 3238 | ; AVX2-NEXT: retq |
| 3239 | ; |
| 3240 | ; AVX512F-LABEL: trunc_packus_v16i16_v16i8: |
| 3241 | ; AVX512F: # %bb.0: |
| 3242 | ; AVX512F-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 |
| 3243 | ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3244 | ; AVX512F-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 |
| 3245 | ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0 |
| 3246 | ; AVX512F-NEXT: vpmovdb %zmm0, %xmm0 |
| 3247 | ; AVX512F-NEXT: vzeroupper |
| 3248 | ; AVX512F-NEXT: retq |
| 3249 | ; |
| 3250 | ; AVX512VL-LABEL: trunc_packus_v16i16_v16i8: |
| 3251 | ; AVX512VL: # %bb.0: |
| 3252 | ; AVX512VL-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 |
| 3253 | ; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3254 | ; AVX512VL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 |
| 3255 | ; AVX512VL-NEXT: vpmovsxwd %ymm0, %zmm0 |
| 3256 | ; AVX512VL-NEXT: vpmovdb %zmm0, %xmm0 |
| 3257 | ; AVX512VL-NEXT: vzeroupper |
| 3258 | ; AVX512VL-NEXT: retq |
| 3259 | ; |
| 3260 | ; AVX512BW-LABEL: trunc_packus_v16i16_v16i8: |
| 3261 | ; AVX512BW: # %bb.0: |
| 3262 | ; AVX512BW-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 |
| 3263 | ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3264 | ; AVX512BW-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 |
| 3265 | ; AVX512BW-NEXT: vpmovwb %zmm0, %ymm0 |
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3266 | ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3267 | ; AVX512BW-NEXT: vzeroupper |
| 3268 | ; AVX512BW-NEXT: retq |
| 3269 | ; |
| 3270 | ; AVX512BWVL-LABEL: trunc_packus_v16i16_v16i8: |
| 3271 | ; AVX512BWVL: # %bb.0: |
| 3272 | ; AVX512BWVL-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 |
| 3273 | ; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3274 | ; AVX512BWVL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 |
| 3275 | ; AVX512BWVL-NEXT: vpmovwb %ymm0, %xmm0 |
| 3276 | ; AVX512BWVL-NEXT: vzeroupper |
| 3277 | ; AVX512BWVL-NEXT: retq |
| 3278 | %1 = icmp slt <16 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> |
| 3279 | %2 = select <16 x i1> %1, <16 x i16> %a0, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> |
| 3280 | %3 = icmp sgt <16 x i16> %2, zeroinitializer |
| 3281 | %4 = select <16 x i1> %3, <16 x i16> %2, <16 x i16> zeroinitializer |
| 3282 | %5 = trunc <16 x i16> %4 to <16 x i8> |
| 3283 | ret <16 x i8> %5 |
| 3284 | } |
| 3285 | |
| 3286 | define <32 x i8> @trunc_packus_v32i16_v32i8(<32 x i16> %a0) { |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 3287 | ; SSE-LABEL: trunc_packus_v32i16_v32i8: |
| 3288 | ; SSE: # %bb.0: |
| 3289 | ; SSE-NEXT: packuswb %xmm1, %xmm0 |
| 3290 | ; SSE-NEXT: packuswb %xmm3, %xmm2 |
| 3291 | ; SSE-NEXT: movdqa %xmm2, %xmm1 |
| 3292 | ; SSE-NEXT: retq |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3293 | ; |
| 3294 | ; AVX1-LABEL: trunc_packus_v32i16_v32i8: |
| 3295 | ; AVX1: # %bb.0: |
| 3296 | ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2 |
| Simon Pilgrim | ae00a71 | 2018-02-06 14:07:46 +0000 | [diff] [blame] | 3297 | ; AVX1-NEXT: vpackuswb %xmm2, %xmm1, %xmm1 |
| 3298 | ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2 |
| 3299 | ; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0 |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3300 | ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 |
| 3301 | ; AVX1-NEXT: retq |
| 3302 | ; |
| 3303 | ; AVX2-LABEL: trunc_packus_v32i16_v32i8: |
| 3304 | ; AVX2: # %bb.0: |
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 3305 | ; AVX2-NEXT: vpackuswb %ymm1, %ymm0, %ymm0 |
| 3306 | ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3307 | ; AVX2-NEXT: retq |
| 3308 | ; |
| 3309 | ; AVX512F-LABEL: trunc_packus_v32i16_v32i8: |
| 3310 | ; AVX512F: # %bb.0: |
| 3311 | ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] |
| 3312 | ; AVX512F-NEXT: vpminsw %ymm2, %ymm0, %ymm0 |
| 3313 | ; AVX512F-NEXT: vpminsw %ymm2, %ymm1, %ymm1 |
| 3314 | ; AVX512F-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| 3315 | ; AVX512F-NEXT: vpmaxsw %ymm2, %ymm1, %ymm1 |
| 3316 | ; AVX512F-NEXT: vpmaxsw %ymm2, %ymm0, %ymm0 |
| 3317 | ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0 |
| 3318 | ; AVX512F-NEXT: vpmovdb %zmm0, %xmm0 |
| 3319 | ; AVX512F-NEXT: vpmovsxwd %ymm1, %zmm1 |
| 3320 | ; AVX512F-NEXT: vpmovdb %zmm1, %xmm1 |
| 3321 | ; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| 3322 | ; AVX512F-NEXT: retq |
| 3323 | ; |
| 3324 | ; AVX512VL-LABEL: trunc_packus_v32i16_v32i8: |
| 3325 | ; AVX512VL: # %bb.0: |
| 3326 | ; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] |
| 3327 | ; AVX512VL-NEXT: vpminsw %ymm2, %ymm0, %ymm0 |
| 3328 | ; AVX512VL-NEXT: vpminsw %ymm2, %ymm1, %ymm1 |
| 3329 | ; AVX512VL-NEXT: vpxor %xmm2, %xmm2, %xmm2 |
| 3330 | ; AVX512VL-NEXT: vpmaxsw %ymm2, %ymm1, %ymm1 |
| 3331 | ; AVX512VL-NEXT: vpmaxsw %ymm2, %ymm0, %ymm0 |
| 3332 | ; AVX512VL-NEXT: vpmovsxwd %ymm0, %zmm0 |
| 3333 | ; AVX512VL-NEXT: vpmovdb %zmm0, %xmm0 |
| 3334 | ; AVX512VL-NEXT: vpmovsxwd %ymm1, %zmm1 |
| 3335 | ; AVX512VL-NEXT: vpmovdb %zmm1, %xmm1 |
| 3336 | ; AVX512VL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| 3337 | ; AVX512VL-NEXT: retq |
| 3338 | ; |
| 3339 | ; AVX512BW-LABEL: trunc_packus_v32i16_v32i8: |
| 3340 | ; AVX512BW: # %bb.0: |
| 3341 | ; AVX512BW-NEXT: vpminsw {{.*}}(%rip), %zmm0, %zmm0 |
| 3342 | ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3343 | ; AVX512BW-NEXT: vpmaxsw %zmm1, %zmm0, %zmm0 |
| 3344 | ; AVX512BW-NEXT: vpmovwb %zmm0, %ymm0 |
| 3345 | ; AVX512BW-NEXT: retq |
| 3346 | ; |
| 3347 | ; AVX512BWVL-LABEL: trunc_packus_v32i16_v32i8: |
| 3348 | ; AVX512BWVL: # %bb.0: |
| 3349 | ; AVX512BWVL-NEXT: vpminsw {{.*}}(%rip), %zmm0, %zmm0 |
| 3350 | ; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1 |
| 3351 | ; AVX512BWVL-NEXT: vpmaxsw %zmm1, %zmm0, %zmm0 |
| 3352 | ; AVX512BWVL-NEXT: vpmovwb %zmm0, %ymm0 |
| 3353 | ; AVX512BWVL-NEXT: retq |
| 3354 | %1 = icmp slt <32 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> |
| 3355 | %2 = select <32 x i1> %1, <32 x i16> %a0, <32 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> |
| 3356 | %3 = icmp sgt <32 x i16> %2, zeroinitializer |
| 3357 | %4 = select <32 x i1> %3, <32 x i16> %2, <32 x i16> zeroinitializer |
| 3358 | %5 = trunc <32 x i16> %4 to <32 x i8> |
| 3359 | ret <32 x i8> %5 |
| 3360 | } |