Home | History | Annotate | Download | only in X86 
      1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
      2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE
      3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX12,AVX1
      4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX12,AVX2
      5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512F
      6 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512BW
      7 
      8 define i8 @v8i64(<8 x i64> %a, <8 x i64> %b, <8 x i64> %c, <8 x i64> %d) {
      9 ; SSE-LABEL: v8i64:
     10 ; SSE:       # %bb.0:
     11 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm8
     12 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm9
     13 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm10
     14 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm11
     15 ; SSE-NEXT:    pcmpgtq %xmm7, %xmm3
     16 ; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
     17 ; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[0,1,0,2,4,5,6,7]
     18 ; SSE-NEXT:    pcmpgtq %xmm6, %xmm2
     19 ; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
     20 ; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,0,2,4,5,6,7]
     21 ; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
     22 ; SSE-NEXT:    pcmpgtq %xmm5, %xmm1
     23 ; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
     24 ; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
     25 ; SSE-NEXT:    pcmpgtq %xmm4, %xmm0
     26 ; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
     27 ; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
     28 ; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
     29 ; SSE-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
     30 ; SSE-NEXT:    pcmpgtq {{[0-9]+}}(%rsp), %xmm11
     31 ; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm11[0,2,2,3]
     32 ; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7]
     33 ; SSE-NEXT:    pcmpgtq {{[0-9]+}}(%rsp), %xmm10
     34 ; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3]
     35 ; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,0,2,4,5,6,7]
     36 ; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
     37 ; SSE-NEXT:    pcmpgtq {{[0-9]+}}(%rsp), %xmm9
     38 ; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm9[0,2,2,3]
     39 ; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
     40 ; SSE-NEXT:    pcmpgtq {{[0-9]+}}(%rsp), %xmm8
     41 ; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm8[0,2,2,3]
     42 ; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
     43 ; SSE-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
     44 ; SSE-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0,1,2,3],xmm2[4,5,6,7]
     45 ; SSE-NEXT:    pand %xmm0, %xmm3
     46 ; SSE-NEXT:    packsswb %xmm0, %xmm3
     47 ; SSE-NEXT:    pmovmskb %xmm3, %eax
     48 ; SSE-NEXT:    # kill: def $al killed $al killed $eax
     49 ; SSE-NEXT:    retq
     50 ;
     51 ; AVX1-LABEL: v8i64:
     52 ; AVX1:       # %bb.0:
     53 ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm8
     54 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm9
     55 ; AVX1-NEXT:    vpcmpgtq %xmm8, %xmm9, %xmm8
     56 ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm1, %xmm1
     57 ; AVX1-NEXT:    vpackssdw %xmm8, %xmm1, %xmm8
     58 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
     59 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
     60 ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm1, %xmm1
     61 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm0
     62 ; AVX1-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
     63 ; AVX1-NEXT:    vpackssdw %xmm8, %xmm0, %xmm0
     64 ; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm1
     65 ; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm2
     66 ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm1
     67 ; AVX1-NEXT:    vpcmpgtq %xmm7, %xmm5, %xmm2
     68 ; AVX1-NEXT:    vpackssdw %xmm1, %xmm2, %xmm1
     69 ; AVX1-NEXT:    vextractf128 $1, %ymm6, %xmm2
     70 ; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm3
     71 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
     72 ; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm4, %xmm3
     73 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm3, %xmm2
     74 ; AVX1-NEXT:    vpackssdw %xmm1, %xmm2, %xmm1
     75 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
     76 ; AVX1-NEXT:    vpsllw $15, %xmm0, %xmm0
     77 ; AVX1-NEXT:    vpsraw $15, %xmm0, %xmm0
     78 ; AVX1-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
     79 ; AVX1-NEXT:    vpmovmskb %xmm0, %eax
     80 ; AVX1-NEXT:    # kill: def $al killed $al killed $eax
     81 ; AVX1-NEXT:    vzeroupper
     82 ; AVX1-NEXT:    retq
     83 ;
     84 ; AVX2-LABEL: v8i64:
     85 ; AVX2:       # %bb.0:
     86 ; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm1, %ymm1
     87 ; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
     88 ; AVX2-NEXT:    vpackssdw %xmm3, %xmm1, %xmm1
     89 ; AVX2-NEXT:    vpcmpgtq %ymm2, %ymm0, %ymm0
     90 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
     91 ; AVX2-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
     92 ; AVX2-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
     93 ; AVX2-NEXT:    vpcmpgtq %ymm7, %ymm5, %ymm1
     94 ; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
     95 ; AVX2-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
     96 ; AVX2-NEXT:    vpcmpgtq %ymm6, %ymm4, %ymm2
     97 ; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm3
     98 ; AVX2-NEXT:    vpackssdw %xmm3, %xmm2, %xmm2
     99 ; AVX2-NEXT:    vpackssdw %xmm1, %xmm2, %xmm1
    100 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
    101 ; AVX2-NEXT:    vpsllw $15, %xmm0, %xmm0
    102 ; AVX2-NEXT:    vpsraw $15, %xmm0, %xmm0
    103 ; AVX2-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
    104 ; AVX2-NEXT:    vpmovmskb %xmm0, %eax
    105 ; AVX2-NEXT:    # kill: def $al killed $al killed $eax
    106 ; AVX2-NEXT:    vzeroupper
    107 ; AVX2-NEXT:    retq
    108 ;
    109 ; AVX512F-LABEL: v8i64:
    110 ; AVX512F:       # %bb.0:
    111 ; AVX512F-NEXT:    vpcmpgtq %zmm1, %zmm0, %k1
    112 ; AVX512F-NEXT:    vpcmpgtq %zmm3, %zmm2, %k0 {%k1}
    113 ; AVX512F-NEXT:    kmovw %k0, %eax
    114 ; AVX512F-NEXT:    # kill: def $al killed $al killed $eax
    115 ; AVX512F-NEXT:    vzeroupper
    116 ; AVX512F-NEXT:    retq
    117 ;
    118 ; AVX512BW-LABEL: v8i64:
    119 ; AVX512BW:       # %bb.0:
    120 ; AVX512BW-NEXT:    vpcmpgtq %zmm1, %zmm0, %k1
    121 ; AVX512BW-NEXT:    vpcmpgtq %zmm3, %zmm2, %k0 {%k1}
    122 ; AVX512BW-NEXT:    kmovd %k0, %eax
    123 ; AVX512BW-NEXT:    # kill: def $al killed $al killed $eax
    124 ; AVX512BW-NEXT:    vzeroupper
    125 ; AVX512BW-NEXT:    retq
    126   %x0 = icmp sgt <8 x i64> %a, %b
    127   %x1 = icmp sgt <8 x i64> %c, %d
    128   %y = and <8 x i1> %x0, %x1
    129   %res = bitcast <8 x i1> %y to i8
    130   ret i8 %res
    131 }
    132 
    133 define i8 @v8f64(<8 x double> %a, <8 x double> %b, <8 x double> %c, <8 x double> %d) {
    134 ; SSE-LABEL: v8f64:
    135 ; SSE:       # %bb.0:
    136 ; SSE-NEXT:    movapd {{[0-9]+}}(%rsp), %xmm8
    137 ; SSE-NEXT:    movapd {{[0-9]+}}(%rsp), %xmm9
    138 ; SSE-NEXT:    movapd {{[0-9]+}}(%rsp), %xmm10
    139 ; SSE-NEXT:    movapd {{[0-9]+}}(%rsp), %xmm11
    140 ; SSE-NEXT:    cmpltpd %xmm3, %xmm7
    141 ; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[0,2,2,3]
    142 ; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm7[0,1,0,2,4,5,6,7]
    143 ; SSE-NEXT:    cmpltpd %xmm2, %xmm6
    144 ; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[0,2,2,3]
    145 ; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm6[0,1,0,2,4,5,6,7]
    146 ; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
    147 ; SSE-NEXT:    cmpltpd %xmm1, %xmm5
    148 ; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,2,2,3]
    149 ; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm5[0,2,2,3,4,5,6,7]
    150 ; SSE-NEXT:    cmpltpd %xmm0, %xmm4
    151 ; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,2,2,3]
    152 ; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm4[0,2,2,3,4,5,6,7]
    153 ; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
    154 ; SSE-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
    155 ; SSE-NEXT:    cmpltpd {{[0-9]+}}(%rsp), %xmm11
    156 ; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[0,2,2,3]
    157 ; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm11[0,1,0,2,4,5,6,7]
    158 ; SSE-NEXT:    cmpltpd {{[0-9]+}}(%rsp), %xmm10
    159 ; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,2,2,3]
    160 ; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm10[0,1,0,2,4,5,6,7]
    161 ; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
    162 ; SSE-NEXT:    cmpltpd {{[0-9]+}}(%rsp), %xmm9
    163 ; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,2,2,3]
    164 ; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm9[0,2,2,3,4,5,6,7]
    165 ; SSE-NEXT:    cmpltpd {{[0-9]+}}(%rsp), %xmm8
    166 ; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,2,2,3]
    167 ; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm8[0,2,2,3,4,5,6,7]
    168 ; SSE-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
    169 ; SSE-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0,1,2,3],xmm2[4,5,6,7]
    170 ; SSE-NEXT:    pand %xmm0, %xmm3
    171 ; SSE-NEXT:    packsswb %xmm0, %xmm3
    172 ; SSE-NEXT:    pmovmskb %xmm3, %eax
    173 ; SSE-NEXT:    # kill: def $al killed $al killed $eax
    174 ; SSE-NEXT:    retq
    175 ;
    176 ; AVX12-LABEL: v8f64:
    177 ; AVX12:       # %bb.0:
    178 ; AVX12-NEXT:    vcmpltpd %ymm1, %ymm3, %ymm1
    179 ; AVX12-NEXT:    vextractf128 $1, %ymm1, %xmm3
    180 ; AVX12-NEXT:    vpackssdw %xmm3, %xmm1, %xmm1
    181 ; AVX12-NEXT:    vcmpltpd %ymm0, %ymm2, %ymm0
    182 ; AVX12-NEXT:    vextractf128 $1, %ymm0, %xmm2
    183 ; AVX12-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
    184 ; AVX12-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
    185 ; AVX12-NEXT:    vcmpltpd %ymm5, %ymm7, %ymm1
    186 ; AVX12-NEXT:    vextractf128 $1, %ymm1, %xmm2
    187 ; AVX12-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
    188 ; AVX12-NEXT:    vcmpltpd %ymm4, %ymm6, %ymm2
    189 ; AVX12-NEXT:    vextractf128 $1, %ymm2, %xmm3
    190 ; AVX12-NEXT:    vpackssdw %xmm3, %xmm2, %xmm2
    191 ; AVX12-NEXT:    vpackssdw %xmm1, %xmm2, %xmm1
    192 ; AVX12-NEXT:    vpand %xmm1, %xmm0, %xmm0
    193 ; AVX12-NEXT:    vpsllw $15, %xmm0, %xmm0
    194 ; AVX12-NEXT:    vpsraw $15, %xmm0, %xmm0
    195 ; AVX12-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
    196 ; AVX12-NEXT:    vpmovmskb %xmm0, %eax
    197 ; AVX12-NEXT:    # kill: def $al killed $al killed $eax
    198 ; AVX12-NEXT:    vzeroupper
    199 ; AVX12-NEXT:    retq
    200 ;
    201 ; AVX512F-LABEL: v8f64:
    202 ; AVX512F:       # %bb.0:
    203 ; AVX512F-NEXT:    vcmpltpd %zmm0, %zmm1, %k1
    204 ; AVX512F-NEXT:    vcmpltpd %zmm2, %zmm3, %k0 {%k1}
    205 ; AVX512F-NEXT:    kmovw %k0, %eax
    206 ; AVX512F-NEXT:    # kill: def $al killed $al killed $eax
    207 ; AVX512F-NEXT:    vzeroupper
    208 ; AVX512F-NEXT:    retq
    209 ;
    210 ; AVX512BW-LABEL: v8f64:
    211 ; AVX512BW:       # %bb.0:
    212 ; AVX512BW-NEXT:    vcmpltpd %zmm0, %zmm1, %k1
    213 ; AVX512BW-NEXT:    vcmpltpd %zmm2, %zmm3, %k0 {%k1}
    214 ; AVX512BW-NEXT:    kmovd %k0, %eax
    215 ; AVX512BW-NEXT:    # kill: def $al killed $al killed $eax
    216 ; AVX512BW-NEXT:    vzeroupper
    217 ; AVX512BW-NEXT:    retq
    218   %x0 = fcmp ogt <8 x double> %a, %b
    219   %x1 = fcmp ogt <8 x double> %c, %d
    220   %y = and <8 x i1> %x0, %x1
    221   %res = bitcast <8 x i1> %y to i8
    222   ret i8 %res
    223 }
    224 
    225 define i32 @v32i16(<32 x i16> %a, <32 x i16> %b, <32 x i16> %c, <32 x i16> %d) {
    226 ; SSE-LABEL: v32i16:
    227 ; SSE:       # %bb.0:
    228 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm8
    229 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm9
    230 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm10
    231 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm11
    232 ; SSE-NEXT:    pcmpgtw %xmm5, %xmm1
    233 ; SSE-NEXT:    pcmpgtw %xmm4, %xmm0
    234 ; SSE-NEXT:    packsswb %xmm1, %xmm0
    235 ; SSE-NEXT:    pcmpgtw %xmm7, %xmm3
    236 ; SSE-NEXT:    pcmpgtw %xmm6, %xmm2
    237 ; SSE-NEXT:    packsswb %xmm3, %xmm2
    238 ; SSE-NEXT:    pcmpgtw {{[0-9]+}}(%rsp), %xmm11
    239 ; SSE-NEXT:    pcmpgtw {{[0-9]+}}(%rsp), %xmm10
    240 ; SSE-NEXT:    packsswb %xmm11, %xmm10
    241 ; SSE-NEXT:    pand %xmm0, %xmm10
    242 ; SSE-NEXT:    pcmpgtw {{[0-9]+}}(%rsp), %xmm9
    243 ; SSE-NEXT:    pcmpgtw {{[0-9]+}}(%rsp), %xmm8
    244 ; SSE-NEXT:    packsswb %xmm9, %xmm8
    245 ; SSE-NEXT:    pand %xmm2, %xmm8
    246 ; SSE-NEXT:    pmovmskb %xmm10, %ecx
    247 ; SSE-NEXT:    pmovmskb %xmm8, %eax
    248 ; SSE-NEXT:    shll $16, %eax
    249 ; SSE-NEXT:    orl %ecx, %eax
    250 ; SSE-NEXT:    retq
    251 ;
    252 ; AVX1-LABEL: v32i16:
    253 ; AVX1:       # %bb.0:
    254 ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm8
    255 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm9
    256 ; AVX1-NEXT:    vpcmpgtw %xmm8, %xmm9, %xmm8
    257 ; AVX1-NEXT:    vpcmpgtw %xmm3, %xmm1, %xmm1
    258 ; AVX1-NEXT:    vpacksswb %xmm8, %xmm1, %xmm8
    259 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
    260 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
    261 ; AVX1-NEXT:    vpcmpgtw %xmm3, %xmm1, %xmm1
    262 ; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm0, %xmm0
    263 ; AVX1-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
    264 ; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm1
    265 ; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm2
    266 ; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm2, %xmm1
    267 ; AVX1-NEXT:    vpcmpgtw %xmm7, %xmm5, %xmm2
    268 ; AVX1-NEXT:    vpacksswb %xmm1, %xmm2, %xmm1
    269 ; AVX1-NEXT:    vpand %xmm1, %xmm8, %xmm1
    270 ; AVX1-NEXT:    vextractf128 $1, %ymm6, %xmm2
    271 ; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm3
    272 ; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm3, %xmm2
    273 ; AVX1-NEXT:    vpcmpgtw %xmm6, %xmm4, %xmm3
    274 ; AVX1-NEXT:    vpacksswb %xmm2, %xmm3, %xmm2
    275 ; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
    276 ; AVX1-NEXT:    vpmovmskb %xmm0, %ecx
    277 ; AVX1-NEXT:    vpmovmskb %xmm1, %eax
    278 ; AVX1-NEXT:    shll $16, %eax
    279 ; AVX1-NEXT:    orl %ecx, %eax
    280 ; AVX1-NEXT:    vzeroupper
    281 ; AVX1-NEXT:    retq
    282 ;
    283 ; AVX2-LABEL: v32i16:
    284 ; AVX2:       # %bb.0:
    285 ; AVX2-NEXT:    vpcmpgtw %ymm3, %ymm1, %ymm1
    286 ; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
    287 ; AVX2-NEXT:    vpacksswb %xmm3, %xmm1, %xmm1
    288 ; AVX2-NEXT:    vpcmpgtw %ymm2, %ymm0, %ymm0
    289 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
    290 ; AVX2-NEXT:    vpacksswb %xmm2, %xmm0, %xmm0
    291 ; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
    292 ; AVX2-NEXT:    vpcmpgtw %ymm7, %ymm5, %ymm1
    293 ; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
    294 ; AVX2-NEXT:    vpacksswb %xmm2, %xmm1, %xmm1
    295 ; AVX2-NEXT:    vpcmpgtw %ymm6, %ymm4, %ymm2
    296 ; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm3
    297 ; AVX2-NEXT:    vpacksswb %xmm3, %xmm2, %xmm2
    298 ; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
    299 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
    300 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
    301 ; AVX2-NEXT:    vzeroupper
    302 ; AVX2-NEXT:    retq
    303 ;
    304 ; AVX512F-LABEL: v32i16:
    305 ; AVX512F:       # %bb.0:
    306 ; AVX512F-NEXT:    vpcmpgtw %ymm3, %ymm1, %ymm1
    307 ; AVX512F-NEXT:    vpmovsxwd %ymm1, %zmm1
    308 ; AVX512F-NEXT:    vptestmd %zmm1, %zmm1, %k1
    309 ; AVX512F-NEXT:    vpcmpgtw %ymm2, %ymm0, %ymm0
    310 ; AVX512F-NEXT:    vpmovsxwd %ymm0, %zmm0
    311 ; AVX512F-NEXT:    vptestmd %zmm0, %zmm0, %k2
    312 ; AVX512F-NEXT:    vpcmpgtw %ymm7, %ymm5, %ymm0
    313 ; AVX512F-NEXT:    vpmovsxwd %ymm0, %zmm0
    314 ; AVX512F-NEXT:    vpcmpgtw %ymm6, %ymm4, %ymm1
    315 ; AVX512F-NEXT:    vpmovsxwd %ymm1, %zmm1
    316 ; AVX512F-NEXT:    vptestmd %zmm1, %zmm1, %k0 {%k2}
    317 ; AVX512F-NEXT:    kmovw %k0, %ecx
    318 ; AVX512F-NEXT:    vptestmd %zmm0, %zmm0, %k0 {%k1}
    319 ; AVX512F-NEXT:    kmovw %k0, %eax
    320 ; AVX512F-NEXT:    shll $16, %eax
    321 ; AVX512F-NEXT:    orl %ecx, %eax
    322 ; AVX512F-NEXT:    vzeroupper
    323 ; AVX512F-NEXT:    retq
    324 ;
    325 ; AVX512BW-LABEL: v32i16:
    326 ; AVX512BW:       # %bb.0:
    327 ; AVX512BW-NEXT:    vpcmpgtw %zmm1, %zmm0, %k1
    328 ; AVX512BW-NEXT:    vpcmpgtw %zmm3, %zmm2, %k0 {%k1}
    329 ; AVX512BW-NEXT:    kmovd %k0, %eax
    330 ; AVX512BW-NEXT:    vzeroupper
    331 ; AVX512BW-NEXT:    retq
    332   %x0 = icmp sgt <32 x i16> %a, %b
    333   %x1 = icmp sgt <32 x i16> %c, %d
    334   %y = and <32 x i1> %x0, %x1
    335   %res = bitcast <32 x i1> %y to i32
    336   ret i32 %res
    337 }
    338 
    339 define i16 @v16i32(<16 x i32> %a, <16 x i32> %b, <16 x i32> %c, <16 x i32> %d) {
    340 ; SSE-LABEL: v16i32:
    341 ; SSE:       # %bb.0:
    342 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm8
    343 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm10
    344 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm9
    345 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm11
    346 ; SSE-NEXT:    pcmpgtd %xmm7, %xmm3
    347 ; SSE-NEXT:    movdqa {{.*#+}} xmm7 = <u,u,u,u,0,4,8,12,u,u,u,u,u,u,u,u>
    348 ; SSE-NEXT:    pshufb %xmm7, %xmm3
    349 ; SSE-NEXT:    pcmpgtd %xmm6, %xmm2
    350 ; SSE-NEXT:    pshufb %xmm7, %xmm2
    351 ; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
    352 ; SSE-NEXT:    pcmpgtd %xmm5, %xmm1
    353 ; SSE-NEXT:    movdqa {{.*#+}} xmm3 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u>
    354 ; SSE-NEXT:    pshufb %xmm3, %xmm1
    355 ; SSE-NEXT:    pcmpgtd %xmm4, %xmm0
    356 ; SSE-NEXT:    pshufb %xmm3, %xmm0
    357 ; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
    358 ; SSE-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
    359 ; SSE-NEXT:    pcmpgtd {{[0-9]+}}(%rsp), %xmm11
    360 ; SSE-NEXT:    pshufb %xmm7, %xmm11
    361 ; SSE-NEXT:    pcmpgtd {{[0-9]+}}(%rsp), %xmm9
    362 ; SSE-NEXT:    pshufb %xmm7, %xmm9
    363 ; SSE-NEXT:    punpckldq {{.*#+}} xmm9 = xmm9[0],xmm11[0],xmm9[1],xmm11[1]
    364 ; SSE-NEXT:    pcmpgtd {{[0-9]+}}(%rsp), %xmm10
    365 ; SSE-NEXT:    pshufb %xmm3, %xmm10
    366 ; SSE-NEXT:    pcmpgtd {{[0-9]+}}(%rsp), %xmm8
    367 ; SSE-NEXT:    pshufb %xmm3, %xmm8
    368 ; SSE-NEXT:    punpckldq {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1]
    369 ; SSE-NEXT:    pblendw {{.*#+}} xmm8 = xmm8[0,1,2,3],xmm9[4,5,6,7]
    370 ; SSE-NEXT:    pand %xmm0, %xmm8
    371 ; SSE-NEXT:    pmovmskb %xmm8, %eax
    372 ; SSE-NEXT:    # kill: def $ax killed $ax killed $eax
    373 ; SSE-NEXT:    retq
    374 ;
    375 ; AVX1-LABEL: v16i32:
    376 ; AVX1:       # %bb.0:
    377 ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm8
    378 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm9
    379 ; AVX1-NEXT:    vpcmpgtd %xmm8, %xmm9, %xmm8
    380 ; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm1, %xmm1
    381 ; AVX1-NEXT:    vpackssdw %xmm8, %xmm1, %xmm8
    382 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
    383 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
    384 ; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm1, %xmm1
    385 ; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm0, %xmm0
    386 ; AVX1-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
    387 ; AVX1-NEXT:    vpacksswb %xmm8, %xmm0, %xmm0
    388 ; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm1
    389 ; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm2
    390 ; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm1
    391 ; AVX1-NEXT:    vpcmpgtd %xmm7, %xmm5, %xmm2
    392 ; AVX1-NEXT:    vpackssdw %xmm1, %xmm2, %xmm1
    393 ; AVX1-NEXT:    vextractf128 $1, %ymm6, %xmm2
    394 ; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm3
    395 ; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm2
    396 ; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm4, %xmm3
    397 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm3, %xmm2
    398 ; AVX1-NEXT:    vpacksswb %xmm1, %xmm2, %xmm1
    399 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
    400 ; AVX1-NEXT:    vpmovmskb %xmm0, %eax
    401 ; AVX1-NEXT:    # kill: def $ax killed $ax killed $eax
    402 ; AVX1-NEXT:    vzeroupper
    403 ; AVX1-NEXT:    retq
    404 ;
    405 ; AVX2-LABEL: v16i32:
    406 ; AVX2:       # %bb.0:
    407 ; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm1, %ymm1
    408 ; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
    409 ; AVX2-NEXT:    vpackssdw %xmm3, %xmm1, %xmm1
    410 ; AVX2-NEXT:    vpcmpgtd %ymm2, %ymm0, %ymm0
    411 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
    412 ; AVX2-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
    413 ; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
    414 ; AVX2-NEXT:    vpcmpgtd %ymm7, %ymm5, %ymm1
    415 ; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
    416 ; AVX2-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
    417 ; AVX2-NEXT:    vpcmpgtd %ymm6, %ymm4, %ymm2
    418 ; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm3
    419 ; AVX2-NEXT:    vpackssdw %xmm3, %xmm2, %xmm2
    420 ; AVX2-NEXT:    vpacksswb %xmm1, %xmm2, %xmm1
    421 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
    422 ; AVX2-NEXT:    vpmovmskb %xmm0, %eax
    423 ; AVX2-NEXT:    # kill: def $ax killed $ax killed $eax
    424 ; AVX2-NEXT:    vzeroupper
    425 ; AVX2-NEXT:    retq
    426 ;
    427 ; AVX512F-LABEL: v16i32:
    428 ; AVX512F:       # %bb.0:
    429 ; AVX512F-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1
    430 ; AVX512F-NEXT:    vpcmpgtd %zmm3, %zmm2, %k0 {%k1}
    431 ; AVX512F-NEXT:    kmovw %k0, %eax
    432 ; AVX512F-NEXT:    # kill: def $ax killed $ax killed $eax
    433 ; AVX512F-NEXT:    vzeroupper
    434 ; AVX512F-NEXT:    retq
    435 ;
    436 ; AVX512BW-LABEL: v16i32:
    437 ; AVX512BW:       # %bb.0:
    438 ; AVX512BW-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1
    439 ; AVX512BW-NEXT:    vpcmpgtd %zmm3, %zmm2, %k0 {%k1}
    440 ; AVX512BW-NEXT:    kmovd %k0, %eax
    441 ; AVX512BW-NEXT:    # kill: def $ax killed $ax killed $eax
    442 ; AVX512BW-NEXT:    vzeroupper
    443 ; AVX512BW-NEXT:    retq
    444   %x0 = icmp sgt <16 x i32> %a, %b
    445   %x1 = icmp sgt <16 x i32> %c, %d
    446   %y = and <16 x i1> %x0, %x1
    447   %res = bitcast <16 x i1> %y to i16
    448   ret i16 %res
    449 }
    450 
    451 define i16 @v16f32(<16 x float> %a, <16 x float> %b, <16 x float> %c, <16 x float> %d) {
    452 ; SSE-LABEL: v16f32:
    453 ; SSE:       # %bb.0:
    454 ; SSE-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm8
    455 ; SSE-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm10
    456 ; SSE-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm9
    457 ; SSE-NEXT:    movaps {{[0-9]+}}(%rsp), %xmm11
    458 ; SSE-NEXT:    cmpltps %xmm3, %xmm7
    459 ; SSE-NEXT:    movdqa {{.*#+}} xmm3 = <u,u,u,u,0,4,8,12,u,u,u,u,u,u,u,u>
    460 ; SSE-NEXT:    pshufb %xmm3, %xmm7
    461 ; SSE-NEXT:    cmpltps %xmm2, %xmm6
    462 ; SSE-NEXT:    pshufb %xmm3, %xmm6
    463 ; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]
    464 ; SSE-NEXT:    cmpltps %xmm1, %xmm5
    465 ; SSE-NEXT:    movdqa {{.*#+}} xmm1 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u>
    466 ; SSE-NEXT:    pshufb %xmm1, %xmm5
    467 ; SSE-NEXT:    cmpltps %xmm0, %xmm4
    468 ; SSE-NEXT:    pshufb %xmm1, %xmm4
    469 ; SSE-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]
    470 ; SSE-NEXT:    pblendw {{.*#+}} xmm4 = xmm4[0,1,2,3],xmm6[4,5,6,7]
    471 ; SSE-NEXT:    cmpltps {{[0-9]+}}(%rsp), %xmm11
    472 ; SSE-NEXT:    pshufb %xmm3, %xmm11
    473 ; SSE-NEXT:    cmpltps {{[0-9]+}}(%rsp), %xmm9
    474 ; SSE-NEXT:    pshufb %xmm3, %xmm9
    475 ; SSE-NEXT:    punpckldq {{.*#+}} xmm9 = xmm9[0],xmm11[0],xmm9[1],xmm11[1]
    476 ; SSE-NEXT:    cmpltps {{[0-9]+}}(%rsp), %xmm10
    477 ; SSE-NEXT:    pshufb %xmm1, %xmm10
    478 ; SSE-NEXT:    cmpltps {{[0-9]+}}(%rsp), %xmm8
    479 ; SSE-NEXT:    pshufb %xmm1, %xmm8
    480 ; SSE-NEXT:    punpckldq {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1]
    481 ; SSE-NEXT:    pblendw {{.*#+}} xmm8 = xmm8[0,1,2,3],xmm9[4,5,6,7]
    482 ; SSE-NEXT:    pand %xmm4, %xmm8
    483 ; SSE-NEXT:    pmovmskb %xmm8, %eax
    484 ; SSE-NEXT:    # kill: def $ax killed $ax killed $eax
    485 ; SSE-NEXT:    retq
    486 ;
    487 ; AVX12-LABEL: v16f32:
    488 ; AVX12:       # %bb.0:
    489 ; AVX12-NEXT:    vcmpltps %ymm1, %ymm3, %ymm1
    490 ; AVX12-NEXT:    vextractf128 $1, %ymm1, %xmm3
    491 ; AVX12-NEXT:    vpackssdw %xmm3, %xmm1, %xmm1
    492 ; AVX12-NEXT:    vcmpltps %ymm0, %ymm2, %ymm0
    493 ; AVX12-NEXT:    vextractf128 $1, %ymm0, %xmm2
    494 ; AVX12-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
    495 ; AVX12-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
    496 ; AVX12-NEXT:    vcmpltps %ymm5, %ymm7, %ymm1
    497 ; AVX12-NEXT:    vextractf128 $1, %ymm1, %xmm2
    498 ; AVX12-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
    499 ; AVX12-NEXT:    vcmpltps %ymm4, %ymm6, %ymm2
    500 ; AVX12-NEXT:    vextractf128 $1, %ymm2, %xmm3
    501 ; AVX12-NEXT:    vpackssdw %xmm3, %xmm2, %xmm2
    502 ; AVX12-NEXT:    vpacksswb %xmm1, %xmm2, %xmm1
    503 ; AVX12-NEXT:    vpand %xmm1, %xmm0, %xmm0
    504 ; AVX12-NEXT:    vpmovmskb %xmm0, %eax
    505 ; AVX12-NEXT:    # kill: def $ax killed $ax killed $eax
    506 ; AVX12-NEXT:    vzeroupper
    507 ; AVX12-NEXT:    retq
    508 ;
    509 ; AVX512F-LABEL: v16f32:
    510 ; AVX512F:       # %bb.0:
    511 ; AVX512F-NEXT:    vcmpltps %zmm0, %zmm1, %k1
    512 ; AVX512F-NEXT:    vcmpltps %zmm2, %zmm3, %k0 {%k1}
    513 ; AVX512F-NEXT:    kmovw %k0, %eax
    514 ; AVX512F-NEXT:    # kill: def $ax killed $ax killed $eax
    515 ; AVX512F-NEXT:    vzeroupper
    516 ; AVX512F-NEXT:    retq
    517 ;
    518 ; AVX512BW-LABEL: v16f32:
    519 ; AVX512BW:       # %bb.0:
    520 ; AVX512BW-NEXT:    vcmpltps %zmm0, %zmm1, %k1
    521 ; AVX512BW-NEXT:    vcmpltps %zmm2, %zmm3, %k0 {%k1}
    522 ; AVX512BW-NEXT:    kmovd %k0, %eax
    523 ; AVX512BW-NEXT:    # kill: def $ax killed $ax killed $eax
    524 ; AVX512BW-NEXT:    vzeroupper
    525 ; AVX512BW-NEXT:    retq
    526   %x0 = fcmp ogt <16 x float> %a, %b
    527   %x1 = fcmp ogt <16 x float> %c, %d
    528   %y = and <16 x i1> %x0, %x1
    529   %res = bitcast <16 x i1> %y to i16
    530   ret i16 %res
    531 }
    532 
    533 define i64 @v64i8(<64 x i8> %a, <64 x i8> %b, <64 x i8> %c, <64 x i8> %d) {
    534 ; SSE-LABEL: v64i8:
    535 ; SSE:       # %bb.0:
    536 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm10
    537 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm11
    538 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm8
    539 ; SSE-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm9
    540 ; SSE-NEXT:    pcmpgtb %xmm7, %xmm3
    541 ; SSE-NEXT:    pcmpgtb %xmm6, %xmm2
    542 ; SSE-NEXT:    pcmpgtb %xmm5, %xmm1
    543 ; SSE-NEXT:    pcmpgtb %xmm4, %xmm0
    544 ; SSE-NEXT:    pcmpgtb {{[0-9]+}}(%rsp), %xmm9
    545 ; SSE-NEXT:    pand %xmm3, %xmm9
    546 ; SSE-NEXT:    pcmpgtb {{[0-9]+}}(%rsp), %xmm8
    547 ; SSE-NEXT:    pand %xmm2, %xmm8
    548 ; SSE-NEXT:    pcmpgtb {{[0-9]+}}(%rsp), %xmm11
    549 ; SSE-NEXT:    pand %xmm1, %xmm11
    550 ; SSE-NEXT:    pcmpgtb {{[0-9]+}}(%rsp), %xmm10
    551 ; SSE-NEXT:    pand %xmm0, %xmm10
    552 ; SSE-NEXT:    pmovmskb %xmm10, %eax
    553 ; SSE-NEXT:    pmovmskb %xmm11, %ecx
    554 ; SSE-NEXT:    shll $16, %ecx
    555 ; SSE-NEXT:    orl %eax, %ecx
    556 ; SSE-NEXT:    pmovmskb %xmm8, %edx
    557 ; SSE-NEXT:    pmovmskb %xmm9, %eax
    558 ; SSE-NEXT:    shll $16, %eax
    559 ; SSE-NEXT:    orl %edx, %eax
    560 ; SSE-NEXT:    shlq $32, %rax
    561 ; SSE-NEXT:    orq %rcx, %rax
    562 ; SSE-NEXT:    retq
    563 ;
    564 ; AVX1-LABEL: v64i8:
    565 ; AVX1:       # %bb.0:
    566 ; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm8
    567 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm9
    568 ; AVX1-NEXT:    vpcmpgtb %xmm8, %xmm9, %xmm8
    569 ; AVX1-NEXT:    vpcmpgtb %xmm3, %xmm1, %xmm9
    570 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
    571 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
    572 ; AVX1-NEXT:    vpcmpgtb %xmm3, %xmm1, %xmm1
    573 ; AVX1-NEXT:    vpcmpgtb %xmm2, %xmm0, %xmm0
    574 ; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm2
    575 ; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm3
    576 ; AVX1-NEXT:    vpcmpgtb %xmm2, %xmm3, %xmm2
    577 ; AVX1-NEXT:    vpand %xmm2, %xmm8, %xmm2
    578 ; AVX1-NEXT:    vpcmpgtb %xmm7, %xmm5, %xmm3
    579 ; AVX1-NEXT:    vpand %xmm3, %xmm9, %xmm3
    580 ; AVX1-NEXT:    vextractf128 $1, %ymm6, %xmm5
    581 ; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm7
    582 ; AVX1-NEXT:    vpcmpgtb %xmm5, %xmm7, %xmm5
    583 ; AVX1-NEXT:    vpand %xmm5, %xmm1, %xmm1
    584 ; AVX1-NEXT:    vpcmpgtb %xmm6, %xmm4, %xmm4
    585 ; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
    586 ; AVX1-NEXT:    vpmovmskb %xmm0, %eax
    587 ; AVX1-NEXT:    vpmovmskb %xmm1, %ecx
    588 ; AVX1-NEXT:    shll $16, %ecx
    589 ; AVX1-NEXT:    orl %eax, %ecx
    590 ; AVX1-NEXT:    vpmovmskb %xmm3, %edx
    591 ; AVX1-NEXT:    vpmovmskb %xmm2, %eax
    592 ; AVX1-NEXT:    shll $16, %eax
    593 ; AVX1-NEXT:    orl %edx, %eax
    594 ; AVX1-NEXT:    shlq $32, %rax
    595 ; AVX1-NEXT:    orq %rcx, %rax
    596 ; AVX1-NEXT:    vzeroupper
    597 ; AVX1-NEXT:    retq
    598 ;
    599 ; AVX2-LABEL: v64i8:
    600 ; AVX2:       # %bb.0:
    601 ; AVX2-NEXT:    vpcmpgtb %ymm3, %ymm1, %ymm1
    602 ; AVX2-NEXT:    vpcmpgtb %ymm2, %ymm0, %ymm0
    603 ; AVX2-NEXT:    vpcmpgtb %ymm7, %ymm5, %ymm2
    604 ; AVX2-NEXT:    vpand %ymm2, %ymm1, %ymm1
    605 ; AVX2-NEXT:    vpcmpgtb %ymm6, %ymm4, %ymm2
    606 ; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
    607 ; AVX2-NEXT:    vpmovmskb %ymm0, %ecx
    608 ; AVX2-NEXT:    vpmovmskb %ymm1, %eax
    609 ; AVX2-NEXT:    shlq $32, %rax
    610 ; AVX2-NEXT:    orq %rcx, %rax
    611 ; AVX2-NEXT:    vzeroupper
    612 ; AVX2-NEXT:    retq
    613 ;
    614 ; AVX512F-LABEL: v64i8:
    615 ; AVX512F:       # %bb.0:
    616 ; AVX512F-NEXT:    vpcmpgtb %ymm3, %ymm1, %ymm1
    617 ; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm3
    618 ; AVX512F-NEXT:    vpmovsxbd %xmm3, %zmm3
    619 ; AVX512F-NEXT:    vptestmd %zmm3, %zmm3, %k1
    620 ; AVX512F-NEXT:    vpmovsxbd %xmm1, %zmm1
    621 ; AVX512F-NEXT:    vptestmd %zmm1, %zmm1, %k2
    622 ; AVX512F-NEXT:    vpcmpgtb %ymm2, %ymm0, %ymm0
    623 ; AVX512F-NEXT:    vextracti128 $1, %ymm0, %xmm1
    624 ; AVX512F-NEXT:    vpmovsxbd %xmm1, %zmm1
    625 ; AVX512F-NEXT:    vptestmd %zmm1, %zmm1, %k3
    626 ; AVX512F-NEXT:    vpmovsxbd %xmm0, %zmm0
    627 ; AVX512F-NEXT:    vptestmd %zmm0, %zmm0, %k4
    628 ; AVX512F-NEXT:    vpcmpgtb %ymm7, %ymm5, %ymm0
    629 ; AVX512F-NEXT:    vextracti128 $1, %ymm0, %xmm1
    630 ; AVX512F-NEXT:    vpmovsxbd %xmm1, %zmm1
    631 ; AVX512F-NEXT:    vpmovsxbd %xmm0, %zmm0
    632 ; AVX512F-NEXT:    vpcmpgtb %ymm6, %ymm4, %ymm2
    633 ; AVX512F-NEXT:    vextracti128 $1, %ymm2, %xmm3
    634 ; AVX512F-NEXT:    vpmovsxbd %xmm3, %zmm3
    635 ; AVX512F-NEXT:    vpmovsxbd %xmm2, %zmm2
    636 ; AVX512F-NEXT:    vptestmd %zmm2, %zmm2, %k0 {%k4}
    637 ; AVX512F-NEXT:    kmovw %k0, %eax
    638 ; AVX512F-NEXT:    vptestmd %zmm3, %zmm3, %k0 {%k3}
    639 ; AVX512F-NEXT:    kmovw %k0, %ecx
    640 ; AVX512F-NEXT:    shll $16, %ecx
    641 ; AVX512F-NEXT:    orl %eax, %ecx
    642 ; AVX512F-NEXT:    vptestmd %zmm0, %zmm0, %k0 {%k2}
    643 ; AVX512F-NEXT:    kmovw %k0, %edx
    644 ; AVX512F-NEXT:    vptestmd %zmm1, %zmm1, %k0 {%k1}
    645 ; AVX512F-NEXT:    kmovw %k0, %eax
    646 ; AVX512F-NEXT:    shll $16, %eax
    647 ; AVX512F-NEXT:    orl %edx, %eax
    648 ; AVX512F-NEXT:    shlq $32, %rax
    649 ; AVX512F-NEXT:    orq %rcx, %rax
    650 ; AVX512F-NEXT:    vzeroupper
    651 ; AVX512F-NEXT:    retq
    652 ;
    653 ; AVX512BW-LABEL: v64i8:
    654 ; AVX512BW:       # %bb.0:
    655 ; AVX512BW-NEXT:    vpcmpgtb %zmm1, %zmm0, %k1
    656 ; AVX512BW-NEXT:    vpcmpgtb %zmm3, %zmm2, %k0 {%k1}
    657 ; AVX512BW-NEXT:    kmovq %k0, %rax
    658 ; AVX512BW-NEXT:    vzeroupper
    659 ; AVX512BW-NEXT:    retq
    660   %x0 = icmp sgt <64 x i8> %a, %b
    661   %x1 = icmp sgt <64 x i8> %c, %d
    662   %y = and <64 x i1> %x0, %x1
    663   %res = bitcast <64 x i1> %y to i64
    664   ret i64 %res
    665 }
    666