.section    __TEXT,__literal4,4byte_literals
    .align  2
LCPI0_0:
    .long   1431655765              ## 0x55555555
LCPI0_1:
    .long   858993459               ## 0x33333333
LCPI0_2:
    .long   252645135               ## 0xf0f0f0f
LCPI0_3:
    .long   63                      ## 0x3f
    .section    __TEXT,__text,regular,pure_instructions
    .globl  v8i32-new-avx2
    .align  4, 0x90
v8i32-new-avx2:                                     ## @A
    .cfi_startproc
## BB#0:                                ## %entry
    vpsrld  $1, %ymm0, %ymm1
    vpbroadcastd    LCPI0_0(%rip), %ymm2
    vpand   %ymm2, %ymm1, %ymm1
    vpsubd  %ymm1, %ymm0, %ymm0
    vpbroadcastd    LCPI0_1(%rip), %ymm1
    vpand   %ymm1, %ymm0, %ymm2
    vpsrld  $2, %ymm0, %ymm0
    vpand   %ymm1, %ymm0, %ymm0
    vpaddd  %ymm0, %ymm2, %ymm0
    vpsrld  $4, %ymm0, %ymm1
    vpaddd  %ymm1, %ymm0, %ymm0
    vpbroadcastd    LCPI0_2(%rip), %ymm1
    vpand   %ymm1, %ymm0, %ymm0
    vpsrld  $8, %ymm0, %ymm1
    vpaddd  %ymm1, %ymm0, %ymm0
    vpsrld  $16, %ymm0, %ymm1
    vpaddd  %ymm1, %ymm0, %ymm0
    vpbroadcastd    LCPI0_3(%rip), %ymm1
    vpand   %ymm1, %ymm0, %ymm0
    retq
    .cfi_endproc

    .globl  v8i32-old-avx2
    .align  4, 0x90
v8i32-old-avx2:                                     ## @A
    .cfi_startproc
## BB#0:                                ## %entry
    vextracti128    $1, %ymm0, %xmm1
    vpextrd $1, %xmm1, %eax
    popcntl %eax, %eax
    vmovd   %xmm1, %ecx
    popcntl %ecx, %ecx
    vmovd   %ecx, %xmm2
    vpinsrd $1, %eax, %xmm2, %xmm2
    vpextrd $2, %xmm1, %eax
    popcntl %eax, %eax
    vpinsrd $2, %eax, %xmm2, %xmm2
    vpextrd $3, %xmm1, %eax
    popcntl %eax, %eax
    vpinsrd $3, %eax, %xmm2, %xmm1
    vpextrd $1, %xmm0, %eax
    popcntl %eax, %eax
    vmovd   %xmm0, %ecx
    popcntl %ecx, %ecx
    vmovd   %ecx, %xmm2
    vpinsrd $1, %eax, %xmm2, %xmm2
    vpextrd $2, %xmm0, %eax
    popcntl %eax, %eax
    vpinsrd $2, %eax, %xmm2, %xmm2
    vpextrd $3, %xmm0, %eax
    popcntl %eax, %eax
    vpinsrd $3, %eax, %xmm2, %xmm0
    vinserti128 $1, %xmm1, %ymm0, %ymm0
    retq
    .cfi_endproc

    .align  4
LCPI0_0:
    .space  16,85
LCPI0_1:
    .space  16,51
LCPI0_2:
    .space  16,15
LCPI0_3:
    .long   63                      ## 0x3f
    .long   63                      ## 0x3f
    .long   63                      ## 0x3f
    .long   63                      ## 0x3f
    .section    __TEXT,__text,regular,pure_instructions
    .globl  v8i32-new-avx1
    .align  4, 0x90
v8i32-new-avx1:                                     ## @A
    .cfi_startproc
## BB#0:                                ## %entry
    vextractf128    $1, %ymm0, %xmm1
    vpsrld  $1, %xmm1, %xmm2
    vmovdqa LCPI0_0(%rip), %xmm3    ## xmm3 = [1431655765,1431655765,1431655765,1431655765]
    vpand   %xmm3, %xmm2, %xmm2
    vpsubd  %xmm2, %xmm1, %xmm1
    vmovdqa LCPI0_1(%rip), %xmm2    ## xmm2 = [858993459,858993459,858993459,858993459]
    vpand   %xmm2, %xmm1, %xmm4
    vpsrld  $2, %xmm1, %xmm1
    vpand   %xmm2, %xmm1, %xmm1
    vpaddd  %xmm1, %xmm4, %xmm1
    vpsrld  $4, %xmm1, %xmm4
    vpaddd  %xmm4, %xmm1, %xmm1
    vmovdqa LCPI0_2(%rip), %xmm4    ## xmm4 = [252645135,252645135,252645135,252645135]
    vpand   %xmm4, %xmm1, %xmm1
    vpsrld  $8, %xmm1, %xmm5
    vpaddd  %xmm5, %xmm1, %xmm1
    vpsrld  $16, %xmm1, %xmm5
    vpaddd  %xmm5, %xmm1, %xmm1
    vmovdqa LCPI0_3(%rip), %xmm5    ## xmm5 = [63,63,63,63]
    vpand   %xmm5, %xmm1, %xmm1
    vpsrld  $1, %xmm0, %xmm6
    vpand   %xmm3, %xmm6, %xmm3
    vpsubd  %xmm3, %xmm0, %xmm0
    vpand   %xmm2, %xmm0, %xmm3
    vpsrld  $2, %xmm0, %xmm0
    vpand   %xmm2, %xmm0, %xmm0
    vpaddd  %xmm0, %xmm3, %xmm0
    vpsrld  $4, %xmm0, %xmm2
    vpaddd  %xmm2, %xmm0, %xmm0
    vpand   %xmm4, %xmm0, %xmm0
    vpsrld  $8, %xmm0, %xmm2
    vpaddd  %xmm2, %xmm0, %xmm0
    vpsrld  $16, %xmm0, %xmm2
    vpaddd  %xmm2, %xmm0, %xmm0
    vpand   %xmm5, %xmm0, %xmm0
    vinsertf128 $1, %xmm1, %ymm0, %ymm0
    retq
    .cfi_endproc

    .globl  v8i32-old-avx1
    .align  4, 0x90
v8i32-old-avx1:                                     ## @A
    .cfi_startproc
## BB#0:                                ## %entry
    vextractf128    $1, %ymm0, %xmm1
    vpextrd $1, %xmm1, %eax
    popcntl %eax, %eax
    vmovd   %xmm1, %ecx
    popcntl %ecx, %ecx
    vmovd   %ecx, %xmm2
    vpinsrd $1, %eax, %xmm2, %xmm2
    vpextrd $2, %xmm1, %eax
    popcntl %eax, %eax
    vpinsrd $2, %eax, %xmm2, %xmm2
    vpextrd $3, %xmm1, %eax
    popcntl %eax, %eax
    vpinsrd $3, %eax, %xmm2, %xmm1
    vpextrd $1, %xmm0, %eax
    popcntl %eax, %eax
    vmovd   %xmm0, %ecx
    popcntl %ecx, %ecx
    vmovd   %ecx, %xmm2
    vpinsrd $1, %eax, %xmm2, %xmm2
    vpextrd $2, %xmm0, %eax
    popcntl %eax, %eax
    vpinsrd $2, %eax, %xmm2, %xmm2
    vpextrd $3, %xmm0, %eax
    popcntl %eax, %eax
    vpinsrd $3, %eax, %xmm2, %xmm0
    vinsertf128 $1, %xmm1, %ymm0, %ymm0
    retq
    .cfi_endproc