| 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] | 
|  | 1385 | ; SSE2-NEXT:    movdqa %xmm1, %xmm5 | 
|  | 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] | 
|  | 1388 | ; SSE2-NEXT:    movdqa %xmm9, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1389 | ; SSE2-NEXT:    pcmpgtd %xmm5, %xmm6 | 
|  | 1390 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] | 
|  | 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] | 
|  | 1393 | ; SSE2-NEXT:    pand %xmm7, %xmm4 | 
|  | 1394 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3] | 
|  | 1395 | ; SSE2-NEXT:    por %xmm4, %xmm5 | 
|  | 1396 | ; SSE2-NEXT:    pand %xmm5, %xmm1 | 
|  | 1397 | ; SSE2-NEXT:    pandn %xmm8, %xmm5 | 
|  | 1398 | ; SSE2-NEXT:    por %xmm1, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1399 | ; SSE2-NEXT:    movdqa %xmm0, %xmm1 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1400 | ; SSE2-NEXT:    pxor %xmm10, %xmm1 | 
|  | 1401 | ; SSE2-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1402 | ; SSE2-NEXT:    pcmpgtd %xmm1, %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 | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1404 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1 | 
|  | 1405 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] | 
|  | 1406 | ; SSE2-NEXT:    pand %xmm6, %xmm7 | 
|  | 1407 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] | 
|  | 1408 | ; SSE2-NEXT:    por %xmm7, %xmm1 | 
|  | 1409 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 1410 | ; SSE2-NEXT:    pandn %xmm8, %xmm1 | 
|  | 1411 | ; SSE2-NEXT:    por %xmm0, %xmm1 | 
|  | 1412 | ; SSE2-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 1413 | ; SSE2-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1414 | ; SSE2-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1415 | ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4 | 
|  | 1416 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
|  | 1417 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1418 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 1419 | ; SSE2-NEXT:    pand %xmm6, %xmm0 | 
|  | 1420 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] | 
|  | 1421 | ; SSE2-NEXT:    por %xmm0, %xmm6 | 
|  | 1422 | ; SSE2-NEXT:    pand %xmm6, %xmm3 | 
|  | 1423 | ; SSE2-NEXT:    pandn %xmm8, %xmm6 | 
|  | 1424 | ; SSE2-NEXT:    por %xmm3, %xmm6 | 
|  | 1425 | ; SSE2-NEXT:    movdqa %xmm2, %xmm0 | 
|  | 1426 | ; SSE2-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1427 | ; SSE2-NEXT:    movdqa %xmm9, %xmm3 | 
|  | 1428 | ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm3 | 
|  | 1429 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 1430 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1431 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 1432 | ; SSE2-NEXT:    pand %xmm4, %xmm0 | 
|  | 1433 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 1434 | ; SSE2-NEXT:    por %xmm0, %xmm3 | 
|  | 1435 | ; SSE2-NEXT:    pand %xmm3, %xmm2 | 
|  | 1436 | ; SSE2-NEXT:    pandn %xmm8, %xmm3 | 
|  | 1437 | ; SSE2-NEXT:    por %xmm2, %xmm3 | 
|  | 1438 | ; SSE2-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 1439 | ; SSE2-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1440 | ; SSE2-NEXT:    movdqa %xmm0, %xmm2 | 
|  | 1441 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm2 | 
|  | 1442 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 1443 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1444 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 1445 | ; SSE2-NEXT:    pand %xmm4, %xmm7 | 
|  | 1446 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 1447 | ; SSE2-NEXT:    por %xmm7, %xmm0 | 
|  | 1448 | ; SSE2-NEXT:    pand %xmm3, %xmm0 | 
|  | 1449 | ; SSE2-NEXT:    movdqa %xmm6, %xmm2 | 
|  | 1450 | ; SSE2-NEXT:    pxor %xmm10, %xmm2 | 
|  | 1451 | ; SSE2-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 1452 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 1453 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 1454 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm2 | 
|  | 1455 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] | 
|  | 1456 | ; SSE2-NEXT:    pand %xmm4, %xmm7 | 
|  | 1457 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3] | 
|  | 1458 | ; SSE2-NEXT:    por %xmm7, %xmm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1459 | ; SSE2-NEXT:    pand %xmm6, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1460 | ; SSE2-NEXT:    movdqa %xmm1, %xmm3 | 
|  | 1461 | ; SSE2-NEXT:    pxor %xmm10, %xmm3 | 
|  | 1462 | ; SSE2-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 1463 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm4 | 
|  | 1464 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
|  | 1465 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm3 | 
|  | 1466 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 1467 | ; SSE2-NEXT:    pand %xmm6, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1468 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1469 | ; SSE2-NEXT:    por %xmm3, %xmm4 | 
|  | 1470 | ; SSE2-NEXT:    pand %xmm1, %xmm4 | 
|  | 1471 | ; SSE2-NEXT:    movdqa %xmm5, %xmm1 | 
|  | 1472 | ; SSE2-NEXT:    pxor %xmm10, %xmm1 | 
|  | 1473 | ; SSE2-NEXT:    movdqa %xmm1, %xmm3 | 
|  | 1474 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 1475 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] | 
|  | 1476 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm1 | 
|  | 1477 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] | 
|  | 1478 | ; SSE2-NEXT:    pand %xmm6, %xmm1 | 
|  | 1479 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 1480 | ; SSE2-NEXT:    por %xmm1, %xmm3 | 
|  | 1481 | ; SSE2-NEXT:    pand %xmm5, %xmm3 | 
|  | 1482 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1483 | ; 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] | 1484 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3] | 
|  | 1485 | ; SSE2-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7] | 
|  | 1486 | ; SSE2-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] | 
|  | 1487 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3] | 
|  | 1488 | ; SSE2-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7] | 
|  | 1489 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1490 | ; 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] | 1491 | ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] | 
|  | 1492 | ; SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1493 | ; SSE2-NEXT:    retq | 
|  | 1494 | ; | 
|  | 1495 | ; SSSE3-LABEL: trunc_packus_v8i64_v8i8: | 
|  | 1496 | ; SSSE3:       # %bb.0: | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1497 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm8 = [255,255] | 
|  | 1498 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] | 
|  | 1499 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm5 | 
|  | 1500 | ; SSSE3-NEXT:    pxor %xmm10, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1501 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm9 = [2147483903,2147483903] | 
|  | 1502 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1503 | ; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm6 | 
|  | 1504 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] | 
|  | 1505 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm5 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1506 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] | 
|  | 1507 | ; SSSE3-NEXT:    pand %xmm7, %xmm4 | 
|  | 1508 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3] | 
|  | 1509 | ; SSSE3-NEXT:    por %xmm4, %xmm5 | 
|  | 1510 | ; SSSE3-NEXT:    pand %xmm5, %xmm1 | 
|  | 1511 | ; SSSE3-NEXT:    pandn %xmm8, %xmm5 | 
|  | 1512 | ; SSSE3-NEXT:    por %xmm1, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1513 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm1 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1514 | ; SSSE3-NEXT:    pxor %xmm10, %xmm1 | 
|  | 1515 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1516 | ; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1517 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1518 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1 | 
|  | 1519 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] | 
|  | 1520 | ; SSSE3-NEXT:    pand %xmm6, %xmm7 | 
|  | 1521 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] | 
|  | 1522 | ; SSSE3-NEXT:    por %xmm7, %xmm1 | 
|  | 1523 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 1524 | ; SSSE3-NEXT:    pandn %xmm8, %xmm1 | 
|  | 1525 | ; SSSE3-NEXT:    por %xmm0, %xmm1 | 
|  | 1526 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 1527 | ; SSSE3-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1528 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1529 | ; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm4 | 
|  | 1530 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
|  | 1531 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1532 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 1533 | ; SSSE3-NEXT:    pand %xmm6, %xmm0 | 
|  | 1534 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] | 
|  | 1535 | ; SSSE3-NEXT:    por %xmm0, %xmm6 | 
|  | 1536 | ; SSSE3-NEXT:    pand %xmm6, %xmm3 | 
|  | 1537 | ; SSSE3-NEXT:    pandn %xmm8, %xmm6 | 
|  | 1538 | ; SSSE3-NEXT:    por %xmm3, %xmm6 | 
|  | 1539 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm0 | 
|  | 1540 | ; SSSE3-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1541 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm3 | 
|  | 1542 | ; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm3 | 
|  | 1543 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 1544 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1545 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 1546 | ; SSSE3-NEXT:    pand %xmm4, %xmm0 | 
|  | 1547 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 1548 | ; SSSE3-NEXT:    por %xmm0, %xmm3 | 
|  | 1549 | ; SSSE3-NEXT:    pand %xmm3, %xmm2 | 
|  | 1550 | ; SSSE3-NEXT:    pandn %xmm8, %xmm3 | 
|  | 1551 | ; SSSE3-NEXT:    por %xmm2, %xmm3 | 
|  | 1552 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 1553 | ; SSSE3-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1554 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm2 | 
|  | 1555 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm2 | 
|  | 1556 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 1557 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1558 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 1559 | ; SSSE3-NEXT:    pand %xmm4, %xmm7 | 
|  | 1560 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 1561 | ; SSSE3-NEXT:    por %xmm7, %xmm0 | 
|  | 1562 | ; SSSE3-NEXT:    pand %xmm3, %xmm0 | 
|  | 1563 | ; SSSE3-NEXT:    movdqa %xmm6, %xmm2 | 
|  | 1564 | ; SSSE3-NEXT:    pxor %xmm10, %xmm2 | 
|  | 1565 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 1566 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 1567 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 1568 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm2 | 
|  | 1569 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] | 
|  | 1570 | ; SSSE3-NEXT:    pand %xmm4, %xmm7 | 
|  | 1571 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3] | 
|  | 1572 | ; SSSE3-NEXT:    por %xmm7, %xmm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1573 | ; SSSE3-NEXT:    pand %xmm6, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1574 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm3 | 
|  | 1575 | ; SSSE3-NEXT:    pxor %xmm10, %xmm3 | 
|  | 1576 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 1577 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm4 | 
|  | 1578 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
|  | 1579 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm3 | 
|  | 1580 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 1581 | ; SSSE3-NEXT:    pand %xmm6, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1582 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1583 | ; SSSE3-NEXT:    por %xmm3, %xmm4 | 
|  | 1584 | ; SSSE3-NEXT:    pand %xmm1, %xmm4 | 
|  | 1585 | ; SSSE3-NEXT:    movdqa %xmm5, %xmm1 | 
|  | 1586 | ; SSSE3-NEXT:    pxor %xmm10, %xmm1 | 
|  | 1587 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm3 | 
|  | 1588 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 1589 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] | 
|  | 1590 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm1 | 
|  | 1591 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] | 
|  | 1592 | ; SSSE3-NEXT:    pand %xmm6, %xmm1 | 
|  | 1593 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 1594 | ; SSSE3-NEXT:    por %xmm1, %xmm3 | 
|  | 1595 | ; SSSE3-NEXT:    pand %xmm5, %xmm3 | 
|  | 1596 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1597 | ; 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] | 1598 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3] | 
|  | 1599 | ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7] | 
|  | 1600 | ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] | 
|  | 1601 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3] | 
|  | 1602 | ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7] | 
|  | 1603 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1604 | ; 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] | 1605 | ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] | 
|  | 1606 | ; SSSE3-NEXT:    movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1607 | ; SSSE3-NEXT:    retq | 
|  | 1608 | ; | 
|  | 1609 | ; SSE41-LABEL: trunc_packus_v8i64_v8i8: | 
|  | 1610 | ; SSE41:       # %bb.0: | 
|  | 1611 | ; SSE41-NEXT:    movdqa %xmm0, %xmm8 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1612 | ; SSE41-NEXT:    movapd {{.*#+}} xmm7 = [255,255] | 
|  | 1613 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,0,2147483648,0] | 
|  | 1614 | ; SSE41-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 1615 | ; SSE41-NEXT:    pxor %xmm9, %xmm0 | 
|  | 1616 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [2147483903,2147483903] | 
|  | 1617 | ; SSE41-NEXT:    movdqa %xmm10, %xmm6 | 
|  | 1618 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm6 | 
|  | 1619 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm6[0,0,2,2] | 
|  | 1620 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1621 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3] | 
|  | 1622 | ; SSE41-NEXT:    pand %xmm5, %xmm4 | 
|  | 1623 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3] | 
|  | 1624 | ; SSE41-NEXT:    por %xmm4, %xmm0 | 
|  | 1625 | ; SSE41-NEXT:    movapd %xmm7, %xmm6 | 
|  | 1626 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm6 | 
|  | 1627 | ; SSE41-NEXT:    movdqa %xmm2, %xmm0 | 
|  | 1628 | ; SSE41-NEXT:    pxor %xmm9, %xmm0 | 
|  | 1629 | ; SSE41-NEXT:    movdqa %xmm10, %xmm3 | 
|  | 1630 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3 | 
|  | 1631 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 1632 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1633 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1634 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1635 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 1636 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 1637 | ; SSE41-NEXT:    movapd %xmm7, %xmm3 | 
|  | 1638 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3 | 
|  | 1639 | ; SSE41-NEXT:    movdqa %xmm1, %xmm0 | 
|  | 1640 | ; SSE41-NEXT:    pxor %xmm9, %xmm0 | 
|  | 1641 | ; SSE41-NEXT:    movdqa %xmm10, %xmm2 | 
|  | 1642 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2 | 
|  | 1643 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 1644 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1645 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 1646 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
|  | 1647 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 1648 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 1649 | ; SSE41-NEXT:    movapd %xmm7, %xmm2 | 
|  | 1650 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2 | 
|  | 1651 | ; SSE41-NEXT:    movdqa %xmm8, %xmm0 | 
|  | 1652 | ; SSE41-NEXT:    pxor %xmm9, %xmm0 | 
|  | 1653 | ; SSE41-NEXT:    movdqa %xmm10, %xmm1 | 
|  | 1654 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 1655 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] | 
|  | 1656 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1657 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 1658 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
|  | 1659 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] | 
|  | 1660 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 1661 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm7 | 
|  | 1662 | ; SSE41-NEXT:    movapd %xmm7, %xmm0 | 
|  | 1663 | ; SSE41-NEXT:    xorpd %xmm9, %xmm0 | 
|  | 1664 | ; SSE41-NEXT:    movapd %xmm0, %xmm1 | 
|  | 1665 | ; SSE41-NEXT:    pcmpgtd %xmm9, %xmm1 | 
|  | 1666 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] | 
|  | 1667 | ; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1668 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 1669 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
|  | 1670 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] | 
|  | 1671 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 1672 | ; SSE41-NEXT:    pxor %xmm8, %xmm8 | 
|  | 1673 | ; SSE41-NEXT:    pxor %xmm1, %xmm1 | 
|  | 1674 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm1 | 
|  | 1675 | ; SSE41-NEXT:    movapd %xmm2, %xmm0 | 
|  | 1676 | ; SSE41-NEXT:    xorpd %xmm9, %xmm0 | 
|  | 1677 | ; SSE41-NEXT:    movapd %xmm0, %xmm5 | 
|  | 1678 | ; SSE41-NEXT:    pcmpgtd %xmm9, %xmm5 | 
|  | 1679 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] | 
|  | 1680 | ; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1681 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3] | 
|  | 1682 | ; SSE41-NEXT:    pand %xmm7, %xmm4 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1683 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] | 
|  | 1684 | ; SSE41-NEXT:    por %xmm4, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1685 | ; SSE41-NEXT:    pxor %xmm7, %xmm7 | 
|  | 1686 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm7 | 
|  | 1687 | ; SSE41-NEXT:    movapd %xmm3, %xmm0 | 
|  | 1688 | ; SSE41-NEXT:    xorpd %xmm9, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1689 | ; SSE41-NEXT:    movapd %xmm0, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1690 | ; SSE41-NEXT:    pcmpgtd %xmm9, %xmm2 | 
|  | 1691 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 1692 | ; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1693 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 1694 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
|  | 1695 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 1696 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 1697 | ; SSE41-NEXT:    pxor %xmm2, %xmm2 | 
|  | 1698 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2 | 
|  | 1699 | ; SSE41-NEXT:    movapd %xmm6, %xmm0 | 
|  | 1700 | ; SSE41-NEXT:    xorpd %xmm9, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1701 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1702 | ; SSE41-NEXT:    pcmpgtd %xmm9, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1703 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1704 | ; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 1705 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 1706 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
|  | 1707 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 1708 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 1709 | ; SSE41-NEXT:    pxor %xmm3, %xmm3 | 
|  | 1710 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm3 | 
|  | 1711 | ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0],xmm8[1,2,3],xmm3[4],xmm8[5,6,7] | 
|  | 1712 | ; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm2[0],xmm8[1,2,3],xmm2[4],xmm8[5,6,7] | 
|  | 1713 | ; SSE41-NEXT:    packusdw %xmm3, %xmm2 | 
|  | 1714 | ; SSE41-NEXT:    pblendw {{.*#+}} xmm7 = xmm7[0],xmm8[1,2,3],xmm7[4],xmm8[5,6,7] | 
|  | 1715 | ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0],xmm8[1,2,3],xmm1[4],xmm8[5,6,7] | 
|  | 1716 | ; SSE41-NEXT:    packusdw %xmm7, %xmm1 | 
|  | 1717 | ; SSE41-NEXT:    packusdw %xmm2, %xmm1 | 
|  | 1718 | ; SSE41-NEXT:    movdqa %xmm1, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1719 | ; SSE41-NEXT:    retq | 
|  | 1720 | ; | 
|  | 1721 | ; AVX1-LABEL: trunc_packus_v8i64_v8i8: | 
|  | 1722 | ; AVX1:       # %bb.0: | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1723 | ; AVX1-NEXT:    vmovapd {{.*#+}} ymm2 = [255,255,255,255] | 
|  | 1724 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3 | 
|  | 1725 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [255,255] | 
|  | 1726 | ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3 | 
|  | 1727 | ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm5 | 
|  | 1728 | ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm5, %ymm3 | 
|  | 1729 | ; AVX1-NEXT:    vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 | 
|  | 1730 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3 | 
|  | 1731 | ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3 | 
|  | 1732 | ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm4, %xmm4 | 
|  | 1733 | ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3 | 
|  | 1734 | ; AVX1-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1735 | ; AVX1-NEXT:    vxorpd %xmm2, %xmm2, %xmm2 | 
|  | 1736 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm8 | 
|  | 1737 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm4 | 
|  | 1738 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm4, %xmm5 | 
|  | 1739 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm1, %xmm6 | 
|  | 1740 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm7 | 
|  | 1741 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm7, %xmm3 | 
|  | 1742 | ; AVX1-NEXT:    vpand %xmm7, %xmm3, %xmm3 | 
|  | 1743 | ; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1,2,3],xmm3[4],xmm2[5,6,7] | 
|  | 1744 | ; AVX1-NEXT:    vpand %xmm1, %xmm6, %xmm1 | 
|  | 1745 | ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3],xmm1[4],xmm2[5,6,7] | 
|  | 1746 | ; AVX1-NEXT:    vpackusdw %xmm3, %xmm1, %xmm1 | 
|  | 1747 | ; AVX1-NEXT:    vpand %xmm4, %xmm5, %xmm3 | 
|  | 1748 | ; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1,2,3],xmm3[4],xmm2[5,6,7] | 
|  | 1749 | ; AVX1-NEXT:    vpand %xmm0, %xmm8, %xmm0 | 
|  | 1750 | ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3],xmm0[4],xmm2[5,6,7] | 
|  | 1751 | ; AVX1-NEXT:    vpackusdw %xmm3, %xmm0, %xmm0 | 
|  | 1752 | ; AVX1-NEXT:    vpackusdw %xmm1, %xmm0, %xmm0 | 
|  | 1753 | ; AVX1-NEXT:    vzeroupper | 
|  | 1754 | ; AVX1-NEXT:    retq | 
|  | 1755 | ; | 
|  | 1756 | ; AVX2-SLOW-LABEL: trunc_packus_v8i64_v8i8: | 
|  | 1757 | ; AVX2-SLOW:       # %bb.0: | 
|  | 1758 | ; AVX2-SLOW-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255] | 
|  | 1759 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1760 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1761 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm1, %ymm2, %ymm3 | 
|  | 1762 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1763 | ; AVX2-SLOW-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1764 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm2, %ymm1, %ymm3 | 
|  | 1765 | ; AVX2-SLOW-NEXT:    vpand %ymm1, %ymm3, %ymm1 | 
|  | 1766 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm2, %ymm0, %ymm2 | 
|  | 1767 | ; AVX2-SLOW-NEXT:    vpand %ymm0, %ymm2, %ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1768 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] | 
|  | 1769 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
|  | 1770 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] | 
|  | 1771 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3] | 
|  | 1772 | ; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 1773 | ; 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] | 
|  | 1774 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 1775 | ; AVX2-SLOW-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1776 | ; AVX2-SLOW-NEXT:    vzeroupper | 
|  | 1777 | ; AVX2-SLOW-NEXT:    retq | 
|  | 1778 | ; | 
|  | 1779 | ; AVX2-FAST-LABEL: trunc_packus_v8i64_v8i8: | 
|  | 1780 | ; AVX2-FAST:       # %bb.0: | 
|  | 1781 | ; AVX2-FAST-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255] | 
|  | 1782 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1783 | ; AVX2-FAST-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1784 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm1, %ymm2, %ymm3 | 
|  | 1785 | ; AVX2-FAST-NEXT:    vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1786 | ; AVX2-FAST-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 1787 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm2, %ymm1, %ymm3 | 
|  | 1788 | ; AVX2-FAST-NEXT:    vpand %ymm1, %ymm3, %ymm1 | 
|  | 1789 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm2, %ymm0, %ymm2 | 
|  | 1790 | ; AVX2-FAST-NEXT:    vpand %ymm0, %ymm2, %ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1791 | ; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,2,4,6,4,6,6,7] | 
|  | 1792 | ; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm2, %ymm0 | 
|  | 1793 | ; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm2, %ymm1 | 
|  | 1794 | ; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 1795 | ; 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] | 
|  | 1796 | ; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 1797 | ; AVX2-FAST-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 1798 | ; AVX2-FAST-NEXT:    vzeroupper | 
|  | 1799 | ; AVX2-FAST-NEXT:    retq | 
|  | 1800 | ; | 
|  | 1801 | ; AVX512-LABEL: trunc_packus_v8i64_v8i8: | 
|  | 1802 | ; AVX512:       # %bb.0: | 
|  | 1803 | ; AVX512-NEXT:    vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0 | 
|  | 1804 | ; AVX512-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 1805 | ; AVX512-NEXT:    vpmaxsq %zmm1, %zmm0, %zmm0 | 
|  | 1806 | ; AVX512-NEXT:    vpmovqw %zmm0, %xmm0 | 
|  | 1807 | ; AVX512-NEXT:    vzeroupper | 
|  | 1808 | ; AVX512-NEXT:    retq | 
|  | 1809 | %1 = icmp slt <8 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> | 
|  | 1810 | %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> | 
|  | 1811 | %3 = icmp sgt <8 x i64> %2, zeroinitializer | 
|  | 1812 | %4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer | 
|  | 1813 | %5 = trunc <8 x i64> %4 to <8 x i8> | 
|  | 1814 | ret <8 x i8> %5 | 
|  | 1815 | } | 
|  | 1816 |  | 
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 1817 | define void @trunc_packus_v8i64_v8i8_store(<8 x i64> %a0, <8 x i8> *%p1) { | 
|  | 1818 | ; SSE2-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 1819 | ; SSE2:       # %bb.0: | 
|  | 1820 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [255,255] | 
|  | 1821 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] | 
|  | 1822 | ; SSE2-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 1823 | ; SSE2-NEXT:    pxor %xmm10, %xmm4 | 
|  | 1824 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [2147483903,2147483903] | 
|  | 1825 | ; SSE2-NEXT:    movdqa %xmm9, %xmm6 | 
|  | 1826 | ; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6 | 
|  | 1827 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] | 
|  | 1828 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4 | 
|  | 1829 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
|  | 1830 | ; SSE2-NEXT:    pand %xmm7, %xmm4 | 
|  | 1831 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm6[1,1,3,3] | 
|  | 1832 | ; SSE2-NEXT:    por %xmm4, %xmm11 | 
|  | 1833 | ; SSE2-NEXT:    pand %xmm11, %xmm3 | 
|  | 1834 | ; SSE2-NEXT:    pandn %xmm8, %xmm11 | 
|  | 1835 | ; SSE2-NEXT:    por %xmm3, %xmm11 | 
|  | 1836 | ; SSE2-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 1837 | ; SSE2-NEXT:    pxor %xmm10, %xmm3 | 
|  | 1838 | ; SSE2-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1839 | ; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4 | 
|  | 1840 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2] | 
|  | 1841 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm3 | 
|  | 1842 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,1,3,3] | 
|  | 1843 | ; SSE2-NEXT:    pand %xmm7, %xmm5 | 
|  | 1844 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3] | 
|  | 1845 | ; SSE2-NEXT:    por %xmm5, %xmm3 | 
|  | 1846 | ; SSE2-NEXT:    pand %xmm3, %xmm2 | 
|  | 1847 | ; SSE2-NEXT:    pandn %xmm8, %xmm3 | 
|  | 1848 | ; SSE2-NEXT:    por %xmm2, %xmm3 | 
|  | 1849 | ; SSE2-NEXT:    movdqa %xmm1, %xmm2 | 
|  | 1850 | ; SSE2-NEXT:    pxor %xmm10, %xmm2 | 
|  | 1851 | ; SSE2-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1852 | ; SSE2-NEXT:    pcmpgtd %xmm2, %xmm4 | 
|  | 1853 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 1854 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm2 | 
|  | 1855 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] | 
|  | 1856 | ; SSE2-NEXT:    pand %xmm5, %xmm7 | 
|  | 1857 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3] | 
|  | 1858 | ; SSE2-NEXT:    por %xmm7, %xmm2 | 
|  | 1859 | ; SSE2-NEXT:    pand %xmm2, %xmm1 | 
|  | 1860 | ; SSE2-NEXT:    pandn %xmm8, %xmm2 | 
|  | 1861 | ; SSE2-NEXT:    por %xmm1, %xmm2 | 
|  | 1862 | ; SSE2-NEXT:    movdqa %xmm0, %xmm1 | 
|  | 1863 | ; SSE2-NEXT:    pxor %xmm10, %xmm1 | 
|  | 1864 | ; SSE2-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1865 | ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4 | 
|  | 1866 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 1867 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1 | 
|  | 1868 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] | 
|  | 1869 | ; SSE2-NEXT:    pand %xmm5, %xmm7 | 
|  | 1870 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] | 
|  | 1871 | ; SSE2-NEXT:    por %xmm7, %xmm1 | 
|  | 1872 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 1873 | ; SSE2-NEXT:    pandn %xmm8, %xmm1 | 
|  | 1874 | ; SSE2-NEXT:    por %xmm0, %xmm1 | 
|  | 1875 | ; SSE2-NEXT:    movdqa %xmm1, %xmm0 | 
|  | 1876 | ; SSE2-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1877 | ; SSE2-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 1878 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm4 | 
|  | 1879 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 1880 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1881 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 1882 | ; SSE2-NEXT:    pand %xmm5, %xmm7 | 
|  | 1883 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] | 
|  | 1884 | ; SSE2-NEXT:    por %xmm7, %xmm0 | 
|  | 1885 | ; SSE2-NEXT:    movdqa %xmm2, %xmm4 | 
|  | 1886 | ; SSE2-NEXT:    pxor %xmm10, %xmm4 | 
|  | 1887 | ; SSE2-NEXT:    movdqa %xmm4, %xmm5 | 
|  | 1888 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm5 | 
|  | 1889 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] | 
|  | 1890 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm4 | 
|  | 1891 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
|  | 1892 | ; SSE2-NEXT:    pand %xmm7, %xmm4 | 
|  | 1893 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[1,1,3,3] | 
|  | 1894 | ; SSE2-NEXT:    por %xmm4, %xmm7 | 
|  | 1895 | ; SSE2-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 1896 | ; SSE2-NEXT:    pxor %xmm10, %xmm4 | 
|  | 1897 | ; SSE2-NEXT:    movdqa %xmm4, %xmm5 | 
|  | 1898 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm5 | 
|  | 1899 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm5[0,0,2,2] | 
|  | 1900 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm4 | 
|  | 1901 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] | 
|  | 1902 | ; SSE2-NEXT:    pand %xmm9, %xmm6 | 
|  | 1903 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] | 
|  | 1904 | ; SSE2-NEXT:    por %xmm6, %xmm4 | 
|  | 1905 | ; SSE2-NEXT:    movdqa %xmm11, %xmm5 | 
|  | 1906 | ; SSE2-NEXT:    pxor %xmm10, %xmm5 | 
|  | 1907 | ; SSE2-NEXT:    movdqa %xmm5, %xmm6 | 
|  | 1908 | ; SSE2-NEXT:    pcmpgtd %xmm10, %xmm6 | 
|  | 1909 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm6[0,0,2,2] | 
|  | 1910 | ; SSE2-NEXT:    pcmpeqd %xmm10, %xmm5 | 
|  | 1911 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 1912 | ; SSE2-NEXT:    pand %xmm9, %xmm5 | 
|  | 1913 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] | 
|  | 1914 | ; SSE2-NEXT:    por %xmm5, %xmm6 | 
|  | 1915 | ; SSE2-NEXT:    pand %xmm8, %xmm6 | 
|  | 1916 | ; SSE2-NEXT:    pand %xmm11, %xmm6 | 
|  | 1917 | ; SSE2-NEXT:    pand %xmm8, %xmm4 | 
|  | 1918 | ; SSE2-NEXT:    pand %xmm3, %xmm4 | 
|  | 1919 | ; SSE2-NEXT:    packuswb %xmm6, %xmm4 | 
|  | 1920 | ; SSE2-NEXT:    pand %xmm8, %xmm7 | 
|  | 1921 | ; SSE2-NEXT:    pand %xmm2, %xmm7 | 
|  | 1922 | ; SSE2-NEXT:    pand %xmm8, %xmm0 | 
|  | 1923 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 1924 | ; SSE2-NEXT:    packuswb %xmm7, %xmm0 | 
|  | 1925 | ; SSE2-NEXT:    packuswb %xmm4, %xmm0 | 
|  | 1926 | ; SSE2-NEXT:    packuswb %xmm0, %xmm0 | 
|  | 1927 | ; SSE2-NEXT:    movq %xmm0, (%rdi) | 
|  | 1928 | ; SSE2-NEXT:    retq | 
|  | 1929 | ; | 
|  | 1930 | ; SSSE3-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 1931 | ; SSSE3:       # %bb.0: | 
|  | 1932 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm8 = [255,255] | 
|  | 1933 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] | 
|  | 1934 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 1935 | ; SSSE3-NEXT:    pxor %xmm10, %xmm4 | 
|  | 1936 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm9 = [2147483903,2147483903] | 
|  | 1937 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm6 | 
|  | 1938 | ; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6 | 
|  | 1939 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] | 
|  | 1940 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4 | 
|  | 1941 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
|  | 1942 | ; SSSE3-NEXT:    pand %xmm7, %xmm4 | 
|  | 1943 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm6[1,1,3,3] | 
|  | 1944 | ; SSSE3-NEXT:    por %xmm4, %xmm11 | 
|  | 1945 | ; SSSE3-NEXT:    pand %xmm11, %xmm3 | 
|  | 1946 | ; SSSE3-NEXT:    pandn %xmm8, %xmm11 | 
|  | 1947 | ; SSSE3-NEXT:    por %xmm3, %xmm11 | 
|  | 1948 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 1949 | ; SSSE3-NEXT:    pxor %xmm10, %xmm3 | 
|  | 1950 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1951 | ; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4 | 
|  | 1952 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2] | 
|  | 1953 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm3 | 
|  | 1954 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,1,3,3] | 
|  | 1955 | ; SSSE3-NEXT:    pand %xmm7, %xmm5 | 
|  | 1956 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3] | 
|  | 1957 | ; SSSE3-NEXT:    por %xmm5, %xmm3 | 
|  | 1958 | ; SSSE3-NEXT:    pand %xmm3, %xmm2 | 
|  | 1959 | ; SSSE3-NEXT:    pandn %xmm8, %xmm3 | 
|  | 1960 | ; SSSE3-NEXT:    por %xmm2, %xmm3 | 
|  | 1961 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm2 | 
|  | 1962 | ; SSSE3-NEXT:    pxor %xmm10, %xmm2 | 
|  | 1963 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1964 | ; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm4 | 
|  | 1965 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 1966 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm2 | 
|  | 1967 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3] | 
|  | 1968 | ; SSSE3-NEXT:    pand %xmm5, %xmm7 | 
|  | 1969 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3] | 
|  | 1970 | ; SSSE3-NEXT:    por %xmm7, %xmm2 | 
|  | 1971 | ; SSSE3-NEXT:    pand %xmm2, %xmm1 | 
|  | 1972 | ; SSSE3-NEXT:    pandn %xmm8, %xmm2 | 
|  | 1973 | ; SSSE3-NEXT:    por %xmm1, %xmm2 | 
|  | 1974 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm1 | 
|  | 1975 | ; SSSE3-NEXT:    pxor %xmm10, %xmm1 | 
|  | 1976 | ; SSSE3-NEXT:    movdqa %xmm9, %xmm4 | 
|  | 1977 | ; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4 | 
|  | 1978 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 1979 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1 | 
|  | 1980 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3] | 
|  | 1981 | ; SSSE3-NEXT:    pand %xmm5, %xmm7 | 
|  | 1982 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3] | 
|  | 1983 | ; SSSE3-NEXT:    por %xmm7, %xmm1 | 
|  | 1984 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 1985 | ; SSSE3-NEXT:    pandn %xmm8, %xmm1 | 
|  | 1986 | ; SSSE3-NEXT:    por %xmm0, %xmm1 | 
|  | 1987 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm0 | 
|  | 1988 | ; SSSE3-NEXT:    pxor %xmm10, %xmm0 | 
|  | 1989 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 1990 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm4 | 
|  | 1991 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 1992 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 1993 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 1994 | ; SSSE3-NEXT:    pand %xmm5, %xmm7 | 
|  | 1995 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] | 
|  | 1996 | ; SSSE3-NEXT:    por %xmm7, %xmm0 | 
|  | 1997 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm4 | 
|  | 1998 | ; SSSE3-NEXT:    pxor %xmm10, %xmm4 | 
|  | 1999 | ; SSSE3-NEXT:    movdqa %xmm4, %xmm5 | 
|  | 2000 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm5 | 
|  | 2001 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] | 
|  | 2002 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm4 | 
|  | 2003 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
|  | 2004 | ; SSSE3-NEXT:    pand %xmm7, %xmm4 | 
|  | 2005 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[1,1,3,3] | 
|  | 2006 | ; SSSE3-NEXT:    por %xmm4, %xmm7 | 
|  | 2007 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 2008 | ; SSSE3-NEXT:    pxor %xmm10, %xmm4 | 
|  | 2009 | ; SSSE3-NEXT:    movdqa %xmm4, %xmm5 | 
|  | 2010 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm5 | 
|  | 2011 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm5[0,0,2,2] | 
|  | 2012 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm4 | 
|  | 2013 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] | 
|  | 2014 | ; SSSE3-NEXT:    pand %xmm9, %xmm6 | 
|  | 2015 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3] | 
|  | 2016 | ; SSSE3-NEXT:    por %xmm6, %xmm4 | 
|  | 2017 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm5 | 
|  | 2018 | ; SSSE3-NEXT:    pxor %xmm10, %xmm5 | 
|  | 2019 | ; SSSE3-NEXT:    movdqa %xmm5, %xmm6 | 
|  | 2020 | ; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm6 | 
|  | 2021 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm6[0,0,2,2] | 
|  | 2022 | ; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm5 | 
|  | 2023 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2024 | ; SSSE3-NEXT:    pand %xmm9, %xmm5 | 
|  | 2025 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] | 
|  | 2026 | ; SSSE3-NEXT:    por %xmm5, %xmm6 | 
|  | 2027 | ; SSSE3-NEXT:    pand %xmm8, %xmm6 | 
|  | 2028 | ; SSSE3-NEXT:    pand %xmm11, %xmm6 | 
|  | 2029 | ; SSSE3-NEXT:    pand %xmm8, %xmm4 | 
|  | 2030 | ; SSSE3-NEXT:    pand %xmm3, %xmm4 | 
|  | 2031 | ; SSSE3-NEXT:    packuswb %xmm6, %xmm4 | 
|  | 2032 | ; SSSE3-NEXT:    pand %xmm8, %xmm7 | 
|  | 2033 | ; SSSE3-NEXT:    pand %xmm2, %xmm7 | 
|  | 2034 | ; SSSE3-NEXT:    pand %xmm8, %xmm0 | 
|  | 2035 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 2036 | ; SSSE3-NEXT:    packuswb %xmm7, %xmm0 | 
|  | 2037 | ; SSSE3-NEXT:    packuswb %xmm4, %xmm0 | 
|  | 2038 | ; SSSE3-NEXT:    packuswb %xmm0, %xmm0 | 
|  | 2039 | ; SSSE3-NEXT:    movq %xmm0, (%rdi) | 
|  | 2040 | ; SSSE3-NEXT:    retq | 
|  | 2041 | ; | 
|  | 2042 | ; SSE41-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 2043 | ; SSE41:       # %bb.0: | 
|  | 2044 | ; SSE41-NEXT:    movdqa %xmm0, %xmm9 | 
|  | 2045 | ; SSE41-NEXT:    movapd {{.*#+}} xmm8 = [255,255] | 
|  | 2046 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] | 
|  | 2047 | ; SSE41-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 2048 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2049 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [2147483903,2147483903] | 
|  | 2050 | ; SSE41-NEXT:    movdqa %xmm5, %xmm4 | 
|  | 2051 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4 | 
|  | 2052 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2] | 
|  | 2053 | ; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0 | 
|  | 2054 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] | 
|  | 2055 | ; SSE41-NEXT:    pand %xmm7, %xmm6 | 
|  | 2056 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] | 
|  | 2057 | ; SSE41-NEXT:    por %xmm6, %xmm0 | 
|  | 2058 | ; SSE41-NEXT:    movapd %xmm8, %xmm11 | 
|  | 2059 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm11 | 
|  | 2060 | ; SSE41-NEXT:    movdqa %xmm2, %xmm0 | 
|  | 2061 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2062 | ; SSE41-NEXT:    movdqa %xmm5, %xmm3 | 
|  | 2063 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3 | 
|  | 2064 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 2065 | ; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0 | 
|  | 2066 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] | 
|  | 2067 | ; SSE41-NEXT:    pand %xmm4, %xmm6 | 
|  | 2068 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2069 | ; SSE41-NEXT:    por %xmm6, %xmm0 | 
|  | 2070 | ; SSE41-NEXT:    movapd %xmm8, %xmm3 | 
|  | 2071 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3 | 
|  | 2072 | ; SSE41-NEXT:    movdqa %xmm1, %xmm0 | 
|  | 2073 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2074 | ; SSE41-NEXT:    movdqa %xmm5, %xmm2 | 
|  | 2075 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2 | 
|  | 2076 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 2077 | ; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0 | 
|  | 2078 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] | 
|  | 2079 | ; SSE41-NEXT:    pand %xmm4, %xmm6 | 
|  | 2080 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 2081 | ; SSE41-NEXT:    por %xmm6, %xmm0 | 
|  | 2082 | ; SSE41-NEXT:    movapd %xmm8, %xmm4 | 
|  | 2083 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm4 | 
|  | 2084 | ; SSE41-NEXT:    movdqa %xmm9, %xmm0 | 
|  | 2085 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2086 | ; SSE41-NEXT:    movdqa %xmm5, %xmm1 | 
|  | 2087 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 2088 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] | 
|  | 2089 | ; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0 | 
|  | 2090 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 2091 | ; SSE41-NEXT:    pand %xmm2, %xmm5 | 
|  | 2092 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] | 
|  | 2093 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 2094 | ; SSE41-NEXT:    movapd %xmm8, %xmm5 | 
|  | 2095 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm9, %xmm5 | 
|  | 2096 | ; SSE41-NEXT:    movapd %xmm5, %xmm0 | 
|  | 2097 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2098 | ; SSE41-NEXT:    movapd %xmm0, %xmm1 | 
|  | 2099 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm1 | 
|  | 2100 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] | 
|  | 2101 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2102 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] | 
|  | 2103 | ; SSE41-NEXT:    pand %xmm2, %xmm6 | 
|  | 2104 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] | 
|  | 2105 | ; SSE41-NEXT:    por %xmm6, %xmm0 | 
|  | 2106 | ; SSE41-NEXT:    pxor %xmm2, %xmm2 | 
|  | 2107 | ; SSE41-NEXT:    pxor %xmm1, %xmm1 | 
|  | 2108 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm1 | 
|  | 2109 | ; SSE41-NEXT:    movapd %xmm4, %xmm0 | 
|  | 2110 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2111 | ; SSE41-NEXT:    movapd %xmm0, %xmm5 | 
|  | 2112 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm5 | 
|  | 2113 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] | 
|  | 2114 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2115 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2116 | ; SSE41-NEXT:    pand %xmm6, %xmm7 | 
|  | 2117 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] | 
|  | 2118 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2119 | ; SSE41-NEXT:    pxor %xmm5, %xmm5 | 
|  | 2120 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm5 | 
|  | 2121 | ; SSE41-NEXT:    movapd %xmm3, %xmm0 | 
|  | 2122 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2123 | ; SSE41-NEXT:    movapd %xmm0, %xmm4 | 
|  | 2124 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm4 | 
|  | 2125 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
|  | 2126 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2127 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2128 | ; SSE41-NEXT:    pand %xmm6, %xmm7 | 
|  | 2129 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] | 
|  | 2130 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2131 | ; SSE41-NEXT:    pxor %xmm4, %xmm4 | 
|  | 2132 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm4 | 
|  | 2133 | ; SSE41-NEXT:    movapd %xmm11, %xmm0 | 
|  | 2134 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2135 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
|  | 2136 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 2137 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] | 
|  | 2138 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2139 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2140 | ; SSE41-NEXT:    pand %xmm6, %xmm7 | 
|  | 2141 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2142 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2143 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm2 | 
|  | 2144 | ; SSE41-NEXT:    andpd %xmm8, %xmm2 | 
|  | 2145 | ; SSE41-NEXT:    andpd %xmm8, %xmm4 | 
|  | 2146 | ; SSE41-NEXT:    packuswb %xmm2, %xmm4 | 
|  | 2147 | ; SSE41-NEXT:    andpd %xmm8, %xmm5 | 
|  | 2148 | ; SSE41-NEXT:    andpd %xmm8, %xmm1 | 
|  | 2149 | ; SSE41-NEXT:    packuswb %xmm5, %xmm1 | 
|  | 2150 | ; SSE41-NEXT:    packuswb %xmm4, %xmm1 | 
|  | 2151 | ; SSE41-NEXT:    packuswb %xmm1, %xmm1 | 
|  | 2152 | ; SSE41-NEXT:    movq %xmm1, (%rdi) | 
|  | 2153 | ; SSE41-NEXT:    retq | 
|  | 2154 | ; | 
|  | 2155 | ; AVX1-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 2156 | ; AVX1:       # %bb.0: | 
|  | 2157 | ; AVX1-NEXT:    vmovapd {{.*#+}} ymm2 = [255,255,255,255] | 
|  | 2158 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3 | 
|  | 2159 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [255,255] | 
|  | 2160 | ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3 | 
|  | 2161 | ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm5 | 
|  | 2162 | ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm5, %ymm3 | 
|  | 2163 | ; AVX1-NEXT:    vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 | 
|  | 2164 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3 | 
|  | 2165 | ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3 | 
|  | 2166 | ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm4, %xmm5 | 
|  | 2167 | ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm5, %ymm3 | 
|  | 2168 | ; AVX1-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 | 
|  | 2169 | ; AVX1-NEXT:    vxorpd %xmm2, %xmm2, %xmm2 | 
|  | 2170 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm8 | 
|  | 2171 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5 | 
|  | 2172 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm5, %xmm6 | 
|  | 2173 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm1, %xmm7 | 
|  | 2174 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3 | 
|  | 2175 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2 | 
|  | 2176 | ; AVX1-NEXT:    vpand %xmm4, %xmm3, %xmm3 | 
|  | 2177 | ; AVX1-NEXT:    vpand %xmm3, %xmm2, %xmm2 | 
|  | 2178 | ; AVX1-NEXT:    vpand %xmm4, %xmm1, %xmm1 | 
|  | 2179 | ; AVX1-NEXT:    vpand %xmm1, %xmm7, %xmm1 | 
|  | 2180 | ; AVX1-NEXT:    vpackuswb %xmm2, %xmm1, %xmm1 | 
|  | 2181 | ; AVX1-NEXT:    vpand %xmm4, %xmm5, %xmm2 | 
|  | 2182 | ; AVX1-NEXT:    vpand %xmm2, %xmm6, %xmm2 | 
|  | 2183 | ; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0 | 
|  | 2184 | ; AVX1-NEXT:    vpand %xmm0, %xmm8, %xmm0 | 
|  | 2185 | ; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0 | 
|  | 2186 | ; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 | 
|  | 2187 | ; AVX1-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0 | 
|  | 2188 | ; AVX1-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 2189 | ; AVX1-NEXT:    vzeroupper | 
|  | 2190 | ; AVX1-NEXT:    retq | 
|  | 2191 | ; | 
|  | 2192 | ; AVX2-SLOW-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 2193 | ; AVX2-SLOW:       # %bb.0: | 
|  | 2194 | ; AVX2-SLOW-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255] | 
|  | 2195 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm1, %ymm2, %ymm3 | 
|  | 2196 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 | 
|  | 2197 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm3 | 
|  | 2198 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 | 
|  | 2199 | ; AVX2-SLOW-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 2200 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm2, %ymm0, %ymm3 | 
|  | 2201 | ; AVX2-SLOW-NEXT:    vpand %ymm0, %ymm3, %ymm0 | 
|  | 2202 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm2, %ymm1, %ymm2 | 
|  | 2203 | ; AVX2-SLOW-NEXT:    vpand %ymm1, %ymm2, %ymm1 | 
|  | 2204 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] | 
|  | 2205 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3] | 
|  | 2206 | ; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 2207 | ; AVX2-SLOW-NEXT:    vpshufb %xmm2, %xmm1, %xmm1 | 
|  | 2208 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] | 
|  | 2209 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
|  | 2210 | ; AVX2-SLOW-NEXT:    vpshufb %xmm2, %xmm0, %xmm0 | 
|  | 2211 | ; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] | 
|  | 2212 | ; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 2213 | ; AVX2-SLOW-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 2214 | ; AVX2-SLOW-NEXT:    vzeroupper | 
|  | 2215 | ; AVX2-SLOW-NEXT:    retq | 
|  | 2216 | ; | 
|  | 2217 | ; AVX2-FAST-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 2218 | ; AVX2-FAST:       # %bb.0: | 
|  | 2219 | ; AVX2-FAST-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255] | 
|  | 2220 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm1, %ymm2, %ymm3 | 
|  | 2221 | ; AVX2-FAST-NEXT:    vblendvpd %ymm3, %ymm1, %ymm2, %ymm1 | 
|  | 2222 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm3 | 
|  | 2223 | ; AVX2-FAST-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0 | 
|  | 2224 | ; AVX2-FAST-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 2225 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm2, %ymm0, %ymm3 | 
|  | 2226 | ; AVX2-FAST-NEXT:    vpand %ymm0, %ymm3, %ymm0 | 
|  | 2227 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm2, %ymm1, %ymm2 | 
|  | 2228 | ; AVX2-FAST-NEXT:    vpand %ymm1, %ymm2, %ymm1 | 
|  | 2229 | ; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,2,4,6,4,6,6,7] | 
|  | 2230 | ; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm2, %ymm1 | 
|  | 2231 | ; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 2232 | ; AVX2-FAST-NEXT:    vpshufb %xmm3, %xmm1, %xmm1 | 
|  | 2233 | ; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm2, %ymm0 | 
|  | 2234 | ; AVX2-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm0 | 
|  | 2235 | ; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] | 
|  | 2236 | ; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 2237 | ; AVX2-FAST-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 2238 | ; AVX2-FAST-NEXT:    vzeroupper | 
|  | 2239 | ; AVX2-FAST-NEXT:    retq | 
|  | 2240 | ; | 
|  | 2241 | ; AVX512-LABEL: trunc_packus_v8i64_v8i8_store: | 
|  | 2242 | ; AVX512:       # %bb.0: | 
|  | 2243 | ; AVX512-NEXT:    vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0 | 
|  | 2244 | ; AVX512-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 2245 | ; AVX512-NEXT:    vpmaxsq %zmm1, %zmm0, %zmm0 | 
|  | 2246 | ; AVX512-NEXT:    vpmovqb %zmm0, (%rdi) | 
|  | 2247 | ; AVX512-NEXT:    vzeroupper | 
|  | 2248 | ; AVX512-NEXT:    retq | 
|  | 2249 | %1 = icmp slt <8 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> | 
|  | 2250 | %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> | 
|  | 2251 | %3 = icmp sgt <8 x i64> %2, zeroinitializer | 
|  | 2252 | %4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer | 
|  | 2253 | %5 = trunc <8 x i64> %4 to <8 x i8> | 
|  | 2254 | store <8 x i8> %5, <8 x i8> *%p1 | 
|  | 2255 | ret void | 
|  | 2256 | } | 
|  | 2257 |  | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2258 | define <16 x i8> @trunc_packus_v16i64_v16i8(<16 x i64> %a0) { | 
|  | 2259 | ; SSE2-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 2260 | ; SSE2:       # %bb.0: | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2261 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm13 = [255,255] | 
|  | 2262 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,0,2147483648,0] | 
|  | 2263 | ; SSE2-NEXT:    movdqa %xmm7, %xmm8 | 
|  | 2264 | ; SSE2-NEXT:    pxor %xmm9, %xmm8 | 
|  | 2265 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm11 = [2147483903,2147483903] | 
|  | 2266 | ; SSE2-NEXT:    movdqa %xmm11, %xmm10 | 
|  | 2267 | ; SSE2-NEXT:    pcmpgtd %xmm8, %xmm10 | 
|  | 2268 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm10[0,0,2,2] | 
|  | 2269 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm8 | 
|  | 2270 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3] | 
|  | 2271 | ; SSE2-NEXT:    pand %xmm12, %xmm8 | 
|  | 2272 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3] | 
|  | 2273 | ; SSE2-NEXT:    por %xmm8, %xmm10 | 
|  | 2274 | ; SSE2-NEXT:    pand %xmm10, %xmm7 | 
|  | 2275 | ; SSE2-NEXT:    pandn %xmm13, %xmm10 | 
|  | 2276 | ; SSE2-NEXT:    por %xmm7, %xmm10 | 
|  | 2277 | ; SSE2-NEXT:    movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
|  | 2278 | ; SSE2-NEXT:    movdqa %xmm6, %xmm7 | 
|  | 2279 | ; SSE2-NEXT:    pxor %xmm9, %xmm7 | 
|  | 2280 | ; SSE2-NEXT:    movdqa %xmm11, %xmm8 | 
|  | 2281 | ; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8 | 
|  | 2282 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm8[0,0,2,2] | 
|  | 2283 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm7 | 
|  | 2284 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] | 
|  | 2285 | ; SSE2-NEXT:    pand %xmm12, %xmm10 | 
|  | 2286 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm14 = xmm8[1,1,3,3] | 
|  | 2287 | ; SSE2-NEXT:    por %xmm10, %xmm14 | 
|  | 2288 | ; SSE2-NEXT:    pand %xmm14, %xmm6 | 
|  | 2289 | ; SSE2-NEXT:    pandn %xmm13, %xmm14 | 
|  | 2290 | ; SSE2-NEXT:    por %xmm6, %xmm14 | 
|  | 2291 | ; SSE2-NEXT:    movdqa %xmm5, %xmm6 | 
|  | 2292 | ; SSE2-NEXT:    pxor %xmm9, %xmm6 | 
|  | 2293 | ; SSE2-NEXT:    movdqa %xmm11, %xmm7 | 
|  | 2294 | ; SSE2-NEXT:    pcmpgtd %xmm6, %xmm7 | 
|  | 2295 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2] | 
|  | 2296 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm6 | 
|  | 2297 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] | 
|  | 2298 | ; SSE2-NEXT:    pand %xmm8, %xmm6 | 
|  | 2299 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3] | 
|  | 2300 | ; SSE2-NEXT:    por %xmm6, %xmm7 | 
|  | 2301 | ; SSE2-NEXT:    pand %xmm7, %xmm5 | 
|  | 2302 | ; SSE2-NEXT:    pandn %xmm13, %xmm7 | 
|  | 2303 | ; SSE2-NEXT:    por %xmm5, %xmm7 | 
|  | 2304 | ; SSE2-NEXT:    movdqa %xmm7, %xmm10 | 
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2305 | ; SSE2-NEXT:    movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2306 | ; SSE2-NEXT:    movdqa %xmm4, %xmm5 | 
|  | 2307 | ; SSE2-NEXT:    pxor %xmm9, %xmm5 | 
|  | 2308 | ; SSE2-NEXT:    movdqa %xmm11, %xmm6 | 
|  | 2309 | ; SSE2-NEXT:    pcmpgtd %xmm5, %xmm6 | 
|  | 2310 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] | 
|  | 2311 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm5 | 
|  | 2312 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2313 | ; SSE2-NEXT:    pand %xmm7, %xmm5 | 
|  | 2314 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] | 
|  | 2315 | ; SSE2-NEXT:    por %xmm5, %xmm6 | 
|  | 2316 | ; SSE2-NEXT:    pand %xmm6, %xmm4 | 
|  | 2317 | ; SSE2-NEXT:    pandn %xmm13, %xmm6 | 
|  | 2318 | ; SSE2-NEXT:    por %xmm4, %xmm6 | 
|  | 2319 | ; SSE2-NEXT:    movdqa %xmm6, %xmm7 | 
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2320 | ; SSE2-NEXT:    movdqa %xmm7, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2321 | ; SSE2-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 2322 | ; SSE2-NEXT:    pxor %xmm9, %xmm4 | 
|  | 2323 | ; SSE2-NEXT:    movdqa %xmm11, %xmm5 | 
|  | 2324 | ; SSE2-NEXT:    pcmpgtd %xmm4, %xmm5 | 
|  | 2325 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] | 
|  | 2326 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm4 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2327 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2328 | ; SSE2-NEXT:    pand %xmm6, %xmm4 | 
|  | 2329 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2330 | ; SSE2-NEXT:    por %xmm4, %xmm5 | 
|  | 2331 | ; SSE2-NEXT:    pand %xmm5, %xmm3 | 
|  | 2332 | ; SSE2-NEXT:    pandn %xmm13, %xmm5 | 
|  | 2333 | ; SSE2-NEXT:    por %xmm3, %xmm5 | 
|  | 2334 | ; SSE2-NEXT:    movdqa %xmm5, %xmm8 | 
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2335 | ; SSE2-NEXT:    movdqa %xmm8, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2336 | ; SSE2-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 2337 | ; SSE2-NEXT:    pxor %xmm9, %xmm3 | 
|  | 2338 | ; SSE2-NEXT:    movdqa %xmm11, %xmm4 | 
|  | 2339 | ; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4 | 
|  | 2340 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2341 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm3 | 
|  | 2342 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 2343 | ; SSE2-NEXT:    pand %xmm5, %xmm3 | 
|  | 2344 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3] | 
|  | 2345 | ; SSE2-NEXT:    por %xmm3, %xmm15 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2346 | ; SSE2-NEXT:    pand %xmm15, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2347 | ; SSE2-NEXT:    pandn %xmm13, %xmm15 | 
|  | 2348 | ; SSE2-NEXT:    movdqa %xmm13, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2349 | ; SSE2-NEXT:    por %xmm2, %xmm15 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2350 | ; SSE2-NEXT:    movdqa %xmm1, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2351 | ; SSE2-NEXT:    pxor %xmm9, %xmm2 | 
|  | 2352 | ; SSE2-NEXT:    movdqa %xmm11, %xmm3 | 
|  | 2353 | ; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3 | 
|  | 2354 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 2355 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2356 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2357 | ; SSE2-NEXT:    pand %xmm4, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2358 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3] | 
|  | 2359 | ; SSE2-NEXT:    por %xmm2, %xmm13 | 
|  | 2360 | ; SSE2-NEXT:    pand %xmm13, %xmm1 | 
|  | 2361 | ; SSE2-NEXT:    pandn %xmm5, %xmm13 | 
|  | 2362 | ; SSE2-NEXT:    por %xmm1, %xmm13 | 
|  | 2363 | ; SSE2-NEXT:    movdqa %xmm0, %xmm1 | 
|  | 2364 | ; SSE2-NEXT:    pxor %xmm9, %xmm1 | 
|  | 2365 | ; SSE2-NEXT:    movdqa %xmm11, %xmm2 | 
|  | 2366 | ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2 | 
|  | 2367 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 2368 | ; SSE2-NEXT:    pcmpeqd %xmm11, %xmm1 | 
|  | 2369 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] | 
|  | 2370 | ; SSE2-NEXT:    pand %xmm4, %xmm1 | 
|  | 2371 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm2[1,1,3,3] | 
|  | 2372 | ; SSE2-NEXT:    por %xmm1, %xmm11 | 
|  | 2373 | ; SSE2-NEXT:    pand %xmm11, %xmm0 | 
|  | 2374 | ; SSE2-NEXT:    pandn %xmm5, %xmm11 | 
|  | 2375 | ; SSE2-NEXT:    por %xmm0, %xmm11 | 
|  | 2376 | ; SSE2-NEXT:    movdqa %xmm11, %xmm0 | 
|  | 2377 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2378 | ; SSE2-NEXT:    movdqa %xmm0, %xmm1 | 
|  | 2379 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm1 | 
|  | 2380 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] | 
|  | 2381 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2382 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2383 | ; SSE2-NEXT:    pand %xmm4, %xmm0 | 
|  | 2384 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3] | 
|  | 2385 | ; SSE2-NEXT:    por %xmm0, %xmm2 | 
|  | 2386 | ; SSE2-NEXT:    movdqa %xmm13, %xmm0 | 
|  | 2387 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2388 | ; SSE2-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 2389 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm4 | 
|  | 2390 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2391 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2392 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2393 | ; SSE2-NEXT:    pand %xmm5, %xmm0 | 
|  | 2394 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm4[1,1,3,3] | 
|  | 2395 | ; SSE2-NEXT:    por %xmm0, %xmm12 | 
|  | 2396 | ; SSE2-NEXT:    movdqa %xmm15, %xmm0 | 
|  | 2397 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2398 | ; SSE2-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 2399 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm4 | 
|  | 2400 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2401 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2402 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2403 | ; SSE2-NEXT:    pand %xmm5, %xmm0 | 
|  | 2404 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] | 
|  | 2405 | ; SSE2-NEXT:    por %xmm0, %xmm6 | 
|  | 2406 | ; SSE2-NEXT:    movdqa %xmm8, %xmm0 | 
|  | 2407 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2408 | ; SSE2-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 2409 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm4 | 
|  | 2410 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2411 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2412 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2413 | ; SSE2-NEXT:    pand %xmm5, %xmm0 | 
|  | 2414 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[1,1,3,3] | 
|  | 2415 | ; SSE2-NEXT:    por %xmm0, %xmm8 | 
|  | 2416 | ; SSE2-NEXT:    movdqa %xmm7, %xmm0 | 
|  | 2417 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2418 | ; SSE2-NEXT:    movdqa %xmm0, %xmm5 | 
|  | 2419 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm5 | 
|  | 2420 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2421 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2422 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] | 
|  | 2423 | ; SSE2-NEXT:    pand %xmm7, %xmm0 | 
|  | 2424 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2425 | ; SSE2-NEXT:    por %xmm0, %xmm5 | 
|  | 2426 | ; SSE2-NEXT:    movdqa %xmm10, %xmm0 | 
|  | 2427 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2428 | ; SSE2-NEXT:    movdqa %xmm0, %xmm7 | 
|  | 2429 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm7 | 
|  | 2430 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2431 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2432 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2] | 
|  | 2433 | ; SSE2-NEXT:    pand %xmm10, %xmm0 | 
|  | 2434 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] | 
|  | 2435 | ; SSE2-NEXT:    por %xmm0, %xmm10 | 
|  | 2436 | ; SSE2-NEXT:    movdqa %xmm14, %xmm0 | 
|  | 2437 | ; SSE2-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2438 | ; SSE2-NEXT:    movdqa %xmm0, %xmm7 | 
|  | 2439 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm7 | 
|  | 2440 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2441 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3] | 
|  | 2442 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2] | 
|  | 2443 | ; SSE2-NEXT:    pand %xmm0, %xmm1 | 
|  | 2444 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3] | 
|  | 2445 | ; SSE2-NEXT:    por %xmm1, %xmm0 | 
|  | 2446 | ; SSE2-NEXT:    movdqa -{{[0-9]+}}(%rsp), %xmm4 # 16-byte Reload | 
|  | 2447 | ; SSE2-NEXT:    movdqa %xmm4, %xmm1 | 
|  | 2448 | ; SSE2-NEXT:    pxor %xmm9, %xmm1 | 
|  | 2449 | ; SSE2-NEXT:    movdqa %xmm1, %xmm7 | 
|  | 2450 | ; SSE2-NEXT:    pcmpgtd %xmm9, %xmm7 | 
|  | 2451 | ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1 | 
|  | 2452 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] | 
|  | 2453 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2454 | ; SSE2-NEXT:    pand %xmm3, %xmm1 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2455 | ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[1,1,3,3] | 
|  | 2456 | ; SSE2-NEXT:    por %xmm1, %xmm3 | 
|  | 2457 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [255,255] | 
|  | 2458 | ; SSE2-NEXT:    pand %xmm1, %xmm3 | 
|  | 2459 | ; SSE2-NEXT:    pand %xmm4, %xmm3 | 
|  | 2460 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 2461 | ; SSE2-NEXT:    pand %xmm14, %xmm0 | 
|  | 2462 | ; SSE2-NEXT:    packuswb %xmm3, %xmm0 | 
|  | 2463 | ; SSE2-NEXT:    pand %xmm1, %xmm10 | 
|  | 2464 | ; SSE2-NEXT:    pand -{{[0-9]+}}(%rsp), %xmm10 # 16-byte Folded Reload | 
|  | 2465 | ; SSE2-NEXT:    pand %xmm1, %xmm5 | 
|  | 2466 | ; SSE2-NEXT:    pand -{{[0-9]+}}(%rsp), %xmm5 # 16-byte Folded Reload | 
|  | 2467 | ; SSE2-NEXT:    packuswb %xmm10, %xmm5 | 
|  | 2468 | ; SSE2-NEXT:    packuswb %xmm0, %xmm5 | 
|  | 2469 | ; SSE2-NEXT:    pand %xmm1, %xmm8 | 
|  | 2470 | ; SSE2-NEXT:    pand -{{[0-9]+}}(%rsp), %xmm8 # 16-byte Folded Reload | 
|  | 2471 | ; SSE2-NEXT:    pand %xmm1, %xmm6 | 
|  | 2472 | ; SSE2-NEXT:    pand %xmm15, %xmm6 | 
|  | 2473 | ; SSE2-NEXT:    packuswb %xmm8, %xmm6 | 
|  | 2474 | ; SSE2-NEXT:    pand %xmm1, %xmm12 | 
|  | 2475 | ; SSE2-NEXT:    pand %xmm13, %xmm12 | 
|  | 2476 | ; SSE2-NEXT:    pand %xmm1, %xmm2 | 
|  | 2477 | ; SSE2-NEXT:    pand %xmm11, %xmm2 | 
|  | 2478 | ; SSE2-NEXT:    packuswb %xmm12, %xmm2 | 
|  | 2479 | ; SSE2-NEXT:    packuswb %xmm6, %xmm2 | 
|  | 2480 | ; SSE2-NEXT:    packuswb %xmm5, %xmm2 | 
|  | 2481 | ; SSE2-NEXT:    movdqa %xmm2, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2482 | ; SSE2-NEXT:    retq | 
|  | 2483 | ; | 
|  | 2484 | ; SSSE3-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 2485 | ; SSSE3:       # %bb.0: | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2486 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm13 = [255,255] | 
|  | 2487 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,0,2147483648,0] | 
|  | 2488 | ; SSSE3-NEXT:    movdqa %xmm7, %xmm8 | 
|  | 2489 | ; SSSE3-NEXT:    pxor %xmm9, %xmm8 | 
|  | 2490 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm11 = [2147483903,2147483903] | 
|  | 2491 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm10 | 
|  | 2492 | ; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm10 | 
|  | 2493 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm10[0,0,2,2] | 
|  | 2494 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm8 | 
|  | 2495 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3] | 
|  | 2496 | ; SSSE3-NEXT:    pand %xmm12, %xmm8 | 
|  | 2497 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3] | 
|  | 2498 | ; SSSE3-NEXT:    por %xmm8, %xmm10 | 
|  | 2499 | ; SSSE3-NEXT:    pand %xmm10, %xmm7 | 
|  | 2500 | ; SSSE3-NEXT:    pandn %xmm13, %xmm10 | 
|  | 2501 | ; SSSE3-NEXT:    por %xmm7, %xmm10 | 
|  | 2502 | ; SSSE3-NEXT:    movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
|  | 2503 | ; SSSE3-NEXT:    movdqa %xmm6, %xmm7 | 
|  | 2504 | ; SSSE3-NEXT:    pxor %xmm9, %xmm7 | 
|  | 2505 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm8 | 
|  | 2506 | ; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm8 | 
|  | 2507 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm8[0,0,2,2] | 
|  | 2508 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm7 | 
|  | 2509 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] | 
|  | 2510 | ; SSSE3-NEXT:    pand %xmm12, %xmm10 | 
|  | 2511 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm14 = xmm8[1,1,3,3] | 
|  | 2512 | ; SSSE3-NEXT:    por %xmm10, %xmm14 | 
|  | 2513 | ; SSSE3-NEXT:    pand %xmm14, %xmm6 | 
|  | 2514 | ; SSSE3-NEXT:    pandn %xmm13, %xmm14 | 
|  | 2515 | ; SSSE3-NEXT:    por %xmm6, %xmm14 | 
|  | 2516 | ; SSSE3-NEXT:    movdqa %xmm5, %xmm6 | 
|  | 2517 | ; SSSE3-NEXT:    pxor %xmm9, %xmm6 | 
|  | 2518 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm7 | 
|  | 2519 | ; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm7 | 
|  | 2520 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2] | 
|  | 2521 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm6 | 
|  | 2522 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] | 
|  | 2523 | ; SSSE3-NEXT:    pand %xmm8, %xmm6 | 
|  | 2524 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3] | 
|  | 2525 | ; SSSE3-NEXT:    por %xmm6, %xmm7 | 
|  | 2526 | ; SSSE3-NEXT:    pand %xmm7, %xmm5 | 
|  | 2527 | ; SSSE3-NEXT:    pandn %xmm13, %xmm7 | 
|  | 2528 | ; SSSE3-NEXT:    por %xmm5, %xmm7 | 
|  | 2529 | ; SSSE3-NEXT:    movdqa %xmm7, %xmm10 | 
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2530 | ; SSSE3-NEXT:    movdqa %xmm10, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2531 | ; SSSE3-NEXT:    movdqa %xmm4, %xmm5 | 
|  | 2532 | ; SSSE3-NEXT:    pxor %xmm9, %xmm5 | 
|  | 2533 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm6 | 
|  | 2534 | ; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm6 | 
|  | 2535 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2] | 
|  | 2536 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm5 | 
|  | 2537 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2538 | ; SSSE3-NEXT:    pand %xmm7, %xmm5 | 
|  | 2539 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3] | 
|  | 2540 | ; SSSE3-NEXT:    por %xmm5, %xmm6 | 
|  | 2541 | ; SSSE3-NEXT:    pand %xmm6, %xmm4 | 
|  | 2542 | ; SSSE3-NEXT:    pandn %xmm13, %xmm6 | 
|  | 2543 | ; SSSE3-NEXT:    por %xmm4, %xmm6 | 
|  | 2544 | ; SSSE3-NEXT:    movdqa %xmm6, %xmm7 | 
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2545 | ; SSSE3-NEXT:    movdqa %xmm7, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2546 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm4 | 
|  | 2547 | ; SSSE3-NEXT:    pxor %xmm9, %xmm4 | 
|  | 2548 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm5 | 
|  | 2549 | ; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm5 | 
|  | 2550 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] | 
|  | 2551 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm4 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2552 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2553 | ; SSSE3-NEXT:    pand %xmm6, %xmm4 | 
|  | 2554 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2555 | ; SSSE3-NEXT:    por %xmm4, %xmm5 | 
|  | 2556 | ; SSSE3-NEXT:    pand %xmm5, %xmm3 | 
|  | 2557 | ; SSSE3-NEXT:    pandn %xmm13, %xmm5 | 
|  | 2558 | ; SSSE3-NEXT:    por %xmm3, %xmm5 | 
|  | 2559 | ; SSSE3-NEXT:    movdqa %xmm5, %xmm8 | 
| Quentin Colombet | 48abac8 | 2018-02-17 03:05:33 +0000 | [diff] [blame] | 2560 | ; SSSE3-NEXT:    movdqa %xmm8, -{{[0-9]+}}(%rsp) # 16-byte Spill | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2561 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 2562 | ; SSSE3-NEXT:    pxor %xmm9, %xmm3 | 
|  | 2563 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm4 | 
|  | 2564 | ; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4 | 
|  | 2565 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2566 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm3 | 
|  | 2567 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3] | 
|  | 2568 | ; SSSE3-NEXT:    pand %xmm5, %xmm3 | 
|  | 2569 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3] | 
|  | 2570 | ; SSSE3-NEXT:    por %xmm3, %xmm15 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2571 | ; SSSE3-NEXT:    pand %xmm15, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2572 | ; SSSE3-NEXT:    pandn %xmm13, %xmm15 | 
|  | 2573 | ; SSSE3-NEXT:    movdqa %xmm13, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2574 | ; SSSE3-NEXT:    por %xmm2, %xmm15 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2575 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2576 | ; SSSE3-NEXT:    pxor %xmm9, %xmm2 | 
|  | 2577 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm3 | 
|  | 2578 | ; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3 | 
|  | 2579 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 2580 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2581 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2582 | ; SSSE3-NEXT:    pand %xmm4, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2583 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3] | 
|  | 2584 | ; SSSE3-NEXT:    por %xmm2, %xmm13 | 
|  | 2585 | ; SSSE3-NEXT:    pand %xmm13, %xmm1 | 
|  | 2586 | ; SSSE3-NEXT:    pandn %xmm5, %xmm13 | 
|  | 2587 | ; SSSE3-NEXT:    por %xmm1, %xmm13 | 
|  | 2588 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm1 | 
|  | 2589 | ; SSSE3-NEXT:    pxor %xmm9, %xmm1 | 
|  | 2590 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm2 | 
|  | 2591 | ; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2 | 
|  | 2592 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2] | 
|  | 2593 | ; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm1 | 
|  | 2594 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] | 
|  | 2595 | ; SSSE3-NEXT:    pand %xmm4, %xmm1 | 
|  | 2596 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm2[1,1,3,3] | 
|  | 2597 | ; SSSE3-NEXT:    por %xmm1, %xmm11 | 
|  | 2598 | ; SSSE3-NEXT:    pand %xmm11, %xmm0 | 
|  | 2599 | ; SSSE3-NEXT:    pandn %xmm5, %xmm11 | 
|  | 2600 | ; SSSE3-NEXT:    por %xmm0, %xmm11 | 
|  | 2601 | ; SSSE3-NEXT:    movdqa %xmm11, %xmm0 | 
|  | 2602 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2603 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm1 | 
|  | 2604 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm1 | 
|  | 2605 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2] | 
|  | 2606 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2607 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2608 | ; SSSE3-NEXT:    pand %xmm4, %xmm0 | 
|  | 2609 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3] | 
|  | 2610 | ; SSSE3-NEXT:    por %xmm0, %xmm2 | 
|  | 2611 | ; SSSE3-NEXT:    movdqa %xmm13, %xmm0 | 
|  | 2612 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2613 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 2614 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm4 | 
|  | 2615 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2616 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2617 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2618 | ; SSSE3-NEXT:    pand %xmm5, %xmm0 | 
|  | 2619 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm4[1,1,3,3] | 
|  | 2620 | ; SSSE3-NEXT:    por %xmm0, %xmm12 | 
|  | 2621 | ; SSSE3-NEXT:    movdqa %xmm15, %xmm0 | 
|  | 2622 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2623 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 2624 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm4 | 
|  | 2625 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2626 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2627 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2628 | ; SSSE3-NEXT:    pand %xmm5, %xmm0 | 
|  | 2629 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3] | 
|  | 2630 | ; SSSE3-NEXT:    por %xmm0, %xmm6 | 
|  | 2631 | ; SSSE3-NEXT:    movdqa %xmm8, %xmm0 | 
|  | 2632 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2633 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm4 | 
|  | 2634 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm4 | 
|  | 2635 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2] | 
|  | 2636 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2637 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2638 | ; SSSE3-NEXT:    pand %xmm5, %xmm0 | 
|  | 2639 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[1,1,3,3] | 
|  | 2640 | ; SSSE3-NEXT:    por %xmm0, %xmm8 | 
|  | 2641 | ; SSSE3-NEXT:    movdqa %xmm7, %xmm0 | 
|  | 2642 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2643 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm5 | 
|  | 2644 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm5 | 
|  | 2645 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2646 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2647 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2] | 
|  | 2648 | ; SSSE3-NEXT:    pand %xmm7, %xmm0 | 
|  | 2649 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3] | 
|  | 2650 | ; SSSE3-NEXT:    por %xmm0, %xmm5 | 
|  | 2651 | ; SSSE3-NEXT:    movdqa %xmm10, %xmm0 | 
|  | 2652 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2653 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm7 | 
|  | 2654 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm7 | 
|  | 2655 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2656 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] | 
|  | 2657 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2] | 
|  | 2658 | ; SSSE3-NEXT:    pand %xmm10, %xmm0 | 
|  | 2659 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[1,1,3,3] | 
|  | 2660 | ; SSSE3-NEXT:    por %xmm0, %xmm10 | 
|  | 2661 | ; SSSE3-NEXT:    movdqa %xmm14, %xmm0 | 
|  | 2662 | ; SSSE3-NEXT:    pxor %xmm9, %xmm0 | 
|  | 2663 | ; SSSE3-NEXT:    movdqa %xmm0, %xmm7 | 
|  | 2664 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm7 | 
|  | 2665 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0 | 
|  | 2666 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3] | 
|  | 2667 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2] | 
|  | 2668 | ; SSSE3-NEXT:    pand %xmm0, %xmm1 | 
|  | 2669 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3] | 
|  | 2670 | ; SSSE3-NEXT:    por %xmm1, %xmm0 | 
|  | 2671 | ; SSSE3-NEXT:    movdqa -{{[0-9]+}}(%rsp), %xmm4 # 16-byte Reload | 
|  | 2672 | ; SSSE3-NEXT:    movdqa %xmm4, %xmm1 | 
|  | 2673 | ; SSSE3-NEXT:    pxor %xmm9, %xmm1 | 
|  | 2674 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm7 | 
|  | 2675 | ; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm7 | 
|  | 2676 | ; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1 | 
|  | 2677 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] | 
|  | 2678 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2679 | ; SSSE3-NEXT:    pand %xmm3, %xmm1 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2680 | ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[1,1,3,3] | 
|  | 2681 | ; SSSE3-NEXT:    por %xmm1, %xmm3 | 
|  | 2682 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [255,255] | 
|  | 2683 | ; SSSE3-NEXT:    pand %xmm1, %xmm3 | 
|  | 2684 | ; SSSE3-NEXT:    pand %xmm4, %xmm3 | 
|  | 2685 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 2686 | ; SSSE3-NEXT:    pand %xmm14, %xmm0 | 
|  | 2687 | ; SSSE3-NEXT:    packuswb %xmm3, %xmm0 | 
|  | 2688 | ; SSSE3-NEXT:    pand %xmm1, %xmm10 | 
|  | 2689 | ; SSSE3-NEXT:    pand -{{[0-9]+}}(%rsp), %xmm10 # 16-byte Folded Reload | 
|  | 2690 | ; SSSE3-NEXT:    pand %xmm1, %xmm5 | 
|  | 2691 | ; SSSE3-NEXT:    pand -{{[0-9]+}}(%rsp), %xmm5 # 16-byte Folded Reload | 
|  | 2692 | ; SSSE3-NEXT:    packuswb %xmm10, %xmm5 | 
|  | 2693 | ; SSSE3-NEXT:    packuswb %xmm0, %xmm5 | 
|  | 2694 | ; SSSE3-NEXT:    pand %xmm1, %xmm8 | 
|  | 2695 | ; SSSE3-NEXT:    pand -{{[0-9]+}}(%rsp), %xmm8 # 16-byte Folded Reload | 
|  | 2696 | ; SSSE3-NEXT:    pand %xmm1, %xmm6 | 
|  | 2697 | ; SSSE3-NEXT:    pand %xmm15, %xmm6 | 
|  | 2698 | ; SSSE3-NEXT:    packuswb %xmm8, %xmm6 | 
|  | 2699 | ; SSSE3-NEXT:    pand %xmm1, %xmm12 | 
|  | 2700 | ; SSSE3-NEXT:    pand %xmm13, %xmm12 | 
|  | 2701 | ; SSSE3-NEXT:    pand %xmm1, %xmm2 | 
|  | 2702 | ; SSSE3-NEXT:    pand %xmm11, %xmm2 | 
|  | 2703 | ; SSSE3-NEXT:    packuswb %xmm12, %xmm2 | 
|  | 2704 | ; SSSE3-NEXT:    packuswb %xmm6, %xmm2 | 
|  | 2705 | ; SSSE3-NEXT:    packuswb %xmm5, %xmm2 | 
|  | 2706 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2707 | ; SSSE3-NEXT:    retq | 
|  | 2708 | ; | 
|  | 2709 | ; SSE41-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 2710 | ; SSE41:       # %bb.0: | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2711 | ; SSE41-NEXT:    movdqa %xmm0, %xmm8 | 
|  | 2712 | ; SSE41-NEXT:    movapd {{.*#+}} xmm9 = [255,255] | 
|  | 2713 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,0,2147483648,0] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2714 | ; SSE41-NEXT:    movdqa %xmm7, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2715 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2716 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm12 = [2147483903,2147483903] | 
|  | 2717 | ; SSE41-NEXT:    movdqa %xmm12, %xmm11 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2718 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm11 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2719 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm11[0,0,2,2] | 
|  | 2720 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2721 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2722 | ; SSE41-NEXT:    pand %xmm13, %xmm14 | 
|  | 2723 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[1,1,3,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2724 | ; SSE41-NEXT:    por %xmm14, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2725 | ; SSE41-NEXT:    movapd %xmm9, %xmm11 | 
|  | 2726 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm11 | 
|  | 2727 | ; SSE41-NEXT:    movdqa %xmm6, %xmm0 | 
|  | 2728 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2729 | ; SSE41-NEXT:    movdqa %xmm12, %xmm7 | 
|  | 2730 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7 | 
|  | 2731 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm7[0,0,2,2] | 
|  | 2732 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2733 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3] | 
|  | 2734 | ; SSE41-NEXT:    pand %xmm13, %xmm14 | 
|  | 2735 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3] | 
|  | 2736 | ; SSE41-NEXT:    por %xmm14, %xmm0 | 
|  | 2737 | ; SSE41-NEXT:    movapd %xmm9, %xmm13 | 
|  | 2738 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm13 | 
|  | 2739 | ; SSE41-NEXT:    movdqa %xmm5, %xmm0 | 
|  | 2740 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2741 | ; SSE41-NEXT:    movdqa %xmm12, %xmm6 | 
|  | 2742 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm6 | 
|  | 2743 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm6[0,0,2,2] | 
|  | 2744 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2745 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2746 | ; SSE41-NEXT:    pand %xmm14, %xmm7 | 
|  | 2747 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3] | 
|  | 2748 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2749 | ; SSE41-NEXT:    movapd %xmm9, %xmm14 | 
|  | 2750 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm14 | 
|  | 2751 | ; SSE41-NEXT:    movdqa %xmm4, %xmm0 | 
|  | 2752 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2753 | ; SSE41-NEXT:    movdqa %xmm12, %xmm5 | 
|  | 2754 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5 | 
|  | 2755 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2] | 
|  | 2756 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2757 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2758 | ; SSE41-NEXT:    pand %xmm6, %xmm7 | 
|  | 2759 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3] | 
|  | 2760 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2761 | ; SSE41-NEXT:    movapd %xmm9, %xmm15 | 
|  | 2762 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm15 | 
|  | 2763 | ; SSE41-NEXT:    movdqa %xmm3, %xmm0 | 
|  | 2764 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2765 | ; SSE41-NEXT:    movdqa %xmm12, %xmm4 | 
|  | 2766 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4 | 
|  | 2767 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2] | 
|  | 2768 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2769 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2770 | ; SSE41-NEXT:    pand %xmm6, %xmm7 | 
|  | 2771 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3] | 
|  | 2772 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2773 | ; SSE41-NEXT:    movapd %xmm9, %xmm4 | 
|  | 2774 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm4 | 
|  | 2775 | ; SSE41-NEXT:    movdqa %xmm2, %xmm0 | 
|  | 2776 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2777 | ; SSE41-NEXT:    movdqa %xmm12, %xmm3 | 
|  | 2778 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3 | 
|  | 2779 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2] | 
|  | 2780 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2781 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2782 | ; SSE41-NEXT:    pand %xmm6, %xmm7 | 
|  | 2783 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2784 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2785 | ; SSE41-NEXT:    movapd %xmm9, %xmm6 | 
|  | 2786 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2787 | ; SSE41-NEXT:    movdqa %xmm1, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2788 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2789 | ; SSE41-NEXT:    movdqa %xmm12, %xmm2 | 
|  | 2790 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2 | 
|  | 2791 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2] | 
|  | 2792 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2793 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2794 | ; SSE41-NEXT:    pand %xmm3, %xmm7 | 
|  | 2795 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 2796 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2797 | ; SSE41-NEXT:    movapd %xmm9, %xmm7 | 
|  | 2798 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm7 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2799 | ; SSE41-NEXT:    movdqa %xmm8, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2800 | ; SSE41-NEXT:    pxor %xmm10, %xmm0 | 
|  | 2801 | ; SSE41-NEXT:    movdqa %xmm12, %xmm1 | 
|  | 2802 | ; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 2803 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2] | 
|  | 2804 | ; SSE41-NEXT:    pcmpeqd %xmm12, %xmm0 | 
|  | 2805 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] | 
|  | 2806 | ; SSE41-NEXT:    pand %xmm2, %xmm3 | 
|  | 2807 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] | 
|  | 2808 | ; SSE41-NEXT:    por %xmm3, %xmm0 | 
|  | 2809 | ; SSE41-NEXT:    movapd %xmm9, %xmm2 | 
|  | 2810 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2811 | ; SSE41-NEXT:    movapd %xmm2, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2812 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2813 | ; SSE41-NEXT:    movapd %xmm0, %xmm1 | 
|  | 2814 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm1 | 
|  | 2815 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2] | 
|  | 2816 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2817 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 2818 | ; SSE41-NEXT:    pand %xmm3, %xmm5 | 
|  | 2819 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3] | 
|  | 2820 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 2821 | ; SSE41-NEXT:    xorpd %xmm8, %xmm8 | 
|  | 2822 | ; SSE41-NEXT:    pxor %xmm1, %xmm1 | 
|  | 2823 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm1 | 
|  | 2824 | ; SSE41-NEXT:    movapd %xmm7, %xmm0 | 
|  | 2825 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2826 | ; SSE41-NEXT:    movapd %xmm0, %xmm2 | 
|  | 2827 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm2 | 
|  | 2828 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2] | 
|  | 2829 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2830 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] | 
|  | 2831 | ; SSE41-NEXT:    pand %xmm5, %xmm3 | 
|  | 2832 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 2833 | ; SSE41-NEXT:    por %xmm3, %xmm0 | 
|  | 2834 | ; SSE41-NEXT:    pxor %xmm12, %xmm12 | 
|  | 2835 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm12 | 
|  | 2836 | ; SSE41-NEXT:    movapd %xmm6, %xmm0 | 
|  | 2837 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2838 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
|  | 2839 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 2840 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] | 
|  | 2841 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2842 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3] | 
|  | 2843 | ; SSE41-NEXT:    pand %xmm5, %xmm7 | 
|  | 2844 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2845 | ; SSE41-NEXT:    por %xmm7, %xmm0 | 
|  | 2846 | ; SSE41-NEXT:    pxor %xmm7, %xmm7 | 
|  | 2847 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm7 | 
|  | 2848 | ; SSE41-NEXT:    movapd %xmm4, %xmm0 | 
|  | 2849 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2850 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
|  | 2851 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 2852 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] | 
|  | 2853 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2854 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3] | 
|  | 2855 | ; SSE41-NEXT:    pand %xmm5, %xmm6 | 
|  | 2856 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2857 | ; SSE41-NEXT:    por %xmm6, %xmm0 | 
|  | 2858 | ; SSE41-NEXT:    pxor %xmm6, %xmm6 | 
|  | 2859 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2860 | ; SSE41-NEXT:    movapd %xmm15, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2861 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2862 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
|  | 2863 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 2864 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2] | 
|  | 2865 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2866 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 2867 | ; SSE41-NEXT:    pand %xmm4, %xmm5 | 
|  | 2868 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2869 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 2870 | ; SSE41-NEXT:    pxor %xmm4, %xmm4 | 
|  | 2871 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm15, %xmm4 | 
|  | 2872 | ; SSE41-NEXT:    movapd %xmm14, %xmm0 | 
|  | 2873 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2874 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
|  | 2875 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 2876 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2] | 
|  | 2877 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2878 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3] | 
|  | 2879 | ; SSE41-NEXT:    pand %xmm5, %xmm2 | 
|  | 2880 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2881 | ; SSE41-NEXT:    por %xmm2, %xmm0 | 
|  | 2882 | ; SSE41-NEXT:    xorpd %xmm15, %xmm15 | 
|  | 2883 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm14, %xmm15 | 
|  | 2884 | ; SSE41-NEXT:    movapd %xmm13, %xmm0 | 
|  | 2885 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2886 | ; SSE41-NEXT:    movapd %xmm0, %xmm2 | 
|  | 2887 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm2 | 
|  | 2888 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[0,0,2,2] | 
|  | 2889 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2890 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3] | 
|  | 2891 | ; SSE41-NEXT:    pand %xmm14, %xmm3 | 
|  | 2892 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3] | 
|  | 2893 | ; SSE41-NEXT:    por %xmm3, %xmm0 | 
|  | 2894 | ; SSE41-NEXT:    pxor %xmm2, %xmm2 | 
|  | 2895 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm13, %xmm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2896 | ; SSE41-NEXT:    movapd %xmm11, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2897 | ; SSE41-NEXT:    xorpd %xmm10, %xmm0 | 
|  | 2898 | ; SSE41-NEXT:    movapd %xmm0, %xmm3 | 
|  | 2899 | ; SSE41-NEXT:    pcmpgtd %xmm10, %xmm3 | 
|  | 2900 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[0,0,2,2] | 
|  | 2901 | ; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0 | 
|  | 2902 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3] | 
|  | 2903 | ; SSE41-NEXT:    pand %xmm13, %xmm5 | 
|  | 2904 | ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3] | 
|  | 2905 | ; SSE41-NEXT:    por %xmm5, %xmm0 | 
|  | 2906 | ; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm8 | 
|  | 2907 | ; SSE41-NEXT:    andpd %xmm9, %xmm8 | 
|  | 2908 | ; SSE41-NEXT:    andpd %xmm9, %xmm2 | 
|  | 2909 | ; SSE41-NEXT:    packuswb %xmm8, %xmm2 | 
|  | 2910 | ; SSE41-NEXT:    andpd %xmm9, %xmm15 | 
|  | 2911 | ; SSE41-NEXT:    andpd %xmm9, %xmm4 | 
|  | 2912 | ; SSE41-NEXT:    packuswb %xmm15, %xmm4 | 
|  | 2913 | ; SSE41-NEXT:    packuswb %xmm2, %xmm4 | 
|  | 2914 | ; SSE41-NEXT:    andpd %xmm9, %xmm6 | 
|  | 2915 | ; SSE41-NEXT:    andpd %xmm9, %xmm7 | 
|  | 2916 | ; SSE41-NEXT:    packuswb %xmm6, %xmm7 | 
|  | 2917 | ; SSE41-NEXT:    andpd %xmm9, %xmm12 | 
|  | 2918 | ; SSE41-NEXT:    andpd %xmm9, %xmm1 | 
|  | 2919 | ; SSE41-NEXT:    packuswb %xmm12, %xmm1 | 
|  | 2920 | ; SSE41-NEXT:    packuswb %xmm7, %xmm1 | 
|  | 2921 | ; SSE41-NEXT:    packuswb %xmm4, %xmm1 | 
|  | 2922 | ; SSE41-NEXT:    movdqa %xmm1, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2923 | ; SSE41-NEXT:    retq | 
|  | 2924 | ; | 
|  | 2925 | ; AVX1-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 2926 | ; AVX1:       # %bb.0: | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2927 | ; AVX1-NEXT:    vmovapd {{.*#+}} ymm5 = [255,255,255,255] | 
|  | 2928 | ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2929 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [255,255] | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2930 | ; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm4, %xmm6 | 
|  | 2931 | ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm7 | 
|  | 2932 | ; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm7, %ymm6 | 
|  | 2933 | ; AVX1-NEXT:    vblendvpd %ymm6, %ymm3, %ymm5, %ymm3 | 
|  | 2934 | ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm6 | 
|  | 2935 | ; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm4, %xmm6 | 
|  | 2936 | ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm4, %xmm7 | 
|  | 2937 | ; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm7, %ymm6 | 
|  | 2938 | ; AVX1-NEXT:    vblendvpd %ymm6, %ymm2, %ymm5, %ymm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2939 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm6 | 
|  | 2940 | ; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm4, %xmm6 | 
|  | 2941 | ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm7 | 
|  | 2942 | ; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm7, %ymm6 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2943 | ; AVX1-NEXT:    vblendvpd %ymm6, %ymm1, %ymm5, %ymm1 | 
|  | 2944 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6 | 
|  | 2945 | ; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm4, %xmm6 | 
|  | 2946 | ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm4, %xmm7 | 
|  | 2947 | ; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm7, %ymm6 | 
|  | 2948 | ; AVX1-NEXT:    vblendvpd %ymm6, %ymm0, %ymm5, %ymm0 | 
|  | 2949 | ; AVX1-NEXT:    vxorpd %xmm5, %xmm5, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2950 | ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm6 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2951 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm6, %xmm7 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2952 | ; AVX1-NEXT:    vpand %xmm4, %xmm6, %xmm6 | 
|  | 2953 | ; AVX1-NEXT:    vpand %xmm6, %xmm7, %xmm6 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2954 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm3, %xmm7 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2955 | ; AVX1-NEXT:    vpand %xmm4, %xmm3, %xmm3 | 
|  | 2956 | ; AVX1-NEXT:    vpand %xmm3, %xmm7, %xmm3 | 
|  | 2957 | ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm7 | 
|  | 2958 | ; AVX1-NEXT:    vpackuswb %xmm6, %xmm3, %xmm3 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2959 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm7, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2960 | ; AVX1-NEXT:    vpand %xmm4, %xmm7, %xmm7 | 
|  | 2961 | ; AVX1-NEXT:    vpand %xmm7, %xmm6, %xmm6 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2962 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm2, %xmm7 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2963 | ; AVX1-NEXT:    vpand %xmm4, %xmm2, %xmm2 | 
|  | 2964 | ; AVX1-NEXT:    vpand %xmm2, %xmm7, %xmm2 | 
|  | 2965 | ; AVX1-NEXT:    vpackuswb %xmm6, %xmm2, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2966 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm6 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2967 | ; AVX1-NEXT:    vpackuswb %xmm3, %xmm2, %xmm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2968 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm6, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2969 | ; AVX1-NEXT:    vpand %xmm4, %xmm6, %xmm6 | 
|  | 2970 | ; AVX1-NEXT:    vpand %xmm6, %xmm3, %xmm3 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2971 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm1, %xmm6 | 
|  | 2972 | ; AVX1-NEXT:    vpand %xmm4, %xmm1, %xmm1 | 
|  | 2973 | ; AVX1-NEXT:    vpand %xmm1, %xmm6, %xmm1 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2974 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2975 | ; AVX1-NEXT:    vpackuswb %xmm3, %xmm1, %xmm1 | 
|  | 2976 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm6, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2977 | ; AVX1-NEXT:    vpand %xmm4, %xmm6, %xmm6 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2978 | ; AVX1-NEXT:    vpand %xmm6, %xmm3, %xmm3 | 
|  | 2979 | ; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm0, %xmm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2980 | ; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2981 | ; AVX1-NEXT:    vpand %xmm0, %xmm5, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2982 | ; AVX1-NEXT:    vpackuswb %xmm3, %xmm0, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2983 | ; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2984 | ; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0 | 
|  | 2985 | ; AVX1-NEXT:    vzeroupper | 
|  | 2986 | ; AVX1-NEXT:    retq | 
|  | 2987 | ; | 
|  | 2988 | ; AVX2-SLOW-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 2989 | ; AVX2-SLOW:       # %bb.0: | 
|  | 2990 | ; AVX2-SLOW-NEXT:    vpbroadcastq {{.*#+}} ymm4 = [255,255,255,255] | 
|  | 2991 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm2, %ymm4, %ymm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2992 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm5, %ymm2, %ymm4, %ymm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 2993 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm5 | 
|  | 2994 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm5, %ymm3, %ymm4, %ymm3 | 
|  | 2995 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm0, %ymm4, %ymm5 | 
|  | 2996 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm5, %ymm0, %ymm4, %ymm0 | 
|  | 2997 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm1, %ymm4, %ymm5 | 
|  | 2998 | ; AVX2-SLOW-NEXT:    vblendvpd %ymm5, %ymm1, %ymm4, %ymm1 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 2999 | ; AVX2-SLOW-NEXT:    vpxor %xmm4, %xmm4, %xmm4 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 3000 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm4, %ymm1, %ymm5 | 
|  | 3001 | ; AVX2-SLOW-NEXT:    vpand %ymm1, %ymm5, %ymm1 | 
|  | 3002 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm4, %ymm0, %ymm5 | 
|  | 3003 | ; AVX2-SLOW-NEXT:    vpand %ymm0, %ymm5, %ymm0 | 
|  | 3004 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm4, %ymm3, %ymm5 | 
|  | 3005 | ; AVX2-SLOW-NEXT:    vpand %ymm3, %ymm5, %ymm3 | 
|  | 3006 | ; AVX2-SLOW-NEXT:    vpcmpgtq %ymm4, %ymm2, %ymm4 | 
|  | 3007 | ; AVX2-SLOW-NEXT:    vpand %ymm2, %ymm4, %ymm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3008 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] | 
|  | 3009 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3] | 
|  | 3010 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] | 
|  | 3011 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,2,2,3] | 
|  | 3012 | ; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2 | 
|  | 3013 | ; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm3 = [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] | 
|  | 3014 | ; AVX2-SLOW-NEXT:    vpshufb %ymm3, %ymm2, %ymm2 | 
|  | 3015 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3] | 
|  | 3016 | ; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} xmm4 = <0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u> | 
|  | 3017 | ; AVX2-SLOW-NEXT:    vpshufb %xmm4, %xmm2, %xmm2 | 
|  | 3018 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] | 
|  | 3019 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
|  | 3020 | ; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] | 
|  | 3021 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3] | 
|  | 3022 | ; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 3023 | ; AVX2-SLOW-NEXT:    vpshufb %ymm3, %ymm0, %ymm0 | 
|  | 3024 | ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
|  | 3025 | ; AVX2-SLOW-NEXT:    vpshufb %xmm4, %xmm0, %xmm0 | 
|  | 3026 | ; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0] | 
|  | 3027 | ; AVX2-SLOW-NEXT:    vzeroupper | 
|  | 3028 | ; AVX2-SLOW-NEXT:    retq | 
|  | 3029 | ; | 
|  | 3030 | ; AVX2-FAST-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 3031 | ; AVX2-FAST:       # %bb.0: | 
|  | 3032 | ; AVX2-FAST-NEXT:    vpbroadcastq {{.*#+}} ymm4 = [255,255,255,255] | 
|  | 3033 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm2, %ymm4, %ymm5 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3034 | ; AVX2-FAST-NEXT:    vblendvpd %ymm5, %ymm2, %ymm4, %ymm2 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 3035 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm5 | 
|  | 3036 | ; AVX2-FAST-NEXT:    vblendvpd %ymm5, %ymm3, %ymm4, %ymm3 | 
|  | 3037 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm0, %ymm4, %ymm5 | 
|  | 3038 | ; AVX2-FAST-NEXT:    vblendvpd %ymm5, %ymm0, %ymm4, %ymm0 | 
|  | 3039 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm1, %ymm4, %ymm5 | 
|  | 3040 | ; AVX2-FAST-NEXT:    vblendvpd %ymm5, %ymm1, %ymm4, %ymm1 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3041 | ; AVX2-FAST-NEXT:    vpxor %xmm4, %xmm4, %xmm4 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 3042 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm4, %ymm1, %ymm5 | 
|  | 3043 | ; AVX2-FAST-NEXT:    vpand %ymm1, %ymm5, %ymm1 | 
|  | 3044 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm4, %ymm0, %ymm5 | 
|  | 3045 | ; AVX2-FAST-NEXT:    vpand %ymm0, %ymm5, %ymm0 | 
|  | 3046 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm4, %ymm3, %ymm5 | 
|  | 3047 | ; AVX2-FAST-NEXT:    vpand %ymm3, %ymm5, %ymm3 | 
|  | 3048 | ; AVX2-FAST-NEXT:    vpcmpgtq %ymm4, %ymm2, %ymm4 | 
|  | 3049 | ; AVX2-FAST-NEXT:    vpand %ymm2, %ymm4, %ymm2 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3050 | ; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,2,4,6,4,6,6,7] | 
|  | 3051 | ; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm4, %ymm2 | 
|  | 3052 | ; AVX2-FAST-NEXT:    vpermd %ymm3, %ymm4, %ymm3 | 
|  | 3053 | ; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2 | 
|  | 3054 | ; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [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] | 
|  | 3055 | ; AVX2-FAST-NEXT:    vpshufb %ymm3, %ymm2, %ymm2 | 
|  | 3056 | ; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3] | 
|  | 3057 | ; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm5 = <0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u> | 
|  | 3058 | ; AVX2-FAST-NEXT:    vpshufb %xmm5, %xmm2, %xmm2 | 
|  | 3059 | ; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm4, %ymm0 | 
|  | 3060 | ; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm4, %ymm1 | 
|  | 3061 | ; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 3062 | ; AVX2-FAST-NEXT:    vpshufb %ymm3, %ymm0, %ymm0 | 
|  | 3063 | ; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
|  | 3064 | ; AVX2-FAST-NEXT:    vpshufb %xmm5, %xmm0, %xmm0 | 
|  | 3065 | ; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0] | 
|  | 3066 | ; AVX2-FAST-NEXT:    vzeroupper | 
|  | 3067 | ; AVX2-FAST-NEXT:    retq | 
|  | 3068 | ; | 
|  | 3069 | ; AVX512-LABEL: trunc_packus_v16i64_v16i8: | 
|  | 3070 | ; AVX512:       # %bb.0: | 
|  | 3071 | ; AVX512-NEXT:    vpbroadcastq {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255] | 
|  | 3072 | ; AVX512-NEXT:    vpminsq %zmm2, %zmm0, %zmm0 | 
|  | 3073 | ; AVX512-NEXT:    vpminsq %zmm2, %zmm1, %zmm1 | 
|  | 3074 | ; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 3075 | ; AVX512-NEXT:    vpmaxsq %zmm2, %zmm1, %zmm1 | 
|  | 3076 | ; AVX512-NEXT:    vpmaxsq %zmm2, %zmm0, %zmm0 | 
|  | 3077 | ; AVX512-NEXT:    vpmovqd %zmm0, %ymm0 | 
|  | 3078 | ; AVX512-NEXT:    vpmovqd %zmm1, %ymm1 | 
|  | 3079 | ; AVX512-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0 | 
|  | 3080 | ; AVX512-NEXT:    vpmovdb %zmm0, %xmm0 | 
|  | 3081 | ; AVX512-NEXT:    vzeroupper | 
|  | 3082 | ; AVX512-NEXT:    retq | 
|  | 3083 | %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> | 
|  | 3084 | %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> | 
|  | 3085 | %3 = icmp sgt <16 x i64> %2, zeroinitializer | 
|  | 3086 | %4 = select <16 x i1> %3, <16 x i64> %2, <16 x i64> zeroinitializer | 
|  | 3087 | %5 = trunc <16 x i64> %4 to <16 x i8> | 
|  | 3088 | ret <16 x i8> %5 | 
|  | 3089 | } | 
|  | 3090 |  | 
|  | 3091 | define <8 x i8> @trunc_packus_v8i32_v8i8(<8 x i32> %a0) { | 
|  | 3092 | ; SSE2-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3093 | ; SSE2:       # %bb.0: | 
|  | 3094 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3095 | ; SSE2-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 3096 | ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3097 | ; SSE2-NEXT:    pand %xmm3, %xmm1 | 
|  | 3098 | ; SSE2-NEXT:    pandn %xmm2, %xmm3 | 
|  | 3099 | ; SSE2-NEXT:    por %xmm1, %xmm3 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 3100 | ; SSE2-NEXT:    movdqa %xmm2, %xmm1 | 
|  | 3101 | ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 3102 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 3103 | ; SSE2-NEXT:    pandn %xmm2, %xmm1 | 
|  | 3104 | ; SSE2-NEXT:    por %xmm0, %xmm1 | 
|  | 3105 | ; SSE2-NEXT:    pxor %xmm2, %xmm2 | 
|  | 3106 | ; SSE2-NEXT:    movdqa %xmm1, %xmm0 | 
|  | 3107 | ; SSE2-NEXT:    pcmpgtd %xmm2, %xmm0 | 
|  | 3108 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 3109 | ; SSE2-NEXT:    movdqa %xmm3, %xmm1 | 
|  | 3110 | ; SSE2-NEXT:    pcmpgtd %xmm2, %xmm1 | 
|  | 3111 | ; SSE2-NEXT:    pand %xmm3, %xmm1 | 
|  | 3112 | ; SSE2-NEXT:    pslld $16, %xmm1 | 
|  | 3113 | ; SSE2-NEXT:    psrad $16, %xmm1 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3114 | ; SSE2-NEXT:    pslld $16, %xmm0 | 
|  | 3115 | ; SSE2-NEXT:    psrad $16, %xmm0 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 3116 | ; SSE2-NEXT:    packssdw %xmm1, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3117 | ; SSE2-NEXT:    retq | 
|  | 3118 | ; | 
|  | 3119 | ; SSSE3-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3120 | ; SSSE3:       # %bb.0: | 
|  | 3121 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3122 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 3123 | ; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm3 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3124 | ; SSSE3-NEXT:    pand %xmm3, %xmm1 | 
|  | 3125 | ; SSSE3-NEXT:    pandn %xmm2, %xmm3 | 
|  | 3126 | ; SSSE3-NEXT:    por %xmm1, %xmm3 | 
| Simon Pilgrim | 0be5567 | 2018-02-11 10:52:37 +0000 | [diff] [blame] | 3127 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm1 | 
|  | 3128 | ; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 3129 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 3130 | ; SSSE3-NEXT:    pandn %xmm2, %xmm1 | 
|  | 3131 | ; SSSE3-NEXT:    por %xmm0, %xmm1 | 
|  | 3132 | ; SSSE3-NEXT:    pxor %xmm2, %xmm2 | 
|  | 3133 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm0 | 
|  | 3134 | ; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm0 | 
|  | 3135 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 3136 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm1 | 
|  | 3137 | ; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm1 | 
|  | 3138 | ; SSSE3-NEXT:    pand %xmm3, %xmm1 | 
|  | 3139 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 3140 | ; SSSE3-NEXT:    pshufb %xmm2, %xmm1 | 
|  | 3141 | ; SSSE3-NEXT:    pshufb %xmm2, %xmm0 | 
|  | 3142 | ; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3143 | ; SSSE3-NEXT:    retq | 
|  | 3144 | ; | 
|  | 3145 | ; SSE41-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3146 | ; SSE41:       # %bb.0: | 
|  | 3147 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3148 | ; SSE41-NEXT:    pminsd %xmm2, %xmm1 | 
|  | 3149 | ; SSE41-NEXT:    pminsd %xmm2, %xmm0 | 
|  | 3150 | ; SSE41-NEXT:    pxor %xmm2, %xmm2 | 
|  | 3151 | ; SSE41-NEXT:    pmaxsd %xmm2, %xmm0 | 
|  | 3152 | ; SSE41-NEXT:    pmaxsd %xmm2, %xmm1 | 
|  | 3153 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 3154 | ; SSE41-NEXT:    pshufb %xmm2, %xmm1 | 
|  | 3155 | ; SSE41-NEXT:    pshufb %xmm2, %xmm0 | 
|  | 3156 | ; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] | 
|  | 3157 | ; SSE41-NEXT:    retq | 
|  | 3158 | ; | 
|  | 3159 | ; AVX1-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3160 | ; AVX1:       # %bb.0: | 
|  | 3161 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1 | 
|  | 3162 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3163 | ; AVX1-NEXT:    vpminsd %xmm2, %xmm1, %xmm1 | 
|  | 3164 | ; AVX1-NEXT:    vpminsd %xmm2, %xmm0, %xmm0 | 
|  | 3165 | ; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 3166 | ; AVX1-NEXT:    vpmaxsd %xmm2, %xmm0, %xmm0 | 
|  | 3167 | ; AVX1-NEXT:    vpmaxsd %xmm2, %xmm1, %xmm1 | 
|  | 3168 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 3169 | ; AVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1 | 
|  | 3170 | ; AVX1-NEXT:    vpshufb %xmm2, %xmm0, %xmm0 | 
|  | 3171 | ; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] | 
|  | 3172 | ; AVX1-NEXT:    vzeroupper | 
|  | 3173 | ; AVX1-NEXT:    retq | 
|  | 3174 | ; | 
|  | 3175 | ; AVX2-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3176 | ; AVX2:       # %bb.0: | 
|  | 3177 | ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] | 
|  | 3178 | ; AVX2-NEXT:    vpminsd %ymm1, %ymm0, %ymm0 | 
|  | 3179 | ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3180 | ; AVX2-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3181 | ; AVX2-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] | 
|  | 3182 | ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3183 | ; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3184 | ; AVX2-NEXT:    vzeroupper | 
|  | 3185 | ; AVX2-NEXT:    retq | 
|  | 3186 | ; | 
|  | 3187 | ; AVX512F-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3188 | ; AVX512F:       # %bb.0: | 
|  | 3189 | ; AVX512F-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] | 
|  | 3190 | ; AVX512F-NEXT:    vpminsd %ymm1, %ymm0, %ymm0 | 
|  | 3191 | ; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3192 | ; AVX512F-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3193 | ; AVX512F-NEXT:    vpmovdw %zmm0, %ymm0 | 
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3194 | ; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3195 | ; AVX512F-NEXT:    vzeroupper | 
|  | 3196 | ; AVX512F-NEXT:    retq | 
|  | 3197 | ; | 
|  | 3198 | ; AVX512VL-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3199 | ; AVX512VL:       # %bb.0: | 
|  | 3200 | ; AVX512VL-NEXT:    vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 | 
|  | 3201 | ; AVX512VL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3202 | ; AVX512VL-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3203 | ; AVX512VL-NEXT:    vpmovdw %ymm0, %xmm0 | 
|  | 3204 | ; AVX512VL-NEXT:    vzeroupper | 
|  | 3205 | ; AVX512VL-NEXT:    retq | 
|  | 3206 | ; | 
|  | 3207 | ; AVX512BW-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3208 | ; AVX512BW:       # %bb.0: | 
|  | 3209 | ; AVX512BW-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] | 
|  | 3210 | ; AVX512BW-NEXT:    vpminsd %ymm1, %ymm0, %ymm0 | 
|  | 3211 | ; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3212 | ; AVX512BW-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3213 | ; AVX512BW-NEXT:    vpmovdw %zmm0, %ymm0 | 
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3214 | ; AVX512BW-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3215 | ; AVX512BW-NEXT:    vzeroupper | 
|  | 3216 | ; AVX512BW-NEXT:    retq | 
|  | 3217 | ; | 
|  | 3218 | ; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i8: | 
|  | 3219 | ; AVX512BWVL:       # %bb.0: | 
|  | 3220 | ; AVX512BWVL-NEXT:    vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 | 
|  | 3221 | ; AVX512BWVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3222 | ; AVX512BWVL-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3223 | ; AVX512BWVL-NEXT:    vpmovdw %ymm0, %xmm0 | 
|  | 3224 | ; AVX512BWVL-NEXT:    vzeroupper | 
|  | 3225 | ; AVX512BWVL-NEXT:    retq | 
|  | 3226 | %1 = icmp slt <8 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> | 
|  | 3227 | %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> | 
|  | 3228 | %3 = icmp sgt <8 x i32> %2, zeroinitializer | 
|  | 3229 | %4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer | 
|  | 3230 | %5 = trunc <8 x i32> %4 to <8 x i8> | 
|  | 3231 | ret <8 x i8> %5 | 
|  | 3232 | } | 
|  | 3233 |  | 
| Simon Pilgrim | 689d813 | 2018-02-15 17:48:34 +0000 | [diff] [blame] | 3234 | define void @trunc_packus_v8i32_v8i8_store(<8 x i32> %a0, <8 x i8> *%p1) { | 
|  | 3235 | ; SSE2-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3236 | ; SSE2:       # %bb.0: | 
|  | 3237 | ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3238 | ; SSE2-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 3239 | ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm3 | 
|  | 3240 | ; SSE2-NEXT:    pand %xmm3, %xmm1 | 
|  | 3241 | ; SSE2-NEXT:    pandn %xmm2, %xmm3 | 
|  | 3242 | ; SSE2-NEXT:    por %xmm1, %xmm3 | 
|  | 3243 | ; SSE2-NEXT:    movdqa %xmm2, %xmm1 | 
|  | 3244 | ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 3245 | ; SSE2-NEXT:    pand %xmm1, %xmm0 | 
|  | 3246 | ; SSE2-NEXT:    pandn %xmm2, %xmm1 | 
|  | 3247 | ; SSE2-NEXT:    por %xmm0, %xmm1 | 
|  | 3248 | ; SSE2-NEXT:    pxor %xmm0, %xmm0 | 
|  | 3249 | ; SSE2-NEXT:    movdqa %xmm1, %xmm4 | 
|  | 3250 | ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4 | 
|  | 3251 | ; SSE2-NEXT:    movdqa %xmm3, %xmm5 | 
|  | 3252 | ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm5 | 
|  | 3253 | ; SSE2-NEXT:    pand %xmm2, %xmm5 | 
|  | 3254 | ; SSE2-NEXT:    pand %xmm3, %xmm5 | 
|  | 3255 | ; SSE2-NEXT:    pand %xmm2, %xmm4 | 
|  | 3256 | ; SSE2-NEXT:    pand %xmm1, %xmm4 | 
|  | 3257 | ; SSE2-NEXT:    packuswb %xmm5, %xmm4 | 
|  | 3258 | ; SSE2-NEXT:    packuswb %xmm4, %xmm4 | 
|  | 3259 | ; SSE2-NEXT:    movq %xmm4, (%rdi) | 
|  | 3260 | ; SSE2-NEXT:    retq | 
|  | 3261 | ; | 
|  | 3262 | ; SSSE3-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3263 | ; SSSE3:       # %bb.0: | 
|  | 3264 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3265 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm3 | 
|  | 3266 | ; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm3 | 
|  | 3267 | ; SSSE3-NEXT:    pand %xmm3, %xmm1 | 
|  | 3268 | ; SSSE3-NEXT:    pandn %xmm2, %xmm3 | 
|  | 3269 | ; SSSE3-NEXT:    por %xmm1, %xmm3 | 
|  | 3270 | ; SSSE3-NEXT:    movdqa %xmm2, %xmm1 | 
|  | 3271 | ; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 3272 | ; SSSE3-NEXT:    pand %xmm1, %xmm0 | 
|  | 3273 | ; SSSE3-NEXT:    pandn %xmm2, %xmm1 | 
|  | 3274 | ; SSSE3-NEXT:    por %xmm0, %xmm1 | 
|  | 3275 | ; SSSE3-NEXT:    pxor %xmm0, %xmm0 | 
|  | 3276 | ; SSSE3-NEXT:    movdqa %xmm1, %xmm2 | 
|  | 3277 | ; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2 | 
|  | 3278 | ; SSSE3-NEXT:    pand %xmm1, %xmm2 | 
|  | 3279 | ; SSSE3-NEXT:    movdqa %xmm3, %xmm1 | 
|  | 3280 | ; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm1 | 
|  | 3281 | ; SSSE3-NEXT:    pand %xmm3, %xmm1 | 
|  | 3282 | ; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 3283 | ; SSSE3-NEXT:    pshufb %xmm0, %xmm1 | 
|  | 3284 | ; SSSE3-NEXT:    pshufb %xmm0, %xmm2 | 
|  | 3285 | ; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm1[0] | 
|  | 3286 | ; SSSE3-NEXT:    pshufb {{.*#+}} xmm2 = xmm2[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 3287 | ; SSSE3-NEXT:    movq %xmm2, (%rdi) | 
|  | 3288 | ; SSSE3-NEXT:    retq | 
|  | 3289 | ; | 
|  | 3290 | ; SSE41-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3291 | ; SSE41:       # %bb.0: | 
|  | 3292 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3293 | ; SSE41-NEXT:    pminsd %xmm2, %xmm1 | 
|  | 3294 | ; SSE41-NEXT:    pminsd %xmm2, %xmm0 | 
|  | 3295 | ; SSE41-NEXT:    pxor %xmm2, %xmm2 | 
|  | 3296 | ; SSE41-NEXT:    pmaxsd %xmm2, %xmm0 | 
|  | 3297 | ; SSE41-NEXT:    pmaxsd %xmm2, %xmm1 | 
|  | 3298 | ; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] | 
|  | 3299 | ; SSE41-NEXT:    pshufb %xmm2, %xmm1 | 
|  | 3300 | ; SSE41-NEXT:    pshufb %xmm2, %xmm0 | 
|  | 3301 | ; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] | 
|  | 3302 | ; SSE41-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 3303 | ; SSE41-NEXT:    movq %xmm0, (%rdi) | 
|  | 3304 | ; SSE41-NEXT:    retq | 
|  | 3305 | ; | 
|  | 3306 | ; AVX1-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3307 | ; AVX1:       # %bb.0: | 
|  | 3308 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1 | 
|  | 3309 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [255,255,255,255] | 
|  | 3310 | ; AVX1-NEXT:    vpminsd %xmm2, %xmm1, %xmm1 | 
|  | 3311 | ; AVX1-NEXT:    vpminsd %xmm2, %xmm0, %xmm0 | 
|  | 3312 | ; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 3313 | ; AVX1-NEXT:    vpmaxsd %xmm2, %xmm0, %xmm0 | 
|  | 3314 | ; AVX1-NEXT:    vpmaxsd %xmm2, %xmm1, %xmm1 | 
|  | 3315 | ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> | 
|  | 3316 | ; AVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1 | 
|  | 3317 | ; AVX1-NEXT:    vpshufb %xmm2, %xmm0, %xmm0 | 
|  | 3318 | ; AVX1-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] | 
|  | 3319 | ; AVX1-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 3320 | ; AVX1-NEXT:    vzeroupper | 
|  | 3321 | ; AVX1-NEXT:    retq | 
|  | 3322 | ; | 
|  | 3323 | ; AVX2-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3324 | ; AVX2:       # %bb.0: | 
|  | 3325 | ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] | 
|  | 3326 | ; AVX2-NEXT:    vpminsd %ymm1, %ymm0, %ymm0 | 
|  | 3327 | ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3328 | ; AVX2-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3329 | ; AVX2-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] | 
|  | 3330 | ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] | 
|  | 3331 | ; AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 3332 | ; AVX2-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 3333 | ; AVX2-NEXT:    vzeroupper | 
|  | 3334 | ; AVX2-NEXT:    retq | 
|  | 3335 | ; | 
|  | 3336 | ; AVX512F-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3337 | ; AVX512F:       # %bb.0: | 
|  | 3338 | ; AVX512F-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] | 
|  | 3339 | ; AVX512F-NEXT:    vpminsd %ymm1, %ymm0, %ymm0 | 
|  | 3340 | ; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3341 | ; AVX512F-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3342 | ; AVX512F-NEXT:    vpmovdw %zmm0, %ymm0 | 
|  | 3343 | ; AVX512F-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 3344 | ; AVX512F-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 3345 | ; AVX512F-NEXT:    vzeroupper | 
|  | 3346 | ; AVX512F-NEXT:    retq | 
|  | 3347 | ; | 
|  | 3348 | ; AVX512VL-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3349 | ; AVX512VL:       # %bb.0: | 
|  | 3350 | ; AVX512VL-NEXT:    vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 | 
|  | 3351 | ; AVX512VL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3352 | ; AVX512VL-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3353 | ; AVX512VL-NEXT:    vpmovdb %ymm0, (%rdi) | 
|  | 3354 | ; AVX512VL-NEXT:    vzeroupper | 
|  | 3355 | ; AVX512VL-NEXT:    retq | 
|  | 3356 | ; | 
|  | 3357 | ; AVX512BW-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3358 | ; AVX512BW:       # %bb.0: | 
|  | 3359 | ; AVX512BW-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] | 
|  | 3360 | ; AVX512BW-NEXT:    vpminsd %ymm1, %ymm0, %ymm0 | 
|  | 3361 | ; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3362 | ; AVX512BW-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3363 | ; AVX512BW-NEXT:    vpmovdw %zmm0, %ymm0 | 
|  | 3364 | ; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] | 
|  | 3365 | ; AVX512BW-NEXT:    vmovq %xmm0, (%rdi) | 
|  | 3366 | ; AVX512BW-NEXT:    vzeroupper | 
|  | 3367 | ; AVX512BW-NEXT:    retq | 
|  | 3368 | ; | 
|  | 3369 | ; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i8_store: | 
|  | 3370 | ; AVX512BWVL:       # %bb.0: | 
|  | 3371 | ; AVX512BWVL-NEXT:    vpminsd {{.*}}(%rip){1to8}, %ymm0, %ymm0 | 
|  | 3372 | ; AVX512BWVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3373 | ; AVX512BWVL-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0 | 
|  | 3374 | ; AVX512BWVL-NEXT:    vpmovdb %ymm0, (%rdi) | 
|  | 3375 | ; AVX512BWVL-NEXT:    vzeroupper | 
|  | 3376 | ; AVX512BWVL-NEXT:    retq | 
|  | 3377 | %1 = icmp slt <8 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> | 
|  | 3378 | %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> | 
|  | 3379 | %3 = icmp sgt <8 x i32> %2, zeroinitializer | 
|  | 3380 | %4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer | 
|  | 3381 | %5 = trunc <8 x i32> %4 to <8 x i8> | 
|  | 3382 | store <8 x i8> %5, <8 x i8> *%p1 | 
|  | 3383 | ret void | 
|  | 3384 | } | 
|  | 3385 |  | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3386 | define <16 x i8> @trunc_packus_v16i32_v16i8(<16 x i32> %a0) { | 
| Simon Pilgrim | 17bb6f0 | 2018-02-15 14:37:59 +0000 | [diff] [blame] | 3387 | ; SSE-LABEL: trunc_packus_v16i32_v16i8: | 
|  | 3388 | ; SSE:       # %bb.0: | 
|  | 3389 | ; SSE-NEXT:    packssdw %xmm3, %xmm2 | 
|  | 3390 | ; SSE-NEXT:    packssdw %xmm1, %xmm0 | 
|  | 3391 | ; SSE-NEXT:    packuswb %xmm2, %xmm0 | 
|  | 3392 | ; SSE-NEXT:    retq | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3393 | ; | 
|  | 3394 | ; AVX1-LABEL: trunc_packus_v16i32_v16i8: | 
|  | 3395 | ; AVX1:       # %bb.0: | 
|  | 3396 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2 | 
| Simon Pilgrim | 17bb6f0 | 2018-02-15 14:37:59 +0000 | [diff] [blame] | 3397 | ; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1 | 
|  | 3398 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2 | 
|  | 3399 | ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3400 | ; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 | 
|  | 3401 | ; AVX1-NEXT:    vzeroupper | 
|  | 3402 | ; AVX1-NEXT:    retq | 
|  | 3403 | ; | 
|  | 3404 | ; AVX2-LABEL: trunc_packus_v16i32_v16i8: | 
|  | 3405 | ; AVX2:       # %bb.0: | 
| Simon Pilgrim | 17bb6f0 | 2018-02-15 14:37:59 +0000 | [diff] [blame] | 3406 | ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0 | 
|  | 3407 | ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3] | 
|  | 3408 | ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1 | 
|  | 3409 | ; AVX2-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3410 | ; AVX2-NEXT:    vzeroupper | 
|  | 3411 | ; AVX2-NEXT:    retq | 
|  | 3412 | ; | 
|  | 3413 | ; AVX512-LABEL: trunc_packus_v16i32_v16i8: | 
|  | 3414 | ; AVX512:       # %bb.0: | 
|  | 3415 | ; AVX512-NEXT:    vpminsd {{.*}}(%rip){1to16}, %zmm0, %zmm0 | 
|  | 3416 | ; AVX512-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3417 | ; AVX512-NEXT:    vpmaxsd %zmm1, %zmm0, %zmm0 | 
|  | 3418 | ; AVX512-NEXT:    vpmovdb %zmm0, %xmm0 | 
|  | 3419 | ; AVX512-NEXT:    vzeroupper | 
|  | 3420 | ; AVX512-NEXT:    retq | 
|  | 3421 | %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> | 
|  | 3422 | %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> | 
|  | 3423 | %3 = icmp sgt <16 x i32> %2, zeroinitializer | 
|  | 3424 | %4 = select <16 x i1> %3, <16 x i32> %2, <16 x i32> zeroinitializer | 
|  | 3425 | %5 = trunc <16 x i32> %4 to <16 x i8> | 
|  | 3426 | ret <16 x i8> %5 | 
|  | 3427 | } | 
|  | 3428 |  | 
|  | 3429 | define <16 x i8> @trunc_packus_v16i16_v16i8(<16 x i16> %a0) { | 
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 3430 | ; SSE-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3431 | ; SSE:       # %bb.0: | 
|  | 3432 | ; SSE-NEXT:    packuswb %xmm1, %xmm0 | 
|  | 3433 | ; SSE-NEXT:    retq | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3434 | ; | 
|  | 3435 | ; AVX1-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3436 | ; AVX1:       # %bb.0: | 
|  | 3437 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1 | 
| Simon Pilgrim | ae00a71 | 2018-02-06 14:07:46 +0000 | [diff] [blame] | 3438 | ; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3439 | ; AVX1-NEXT:    vzeroupper | 
|  | 3440 | ; AVX1-NEXT:    retq | 
|  | 3441 | ; | 
|  | 3442 | ; AVX2-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3443 | ; AVX2:       # %bb.0: | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3444 | ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1 | 
| Simon Pilgrim | ae00a71 | 2018-02-06 14:07:46 +0000 | [diff] [blame] | 3445 | ; AVX2-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3446 | ; AVX2-NEXT:    vzeroupper | 
|  | 3447 | ; AVX2-NEXT:    retq | 
|  | 3448 | ; | 
|  | 3449 | ; AVX512F-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3450 | ; AVX512F:       # %bb.0: | 
|  | 3451 | ; AVX512F-NEXT:    vpminsw {{.*}}(%rip), %ymm0, %ymm0 | 
|  | 3452 | ; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3453 | ; AVX512F-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0 | 
|  | 3454 | ; AVX512F-NEXT:    vpmovsxwd %ymm0, %zmm0 | 
|  | 3455 | ; AVX512F-NEXT:    vpmovdb %zmm0, %xmm0 | 
|  | 3456 | ; AVX512F-NEXT:    vzeroupper | 
|  | 3457 | ; AVX512F-NEXT:    retq | 
|  | 3458 | ; | 
|  | 3459 | ; AVX512VL-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3460 | ; AVX512VL:       # %bb.0: | 
|  | 3461 | ; AVX512VL-NEXT:    vpminsw {{.*}}(%rip), %ymm0, %ymm0 | 
|  | 3462 | ; AVX512VL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3463 | ; AVX512VL-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0 | 
|  | 3464 | ; AVX512VL-NEXT:    vpmovsxwd %ymm0, %zmm0 | 
|  | 3465 | ; AVX512VL-NEXT:    vpmovdb %zmm0, %xmm0 | 
|  | 3466 | ; AVX512VL-NEXT:    vzeroupper | 
|  | 3467 | ; AVX512VL-NEXT:    retq | 
|  | 3468 | ; | 
|  | 3469 | ; AVX512BW-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3470 | ; AVX512BW:       # %bb.0: | 
|  | 3471 | ; AVX512BW-NEXT:    vpminsw {{.*}}(%rip), %ymm0, %ymm0 | 
|  | 3472 | ; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3473 | ; AVX512BW-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0 | 
|  | 3474 | ; AVX512BW-NEXT:    vpmovwb %zmm0, %ymm0 | 
| Puyan Lotfi | 43e94b1 | 2018-01-31 22:04:26 +0000 | [diff] [blame] | 3475 | ; AVX512BW-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3476 | ; AVX512BW-NEXT:    vzeroupper | 
|  | 3477 | ; AVX512BW-NEXT:    retq | 
|  | 3478 | ; | 
|  | 3479 | ; AVX512BWVL-LABEL: trunc_packus_v16i16_v16i8: | 
|  | 3480 | ; AVX512BWVL:       # %bb.0: | 
|  | 3481 | ; AVX512BWVL-NEXT:    vpminsw {{.*}}(%rip), %ymm0, %ymm0 | 
|  | 3482 | ; AVX512BWVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3483 | ; AVX512BWVL-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0 | 
|  | 3484 | ; AVX512BWVL-NEXT:    vpmovwb %ymm0, %xmm0 | 
|  | 3485 | ; AVX512BWVL-NEXT:    vzeroupper | 
|  | 3486 | ; AVX512BWVL-NEXT:    retq | 
|  | 3487 | %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> | 
|  | 3488 | %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> | 
|  | 3489 | %3 = icmp sgt <16 x i16> %2, zeroinitializer | 
|  | 3490 | %4 = select <16 x i1> %3, <16 x i16> %2, <16 x i16> zeroinitializer | 
|  | 3491 | %5 = trunc <16 x i16> %4 to <16 x i8> | 
|  | 3492 | ret <16 x i8> %5 | 
|  | 3493 | } | 
|  | 3494 |  | 
|  | 3495 | define <32 x i8> @trunc_packus_v32i16_v32i8(<32 x i16> %a0) { | 
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 3496 | ; SSE-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3497 | ; SSE:       # %bb.0: | 
|  | 3498 | ; SSE-NEXT:    packuswb %xmm1, %xmm0 | 
|  | 3499 | ; SSE-NEXT:    packuswb %xmm3, %xmm2 | 
|  | 3500 | ; SSE-NEXT:    movdqa %xmm2, %xmm1 | 
|  | 3501 | ; SSE-NEXT:    retq | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3502 | ; | 
|  | 3503 | ; AVX1-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3504 | ; AVX1:       # %bb.0: | 
|  | 3505 | ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2 | 
| Simon Pilgrim | ae00a71 | 2018-02-06 14:07:46 +0000 | [diff] [blame] | 3506 | ; AVX1-NEXT:    vpackuswb %xmm2, %xmm1, %xmm1 | 
|  | 3507 | ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2 | 
|  | 3508 | ; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0 | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3509 | ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 3510 | ; AVX1-NEXT:    retq | 
|  | 3511 | ; | 
|  | 3512 | ; AVX2-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3513 | ; AVX2:       # %bb.0: | 
| Simon Pilgrim | 86d15bf | 2018-02-14 14:14:29 +0000 | [diff] [blame] | 3514 | ; AVX2-NEXT:    vpackuswb %ymm1, %ymm0, %ymm0 | 
|  | 3515 | ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3] | 
| Simon Pilgrim | 65ec923 | 2018-01-26 14:58:50 +0000 | [diff] [blame] | 3516 | ; AVX2-NEXT:    retq | 
|  | 3517 | ; | 
|  | 3518 | ; AVX512F-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3519 | ; AVX512F:       # %bb.0: | 
|  | 3520 | ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] | 
|  | 3521 | ; AVX512F-NEXT:    vpminsw %ymm2, %ymm0, %ymm0 | 
|  | 3522 | ; AVX512F-NEXT:    vpminsw %ymm2, %ymm1, %ymm1 | 
|  | 3523 | ; AVX512F-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 3524 | ; AVX512F-NEXT:    vpmaxsw %ymm2, %ymm1, %ymm1 | 
|  | 3525 | ; AVX512F-NEXT:    vpmaxsw %ymm2, %ymm0, %ymm0 | 
|  | 3526 | ; AVX512F-NEXT:    vpmovsxwd %ymm0, %zmm0 | 
|  | 3527 | ; AVX512F-NEXT:    vpmovdb %zmm0, %xmm0 | 
|  | 3528 | ; AVX512F-NEXT:    vpmovsxwd %ymm1, %zmm1 | 
|  | 3529 | ; AVX512F-NEXT:    vpmovdb %zmm1, %xmm1 | 
|  | 3530 | ; AVX512F-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 3531 | ; AVX512F-NEXT:    retq | 
|  | 3532 | ; | 
|  | 3533 | ; AVX512VL-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3534 | ; AVX512VL:       # %bb.0: | 
|  | 3535 | ; AVX512VL-NEXT:    vmovdqa {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] | 
|  | 3536 | ; AVX512VL-NEXT:    vpminsw %ymm2, %ymm0, %ymm0 | 
|  | 3537 | ; AVX512VL-NEXT:    vpminsw %ymm2, %ymm1, %ymm1 | 
|  | 3538 | ; AVX512VL-NEXT:    vpxor %xmm2, %xmm2, %xmm2 | 
|  | 3539 | ; AVX512VL-NEXT:    vpmaxsw %ymm2, %ymm1, %ymm1 | 
|  | 3540 | ; AVX512VL-NEXT:    vpmaxsw %ymm2, %ymm0, %ymm0 | 
|  | 3541 | ; AVX512VL-NEXT:    vpmovsxwd %ymm0, %zmm0 | 
|  | 3542 | ; AVX512VL-NEXT:    vpmovdb %zmm0, %xmm0 | 
|  | 3543 | ; AVX512VL-NEXT:    vpmovsxwd %ymm1, %zmm1 | 
|  | 3544 | ; AVX512VL-NEXT:    vpmovdb %zmm1, %xmm1 | 
|  | 3545 | ; AVX512VL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0 | 
|  | 3546 | ; AVX512VL-NEXT:    retq | 
|  | 3547 | ; | 
|  | 3548 | ; AVX512BW-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3549 | ; AVX512BW:       # %bb.0: | 
|  | 3550 | ; AVX512BW-NEXT:    vpminsw {{.*}}(%rip), %zmm0, %zmm0 | 
|  | 3551 | ; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3552 | ; AVX512BW-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0 | 
|  | 3553 | ; AVX512BW-NEXT:    vpmovwb %zmm0, %ymm0 | 
|  | 3554 | ; AVX512BW-NEXT:    retq | 
|  | 3555 | ; | 
|  | 3556 | ; AVX512BWVL-LABEL: trunc_packus_v32i16_v32i8: | 
|  | 3557 | ; AVX512BWVL:       # %bb.0: | 
|  | 3558 | ; AVX512BWVL-NEXT:    vpminsw {{.*}}(%rip), %zmm0, %zmm0 | 
|  | 3559 | ; AVX512BWVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1 | 
|  | 3560 | ; AVX512BWVL-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0 | 
|  | 3561 | ; AVX512BWVL-NEXT:    vpmovwb %zmm0, %ymm0 | 
|  | 3562 | ; AVX512BWVL-NEXT:    retq | 
|  | 3563 | %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> | 
|  | 3564 | %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> | 
|  | 3565 | %3 = icmp sgt <32 x i16> %2, zeroinitializer | 
|  | 3566 | %4 = select <32 x i1> %3, <32 x i16> %2, <32 x i16> zeroinitializer | 
|  | 3567 | %5 = trunc <32 x i16> %4 to <32 x i8> | 
|  | 3568 | ret <32 x i8> %5 | 
|  | 3569 | } |