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-apple-darwin -mcpu=knl -mattr=+avx512f | FileCheck %s --check-prefix=ALL --check-prefix=AVX512 --check-prefix=AVX512F
      3 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512 --check-prefix=AVX512BW
      4 
      5 define <8 x i64> @testv8i64(<8 x i64> %in) nounwind {
      6 ; AVX512F-LABEL: testv8i64:
      7 ; AVX512F:       ## BB#0:
      8 ; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
      9 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
     10 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
     11 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
     12 ; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
     13 ; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
     14 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
     15 ; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
     16 ; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
     17 ; AVX512F-NEXT:    vpxor %ymm3, %ymm3, %ymm3
     18 ; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
     19 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
     20 ; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
     21 ; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
     22 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
     23 ; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
     24 ; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
     25 ; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
     26 ; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
     27 ; AVX512F-NEXT:    retq
     28 ;
     29 ; AVX512BW-LABEL: testv8i64:
     30 ; AVX512BW:       ## BB#0:
     31 ; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
     32 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
     33 ; AVX512BW-NEXT:    vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
     34 ; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
     35 ; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
     36 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
     37 ; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
     38 ; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
     39 ; AVX512BW-NEXT:    vpxord %zmm1, %zmm1, %zmm1
     40 ; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
     41 ; AVX512BW-NEXT:    retq
     42   %out = call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %in)
     43   ret <8 x i64> %out
     44 }
     45 
     46 define <16 x i32> @testv16i32(<16 x i32> %in) nounwind {
     47 ; AVX512F-LABEL: testv16i32:
     48 ; AVX512F:       ## BB#0:
     49 ; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
     50 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
     51 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
     52 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
     53 ; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
     54 ; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
     55 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
     56 ; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
     57 ; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
     58 ; AVX512F-NEXT:    vpxor %ymm3, %ymm3, %ymm3
     59 ; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
     60 ; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
     61 ; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
     62 ; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
     63 ; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
     64 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
     65 ; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
     66 ; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
     67 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
     68 ; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
     69 ; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
     70 ; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
     71 ; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
     72 ; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
     73 ; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
     74 ; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
     75 ; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
     76 ; AVX512F-NEXT:    retq
     77 ;
     78 ; AVX512BW-LABEL: testv16i32:
     79 ; AVX512BW:       ## BB#0:
     80 ; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
     81 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
     82 ; AVX512BW-NEXT:    vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
     83 ; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
     84 ; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
     85 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
     86 ; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
     87 ; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
     88 ; AVX512BW-NEXT:    vpxord %zmm1, %zmm1, %zmm1
     89 ; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
     90 ; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
     91 ; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
     92 ; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
     93 ; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
     94 ; AVX512BW-NEXT:    retq
     95   %out = call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %in)
     96   ret <16 x i32> %out
     97 }
     98 
     99 define <32 x i16> @testv32i16(<32 x i16> %in) nounwind {
    100 ; AVX512F-LABEL: testv32i16:
    101 ; AVX512F:       ## BB#0:
    102 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
    103 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm3
    104 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
    105 ; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
    106 ; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
    107 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
    108 ; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
    109 ; AVX512F-NEXT:    vpaddb %ymm3, %ymm0, %ymm0
    110 ; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm3
    111 ; AVX512F-NEXT:    vpaddb %ymm0, %ymm3, %ymm0
    112 ; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
    113 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
    114 ; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
    115 ; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
    116 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
    117 ; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
    118 ; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
    119 ; AVX512F-NEXT:    vpsllw $8, %ymm1, %ymm2
    120 ; AVX512F-NEXT:    vpaddb %ymm1, %ymm2, %ymm1
    121 ; AVX512F-NEXT:    vpsrlw $8, %ymm1, %ymm1
    122 ; AVX512F-NEXT:    retq
    123 ;
    124 ; AVX512BW-LABEL: testv32i16:
    125 ; AVX512BW:       ## BB#0:
    126 ; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
    127 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
    128 ; AVX512BW-NEXT:    vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
    129 ; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
    130 ; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
    131 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
    132 ; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
    133 ; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
    134 ; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
    135 ; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
    136 ; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
    137 ; AVX512BW-NEXT:    retq
    138   %out = call <32 x i16> @llvm.ctpop.v32i16(<32 x i16> %in)
    139   ret <32 x i16> %out
    140 }
    141 
    142 define <64 x i8> @testv64i8(<64 x i8> %in) nounwind {
    143 ; AVX512F-LABEL: testv64i8:
    144 ; AVX512F:       ## BB#0:
    145 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
    146 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm3
    147 ; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
    148 ; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
    149 ; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
    150 ; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
    151 ; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
    152 ; AVX512F-NEXT:    vpaddb %ymm3, %ymm0, %ymm0
    153 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
    154 ; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
    155 ; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
    156 ; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
    157 ; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
    158 ; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
    159 ; AVX512F-NEXT:    retq
    160 ;
    161 ; AVX512BW-LABEL: testv64i8:
    162 ; AVX512BW:       ## BB#0:
    163 ; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
    164 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
    165 ; AVX512BW-NEXT:    vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
    166 ; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
    167 ; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
    168 ; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
    169 ; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
    170 ; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
    171 ; AVX512BW-NEXT:    retq
    172   %out = call <64 x i8> @llvm.ctpop.v64i8(<64 x i8> %in)
    173   ret <64 x i8> %out
    174 }
    175 
    176 declare <8 x i64> @llvm.ctpop.v8i64(<8 x i64>)
    177 declare <16 x i32> @llvm.ctpop.v16i32(<16 x i32>)
    178 declare <32 x i16> @llvm.ctpop.v32i16(<32 x i16>)
    179 declare <64 x i8> @llvm.ctpop.v64i8(<64 x i8>)
    180