test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,FALLBACK0
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx  | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-ONLY,FALLBACK1
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-SLOW,FALLBACK2
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST,FALLBACK3
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST-PERLANE,FALLBACK4
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512F-ONLY-SLOW,FALLBACK5
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512F-ONLY-FAST,FALLBACK6
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512DQ-SLOW,FALLBACK7
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512DQ-FAST,FALLBACK8
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512BW-ONLY-SLOW,FALLBACK9
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512BW-ONLY-FAST,FALLBACK10
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512DQBW-SLOW,FALLBACK11
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512DQBW-FAST,FALLBACK12

; These patterns are produced by LoopVectorizer for interleaved stores.

define void @store_i16_stride6_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i16_stride6_vf2:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa (%rdi), %xmm0
; SSE-NEXT:    movdqa (%rdx), %xmm1
; SSE-NEXT:    movdqa (%r8), %xmm2
; SSE-NEXT:    movdqa (%r9), %xmm3
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,5,4,6]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm4[3,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,1,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[2,0,3,3,4,5,6,7]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[1,2,0,3,4,5,6,7]
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movq %xmm1, 16(%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i16_stride6_vf2:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm3
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,5,7,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,2]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
; AVX1-ONLY-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13,u,u,u,u,2,3,6,7]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,0,0,0]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, (%rax)
; AVX1-ONLY-NEXT:    vmovq %xmm1, 16(%rax)
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i16_stride6_vf2:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-SLOW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX2-SLOW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-SLOW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,2,3,6,7,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[1,3,2,3,4,5,6,7,9,11,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
; AVX2-SLOW-NEXT:    vextracti128 $1, %ymm0, %xmm1
; AVX2-SLOW-NEXT:    vmovq %xmm1, 16(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, (%rax)
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i16_stride6_vf2:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX2-FAST-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-FAST-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,2,3,6,7,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,8,9,12,13,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
; AVX2-FAST-NEXT:    vextracti128 $1, %ymm0, %xmm1
; AVX2-FAST-NEXT:    vmovq %xmm1, 16(%rax)
; AVX2-FAST-NEXT:    vmovdqa %xmm0, (%rax)
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i16_stride6_vf2:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX2-FAST-PERLANE-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-FAST-PERLANE-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,2,3,6,7,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,8,9,12,13,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vextracti128 $1, %ymm0, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovq %xmm1, 16(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm0, (%rax)
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-SLOW-LABEL: store_i16_stride6_vf2:
; AVX512F-SLOW:       # %bb.0:
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-SLOW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX512F-SLOW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX512F-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512F-SLOW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,2,3,6,7,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[1,3,2,3,4,5,6,7,9,11,10,11,12,13,14,15]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
; AVX512F-SLOW-NEXT:    vextracti128 $1, %ymm0, %xmm1
; AVX512F-SLOW-NEXT:    vmovq %xmm1, 16(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512F-SLOW-NEXT:    vzeroupper
; AVX512F-SLOW-NEXT:    retq
;
; AVX512F-FAST-LABEL: store_i16_stride6_vf2:
; AVX512F-FAST:       # %bb.0:
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-FAST-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX512F-FAST-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX512F-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512F-FAST-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,2,3,6,7,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,8,9,12,13,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
; AVX512F-FAST-NEXT:    vextracti128 $1, %ymm0, %xmm1
; AVX512F-FAST-NEXT:    vmovq %xmm1, 16(%rax)
; AVX512F-FAST-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512F-FAST-NEXT:    vzeroupper
; AVX512F-FAST-NEXT:    retq
;
; AVX512BW-LABEL: store_i16_stride6_vf2:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm1 = <0,2,4,6,8,10,1,3,5,7,9,11,u,u,u,u>
; AVX512BW-NEXT:    vpermw %ymm0, %ymm1, %ymm0
; AVX512BW-NEXT:    vextracti128 $1, %ymm0, %xmm1
; AVX512BW-NEXT:    vmovq %xmm1, 16(%rax)
; AVX512BW-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %in.vec0 = load <2 x i16>, ptr %in.vecptr0, align 64
  %in.vec1 = load <2 x i16>, ptr %in.vecptr1, align 64
  %in.vec2 = load <2 x i16>, ptr %in.vecptr2, align 64
  %in.vec3 = load <2 x i16>, ptr %in.vecptr3, align 64
  %in.vec4 = load <2 x i16>, ptr %in.vecptr4, align 64
  %in.vec5 = load <2 x i16>, ptr %in.vecptr5, align 64
  %1 = shufflevector <2 x i16> %in.vec0, <2 x i16> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %2 = shufflevector <2 x i16> %in.vec2, <2 x i16> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %3 = shufflevector <2 x i16> %in.vec4, <2 x i16> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = shufflevector <4 x i16> %1, <4 x i16> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %5 = shufflevector <4 x i16> %3, <4 x i16> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
  %6 = shufflevector <8 x i16> %4, <8 x i16> %5, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>
  %interleaved.vec = shufflevector <12 x i16> %6, <12 x i16> poison, <12 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11>
  store <12 x i16> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i16_stride6_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i16_stride6_vf4:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    movq {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm3 = mem[0],zero
; SSE-NEXT:    movdqa %xmm3, %xmm4
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
; SSE-NEXT:    movdqa %xmm1, %xmm5
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm5 = xmm5[0],xmm0[0]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,4,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,1],xmm0[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,0],xmm6[0,2]
; SSE-NEXT:    movdqa %xmm0, %xmm6
; SSE-NEXT:    punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm1[1]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1],xmm1[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,0],xmm6[0,2]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm1[3,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[3,1,1,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,5,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,2],xmm1[0,3]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0,1,3]
; SSE-NEXT:    movaps %xmm3, 16(%rax)
; SSE-NEXT:    movaps %xmm5, (%rax)
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i16_stride6_vf4:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm5 = mem[0],zero
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX1-ONLY-NEXT:    vpshufb {{.*#+}} xmm7 = xmm3[2,3,10,11,u,u,u,u,u,u,u,u,4,5,12,13]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm0[0,1,1,3]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0,1,2,3],xmm8[4,5],xmm7[6,7]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm7[0,1],xmm4[2,3],xmm7[4,5,6,7]
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm6[0,1,2,0]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,4,6,7]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm2[4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm3[3,1,2,3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[3,1,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
; AVX1-ONLY-NEXT:    vpshufb {{.*#+}} xmm2 = xmm6[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3,4,5],xmm2[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i16_stride6_vf4:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm2
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm5 = xmm4[0],xmm3[0]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3,4,5],ymm2[6],ymm6[7]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm5[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i16_stride6_vf4:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm2
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm5 = xmm4[0],xmm3[0]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [4,6,1,3,4,6,1,3]
; AVX2-FAST-NEXT:    # ymm7 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm7, %ymm2
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,0,1,4,5,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,28,29,u,u,u,u]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3,4,5],ymm2[6],ymm6[7]
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[6,7,14,15,4,5,6,7,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[12,13,14,15,6,7,14,15,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm1 = xmm5[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX2-FAST-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i16_stride6_vf4:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm5 = xmm4[0],xmm3[0]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3,4,5],ymm2[6],ymm6[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[6,7,14,15,4,5,6,7,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[12,13,14,15,6,7,14,15,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm1 = xmm5[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-SLOW-LABEL: store_i16_stride6_vf4:
; AVX512F-SLOW:       # %bb.0:
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512F-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512F-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512F-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm3
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm3[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4,5],ymm3[6],ymm4[7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm2[0,2,2,3]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm4[0,2,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm2[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX512F-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm3, %zmm1
; AVX512F-SLOW-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512F-SLOW-NEXT:    vzeroupper
; AVX512F-SLOW-NEXT:    retq
;
; AVX512F-FAST-LABEL: store_i16_stride6_vf4:
; AVX512F-FAST:       # %bb.0:
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512F-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512F-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512F-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm5
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm6 = ymm5[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [4,6,1,3,4,6,1,3]
; AVX512F-FAST-NEXT:    # ymm7 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm5, %ymm7, %ymm5
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,0,1,4,5,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,28,29,u,u,u,u]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3,4,5],ymm5[6],ymm6[7]
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,8,3,4,9,6,7]
; AVX512F-FAST-NEXT:    vpermi2d %ymm2, %ymm5, %ymm3
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[6,7,14,15,4,5,6,7,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[12,13,14,15,6,7,14,15,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} xmm1 = xmm4[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX512F-FAST-NEXT:    vinserti32x4 $2, %xmm0, %zmm3, %zmm1
; AVX512F-FAST-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX512F-FAST-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512F-FAST-NEXT:    vzeroupper
; AVX512F-FAST-NEXT:    retq
;
; AVX512BW-LABEL: store_i16_stride6_vf4:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512BW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512BW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512BW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm0
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = <0,4,8,12,16,20,1,5,9,13,17,21,2,6,10,14,18,22,3,7,11,15,19,23,u,u,u,u,u,u,u,u>
; AVX512BW-NEXT:    vpermw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vextracti32x4 $2, %zmm0, 32(%rax)
; AVX512BW-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %in.vec0 = load <4 x i16>, ptr %in.vecptr0, align 64
  %in.vec1 = load <4 x i16>, ptr %in.vecptr1, align 64
  %in.vec2 = load <4 x i16>, ptr %in.vecptr2, align 64
  %in.vec3 = load <4 x i16>, ptr %in.vecptr3, align 64
  %in.vec4 = load <4 x i16>, ptr %in.vecptr4, align 64
  %in.vec5 = load <4 x i16>, ptr %in.vecptr5, align 64
  %1 = shufflevector <4 x i16> %in.vec0, <4 x i16> %in.vec1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %2 = shufflevector <4 x i16> %in.vec2, <4 x i16> %in.vec3, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %3 = shufflevector <4 x i16> %in.vec4, <4 x i16> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = shufflevector <8 x i16> %1, <8 x i16> %2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %5 = shufflevector <8 x i16> %3, <8 x i16> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %6 = shufflevector <16 x i16> %4, <16 x i16> %5, <24 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23>
  %interleaved.vec = shufflevector <24 x i16> %6, <24 x i16> poison, <24 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23>
  store <24 x i16> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i16_stride6_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i16_stride6_vf8:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa (%rdi), %xmm0
; SSE-NEXT:    movdqa (%rsi), %xmm8
; SSE-NEXT:    movdqa (%rdx), %xmm1
; SSE-NEXT:    movdqa (%rcx), %xmm9
; SSE-NEXT:    movdqa (%r8), %xmm6
; SSE-NEXT:    movdqa (%r9), %xmm5
; SSE-NEXT:    movdqa %xmm1, %xmm4
; SSE-NEXT:    punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm9[4],xmm4[5],xmm9[5],xmm4[6],xmm9[6],xmm4[7],xmm9[7]
; SSE-NEXT:    movdqa %xmm0, %xmm7
; SSE-NEXT:    punpckhwd {{.*#+}} xmm7 = xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
; SSE-NEXT:    movdqa %xmm7, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm4[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[1,2],xmm2[2,3]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,0,1,3]
; SSE-NEXT:    movaps {{.*#+}} xmm2 = [65535,0,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    andps %xmm2, %xmm10
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm5[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm3[2,1,2,3]
; SSE-NEXT:    movaps %xmm2, %xmm3
; SSE-NEXT:    andnps %xmm11, %xmm3
; SSE-NEXT:    orps %xmm10, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1],xmm1[2],xmm9[2],xmm1[3],xmm9[3]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
; SSE-NEXT:    movdqa %xmm0, %xmm8
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm1[3,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm9 = xmm6[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,2],xmm9[0,1]
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,0,1,3]
; SSE-NEXT:    andps %xmm2, %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm9 = xmm5[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[0,1,2,1]
; SSE-NEXT:    andnps %xmm9, %xmm2
; SSE-NEXT:    orps %xmm8, %xmm2
; SSE-NEXT:    movdqa %xmm1, %xmm10
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm10 = xmm10[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm6, %xmm8
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm0[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,0],xmm8[0,2]
; SSE-NEXT:    movaps {{.*#+}} xmm8 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    andps %xmm8, %xmm10
; SSE-NEXT:    movdqa %xmm5, %xmm11
; SSE-NEXT:    pslldq {{.*#+}} xmm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm11[0,1,2,3,4,5]
; SSE-NEXT:    movaps %xmm8, %xmm9
; SSE-NEXT:    andnps %xmm11, %xmm9
; SSE-NEXT:    orps %xmm10, %xmm9
; SSE-NEXT:    movdqa %xmm7, %xmm10
; SSE-NEXT:    punpckhqdq {{.*#+}} xmm10 = xmm10[1],xmm4[1]
; SSE-NEXT:    movdqa %xmm6, %xmm12
; SSE-NEXT:    psrldq {{.*#+}} xmm12 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[1,1],xmm4[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[2,0],xmm10[0,2]
; SSE-NEXT:    movaps {{.*#+}} xmm10 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    andps %xmm10, %xmm12
; SSE-NEXT:    pshufd {{.*#+}} xmm13 = xmm5[2,2,3,3]
; SSE-NEXT:    movaps %xmm10, %xmm11
; SSE-NEXT:    andnps %xmm13, %xmm11
; SSE-NEXT:    orps %xmm12, %xmm11
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm7[0]
; SSE-NEXT:    pshuflw {{.*#+}} xmm12 = xmm6[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,1],xmm7[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,0],xmm6[0,2]
; SSE-NEXT:    andps %xmm8, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
; SSE-NEXT:    pslld $16, %xmm5
; SSE-NEXT:    andnps %xmm5, %xmm8
; SSE-NEXT:    orps %xmm4, %xmm8
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[1,1],xmm1[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[2,0],xmm0[0,2]
; SSE-NEXT:    andps %xmm10, %xmm12
; SSE-NEXT:    andnps %xmm6, %xmm10
; SSE-NEXT:    orps %xmm12, %xmm10
; SSE-NEXT:    movaps %xmm10, 16(%rax)
; SSE-NEXT:    movaps %xmm8, 48(%rax)
; SSE-NEXT:    movaps %xmm11, 64(%rax)
; SSE-NEXT:    movaps %xmm9, (%rax)
; SSE-NEXT:    movaps %xmm2, 32(%rax)
; SSE-NEXT:    movaps %xmm3, 80(%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i16_stride6_vf8:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm4
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1,2,3],xmm6[4,5],xmm8[6,7]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm9[2,3],xmm8[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm7[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm6[0,1,0,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3],xmm11[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm9[0,1,0,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm10[0,1,2,3],xmm11[4,5],xmm10[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm8, %ymm10, %ymm8
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[0,1,0,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm3[0,1,0,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm4[4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm7[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhqdq {{.*#+}} xmm4 = xmm6[1],xmm4[1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm9[2,3,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3,4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm2[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhqdq {{.*#+}} xmm4 = xmm0[1],xmm4[1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm3[2,3,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3,4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3],xmm0[4,5],xmm2[6,7]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm3[2,3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm8, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i16_stride6_vf8:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-SLOW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,18,19,26,27,u,u,u,u,u,u,u,u]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[0,2,1,3]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[0,2,1,3]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u,18,19,26,27]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1,2],ymm5[3],ymm4[4,5],ymm5[6],ymm4[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u,u,u,u,u]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[2,3,10,11,u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u,u,u,u,u,22,23,30,31]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 64(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm4, 32(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm3, (%rax)
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i16_stride6_vf8:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX2-FAST-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX2-FAST-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,4,1,5,0,4,1,5]
; AVX2-FAST-NEXT:    # ymm3 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,0,1,4,5,u,u,u,u,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u,24,25,28,29]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [0,4,4,0,0,4,4,0]
; AVX2-FAST-NEXT:    # ymm4 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm4, %ymm4
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,12,13,8,9,u,u,u,u,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [2,6,1,5,2,6,1,5]
; AVX2-FAST-NEXT:    # ymm4 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm4, %ymm4
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,10,11,14,15,u,u,u,u,u,u,u,u,16,17,20,21,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[0,2,1,3]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u,18,19,26,27]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1,2],ymm5[3],ymm4[4,5],ymm5[6],ymm4[7]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [7,3,3,7,7,3,3,7]
; AVX2-FAST-NEXT:    # ymm5 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm5, %ymm0
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,8,9,12,13,u,u,u,u,u,u,u,u,22,23,18,19,u,u,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[2,3,10,11,u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u,u,u,u,u,22,23,30,31]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 64(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm4, 32(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm3, (%rax)
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i16_stride6_vf8:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,18,19,26,27,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[0,2,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[0,2,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u,18,19,26,27]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1,2],ymm5[3],ymm4[4,5],ymm5[6],ymm4[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[2,3,10,11,u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u,u,u,u,u,22,23,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm0, 64(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm4, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm3, (%rax)
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-SLOW-LABEL: store_i16_stride6_vf8:
; AVX512F-SLOW:       # %bb.0:
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-SLOW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, (%r9), %zmm2, %zmm2
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[0,2,1,3]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u,18,19,26,27]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm2[0,2,1,3]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm1[0,2,0,2]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,18,19,26,27,u,u,u,u,u,u,u,u,20,21,28,29]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[0,2,0,2]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6],ymm4[7]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm2[0,2,0,2]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,0,1,8,9,u,u,u,u,u,u,u,u,18,19,26,27,u,u,u,u,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[2,3,10,11,u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u,u,u,u,u,22,23,30,31]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX512F-SLOW-NEXT:    vmovdqa %ymm0, 64(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512F-SLOW-NEXT:    vzeroupper
; AVX512F-SLOW-NEXT:    retq
;
; AVX512F-FAST-LABEL: store_i16_stride6_vf8:
; AVX512F-FAST:       # %bb.0:
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-FAST-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti32x4 $1, (%r9), %zmm2, %zmm2
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [2,6,1,5,2,6,1,5]
; AVX512F-FAST-NEXT:    # ymm3 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm1, %ymm3, %ymm3
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,10,11,14,15,u,u,u,u,u,u,u,u,16,17,20,21,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u,u,u,u,u,18,19,26,27]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7]
; AVX512F-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm2[0,2,1,3]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[4,5,12,13,u,u,u,u,u,u,u,u,6,7,14,15,u,u,u,u,u,u,u,u,16,17,24,25,u,u,u,u]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7]
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [0,4,1,5,0,4,1,5]
; AVX512F-FAST-NEXT:    # ymm4 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm1, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,0,1,4,5,u,u,u,u,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u,24,25,28,29]
; AVX512F-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[0,2,0,2]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[0,1,8,9,u,u,u,u,u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6],ymm4[7]
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [0,4,4,0,0,4,4,0]
; AVX512F-FAST-NEXT:    # ymm5 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm2, %ymm5, %ymm5
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,12,13,8,9,u,u,u,u,u,u,u,u,18,19,22,23,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [7,3,3,7,7,3,3,7]
; AVX512F-FAST-NEXT:    # ymm4 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm0, %ymm4, %ymm0
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,8,9,12,13,u,u,u,u,u,u,u,u,22,23,18,19,u,u,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[2,3,10,11,u,u,u,u,u,u,u,u,4,5,12,13,u,u,u,u,u,u,u,u,22,23,30,31,u,u,u,u]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX512F-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX512F-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,2,3,10,11,u,u,u,u,u,u,u,u,20,21,28,29,u,u,u,u,u,u,u,u,22,23,30,31]
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX512F-FAST-NEXT:    vmovdqa %ymm0, 64(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512F-FAST-NEXT:    vzeroupper
; AVX512F-FAST-NEXT:    retq
;
; AVX512BW-LABEL: store_i16_stride6_vf8:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512BW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512BW-NEXT:    vinserti32x4 $1, (%r9), %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm2 = [21,29,37,45,6,14,22,30,38,46,7,15,23,31,39,47]
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,8,16,24,32,40,1,9,17,25,33,41,2,10,18,26,34,42,3,11,19,27,35,43,4,12,20,28,36,44,5,13]
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm0, %zmm3
; AVX512BW-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512BW-NEXT:    vmovdqa %ymm2, 64(%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %in.vec0 = load <8 x i16>, ptr %in.vecptr0, align 64
  %in.vec1 = load <8 x i16>, ptr %in.vecptr1, align 64
  %in.vec2 = load <8 x i16>, ptr %in.vecptr2, align 64
  %in.vec3 = load <8 x i16>, ptr %in.vecptr3, align 64
  %in.vec4 = load <8 x i16>, ptr %in.vecptr4, align 64
  %in.vec5 = load <8 x i16>, ptr %in.vecptr5, align 64
  %1 = shufflevector <8 x i16> %in.vec0, <8 x i16> %in.vec1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %2 = shufflevector <8 x i16> %in.vec2, <8 x i16> %in.vec3, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %3 = shufflevector <8 x i16> %in.vec4, <8 x i16> %in.vec5, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = shufflevector <16 x i16> %1, <16 x i16> %2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %5 = shufflevector <16 x i16> %3, <16 x i16> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %6 = shufflevector <32 x i16> %4, <32 x i16> %5, <48 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47>
  %interleaved.vec = shufflevector <48 x i16> %6, <48 x i16> poison, <48 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47>
  store <48 x i16> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i16_stride6_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i16_stride6_vf16:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $24, %rsp
; SSE-NEXT:    movdqa (%rdi), %xmm15
; SSE-NEXT:    movdqa 16(%rdi), %xmm11
; SSE-NEXT:    movdqa (%rsi), %xmm10
; SSE-NEXT:    movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 16(%rsi), %xmm4
; SSE-NEXT:    movdqa (%rdx), %xmm14
; SSE-NEXT:    movdqa 16(%rdx), %xmm12
; SSE-NEXT:    movdqa (%rcx), %xmm3
; SSE-NEXT:    movdqa 16(%rcx), %xmm2
; SSE-NEXT:    movdqa 16(%r8), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 16(%r9), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm12, %xmm9
; SSE-NEXT:    punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm11, %xmm5
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,3],xmm9[3,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,2],xmm0[0,1]
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,0,1,3]
; SSE-NEXT:    movaps {{.*#+}} xmm7 = [65535,0,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    andps %xmm7, %xmm5
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm1[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,1,2,1]
; SSE-NEXT:    movaps %xmm7, %xmm0
; SSE-NEXT:    andnps %xmm8, %xmm0
; SSE-NEXT:    orps %xmm5, %xmm0
; SSE-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm14, %xmm6
; SSE-NEXT:    punpckhwd {{.*#+}} xmm6 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
; SSE-NEXT:    movdqa %xmm15, %xmm13
; SSE-NEXT:    punpckhwd {{.*#+}} xmm13 = xmm13[4],xmm10[4],xmm13[5],xmm10[5],xmm13[6],xmm10[6],xmm13[7],xmm10[7]
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm6[3,3]
; SSE-NEXT:    movdqa (%r8), %xmm10
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm10[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[1,2],xmm8[2,3]
; SSE-NEXT:    movdqa (%r9), %xmm0
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,6,6,7]
; SSE-NEXT:    movdqa %xmm0, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; SSE-NEXT:    movaps %xmm7, %xmm0
; SSE-NEXT:    andnps %xmm1, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,0,1,3]
; SSE-NEXT:    andps %xmm7, %xmm13
; SSE-NEXT:    orps %xmm13, %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm12 = xmm12[4],xmm2[4],xmm12[5],xmm2[5],xmm12[6],xmm2[6],xmm12[7],xmm2[7]
; SSE-NEXT:    punpckhwd {{.*#+}} xmm11 = xmm11[4],xmm4[4],xmm11[5],xmm4[5],xmm11[6],xmm4[6],xmm11[7],xmm4[7]
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm12[3,3]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm5[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm4[2,3]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm13[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[2,1,2,3]
; SSE-NEXT:    movaps %xmm7, %xmm0
; SSE-NEXT:    andnps %xmm4, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm7, %xmm1
; SSE-NEXT:    orps %xmm1, %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm14 = xmm14[0],xmm3[0],xmm14[1],xmm3[1],xmm14[2],xmm3[2],xmm14[3],xmm3[3]
; SSE-NEXT:    punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
; SSE-NEXT:    # xmm15 = xmm15[0],mem[0],xmm15[1],mem[1],xmm15[2],mem[2],xmm15[3],mem[3]
; SSE-NEXT:    movdqa %xmm15, %xmm1
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm14[3,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm10[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm3[0,1]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm7, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm8[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
; SSE-NEXT:    andnps %xmm3, %xmm7
; SSE-NEXT:    orps %xmm1, %xmm7
; SSE-NEXT:    movdqa %xmm12, %xmm3
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm11[0]
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,1],xmm11[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,0],xmm5[0,2]
; SSE-NEXT:    movdqa %xmm13, %xmm4
; SSE-NEXT:    pslld $16, %xmm4
; SSE-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm1, %xmm13
; SSE-NEXT:    pandn %xmm4, %xmm13
; SSE-NEXT:    andps %xmm1, %xmm3
; SSE-NEXT:    por %xmm3, %xmm13
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm10, %xmm4
; SSE-NEXT:    psrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm6[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa {{.*#+}} xmm3 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[2,2,3,3]
; SSE-NEXT:    movdqa %xmm3, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    andps %xmm3, %xmm4
; SSE-NEXT:    por %xmm4, %xmm5
; SSE-NEXT:    movdqa %xmm15, %xmm0
; SSE-NEXT:    punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm14[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm10[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm14[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,0],xmm0[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,1,1]
; SSE-NEXT:    movdqa %xmm3, %xmm8
; SSE-NEXT:    pandn %xmm0, %xmm8
; SSE-NEXT:    andps %xmm3, %xmm4
; SSE-NEXT:    por %xmm4, %xmm8
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm14 = xmm14[0],xmm15[0]
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm15[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,0],xmm0[0,2]
; SSE-NEXT:    movdqa %xmm2, %xmm0
; SSE-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm0, %xmm15
; SSE-NEXT:    andps %xmm1, %xmm14
; SSE-NEXT:    por %xmm14, %xmm15
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm12[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm11[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm12[2,2,3,3]
; SSE-NEXT:    movdqa %xmm3, %xmm11
; SSE-NEXT:    pandn %xmm4, %xmm11
; SSE-NEXT:    andps %xmm3, %xmm0
; SSE-NEXT:    por %xmm0, %xmm11
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movaps %xmm2, %xmm0
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm14[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm9[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,0],xmm0[0,2]
; SSE-NEXT:    andps %xmm3, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm12[0,0,1,1]
; SSE-NEXT:    pandn %xmm0, %xmm3
; SSE-NEXT:    por %xmm4, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm9 = xmm9[0],xmm2[0]
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,0],xmm0[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm4
; SSE-NEXT:    pslldq {{.*#+}} xmm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm4[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm1, %xmm0
; SSE-NEXT:    pandn %xmm4, %xmm0
; SSE-NEXT:    andps %xmm1, %xmm9
; SSE-NEXT:    por %xmm9, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm2[0]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,1],xmm2[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,0],xmm10[0,2]
; SSE-NEXT:    andps %xmm1, %xmm6
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    pslld $16, %xmm2
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    por %xmm6, %xmm1
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa %xmm1, 48(%rax)
; SSE-NEXT:    movdqa %xmm0, 96(%rax)
; SSE-NEXT:    movdqa %xmm3, 112(%rax)
; SSE-NEXT:    movdqa %xmm11, 160(%rax)
; SSE-NEXT:    movdqa %xmm15, (%rax)
; SSE-NEXT:    movdqa %xmm8, 16(%rax)
; SSE-NEXT:    movdqa %xmm5, 64(%rax)
; SSE-NEXT:    movdqa %xmm13, 144(%rax)
; SSE-NEXT:    movaps %xmm7, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rax)
; SSE-NEXT:    addq $24, %rsp
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i16_stride6_vf16:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa 16(%rcx), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm4
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdx), %xmm1
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm8[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm5
; AVX1-ONLY-NEXT:    vmovdqa 16(%rsi), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdi), %xmm7
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm9[2,3,2,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm3[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm10, %ymm7
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm7[0,1],ymm1[2],ymm7[3,4],ymm1[5],ymm7[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm7
; AVX1-ONLY-NEXT:    vmovdqa 16(%r8), %xmm11
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0,1,2,3],xmm11[4,5],xmm7[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 16(%r9), %xmm12
; AVX1-ONLY-NEXT:    vpslld $16, %xmm12, %xmm10
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0,1,2,3,4],xmm10[5],xmm7[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm11[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm7[0],xmm1[1,2],xmm7[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm12[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm7[1],xmm1[2,3,4,5,6],xmm7[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm12[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm3[3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3,4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm12[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm0[0],xmm1[1],xmm0[2,3,4,5,6],xmm1[7]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm14 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm1[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm14[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm14, %ymm10
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm0[0,1],ymm10[2],ymm0[3,4],ymm10[5],ymm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm13
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm13[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,1,2,3]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm15, %xmm10
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm0[0,1],xmm10[2,3,4,5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm0
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm0[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm10[0],xmm3[1],xmm10[2,3,4,5,6],xmm3[7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm3 = xmm15[0],xmm3[1],xmm15[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm15 = xmm3[0,1,2],xmm15[3],xmm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm8[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm8, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm9[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0],ymm3[1],ymm8[2,3],ymm3[4],ymm8[5,6],ymm3[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm8 = xmm3[0,1],xmm11[0],xmm3[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm9 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm12[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1,2,3,4],xmm9[5],xmm8[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm3, %xmm3
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm9 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm9[2,3],xmm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm12[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm9 = xmm3[0,1,2],xmm9[3],xmm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm2[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm3, %xmm5
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm6[2,3],xmm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm6[3],xmm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm3 = xmm3[0,1],xmm13[0],xmm3[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm6[5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm4[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm14[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm13[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm1[1,2],xmm2[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3,4,5,6],xmm4[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm1
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm13[4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm0, %xmm0
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4],xmm0[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, 48(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, 32(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, (%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, 16(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm9, 112(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm8, 96(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm15, 64(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm10, 80(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm7, 176(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 160(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 128(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 144(%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i16_stride6_vf16:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %ymm13
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %ymm2
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %ymm3
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %ymm4
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %xmm5
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm6
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm7, %ymm9
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %xmm7
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm8 = xmm7[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm8[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm8
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm11 = xmm8[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm10 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1],ymm9[2],ymm10[3,4],ymm9[5],ymm10[6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm10
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm10[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0],ymm9[1,2],ymm11[3],ymm9[4,5],ymm11[6],ymm9[7]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %xmm11
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm11[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm9, %ymm12, %ymm9
; AVX2-SLOW-NEXT:    vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm14 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm14 = ymm14[0],ymm12[0],ymm14[1],ymm12[1],ymm14[2],ymm12[2],ymm14[3],ymm12[3],ymm14[8],ymm12[8],ymm14[9],ymm12[9],ymm14[10],ymm12[10],ymm14[11],ymm12[11]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm2[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm13[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm15[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm15 = ymm15[0],ymm12[0],ymm15[1],ymm12[1],ymm15[2],ymm12[2],ymm15[3],ymm12[3],ymm15[8],ymm12[8],ymm15[9],ymm12[9],ymm15[10],ymm12[10],ymm15[11],ymm12[11]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %ymm12
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4],ymm14[5],ymm15[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0],ymm14[1,2],ymm15[3],ymm14[4,5],ymm15[6],ymm14[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm12[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm14, %ymm15, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm14 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[1,1,1,1]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4],ymm14[5],ymm15[6,7]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} xmm15 = xmm10[12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0],ymm15[1],ymm14[2,3],ymm15[4],ymm14[5,6],ymm15[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm15 = xmm11[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm15[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm14, %ymm15, %ymm14
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm15 = ymm13[4],ymm2[4],ymm13[5],ymm2[5],ymm13[6],ymm2[6],ymm13[7],ymm2[7],ymm13[12],ymm2[12],ymm13[13],ymm2[13],ymm13[14],ymm2[14],ymm13[15],ymm2[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1],ymm15[2],ymm9[3,4],ymm15[5],ymm9[6,7]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm15 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0],ymm15[1],ymm9[2,3],ymm15[4],ymm9[5,6],ymm15[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm12[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm15[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm9, %ymm15, %ymm0
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm7[0,0,2,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3],ymm5[4],ymm6[5,6],ymm5[7]
; AVX2-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm11[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm7 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-SLOW-NEXT:    vpblendvb %ymm7, %ymm5, %ymm6, %ymm5
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm13[0],ymm2[0],ymm13[1],ymm2[1],ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[8],ymm2[8],ymm13[9],ymm2[9],ymm13[10],ymm2[10],ymm13[11],ymm2[11]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6],ymm3[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm7, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa %ymm1, 96(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 160(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm14, 64(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm5, (%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i16_stride6_vf16:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %ymm13
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %ymm3
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %ymm4
; AVX2-FAST-NEXT:    vmovdqa (%r8), %ymm2
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %xmm5
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX2-FAST-NEXT:    vpshufb %xmm7, %xmm5, %xmm8
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm6
; AVX2-FAST-NEXT:    vpshufb %xmm7, %xmm6, %xmm7
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %xmm8
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm10 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm9
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm11 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm10, %ymm10
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm7 = ymm7[0,1],ymm10[2],ymm7[3,4],ymm10[5],ymm7[6,7]
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm10
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm10[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm7 = ymm11[0],ymm7[1,2],ymm11[3],ymm7[4,5],ymm11[6],ymm7[7]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %xmm11
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm12 = xmm11[0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm7, %ymm12, %ymm7
; AVX2-FAST-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm1, %ymm14
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm13, %ymm12
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm14 = ymm12[0],ymm14[0],ymm12[1],ymm14[1],ymm12[2],ymm14[2],ymm12[3],ymm14[3],ymm12[8],ymm14[8],ymm12[9],ymm14[9],ymm12[10],ymm14[10],ymm12[11],ymm14[11]
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm15 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm15 = ymm15[0],ymm12[0],ymm15[1],ymm12[1],ymm15[2],ymm12[2],ymm15[3],ymm12[3],ymm15[8],ymm12[8],ymm15[9],ymm12[9],ymm15[10],ymm12[10],ymm15[11],ymm12[11]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %ymm12
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0],ymm14[1,2],ymm15[3],ymm14[4,5],ymm15[6],ymm14[7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm15 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm14, %ymm15, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm14 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm15 = [1,0,3,2,1,0,3,2]
; AVX2-FAST-NEXT:    # ymm15 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm14, %ymm15, %ymm14
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm15 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm15 = xmm10[12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0],ymm15[1],ymm14[2,3],ymm15[4],ymm14[5,6],ymm15[7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm15 = xmm11[8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm14, %ymm15, %ymm14
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm15 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [5,0,7,6,5,0,7,6]
; AVX2-FAST-NEXT:    # ymm7 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm15, %ymm7, %ymm7
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm15 = ymm13[4],ymm1[4],ymm13[5],ymm1[5],ymm13[6],ymm1[6],ymm13[7],ymm1[7],ymm13[12],ymm1[12],ymm13[13],ymm1[13],ymm13[14],ymm1[14],ymm13[15],ymm1[15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm7 = ymm7[0,1],ymm15[2],ymm7[3,4],ymm15[5],ymm7[6,7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm15 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm7 = ymm7[0],ymm15[1],ymm7[2,3],ymm15[4],ymm7[5,6],ymm15[7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm15 = ymm12[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm7, %ymm15, %ymm0
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [1,0,2,2,1,0,2,2]
; AVX2-FAST-NEXT:    # ymm8 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm7, %ymm8, %ymm7
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2,3],ymm7[4],ymm5[5,6],ymm7[7]
; AVX2-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm11[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-FAST-NEXT:    vpblendvb %ymm7, %ymm5, %ymm6, %ymm5
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [5,4,0,6,5,4,0,6]
; AVX2-FAST-NEXT:    # ymm4 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm3, %ymm4, %ymm3
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm13[0],ymm1[0],ymm13[1],ymm1[1],ymm13[2],ymm1[2],ymm13[3],ymm1[3],ymm13[8],ymm1[8],ymm13[9],ymm1[9],ymm13[10],ymm1[10],ymm13[11],ymm1[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3],ymm3[4],ymm1[5,6],ymm3[7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm7, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa %ymm1, 96(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 160(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm14, 64(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm5, (%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i16_stride6_vf16:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %ymm5
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovaps (%r9), %ymm4
; AVX2-FAST-PERLANE-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %xmm7
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %xmm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %xmm14
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r9), %xmm9
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm6, %xmm8
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1],ymm12[2],ymm13[3,4],ymm12[5],ymm13[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm14, %xmm6
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm13 = xmm14[12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0],ymm13[1],ymm12[2,3],ymm13[4],ymm12[5,6],ymm13[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm13 = xmm9[8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm14 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm14, %ymm12, %ymm13, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm0, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} ymm13 = ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[12],ymm2[12],ymm0[13],ymm2[13],ymm0[14],ymm2[14],ymm0[15],ymm2[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} ymm15 = ymm3[4],ymm5[4],ymm3[5],ymm5[5],ymm3[6],ymm5[6],ymm3[7],ymm5[7],ymm3[12],ymm5[12],ymm3[13],ymm5[13],ymm3[14],ymm5[14],ymm3[15],ymm5[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm15 = ymm15[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm13 = ymm15[0,1],ymm13[2],ymm15[3,4],ymm13[5],ymm15[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm15 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm13 = ymm13[0],ymm15[1],ymm13[2,3],ymm15[4],ymm13[5,6],ymm15[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm15 = ymm0[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm14, %ymm13, %ymm15, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0],ymm15[1],ymm14[2,3],ymm15[4],ymm14[5,6],ymm15[7]
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{.*#+}} xmm15 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm6, %xmm11
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm9[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm6 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm6, %ymm14, %ymm15, %ymm14
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm15 = ymm12[0],ymm2[0],ymm12[1],ymm2[1],ymm12[2],ymm2[2],ymm12[3],ymm2[3],ymm12[8],ymm2[8],ymm12[9],ymm2[9],ymm12[10],ymm2[10],ymm12[11],ymm2[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm2, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm12, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm3[0],ymm5[0],ymm3[1],ymm5[1],ymm3[2],ymm5[2],ymm3[3],ymm5[3],ymm3[8],ymm5[8],ymm3[9],ymm5[9],ymm3[10],ymm5[10],ymm3[11],ymm5[11]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm12 = ymm12[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm12 = ymm15[0],ymm12[1],ymm15[2,3],ymm12[4],ymm15[5,6],ymm12[7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm6, %ymm12, %ymm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm12, %xmm7, %xmm6
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm12, %xmm10, %xmm7
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm7, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1],ymm7[2],ymm6[3,4],ymm7[5],ymm6[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm11[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1,2],ymm7[3],ymm6[4,5],ymm7[6],ymm6[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm7 = xmm9[0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm8 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm6, %ymm7, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm7 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm13, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11]
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm1, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm1, 128(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm15, 96(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm6, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm14, (%rax)
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-SLOW-LABEL: store_i16_stride6_vf16:
; AVX512F-SLOW:       # %bb.0:
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512F-SLOW-NEXT:    vmovdqa (%rsi), %ymm2
; AVX512F-SLOW-NEXT:    vmovdqa (%rdx), %ymm4
; AVX512F-SLOW-NEXT:    vmovdqa (%rcx), %ymm5
; AVX512F-SLOW-NEXT:    vmovdqa (%r8), %ymm1
; AVX512F-SLOW-NEXT:    vmovdqa (%r9), %ymm3
; AVX512F-SLOW-NEXT:    vpsrldq {{.*#+}} ymm6 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-SLOW-NEXT:    vpsrldq {{.*#+}} ymm7 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm2[2,1,2,3,6,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm7[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm0[2,1,2,3,6,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm8 = <u,5,12,u,4,13,u,7>
; AVX512F-SLOW-NEXT:    vpermi2d %ymm6, %ymm7, %ymm8
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm6 = ymm3[1,2,2,3,5,6,6,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm1[1,2,2,3,5,6,6,7]
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0],ymm8[1,2],ymm6[3],ymm8[4,5],ymm6[6],ymm8[7]
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm7 = ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[12],ymm2[12],ymm0[13],ymm2[13],ymm0[14],ymm2[14],ymm0[15],ymm2[15]
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm9 = <5,u,14,6,u,15,7,u>
; AVX512F-SLOW-NEXT:    vpermi2d %ymm7, %ymm8, %ymm9
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm7
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm8 = [8,21,10,11,22,13,14,23]
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm1[4],ymm3[4],ymm1[5],ymm3[5],ymm1[6],ymm3[6],ymm1[7],ymm3[7],ymm1[12],ymm3[12],ymm1[13],ymm3[13],ymm1[14],ymm3[14],ymm1[15],ymm3[15]
; AVX512F-SLOW-NEXT:    vpermi2d %zmm9, %zmm7, %zmm8
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm8, %zmm6, %zmm16
; AVX512F-SLOW-NEXT:    vmovdqa (%rcx), %xmm6
; AVX512F-SLOW-NEXT:    vmovdqa (%rdx), %xmm7
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512F-SLOW-NEXT:    vmovdqa (%rsi), %xmm8
; AVX512F-SLOW-NEXT:    vmovdqa (%rdi), %xmm10
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm13 = <0,8,u,1,9,u,2,10>
; AVX512F-SLOW-NEXT:    vpermi2d %ymm9, %ymm11, %ymm13
; AVX512F-SLOW-NEXT:    vmovdqa (%r9), %xmm9
; AVX512F-SLOW-NEXT:    vmovdqa (%r8), %xmm11
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512F-SLOW-NEXT:    vpbroadcastq %xmm14, %ymm14
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm13 = ymm13[0,1],ymm14[2],ymm13[3,4],ymm14[5],ymm13[6,7]
; AVX512F-SLOW-NEXT:    vpsrldq {{.*#+}} xmm14 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-SLOW-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm15 = xmm8[0,1,2,1]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,7,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm10[0,1,2,1]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,7,6,5]
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm15[4],xmm12[5],xmm15[5],xmm12[6],xmm15[6],xmm12[7],xmm15[7]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm15 = <u,1,8,u,0,9,u,3>
; AVX512F-SLOW-NEXT:    vpermi2d %ymm14, %ymm12, %ymm15
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm9[1,2,2,3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm14 = xmm11[1,2,2,3]
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0],ymm15[1,2],ymm12[3],ymm15[4,5],ymm12[6],ymm15[7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm13, %zmm12
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm2 = <4,12,u,5,13,u,6,14>
; AVX512F-SLOW-NEXT:    vpermi2d %ymm4, %ymm0, %ymm2
; AVX512F-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm3 = <1,u,10,2,u,11,3,u>
; AVX512F-SLOW-NEXT:    vpermi2d %ymm1, %ymm2, %ymm3
; AVX512F-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
; AVX512F-SLOW-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,9,2,3,10,5,6,11]
; AVX512F-SLOW-NEXT:    vpermi2d %ymm1, %ymm3, %ymm2
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm0, 64(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm12, (%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm16, 128(%rax)
; AVX512F-SLOW-NEXT:    vzeroupper
; AVX512F-SLOW-NEXT:    retq
;
; AVX512F-FAST-LABEL: store_i16_stride6_vf16:
; AVX512F-FAST:       # %bb.0:
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512F-FAST-NEXT:    vmovdqa (%rsi), %ymm3
; AVX512F-FAST-NEXT:    vmovdqa (%rdx), %ymm4
; AVX512F-FAST-NEXT:    vmovdqa (%rcx), %ymm5
; AVX512F-FAST-NEXT:    vmovdqa64 (%r8), %ymm16
; AVX512F-FAST-NEXT:    vmovdqa (%r9), %ymm1
; AVX512F-FAST-NEXT:    vmovdqa (%rcx), %xmm6
; AVX512F-FAST-NEXT:    vmovdqa (%rdx), %xmm7
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512F-FAST-NEXT:    vmovdqa (%rsi), %xmm8
; AVX512F-FAST-NEXT:    vmovdqa (%rdi), %xmm10
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = <0,8,u,1,9,u,2,10>
; AVX512F-FAST-NEXT:    vpermi2d %ymm9, %ymm11, %ymm12
; AVX512F-FAST-NEXT:    vmovdqa (%r9), %xmm9
; AVX512F-FAST-NEXT:    vmovdqa (%r8), %xmm11
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm14 = [0,1,8,3,4,9,6,7]
; AVX512F-FAST-NEXT:    vpermi2d %ymm13, %ymm12, %ymm14
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512F-FAST-NEXT:    vpshufb %xmm12, %xmm8, %xmm13
; AVX512F-FAST-NEXT:    vpshufb %xmm12, %xmm10, %xmm12
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
; AVX512F-FAST-NEXT:    vpsrldq {{.*#+}} xmm13 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-FAST-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm15 = <u,1,8,u,0,9,u,3>
; AVX512F-FAST-NEXT:    vpermi2d %ymm13, %ymm12, %ymm15
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm15, %zmm0, %zmm12
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm13 = [16,9,10,17,12,13,18,15]
; AVX512F-FAST-NEXT:    vpshufd {{.*#+}} xmm15 = xmm9[1,2,2,3]
; AVX512F-FAST-NEXT:    vpshufd {{.*#+}} xmm0 = xmm11[1,2,2,3]
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
; AVX512F-FAST-NEXT:    vpermi2d %zmm0, %zmm12, %zmm13
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm13, %zmm14, %zmm0
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = <1,u,10,2,u,11,3,u>
; AVX512F-FAST-NEXT:    vpermi2d %ymm8, %ymm6, %ymm7
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,9,2,3,10,5,6,11]
; AVX512F-FAST-NEXT:    vpermi2d %ymm6, %ymm7, %ymm8
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm9 = <4,12,u,5,13,u,6,14>
; AVX512F-FAST-NEXT:    vpermi2d %ymm6, %ymm7, %ymm9
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm6
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [8,9,20,11,12,21,14,15]
; AVX512F-FAST-NEXT:    vmovdqa64 %ymm16, %ymm10
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm10[0],ymm1[0],ymm10[1],ymm1[1],ymm10[2],ymm1[2],ymm10[3],ymm1[3],ymm10[8],ymm1[8],ymm10[9],ymm1[9],ymm10[10],ymm1[10],ymm10[11],ymm1[11]
; AVX512F-FAST-NEXT:    vpermi2d %zmm9, %zmm6, %zmm7
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm7, %zmm8, %zmm6
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX512F-FAST-NEXT:    vpshufb %ymm7, %ymm3, %ymm8
; AVX512F-FAST-NEXT:    vpshufb %ymm7, %ymm2, %ymm7
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11]
; AVX512F-FAST-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-FAST-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm9 = <u,5,12,u,4,13,u,7>
; AVX512F-FAST-NEXT:    vpermi2d %ymm8, %ymm7, %ymm9
; AVX512F-FAST-NEXT:    vpshufd {{.*#+}} ymm7 = ymm1[1,2,2,3,5,6,6,7]
; AVX512F-FAST-NEXT:    vpshufd {{.*#+}} ymm8 = ymm16[1,2,2,3,5,6,6,7]
; AVX512F-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = [12,1,2,13,4,5,14,7]
; AVX512F-FAST-NEXT:    vpermi2d %ymm7, %ymm9, %ymm8
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = <5,u,14,6,u,15,7,u>
; AVX512F-FAST-NEXT:    vpermi2d %ymm2, %ymm3, %ymm4
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm2
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [8,21,10,11,22,13,14,23]
; AVX512F-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm10[4],ymm1[4],ymm10[5],ymm1[5],ymm10[6],ymm1[6],ymm10[7],ymm1[7],ymm10[12],ymm1[12],ymm10[13],ymm1[13],ymm10[14],ymm1[14],ymm10[15],ymm1[15]
; AVX512F-FAST-NEXT:    vpermi2d %zmm1, %zmm2, %zmm3
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm8, %zmm1
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm1, 128(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm6, 64(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512F-FAST-NEXT:    vzeroupper
; AVX512F-FAST-NEXT:    retq
;
; AVX512BW-LABEL: store_i16_stride6_vf16:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512BW-NEXT:    vmovdqa (%rdx), %ymm1
; AVX512BW-NEXT:    vmovdqa (%r8), %ymm2
; AVX512BW-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm0
; AVX512BW-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm1
; AVX512BW-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm2
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = <0,16,32,48,u,u,1,17,33,49,u,u,2,18,34,50,u,u,3,19,35,51,u,u,4,20,36,52,u,u,5,21>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm0, %zmm3
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [0,1,2,3,32,48,6,7,8,9,33,49,12,13,14,15,34,50,18,19,20,21,35,51,24,25,26,27,36,52,30,31]
; AVX512BW-NEXT:    vpermi2w %zmm2, %zmm3, %zmm4
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = <5,21,u,u,38,54,6,22,u,u,39,55,7,23,u,u,40,56,8,24,u,u,41,57,9,25,u,u,42,58,10,26>
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm1, %zmm3
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [0,1,37,53,4,5,6,7,38,54,10,11,12,13,39,55,16,17,18,19,40,56,22,23,24,25,41,57,28,29,30,31]
; AVX512BW-NEXT:    vpermi2w %zmm2, %zmm3, %zmm5
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = <u,u,11,27,43,59,u,u,12,28,44,60,u,u,13,29,45,61,u,u,14,30,46,62,u,u,15,31,47,63,u,u>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm0, %zmm3
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm0 = [42,58,2,3,4,5,43,59,8,9,10,11,44,60,14,15,16,17,45,61,20,21,22,23,46,62,26,27,28,29,47,63]
; AVX512BW-NEXT:    vpermi2w %zmm2, %zmm3, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm0, 128(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %in.vec0 = load <16 x i16>, ptr %in.vecptr0, align 64
  %in.vec1 = load <16 x i16>, ptr %in.vecptr1, align 64
  %in.vec2 = load <16 x i16>, ptr %in.vecptr2, align 64
  %in.vec3 = load <16 x i16>, ptr %in.vecptr3, align 64
  %in.vec4 = load <16 x i16>, ptr %in.vecptr4, align 64
  %in.vec5 = load <16 x i16>, ptr %in.vecptr5, align 64
  %1 = shufflevector <16 x i16> %in.vec0, <16 x i16> %in.vec1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %2 = shufflevector <16 x i16> %in.vec2, <16 x i16> %in.vec3, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %3 = shufflevector <16 x i16> %in.vec4, <16 x i16> %in.vec5, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  %4 = shufflevector <32 x i16> %1, <32 x i16> %2, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %5 = shufflevector <32 x i16> %3, <32 x i16> poison, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %6 = shufflevector <64 x i16> %4, <64 x i16> %5, <96 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95>
  %interleaved.vec = shufflevector <96 x i16> %6, <96 x i16> poison, <96 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95>
  store <96 x i16> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i16_stride6_vf32:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $280, %rsp # imm = 0x118
; SSE-NEXT:    movdqa (%rdi), %xmm1
; SSE-NEXT:    movdqa 16(%rdi), %xmm13
; SSE-NEXT:    movdqa (%rsi), %xmm2
; SSE-NEXT:    movdqa 16(%rsi), %xmm9
; SSE-NEXT:    movdqa (%rdx), %xmm12
; SSE-NEXT:    movdqa 16(%rdx), %xmm14
; SSE-NEXT:    movdqa (%rcx), %xmm4
; SSE-NEXT:    movdqa 16(%rcx), %xmm10
; SSE-NEXT:    movdqa (%r8), %xmm7
; SSE-NEXT:    movdqa (%r9), %xmm11
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm1, %xmm5
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm5, %xmm8
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm0[3,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm7[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,2],xmm3[0,1]
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,0,1,3]
; SSE-NEXT:    movaps {{.*#+}} xmm6 = [65535,0,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    andps %xmm6, %xmm8
; SSE-NEXT:    movdqa %xmm11, %xmm3
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm11 = xmm11[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[0,1,2,1]
; SSE-NEXT:    movaps %xmm6, %xmm0
; SSE-NEXT:    andnps %xmm11, %xmm0
; SSE-NEXT:    orps %xmm8, %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm12 = xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,3],xmm12[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,2],xmm4[2,3]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm2
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[2,1,2,3]
; SSE-NEXT:    movaps %xmm6, %xmm0
; SSE-NEXT:    andnps %xmm4, %xmm0
; SSE-NEXT:    orps %xmm2, %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm11
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm0[3,3]
; SSE-NEXT:    movdqa 16(%r8), %xmm0
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm0[2,1,3,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[1,2],xmm8[0,1]
; SSE-NEXT:    movdqa 16(%r9), %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm12 = xmm8[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm12[0,1,2,1]
; SSE-NEXT:    movaps %xmm6, %xmm0
; SSE-NEXT:    andnps %xmm12, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm11
; SSE-NEXT:    orps %xmm11, %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm14 = xmm14[4],xmm10[4],xmm14[5],xmm10[5],xmm14[6],xmm10[6],xmm14[7],xmm10[7]
; SSE-NEXT:    movdqa %xmm14, (%rsp) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm13 = xmm13[4],xmm9[4],xmm13[5],xmm9[5],xmm13[6],xmm9[6],xmm13[7],xmm9[7]
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm9
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm14[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm1[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[1,2],xmm10[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm8[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[2,1,2,3]
; SSE-NEXT:    movaps %xmm6, %xmm0
; SSE-NEXT:    andnps %xmm10, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm9
; SSE-NEXT:    orps %xmm9, %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 32(%rdx), %xmm1
; SSE-NEXT:    movdqa 32(%rcx), %xmm9
; SSE-NEXT:    movdqa %xmm1, %xmm0
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 32(%rdi), %xmm12
; SSE-NEXT:    movdqa 32(%rsi), %xmm10
; SSE-NEXT:    movdqa %xmm12, %xmm11
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; SSE-NEXT:    movdqa %xmm11, %xmm15
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm0[3,3]
; SSE-NEXT:    movdqa 32(%r8), %xmm13
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm13[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[1,2],xmm14[0,1]
; SSE-NEXT:    movdqa 32(%r9), %xmm2
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm2[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
; SSE-NEXT:    movaps %xmm6, %xmm14
; SSE-NEXT:    andnps %xmm0, %xmm14
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm15
; SSE-NEXT:    orps %xmm15, %xmm14
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm9[4],xmm1[5],xmm9[5],xmm1[6],xmm9[6],xmm1[7],xmm9[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm12 = xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm1[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm13[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,2],xmm9[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm2[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[2,1,2,3]
; SSE-NEXT:    movaps %xmm6, %xmm10
; SSE-NEXT:    andnps %xmm9, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm0
; SSE-NEXT:    orps %xmm0, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 48(%rdx), %xmm0
; SSE-NEXT:    movdqa 48(%rcx), %xmm4
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 48(%rdi), %xmm2
; SSE-NEXT:    movdqa 48(%rsi), %xmm9
; SSE-NEXT:    movdqa %xmm2, %xmm15
; SSE-NEXT:    punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 48(%r8), %xmm10
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm10[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[1,2],xmm14[0,1]
; SSE-NEXT:    movdqa 48(%r9), %xmm1
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm1[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,1,2,1]
; SSE-NEXT:    movaps %xmm6, %xmm3
; SSE-NEXT:    andnps %xmm14, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm15
; SSE-NEXT:    orps %xmm15, %xmm3
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,3],xmm3[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm10[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,2],xmm9[2,3]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0,1,3]
; SSE-NEXT:    andps %xmm6, %xmm2
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    movdqa %xmm1, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[2,1,2,3]
; SSE-NEXT:    andnps %xmm9, %xmm6
; SSE-NEXT:    orps %xmm2, %xmm6
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm0
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
; SSE-NEXT:    movdqa %xmm7, %xmm9
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm9[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm2, %xmm14
; SSE-NEXT:    pslldq {{.*#+}} xmm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm14[0,1,2,3,4,5]
; SSE-NEXT:    movdqa {{.*#+}} xmm9 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    pandn %xmm14, %xmm1
; SSE-NEXT:    andps %xmm9, %xmm0
; SSE-NEXT:    por %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm3[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm7[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[1,1],xmm3[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,0],xmm5[0,2]
; SSE-NEXT:    movdqa {{.*#+}} xmm5 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,1,1]
; SSE-NEXT:    movdqa %xmm5, %xmm0
; SSE-NEXT:    pandn %xmm6, %xmm0
; SSE-NEXT:    andps %xmm5, %xmm14
; SSE-NEXT:    por %xmm14, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; SSE-NEXT:    movaps %xmm15, %xmm6
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm7, %xmm14
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,1],xmm0[1,3]
; SSE-NEXT:    movaps %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,0],xmm14[0,2]
; SSE-NEXT:    movdqa %xmm2, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm0
; SSE-NEXT:    movdqa %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm9, %xmm0
; SSE-NEXT:    pandn %xmm1, %xmm0
; SSE-NEXT:    andps %xmm9, %xmm6
; SSE-NEXT:    por %xmm6, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm15[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[1,1],xmm15[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa %xmm5, %xmm0
; SSE-NEXT:    pandn %xmm14, %xmm0
; SSE-NEXT:    andps %xmm5, %xmm7
; SSE-NEXT:    por %xmm7, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movaps %xmm3, %xmm6
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm2[0]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, %xmm1
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm2[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,0],xmm1[0,2]
; SSE-NEXT:    movdqa %xmm8, %xmm7
; SSE-NEXT:    pslldq {{.*#+}} xmm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm7[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    pandn %xmm7, %xmm1
; SSE-NEXT:    andps %xmm9, %xmm6
; SSE-NEXT:    por %xmm6, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm0[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[1,1],xmm3[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,0],xmm2[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[0,0,1,1]
; SSE-NEXT:    movdqa %xmm5, %xmm1
; SSE-NEXT:    pandn %xmm7, %xmm1
; SSE-NEXT:    andps %xmm5, %xmm6
; SSE-NEXT:    por %xmm6, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps (%rsp), %xmm2 # 16-byte Reload
; SSE-NEXT:    movaps %xmm2, %xmm7
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
; SSE-NEXT:    movaps %xmm0, %xmm6
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,1],xmm1[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,0],xmm6[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm8[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm8
; SSE-NEXT:    movdqa %xmm9, %xmm6
; SSE-NEXT:    pandn %xmm8, %xmm6
; SSE-NEXT:    andps %xmm9, %xmm7
; SSE-NEXT:    por %xmm7, %xmm6
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[0,2]
; SSE-NEXT:    movdqa %xmm5, %xmm7
; SSE-NEXT:    pandn %xmm14, %xmm7
; SSE-NEXT:    andps %xmm5, %xmm0
; SSE-NEXT:    por %xmm0, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm8 = xmm8[0],xmm11[0]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm11[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm0, %xmm14
; SSE-NEXT:    pslldq {{.*#+}} xmm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm14[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm14, %xmm3
; SSE-NEXT:    andps %xmm9, %xmm8
; SSE-NEXT:    por %xmm8, %xmm3
; SSE-NEXT:    punpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm1[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm13[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[1,1],xmm1[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,0],xmm11[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,1,1]
; SSE-NEXT:    movdqa %xmm5, %xmm8
; SSE-NEXT:    pandn %xmm11, %xmm8
; SSE-NEXT:    andps %xmm5, %xmm14
; SSE-NEXT:    por %xmm14, %xmm8
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm14
; SSE-NEXT:    punpcklqdq {{.*#+}} xmm14 = xmm14[0],xmm12[0]
; SSE-NEXT:    movdqa %xmm13, %xmm11
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,1],xmm12[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,0],xmm11[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm0[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm0
; SSE-NEXT:    movdqa %xmm9, %xmm11
; SSE-NEXT:    pandn %xmm0, %xmm11
; SSE-NEXT:    andps %xmm9, %xmm14
; SSE-NEXT:    por %xmm14, %xmm11
; SSE-NEXT:    movdqa %xmm1, %xmm0
; SSE-NEXT:    punpckhqdq {{.*#+}} xmm12 = xmm12[1],xmm1[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm13 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[1,1],xmm0[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,0],xmm12[0,2]
; SSE-NEXT:    movdqa %xmm5, %xmm12
; SSE-NEXT:    pandn %xmm15, %xmm12
; SSE-NEXT:    andps %xmm5, %xmm13
; SSE-NEXT:    por %xmm13, %xmm12
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, %xmm14
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm10, %xmm13
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[0,1],xmm0[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,0],xmm13[0,2]
; SSE-NEXT:    movdqa %xmm4, %xmm15
; SSE-NEXT:    pslldq {{.*#+}} xmm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm15[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm9, %xmm13
; SSE-NEXT:    pandn %xmm15, %xmm13
; SSE-NEXT:    andps %xmm9, %xmm14
; SSE-NEXT:    por %xmm14, %xmm13
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm10[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[1,1],xmm1[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,0],xmm0[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,1,1]
; SSE-NEXT:    movdqa %xmm5, %xmm14
; SSE-NEXT:    pandn %xmm1, %xmm14
; SSE-NEXT:    andps %xmm5, %xmm15
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movaps %xmm2, %xmm1
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm10, %xmm15
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,1],xmm0[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0],xmm15[0,2]
; SSE-NEXT:    andps %xmm9, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm4
; SSE-NEXT:    pandn %xmm4, %xmm9
; SSE-NEXT:    por %xmm1, %xmm9
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm10 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[1,1],xmm2[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,0],xmm0[0,2]
; SSE-NEXT:    andps %xmm5, %xmm10
; SSE-NEXT:    pandn %xmm15, %xmm5
; SSE-NEXT:    por %xmm10, %xmm5
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa %xmm5, 352(%rax)
; SSE-NEXT:    movdqa %xmm9, 336(%rax)
; SSE-NEXT:    movdqa %xmm14, 304(%rax)
; SSE-NEXT:    movdqa %xmm13, 288(%rax)
; SSE-NEXT:    movdqa %xmm12, 256(%rax)
; SSE-NEXT:    movdqa %xmm11, 240(%rax)
; SSE-NEXT:    movdqa %xmm8, 208(%rax)
; SSE-NEXT:    movdqa %xmm3, 192(%rax)
; SSE-NEXT:    movdqa %xmm7, 160(%rax)
; SSE-NEXT:    movdqa %xmm6, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    addq $280, %rsp # imm = 0x118
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i16_stride6_vf32:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $120, %rsp
; AVX1-ONLY-NEXT:    vmovdqa 32(%rcx), %xmm8
; AVX1-ONLY-NEXT:    vmovdqa 48(%rcx), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa 32(%rdx), %xmm9
; AVX1-ONLY-NEXT:    vmovdqa 48(%rdx), %xmm1
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm6[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa 32(%rsi), %xmm10
; AVX1-ONLY-NEXT:    vmovdqa 48(%rsi), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa 32(%rdi), %xmm11
; AVX1-ONLY-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm7[2,3,2,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm3, %xmm0
; AVX1-ONLY-NEXT:    vmovdqa 48(%r8), %xmm12
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm0[0,1,2,3],xmm12[4,5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 48(%r9), %xmm0
; AVX1-ONLY-NEXT:    vpslld $16, %xmm0, %xmm5
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3,4],xmm5[5],xmm4[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm12[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm3 = xmm4[0],xmm3[1,2],xmm4[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2,3,4,5,6],xmm4[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm1[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm12[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3,4,5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm0[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3,4,5,6],xmm3[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm4[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 32(%r8), %xmm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,1,2,3]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm5, %xmm13
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm3[0,1],xmm13[2,3,4,5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 32(%r9), %xmm3
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm14[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0],xmm14[1],xmm13[2,3,4,5,6],xmm14[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm13 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0],xmm13[1],xmm5[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm3[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm13[3],xmm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm6[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm7[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3],ymm5[4],ymm6[5,6],ymm5[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm5, %xmm6
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm7 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3],xmm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2],xmm7[3],xmm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0,1],xmm12[0],xmm5[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm5[0,1,2,3,4],xmm0[5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm6[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2,3],ymm5[4],ymm7[5,6],ymm5[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm5, %xmm7
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm8 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3],xmm7[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0,1,2],xmm8[3],xmm7[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0,1],xmm2[0],xmm5[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa 16(%rcx), %xmm4
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm6[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm5, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm1
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm2[4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm3, %xmm5
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm5[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdx), %xmm1
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm0 = xmm2[0],xmm0[1,2],xmm2[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3,4,5,6],xmm2[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm11[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa 16(%rsi), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm10[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 16(%r8), %xmm3
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm2
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm2[0,1,2,3],xmm3[4,5],xmm2[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 16(%r9), %xmm2
; AVX1-ONLY-NEXT:    vpslld $16, %xmm2, %xmm6
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm6[5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm3[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm0 = xmm5[0],xmm0[1,2],xmm5[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm2[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm5[1],xmm0[2,3,4,5,6],xmm5[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm1[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm4
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2,3,4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3,4,5,6],xmm4[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm2[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm15 = xmm0[0,1,2],xmm1[3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm9
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm8
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm7[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm7[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm5
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm12 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm0
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm12, %xmm13
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm1[0,1],xmm13[2,3,4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm1[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm14[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0],xmm14[1],xmm13[2,3,4,5,6],xmm14[7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm14 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm14[1],xmm12[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm1[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0,1,2],xmm14[3],xmm12[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm11[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm14, %ymm11
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm10[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm14, %ymm10
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm11 = ymm10[0],ymm11[1],ymm10[2,3],ymm11[4],ymm10[5,6],ymm11[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm11, %xmm10
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm14 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm10[0,1],xmm14[2,3],xmm10[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm2[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm10[0,1,2],xmm14[3],xmm10[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm3 = xmm11[0,1],xmm3[0],xmm11[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1,2,3,4],xmm2[5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm3[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0],ymm8[1],ymm6[2,3],ymm8[4],ymm6[5,6],ymm8[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm6, %xmm8
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm9 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm9[2,3],xmm8[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1,2],xmm9[3],xmm8[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm6 = xmm6[0,1],xmm0[0],xmm6[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm9 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm1[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3,4],xmm9[5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm3, %xmm4
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3],xmm0[4,5],xmm4[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm1, %xmm5
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3,4],xmm5[5],xmm4[6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm3[1,2],xmm0[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3,4,5,6],xmm1[7]
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, 32(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, 48(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, (%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm8, 16(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, 96(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm10, 112(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm12, 64(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, 80(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm15, 160(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 176(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 128(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 144(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 224(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 240(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 192(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 208(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 288(%rax)
; AVX1-ONLY-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 304(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 256(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 272(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 352(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 368(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 320(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 336(%rax)
; AVX1-ONLY-NEXT:    addq $120, %rsp
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i16_stride6_vf32:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    subq $680, %rsp # imm = 0x2A8
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm3
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm4
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm5
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%r8), %xmm7
; AVX2-SLOW-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm2 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-SLOW-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa %ymm2, %ymm9
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa %xmm3, %xmm10
; AVX2-SLOW-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa %xmm4, %xmm8
; AVX2-SLOW-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm5[0,1,2,1]
; AVX2-SLOW-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm6[0,1,2,1]
; AVX2-SLOW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r9), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm7[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm9, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdx), %ymm2
; AVX2-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rcx), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-SLOW-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm3[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r8), %ymm14
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm14[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r9), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm9, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa %ymm9, %ymm7
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %ymm15
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %ymm13
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm13[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %ymm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm2[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm4 = ymm4[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm4[0],ymm1[0],ymm4[1],ymm1[1],ymm4[2],ymm1[2],ymm4[3],ymm1[3],ymm4[8],ymm1[8],ymm4[9],ymm1[9],ymm4[10],ymm1[10],ymm4[11],ymm1[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm4 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0],ymm0[1,2],ymm4[3],ymm0[4,5],ymm4[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm7, %ymm4, %ymm9, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1],ymm4[2],ymm9[3,4],ymm4[5],ymm9[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} xmm9 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %xmm9, %xmm0, %xmm12
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm4[0],ymm12[1],ymm4[2,3],ymm12[4],ymm4[5,6],ymm12[7]
; AVX2-SLOW-NEXT:    vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm4 = mem[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm4[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm4[0,1,0,1]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm1 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-SLOW-NEXT:    vpblendvb %ymm1, %ymm12, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa %ymm1, %ymm4
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm11[4],xmm8[5],xmm11[5],xmm8[6],xmm11[6],xmm8[7],xmm11[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm12 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm12 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm12[0,1],ymm0[2],ymm12[3,4],ymm0[5],ymm12[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %xmm9, %xmm5, %xmm9
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3],ymm9[4],ymm0[5,6],ymm9[7]
; AVX2-SLOW-NEXT:    vmovdqa (%rsp), %xmm10 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm10[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm9[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm0, %ymm9, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm0 = ymm3[4],mem[4],ymm3[5],mem[5],ymm3[6],mem[6],ymm3[7],mem[7],ymm3[12],mem[12],ymm3[13],mem[13],ymm3[14],mem[14],ymm3[15],mem[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[12],ymm7[12],ymm6[13],ymm7[13],ymm6[14],ymm7[14],ymm6[15],ymm7[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1],ymm0[2],ymm9[3,4],ymm0[5],ymm9[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm9 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX2-SLOW-NEXT:    vpshufb %ymm9, %ymm14, %ymm12
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3],ymm12[4],ymm0[5,6],ymm12[7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm3[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm0, %ymm12, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm12 = ymm2[4],ymm13[4],ymm2[5],ymm13[5],ymm2[6],ymm13[6],ymm2[7],ymm13[7],ymm2[12],ymm13[12],ymm2[13],ymm13[13],ymm2[14],ymm13[14],ymm2[15],ymm13[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm14 = ymm15[4],ymm4[4],ymm15[5],ymm4[5],ymm15[6],ymm4[6],ymm15[7],ymm4[7],ymm15[12],ymm4[12],ymm15[13],ymm4[13],ymm15[14],ymm4[14],ymm15[15],ymm4[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm14 = ymm14[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0,1],ymm12[2],ymm14[3,4],ymm12[5],ymm14[6,7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %ymm9, %ymm2, %ymm9
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3],ymm9[4],ymm12[5,6],ymm9[7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm9, %ymm12, %ymm15
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm12 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm12 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0],ymm12[1],ymm9[2,3],ymm12[4],ymm9[5,6],ymm12[7]
; AVX2-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm12 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1],ymm12[2],ymm9[3,4],ymm12[5],ymm9[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm10[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm14 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-SLOW-NEXT:    vpblendvb %ymm14, %ymm9, %ymm12, %ymm9
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm12 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm12 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm12[0],ymm0[1],ymm12[2,3],ymm0[4],ymm12[5,6],ymm0[7]
; AVX2-SLOW-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm12 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm12[2],ymm0[3,4],ymm12[5],ymm0[6,7]
; AVX2-SLOW-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm12 = mem[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX2-SLOW-NEXT:    vpblendvb %ymm14, %ymm0, %ymm12, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm12 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm12 = ymm5[0],mem[0],ymm5[1],mem[1],ymm5[2],mem[2],ymm5[3],mem[3],ymm5[8],mem[8],ymm5[9],mem[9],ymm5[10],mem[10],ymm5[11],mem[11]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm13 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[8],ymm7[8],ymm6[9],ymm7[9],ymm6[10],ymm7[10],ymm6[11],ymm7[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm13[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0],ymm13[1],ymm12[2,3],ymm13[4],ymm12[5,6],ymm13[7]
; AVX2-SLOW-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm10 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm12[0,1],ymm10[2],ymm12[3,4],ymm10[5],ymm12[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm14, %ymm10, %ymm8, %ymm8
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm3 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[2],mem[2],ymm3[3],mem[3],ymm3[8],mem[8],ymm3[9],mem[9],ymm3[10],mem[10],ymm3[11],mem[11]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3],ymm5[4],ymm3[5,6],ymm5[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm14, %ymm2, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa %ymm1, 96(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm15, 160(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm8, 288(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 192(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm9, (%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-SLOW-NEXT:    addq $680, %rsp # imm = 0x2A8
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i16_stride6_vf32:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    subq $616, %rsp # imm = 0x268
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rsi), %xmm4
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rcx), %xmm6
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rdx), %xmm7
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, (%rsp) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX2-FAST-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm11 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-NEXT:    vpblendvb %ymm11, %ymm1, %ymm3, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm4, %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm4, %xmm10
; AVX2-FAST-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-FAST-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX2-FAST-NEXT:    vmovdqa 32(%r8), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm3[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r9), %xmm3
; AVX2-FAST-NEXT:    vpshufb %xmm2, %xmm3, %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm3, %xmm8
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm11, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rdi), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-NEXT:    vmovdqa 32(%rdx), %ymm15
; AVX2-FAST-NEXT:    vmovdqa 32(%rcx), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r8), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r9), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX2-FAST-NEXT:    # ymm4 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm11, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %ymm2
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %ymm3
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm9[0],ymm1[0],ymm9[1],ymm1[1],ymm9[2],ymm1[2],ymm9[3],ymm1[3],ymm9[8],ymm1[8],ymm9[9],ymm1[9],ymm9[10],ymm1[10],ymm9[11],ymm1[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0],ymm0[1,2],ymm9[3],ymm0[4,5],ymm9[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm0, %ymm4
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm11, %ymm9, %ymm4, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm9 = <1,2,1,2,u,u,3,3>
; AVX2-FAST-NEXT:    vpermd %ymm4, %ymm9, %ymm4
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm5[4],xmm10[4],xmm5[5],xmm10[5],xmm5[6],xmm10[6],xmm5[7],xmm10[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm12[2],ymm4[3,4],ymm12[5],ymm4[6,7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm12 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm12, %xmm0, %xmm13
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm13 = ymm4[0],ymm13[1],ymm4[2,3],ymm13[4],ymm4[5,6],ymm13[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %xmm1, %xmm8, %xmm4
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm4[0,1,0,1]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-NEXT:    vpblendvb %ymm4, %ymm13, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vmovdqa %ymm4, %ymm6
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm9, %ymm0
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm4[4],xmm13[4],xmm4[5],xmm13[5],xmm4[6],xmm13[6],xmm4[7],xmm13[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3,4],ymm9[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm12, %xmm8, %xmm9
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3],ymm9[4],ymm0[5,6],ymm9[7]
; AVX2-FAST-NEXT:    vmovdqa (%rsp), %xmm5 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm6, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm15[4],ymm14[4],ymm15[5],ymm14[5],ymm15[6],ymm14[6],ymm15[7],ymm14[7],ymm15[12],ymm14[12],ymm15[13],ymm14[13],ymm15[14],ymm14[14],ymm15[15],ymm14[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = [5,6,5,6,5,6,7,7]
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[12],ymm11[12],ymm10[13],ymm11[13],ymm10[14],ymm11[14],ymm10[15],ymm11[15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3,4],ymm9[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3],ymm9[4],ymm0[5,6],ymm9[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm9, %ymm9
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm6, %ymm0, %ymm9, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
; AVX2-FAST-NEXT:    vmovdqa %ymm3, %ymm9
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[12],ymm7[12],ymm6[13],ymm7[13],ymm6[14],ymm7[14],ymm6[15],ymm7[15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-NEXT:    vpblendvb %ymm12, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm13[0],xmm4[1],xmm13[1],xmm4[2],xmm13[2],xmm4[3],xmm13[3]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm12 = [1,0,2,2,1,0,2,2]
; AVX2-FAST-NEXT:    # ymm12 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm12, %ymm1
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2,3],ymm1[4],ymm4[5,6],ymm1[7]
; AVX2-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm5[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm13 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-FAST-NEXT:    vpblendvb %ymm13, %ymm1, %ymm4, %ymm1
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm4 = xmm4[0],mem[0],xmm4[1],mem[1],xmm4[2],mem[2],xmm4[3],mem[3]
; AVX2-FAST-NEXT:    vpermd %ymm4, %ymm12, %ymm4
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm12 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm12 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm12[0],ymm4[1],ymm12[2,3],ymm4[4],ymm12[5,6],ymm4[7]
; AVX2-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm12 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm12[2],ymm4[3,4],ymm12[5],ymm4[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm12 = mem[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX2-FAST-NEXT:    vpblendvb %ymm13, %ymm4, %ymm12, %ymm4
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm15[0],ymm14[0],ymm15[1],ymm14[1],ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[8],ymm14[8],ymm15[9],ymm14[9],ymm15[10],ymm14[10],ymm15[11],ymm14[11]
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm14 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[8],ymm11[8],ymm10[9],ymm11[9],ymm10[10],ymm11[10],ymm10[11],ymm11[11]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm15 = [5,4,2,2,5,4,6,6]
; AVX2-FAST-NEXT:    vpermd %ymm12, %ymm15, %ymm12
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0],ymm12[1],ymm14[2,3],ymm12[4],ymm14[5,6],ymm12[7]
; AVX2-FAST-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm11 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm10 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm13, %ymm11, %ymm10, %ymm10
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm5[0],ymm9[0],ymm5[1],ymm9[1],ymm5[2],ymm9[2],ymm5[3],ymm9[3],ymm5[8],ymm9[8],ymm5[9],ymm9[9],ymm5[10],ymm9[10],ymm5[11],ymm9[11]
; AVX2-FAST-NEXT:    vpermd %ymm5, %ymm15, %ymm5
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[8],ymm7[8],ymm6[9],ymm7[9],ymm6[10],ymm7[10],ymm6[11],ymm7[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3],ymm5[4],ymm6[5,6],ymm5[7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm13, %ymm3, %ymm2, %ymm2
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa %ymm2, 96(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 160(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm10, 288(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 352(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm4, 192(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm1, (%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-NEXT:    addq $616, %rsp # imm = 0x268
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i16_stride6_vf32:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    subq $680, %rsp # imm = 0x2A8
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rsi), %xmm10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rcx), %xmm6
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm3, (%rsp) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdx), %xmm7
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r9), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm12 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm12, %ymm2, %ymm3, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm10, %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm6, %xmm13
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r8), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm3, %xmm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r9), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm12, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rsi), %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm9, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm3, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdx), %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rcx), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r8), %ymm15
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r9), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX2-FAST-PERLANE-NEXT:    # ymm8 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm8, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm12, %ymm1, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm2, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm11[0],ymm1[0],ymm11[1],ymm1[1],ymm11[2],ymm1[2],ymm11[3],ymm1[3],ymm11[8],ymm1[8],ymm11[9],ymm1[9],ymm11[10],ymm1[10],ymm11[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0],ymm0[1,2],ymm11[3],ymm0[4,5],ymm11[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r9), %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm8, %ymm0, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm12, %ymm11, %ymm8, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm5[4],xmm10[4],xmm5[5],xmm10[5],xmm5[6],xmm10[6],xmm5[7],xmm10[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm7[4],xmm13[4],xmm7[5],xmm13[5],xmm7[6],xmm13[6],xmm7[7],xmm13[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm11[0,1],ymm8[2],ymm11[3,4],ymm8[5],ymm11[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm11 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm11, %xmm6, %xmm13
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0],ymm13[1],ymm8[2,3],ymm13[4],ymm8[5,6],ymm13[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm13 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm13, %xmm4, %xmm14
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm7 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm7, %ymm8, %ymm14, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm14 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm14 = xmm14[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1],ymm8[2],ymm14[3,4],ymm8[5],ymm14[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0],ymm11[1],ymm8[2,3],ymm11[4],ymm8[5,6],ymm11[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm13, %xmm4, %xmm11
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm7, %ymm8, %ymm11, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm7, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm9, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm12[4],ymm9[4],ymm12[5],ymm9[5],ymm12[6],ymm9[6],ymm12[7],ymm9[7],ymm12[12],ymm9[12],ymm12[13],ymm9[13],ymm12[14],ymm9[14],ymm12[15],ymm9[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} ymm13 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm13 = ymm13[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm13[0,1],ymm8[2],ymm13[3,4],ymm8[5],ymm13[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm13 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm13, %ymm15, %ymm14
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0],ymm14[1],ymm8[2,3],ymm14[4],ymm8[5,6],ymm14[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm15 = ymm6[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm11, %ymm8, %ymm15, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} ymm15 = ymm14[4],ymm8[4],ymm14[5],ymm8[5],ymm14[6],ymm8[6],ymm14[7],ymm8[7],ymm14[12],ymm8[12],ymm14[13],ymm8[13],ymm14[14],ymm8[14],ymm14[15],ymm8[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm11 = ymm11[4],mem[4],ymm11[5],mem[5],ymm11[6],mem[6],ymm11[7],mem[7],ymm11[12],mem[12],ymm11[13],mem[13],ymm11[14],mem[14],ymm11[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm11 = ymm11[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm15[2],ymm11[3,4],ymm15[5],ymm11[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm13, %ymm15, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0],ymm13[1],ymm11[2,3],ymm13[4],ymm11[5,6],ymm13[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm13 = ymm13[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm15 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm15, %ymm11, %ymm13, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0],ymm13[1],ymm11[2,3],ymm13[4],ymm11[5,6],ymm13[7]
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{.*#+}} xmm13 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm13, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm13[2],ymm11[3,4],ymm13[5],ymm11[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm4[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm13, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm0, %ymm11, %ymm13, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm13 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm13 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm15 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm15 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm13 = ymm13[0],ymm15[1],ymm13[2,3],ymm15[4],ymm13[5,6],ymm15[7]
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm15 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4],ymm15[5],ymm13[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm15 = mem[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm0, %ymm13, %ymm15, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm15 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[8],ymm10[8],ymm12[9],ymm10[9],ymm12[10],ymm10[10],ymm12[11],ymm10[11]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm10 = ymm10[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6],ymm10[7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm9 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1],ymm9[2],ymm10[3,4],ymm9[5],ymm10[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm0, %ymm9, %ymm7, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm14[0],ymm8[0],ymm14[1],ymm8[1],ymm14[2],ymm8[2],ymm14[3],ymm8[3],ymm14[8],ymm8[8],ymm14[9],ymm8[9],ymm14[10],ymm8[10],ymm14[11],ymm8[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm3 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm5[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm2 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm1 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm0, %ymm2, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm1, 96(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm7, 288(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 352(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm13, 192(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm11, (%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    addq $680, %rsp # imm = 0x2A8
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-ONLY-SLOW-LABEL: store_i16_stride6_vf32:
; AVX512F-ONLY-SLOW:       # %bb.0:
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rsi), %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm24
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm23
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rsi), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdi), %ymm14
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm14[0],ymm2[0],ymm14[1],ymm2[1],ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[8],ymm2[8],ymm14[9],ymm2[9],ymm14[10],ymm2[10],ymm14[11],ymm2[11]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %ymm2, %ymm31
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rcx), %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdx), %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm3, %xmm25
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm26
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rcx), %ymm8
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdx), %ymm11
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm11[0],ymm8[0],ymm11[1],ymm8[1],ymm11[2],ymm8[2],ymm11[3],ymm8[3],ymm11[8],ymm8[8],ymm11[9],ymm8[9],ymm11[10],ymm8[10],ymm11[11],ymm8[11]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-SLOW-NEXT:    kmovw %eax, %k1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm1, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r8), %ymm6
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r8), %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{.*#+}} xmm13 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %xmm13, %xmm3, %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm3, %xmm27
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm17 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r9), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm28
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm18 = ymm0[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r9), %ymm12
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm19 = ymm0[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm5, %xmm30
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm4, %xmm29
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rsi), %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm0[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm15
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm5[4],xmm15[5],xmm5[5],xmm15[6],xmm5[6],xmm15[7],xmm5[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm16 = xmm0[1,2,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rcx), %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdx), %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm10[1,0,2,2,5,4,6,6]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm16[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm10, %zmm7, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm10 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r8), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm10, %ymm9
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm9[0,1],ymm7[2],ymm9[3,4],ymm7[5],ymm9[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm9
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r8), %xmm7
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %xmm13, %xmm7, %xmm13
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm10[0],ymm13[1],ymm10[2,3],ymm13[4],ymm10[5,6],ymm13[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm20 = zmm10[0,1,2,3],zmm9[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm4[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm3[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm10[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm10[0],ymm0[0],ymm10[1],ymm0[1],ymm10[2],ymm0[2],ymm10[3],ymm0[3],ymm10[8],ymm0[8],ymm10[9],ymm0[9],ymm10[10],ymm0[10],ymm10[11],ymm0[11]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r9), %xmm0
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm0[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm21 = ymm13[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm10, %zmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm1, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %ymm1, %ymm2, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm4, %ymm9
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0],ymm3[1],ymm9[2,3],ymm3[4],ymm9[5,6],ymm3[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r9), %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm22 = ymm10[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5],ymm2[6],ymm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm16 = zmm2[0,1,2,3],zmm9[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm4[0],ymm2[0],ymm4[1],ymm2[1],ymm4[2],ymm2[2],ymm4[3],ymm2[3],ymm4[8],ymm2[8],ymm4[9],ymm2[9],ymm4[10],ymm2[10],ymm4[11],ymm2[11]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm11[4],ymm8[4],ymm11[5],ymm8[5],ymm11[6],ymm8[6],ymm11[7],ymm8[7],ymm11[12],ymm8[12],ymm11[13],ymm8[13],ymm11[14],ymm8[14],ymm11[15],ymm8[15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %ymm31, %ymm10
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm31[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm14[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm9[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} ymm8 = ymm8[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm3[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm4[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm9[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm14[4],ymm10[4],ymm14[5],ymm10[5],ymm14[6],ymm10[6],ymm14[7],ymm10[7],ymm14[12],ymm10[12],ymm14[13],ymm10[13],ymm14[14],ymm10[14],ymm14[15],ymm10[15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm9[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %ymm1, %ymm6, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm3, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm12[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm2[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm12[2,3,2,3,6,7,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm2[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,0,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm29, %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm30, %xmm5
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm29[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm30[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512F-ONLY-SLOW-NEXT:    kmovw %eax, %k1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm3, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm7[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0],ymm2[1,2],ymm4[3],ymm2[4,5],ymm4[6],ymm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm25, %xmm6
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm26, %xmm5
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,0,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm5, %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm24, %xmm6
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm23, %xmm7
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm24[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm7 = xmm23[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm5, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm4, %zmm5 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm5, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm27, %xmm7
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm7[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm6[0],ymm4[1,2],ymm6[3],ymm4[4,5],ymm6[6],ymm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm4
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm27[0],zero,xmm27[1],zero,xmm27[2],zero,xmm27[3],zero
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm5[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm28, %xmm6
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm6[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm5, %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm19, %zmm18, %zmm7
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm17, %zmm9, %zmm7
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm22, %zmm21, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm20, %zmm9, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm8, %zmm8
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm16, %zmm9, %zmm8
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm13, %zmm11
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm1, %zmm9, %zmm11
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm2, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm5, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm4, %zmm1, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm8, 320(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm10, 256(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm7, 64(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vzeroupper
; AVX512F-ONLY-SLOW-NEXT:    retq
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride6_vf32:
; AVX512F-ONLY-FAST:       # %bb.0:
; AVX512F-ONLY-FAST-NEXT:    pushq %rax
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rcx), %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdx), %ymm12
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm12[0],ymm1[0],ymm12[1],ymm1[1],ymm12[2],ymm1[2],ymm12[3],ymm1[3],ymm12[8],ymm1[8],ymm12[9],ymm1[9],ymm12[10],ymm1[10],ymm12[11],ymm1[11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm1, %ymm26
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rcx), %xmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rcx), %xmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdx), %xmm6
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm20 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
; AVX512F-ONLY-FAST-NEXT:    vpermd %zmm0, %zmm20, %zmm23
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rsi), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rsi), %ymm10
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdi), %ymm5
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm10[0],ymm5[1],ymm10[1],ymm5[2],ymm10[2],ymm5[3],ymm10[3],ymm5[8],ymm10[8],ymm5[9],ymm10[9],ymm5[10],ymm10[10],ymm5[11],ymm10[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-ONLY-FAST-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-FAST-NEXT:    kmovw %eax, %k1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm0, %zmm23 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm22 = [8,9,20,11,12,21,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r8), %ymm9
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm23, %zmm21
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm0, %zmm22, %zmm21
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r8), %xmm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rcx), %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdx), %ymm15
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[1],ymm1[1],ymm15[2],ymm1[2],ymm15[3],ymm1[3],ymm15[8],ymm1[8],ymm15[9],ymm1[9],ymm15[10],ymm1[10],ymm15[11],ymm1[11]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm6, %xmm18
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm4, %xmm17
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm14 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm14, %xmm3, %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm3, %xmm27
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm16 = [0,9,2,3,8,5,6,11]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm0, %ymm16, %ymm23
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r9), %xmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm3, %xmm4, %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm4, %xmm28
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r9), %ymm4
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm30 = ymm0[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rsi), %xmm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm19 = ymm0[1,1,1,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rsi), %ymm11
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm0[0],ymm11[0],ymm0[1],ymm11[1],ymm0[2],ymm11[2],ymm0[3],ymm11[3],ymm0[8],ymm11[8],ymm0[9],ymm11[9],ymm0[10],ymm11[10],ymm0[11],ymm11[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermd %zmm2, %zmm20, %zmm25
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm19, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm2, %zmm25 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r8), %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermi2d %zmm8, %zmm25, %zmm22
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r8), %xmm8
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm14, %xmm8, %xmm14
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm14, %ymm16, %ymm25
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r9), %xmm14
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm20 = ymm3[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm24 = [5,6,5,6,5,6,7,7]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm15[4],ymm1[4],ymm15[5],ymm1[5],ymm15[6],ymm1[6],ymm15[7],ymm1[7],ymm15[12],ymm1[12],ymm15[13],ymm1[13],ymm15[14],ymm1[14],ymm15[15],ymm1[15]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm3, %ymm24, %ymm16
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm15 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm15[0],ymm1[0],ymm15[1],ymm1[1],ymm15[2],ymm1[2],ymm15[3],ymm1[3],ymm15[8],ymm1[8],ymm15[9],ymm1[9],ymm15[10],ymm1[10],ymm15[11],ymm1[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm1[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm1, %ymm11, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm1, %ymm0, %ymm13
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm13[0],ymm3[0],ymm13[1],ymm3[1],ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[8],ymm3[8],ymm13[9],ymm3[9],ymm13[10],ymm3[10],ymm13[11],ymm3[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm11[4],ymm0[5],ymm11[5],ymm0[6],ymm11[6],ymm0[7],ymm11[7],ymm0[12],ymm11[12],ymm0[13],ymm11[13],ymm0[14],ymm11[14],ymm0[15],ymm11[15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm16, %zmm15, %zmm11
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm3, %zmm15
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm11, %zmm15 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm16 = [8,21,10,11,20,13,14,23]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm0, %ymm2, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm3, %zmm16, %zmm19
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm31 = [12,1,2,13,4,5,14,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm2, %ymm31, %ymm15
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm26, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm12[4],ymm3[4],ymm12[5],ymm3[5],ymm12[6],ymm3[6],ymm12[7],ymm3[7],ymm12[12],ymm3[12],ymm12[13],ymm3[13],ymm12[14],ymm3[14],ymm12[15],ymm3[15]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm2, %ymm24, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r9), %ymm11
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm13 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm24 = ymm13[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm13 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[8],ymm13[8],ymm12[9],ymm13[9],ymm12[10],ymm13[10],ymm12[11],ymm13[11]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm1, %ymm10, %ymm13
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm1, %ymm5, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm13[0],ymm1[1],ymm13[1],ymm1[2],ymm13[2],ymm1[3],ymm13[3],ymm1[8],ymm13[8],ymm1[9],ymm13[9],ymm1[10],ymm13[10],ymm1[11],ymm13[11]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm5[4],ymm10[4],ymm5[5],ymm10[5],ymm5[6],ymm10[6],ymm5[7],ymm10[7],ymm5[12],ymm10[12],ymm5[13],ymm10[13],ymm5[14],ymm10[14],ymm5[15],ymm10[15]
; AVX512F-ONLY-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm13 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512F-ONLY-FAST-NEXT:    # ymm13 = mem[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm13, %ymm11, %ymm10
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm26 = ymm10[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm12, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm1, %zmm5
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm2, %zmm5 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm0, %ymm9, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpermi2d %zmm0, %zmm5, %zmm16
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm0, %ymm11, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm29 = ymm1[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm1, %ymm31, %ymm5
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm13, %ymm4, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm31 = ymm1[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm17, %xmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm18, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm7, %xmm4
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm6, %xmm9
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm9[4],xmm4[4],xmm9[5],xmm4[5],xmm9[6],xmm4[6],xmm9[7],xmm4[7]
; AVX512F-ONLY-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm13 = [1,0,2,2,1,0,2,2]
; AVX512F-ONLY-FAST-NEXT:    # ymm13 = mem[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm1, %ymm13, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm6[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm1, %zmm1
; AVX512F-ONLY-FAST-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512F-ONLY-FAST-NEXT:    kmovw %eax, %k1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [16,9,10,17,12,13,18,15]
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm8[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm1, %zmm9
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm4, %zmm2, %zmm9
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm14[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm7, %ymm13, %ymm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm8 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm8, %xmm14, %xmm13
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm14 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm10 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm10, %ymm10
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm3, %xmm14
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm11, %xmm12
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm14[4],xmm12[5],xmm14[5],xmm12[6],xmm14[6],xmm12[7],xmm14[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm10, %zmm7, %zmm7
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm10 = ymm14[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm12, %zmm10, %zmm10
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm7, %zmm10 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm27, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm3[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermi2d %zmm7, %zmm10, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm7 = xmm27[0],zero,xmm27[1],zero,xmm27[2],zero,xmm27[3],zero
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm28, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm14 = zmm23[0,1,2,3],zmm21[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm30, %zmm3, %zmm17
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm14, %zmm18, %zmm17
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm14 = zmm25[0,1,2,3],zmm22[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm24, %zmm20, %zmm20
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm14, %zmm18, %zmm20
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm14 = zmm15[0,1,2,3],zmm19[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm29, %zmm26, %zmm11
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm15 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm14, %zmm15, %zmm11
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm16[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm31, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm5, %zmm15, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,8,3,4,9,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm4, %ymm3, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm9[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm13, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm1, %zmm5, %zmm4
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm7, %ymm3, %ymm10
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm10[0,1,2,3],zmm2[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm12, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm1, %zmm5, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm4, 192(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm0, 128(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm11, 320(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm20, 256(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm17, 64(%rax)
; AVX512F-ONLY-FAST-NEXT:    popq %rax
; AVX512F-ONLY-FAST-NEXT:    vzeroupper
; AVX512F-ONLY-FAST-NEXT:    retq
;
; AVX512DQ-SLOW-LABEL: store_i16_stride6_vf32:
; AVX512DQ-SLOW:       # %bb.0:
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rcx), %xmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm4
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdx), %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm5
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm23
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm24
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rcx), %ymm2
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdx), %ymm14
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm14[0],ymm2[0],ymm14[1],ymm2[1],ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[8],ymm2[8],ymm14[9],ymm2[9],ymm14[10],ymm2[10],ymm14[11],ymm2[11]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %ymm2, %ymm31
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rsi), %xmm3
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdi), %xmm2
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm3, %xmm28
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm27
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rsi), %ymm8
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdi), %ymm11
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm11[0],ymm8[0],ymm11[1],ymm8[1],ymm11[2],ymm8[2],ymm11[3],ymm8[3],ymm11[8],ymm8[8],ymm11[9],ymm8[9],ymm11[10],ymm8[10],ymm11[11],ymm8[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512DQ-SLOW-NEXT:    kmovw %eax, %k1
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r8), %ymm6
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r8), %xmm3
; AVX512DQ-SLOW-NEXT:    vmovdqa {{.*#+}} xmm13 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufb %xmm13, %xmm3, %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm3, %xmm25
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm17 = zmm0[0,1,2,3],zmm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r9), %xmm1
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm26
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm18 = ymm0[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r9), %ymm12
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm19 = ymm0[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm5, %xmm30
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm4, %xmm29
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,2,3,3]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rcx), %ymm4
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdx), %ymm3
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm0[1,0,2,2,5,4,6,6]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm5
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm15
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm5[4],xmm15[5],xmm5[5],xmm15[6],xmm5[6],xmm15[7],xmm5[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm16 = ymm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm7
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm10, %zmm16, %zmm7 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r8), %ymm2
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm7, %ymm9
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm9
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r8), %xmm10
; AVX512DQ-SLOW-NEXT:    vpshufb %xmm13, %xmm10, %xmm13
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm7[0],ymm13[1],ymm7[2,3],ymm13[4],ymm7[5,6],ymm13[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm20 = zmm7[0,1,2,3],zmm9[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm1[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm7[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm0[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm9[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[1],ymm0[1],ymm9[2],ymm0[2],ymm9[3],ymm0[3],ymm9[8],ymm0[8],ymm9[9],ymm0[9],ymm9[10],ymm0[10],ymm9[11],ymm0[11]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r9), %xmm0
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm0[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm21 = ymm13[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm7, %zmm4
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm3, %zmm9, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512DQ-SLOW-NEXT:    vpshufb %ymm1, %ymm2, %ymm3
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm4, %ymm7
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm7[0],ymm3[1],ymm7[2,3],ymm3[4],ymm7[5,6],ymm3[7]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r9), %ymm7
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm22 = ymm9[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5],ymm2[6],ymm4[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm16 = zmm2[0,1,2,3],zmm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm8[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm11[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm11[4],ymm8[4],ymm11[5],ymm8[5],ymm11[6],ymm8[6],ymm11[7],ymm8[7],ymm11[12],ymm8[12],ymm11[13],ymm8[13],ymm11[14],ymm8[14],ymm11[15],ymm8[15]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %ymm31, %ymm9
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm8[0],ymm4[0],ymm8[1],ymm4[1],ymm8[2],ymm4[2],ymm8[3],ymm4[3],ymm8[8],ymm4[8],ymm8[9],ymm4[9],ymm8[10],ymm4[10],ymm8[11],ymm4[11]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} ymm8 = ymm8[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[2,3,2,3,6,7,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm7[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm7[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm7 = ymm14[4],ymm9[4],ymm14[5],ymm9[5],ymm14[6],ymm9[6],ymm14[7],ymm9[7],ymm14[12],ymm9[12],ymm14[13],ymm9[13],ymm14[14],ymm9[14],ymm14[15],ymm9[15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm7[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm3, %zmm4, %zmm2 {%k1}
; AVX512DQ-SLOW-NEXT:    vpshufb %ymm1, %ymm6, %ymm1
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm2, %ymm3
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6],ymm1[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm12[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm12[2,3,2,3,6,7,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm2[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm15[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm29, %xmm5
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm30, %xmm6
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,0,2,2]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm4[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX512DQ-SLOW-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512DQ-SLOW-NEXT:    kmovw %eax, %k1
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm4, %zmm3, %zmm2 {%k1}
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm2, %ymm3
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm10[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm0[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm28, %xmm5
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm27, %xmm6
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm28[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm27[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm23, %xmm7
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm24, %xmm9
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,0,2,2]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm6[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm6, %zmm5, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm4, %ymm5
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm25, %xmm7
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm7[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1,2],ymm6[3],ymm5[4,5],ymm6[6],ymm5[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
; AVX512DQ-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm25[0],zero,xmm25[1],zero,xmm25[2],zero,xmm25[3],zero
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm26, %xmm6
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm6[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm5, %ymm5
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm19, %zmm18, %zmm7
; AVX512DQ-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm17, %zmm9, %zmm7
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm22, %zmm21, %zmm10
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm20, %zmm9, %zmm10
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm8, %zmm8
; AVX512DQ-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm16, %zmm9, %zmm8
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm13, %zmm11
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm1, %zmm9, %zmm11
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512DQ-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm2, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm5, %zmm2
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm4, %zmm1, %zmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm8, 320(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm10, 256(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm7, 64(%rax)
; AVX512DQ-SLOW-NEXT:    vzeroupper
; AVX512DQ-SLOW-NEXT:    retq
;
; AVX512DQ-FAST-LABEL: store_i16_stride6_vf32:
; AVX512DQ-FAST:       # %bb.0:
; AVX512DQ-FAST-NEXT:    subq $40, %rsp
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rsi), %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX512DQ-FAST-NEXT:    vpshufb %ymm0, %ymm3, %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdi), %ymm4
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm0, %ymm4, %ymm2
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm22
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rcx), %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdx), %ymm5
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm3[4],ymm5[5],ymm3[5],ymm5[6],ymm3[6],ymm5[7],ymm3[7],ymm5[12],ymm3[12],ymm5[13],ymm3[13],ymm5[14],ymm3[14],ymm5[15],ymm3[15]
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [5,6,5,6,5,6,7,7]
; AVX512DQ-FAST-NEXT:    vpermd %ymm1, %ymm2, %ymm1
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm5, %ymm29
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX512DQ-FAST-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512DQ-FAST-NEXT:    kmovw %eax, %k1
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm1, %zmm3, %zmm22 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm21 = [8,21,10,11,20,13,14,23]
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r8), %ymm4
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512DQ-FAST-NEXT:    vpshufb %ymm1, %ymm4, %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm22, %zmm20
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm3, %zmm21, %zmm20
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm4, %ymm31
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm9 = [12,1,2,13,4,5,14,7]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm3, %ymm9, %ymm22
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r9), %ymm5
; AVX512DQ-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512DQ-FAST-NEXT:    # ymm3 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm3, %ymm5, %ymm4
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX512DQ-FAST-NEXT:    vpshufb %ymm12, %ymm5, %ymm4
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm5, %ymm17
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm30 = ymm4[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa (%rsi), %ymm5
; AVX512DQ-FAST-NEXT:    vpshufb %ymm0, %ymm5, %ymm4
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512DQ-FAST-NEXT:    vpshufb %ymm0, %ymm15, %ymm0
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[8],ymm4[8],ymm0[9],ymm4[9],ymm0[10],ymm4[10],ymm0[11],ymm4[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm15[4],ymm5[4],ymm15[5],ymm5[5],ymm15[6],ymm5[6],ymm15[7],ymm5[7],ymm15[12],ymm5[12],ymm15[13],ymm5[13],ymm15[14],ymm5[14],ymm15[15],ymm5[15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm5, %ymm16
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[3,3,3,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm23
; AVX512DQ-FAST-NEXT:    vmovdqa (%rcx), %ymm13
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdx), %ymm4
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm4[4],ymm13[4],ymm4[5],ymm13[5],ymm4[6],ymm13[6],ymm4[7],ymm13[7],ymm4[12],ymm13[12],ymm4[13],ymm13[13],ymm4[14],ymm13[14],ymm4[15],ymm13[15]
; AVX512DQ-FAST-NEXT:    vpermd %ymm0, %ymm2, %ymm0
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm10[0],ymm2[0],ymm10[1],ymm2[1],ymm10[2],ymm2[2],ymm10[3],ymm2[3],ymm10[8],ymm2[8],ymm10[9],ymm2[9],ymm10[10],ymm2[10],ymm10[11],ymm2[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm0, %zmm2, %zmm23 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa (%r8), %ymm10
; AVX512DQ-FAST-NEXT:    vpshufb %ymm1, %ymm10, %ymm0
; AVX512DQ-FAST-NEXT:    vpermi2d %zmm0, %zmm23, %zmm21
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm0, %ymm9, %ymm23
; AVX512DQ-FAST-NEXT:    vmovdqa (%r9), %ymm9
; AVX512DQ-FAST-NEXT:    vpshufb %ymm3, %ymm9, %ymm0
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm12, %ymm9, %ymm0
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm19 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rsi), %xmm12
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm6 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512DQ-FAST-NEXT:    vmovdqa (%rcx), %xmm8
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdx), %xmm3
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm4[0],ymm13[0],ymm4[1],ymm13[1],ymm4[2],ymm13[2],ymm4[3],ymm13[3],ymm4[8],ymm13[8],ymm4[9],ymm13[9],ymm4[10],ymm13[10],ymm4[11],ymm13[11]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm4, %zmm24
; AVX512DQ-FAST-NEXT:    vpshufb %xmm6, %xmm12, %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdi), %xmm13
; AVX512DQ-FAST-NEXT:    vpshufb %xmm6, %xmm13, %xmm4
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm25 = ymm0[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rcx), %xmm11
; AVX512DQ-FAST-NEXT:    vbroadcasti32x4 {{.*#+}} ymm26 = [1,0,2,2,1,0,2,2]
; AVX512DQ-FAST-NEXT:    # ymm26 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdx), %xmm2
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3]
; AVX512DQ-FAST-NEXT:    vpermd %ymm5, %ymm26, %ymm5
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm25, %zmm7
; AVX512DQ-FAST-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512DQ-FAST-NEXT:    kmovw %eax, %k2
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm0, %zmm5, %zmm7 {%k2}
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm27 = [16,9,10,17,12,13,18,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm7, %zmm25
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r8), %xmm4
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm4[2,1,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm0, %zmm27, %zmm25
; AVX512DQ-FAST-NEXT:    vmovdqa (%rsi), %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdi), %xmm5
; AVX512DQ-FAST-NEXT:    vpshufb %xmm6, %xmm0, %xmm1
; AVX512DQ-FAST-NEXT:    vpshufb %xmm6, %xmm5, %xmm6
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3]
; AVX512DQ-FAST-NEXT:    vpermd %ymm6, %ymm26, %ymm26
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r9), %xmm6
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm6[0,0,2,1,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm14, %ymm28
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm8, %zmm1
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm3, %zmm26, %zmm1 {%k2}
; AVX512DQ-FAST-NEXT:    vmovdqa (%r8), %xmm3
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm3[2,1,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpermi2d %zmm8, %zmm1, %zmm27
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm16, %ymm5
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm15[0],ymm5[0],ymm15[1],ymm5[1],ymm15[2],ymm5[2],ymm15[3],ymm5[3],ymm15[8],ymm5[8],ymm15[9],ymm5[9],ymm15[10],ymm5[10],ymm15[11],ymm5[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm8 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
; AVX512DQ-FAST-NEXT:    vpermd %zmm24, %zmm8, %zmm15
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm5, %zmm0, %zmm15 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = [8,9,20,11,12,21,14,15]
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm5 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm15, %zmm10
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm5, %zmm0, %zmm10
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm29, %ymm14
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm5 = ymm14[0],mem[0],ymm14[1],mem[1],ymm14[2],mem[2],ymm14[3],mem[3],ymm14[8],mem[8],ymm14[9],mem[9],ymm14[10],mem[10],ymm14[11],mem[11]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm11 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm2, %zmm2
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm5, %xmm3, %xmm14
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm24 = [0,9,2,3,8,5,6,11]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm14, %ymm24, %ymm15
; AVX512DQ-FAST-NEXT:    vpshufb %xmm11, %xmm6, %xmm14
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm26 = ymm14[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
; AVX512DQ-FAST-NEXT:    vpermd %zmm2, %zmm8, %zmm2
; AVX512DQ-FAST-NEXT:    vmovdqa (%r9), %xmm8
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm8[0,0,2,1,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm13, %ymm29
; AVX512DQ-FAST-NEXT:    vpshufb %xmm11, %xmm8, %xmm11
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm14 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm14 = ymm13[0],mem[0],ymm13[1],mem[1],ymm13[2],mem[2],ymm13[3],mem[3],ymm13[8],mem[8],ymm13[9],mem[9],ymm13[10],mem[10],ymm13[11],mem[11]
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm13 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm13, %xmm8, %xmm8
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm9[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[1,1,1,1]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm14, %zmm12, %zmm2 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm31, %ymm12
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermi2d %zmm12, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm12 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
; AVX512DQ-FAST-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm4, %ymm24, %ymm2
; AVX512DQ-FAST-NEXT:    vpshufb %xmm13, %xmm6, %xmm4
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm17, %ymm5
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm5 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm22[0,1,2,3],zmm20[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm30, %zmm13, %zmm13
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm14 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm6, %zmm14, %zmm13
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm6 = zmm23[0,1,2,3],zmm21[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm19, %zmm18, %zmm16
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm6, %zmm14, %zmm16
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,8,3,4,9,6,7]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm12, %ymm6, %ymm7
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm3, %ymm6, %ymm1
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm7[0,1,2,3],zmm25[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm26, %zmm28, %zmm6
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm3, %zmm7, %zmm6
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm27[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm11, %zmm29, %zmm3
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm1, %zmm7, %zmm3
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm15[0,1,2,3],zmm10[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm7
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm1, %zmm8, %zmm7
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm0[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm1
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm0, %zmm8, %zmm1
; AVX512DQ-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm1, 256(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm7, 64(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm16, 128(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm13, 320(%rax)
; AVX512DQ-FAST-NEXT:    addq $40, %rsp
; AVX512DQ-FAST-NEXT:    vzeroupper
; AVX512DQ-FAST-NEXT:    retq
;
; AVX512BW-LABEL: store_i16_stride6_vf32:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 (%rdx), %zmm4
; AVX512BW-NEXT:    vmovdqa64 (%rcx), %zmm5
; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm1
; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm0
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,32,3,35,0,0,1,33,4,36,0,0,2,34,0,0,0,32,3,35,0,0,1,33,4,36,0,0,2,34]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm5, %zmm4, %zmm6
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,32,3,35,0,0,1,33,4,36,0,0,2,34,5,37,0,32,3,35,0,0,1,33,4,36,0,0,2,34,5,37]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm2, %zmm7
; AVX512BW-NEXT:    movw $9362, %cx # imm = 0x2492
; AVX512BW-NEXT:    kmovd %ecx, %k2
; AVX512BW-NEXT:    vmovdqa32 %zmm6, %zmm7 {%k2}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm8 = <0,1,2,3,32,u,6,7,8,9,33,u,12,13,14,15,34,u,18,19,20,21,35,u,24,25,26,27,36,u,30,31>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [0,1,2,3,4,32,6,7,8,9,10,33,12,13,14,15,16,34,18,19,20,21,22,35,24,25,26,27,28,36,30,31]
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm8, %zmm6
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [8,40,0,0,6,38,9,41,0,0,7,39,10,42,0,0,8,40,0,0,6,38,9,41,0,0,7,39,10,42,0,0]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm2, %zmm7
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [5,37,8,40,0,0,6,38,9,41,0,0,7,39,10,42,5,37,8,40,0,0,6,38,9,41,0,0,7,39,10,42]
; AVX512BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm5, %zmm4, %zmm8
; AVX512BW-NEXT:    movw $18724, %cx # imm = 0x4924
; AVX512BW-NEXT:    kmovd %ecx, %k1
; AVX512BW-NEXT:    vmovdqa32 %zmm7, %zmm8 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = <0,1,37,u,4,5,6,7,38,u,10,11,12,13,39,u,16,17,18,19,40,u,22,23,24,25,41,u,28,29,30,31>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm8, %zmm7
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [0,1,2,37,4,5,6,7,8,38,10,11,12,13,14,39,16,17,18,19,20,40,22,23,24,25,26,41,28,29,30,31]
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm7, %zmm8
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [13,45,0,0,11,43,14,46,0,0,12,44,15,47,0,0,13,45,0,0,11,43,14,46,0,0,12,44,15,47,0,0]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm5, %zmm4, %zmm7
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,11,43,14,46,0,0,12,44,15,47,0,0,13,45,0,0,11,43,14,46,0,0,12,44,15,47,0,0,13,45]
; AVX512BW-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm2, %zmm9
; AVX512BW-NEXT:    vmovdqa32 %zmm7, %zmm9 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = <42,u,2,3,4,5,43,u,8,9,10,11,44,u,14,15,16,17,45,u,20,21,22,23,46,u,26,27,28,29,47,u>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm9, %zmm7
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [0,42,2,3,4,5,6,43,8,9,10,11,12,44,14,15,16,17,18,45,20,21,22,23,24,46,26,27,28,29,30,47]
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm7, %zmm9
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,16,48,19,51,0,0,17,49,20,52,0,0,18,50,0,0,16,48,19,51,0,0,17,49,20,52,0,0,18,50]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm5, %zmm4, %zmm7
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [16,48,19,51,0,0,17,49,20,52,0,0,18,50,21,53,16,48,19,51,0,0,17,49,20,52,0,0,18,50,21,53]
; AVX512BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm2, %zmm10
; AVX512BW-NEXT:    vmovdqa32 %zmm7, %zmm10 {%k2}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = <0,1,2,3,48,u,6,7,8,9,49,u,12,13,14,15,50,u,18,19,20,21,51,u,24,25,26,27,52,u,30,31>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm10, %zmm7
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [0,1,2,3,4,48,6,7,8,9,10,49,12,13,14,15,16,50,18,19,20,21,22,51,24,25,26,27,28,52,30,31]
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm7, %zmm10
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [24,56,0,0,22,54,25,57,0,0,23,55,26,58,0,0,24,56,0,0,22,54,25,57,0,0,23,55,26,58,0,0]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm2, %zmm7
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [21,53,24,56,0,0,22,54,25,57,0,0,23,55,26,58,21,53,24,56,0,0,22,54,25,57,0,0,23,55,26,58]
; AVX512BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm5, %zmm4, %zmm11
; AVX512BW-NEXT:    vmovdqa32 %zmm7, %zmm11 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = <0,1,53,u,4,5,6,7,54,u,10,11,12,13,55,u,16,17,18,19,56,u,22,23,24,25,57,u,28,29,30,31>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm11, %zmm7
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm11 = [0,1,2,53,4,5,6,7,8,54,10,11,12,13,14,55,16,17,18,19,20,56,22,23,24,25,26,57,28,29,30,31]
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm7, %zmm11
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [29,61,0,0,27,59,30,62,0,0,28,60,31,63,0,0,29,61,0,0,27,59,30,62,0,0,28,60,31,63,0,0]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm5, %zmm4, %zmm7
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm4 = [0,0,27,59,30,62,0,0,28,60,31,63,0,0,29,61,0,0,27,59,30,62,0,0,28,60,31,63,0,0,29,61]
; AVX512BW-NEXT:    # zmm4 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT:    vmovdqa32 %zmm7, %zmm4 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm2 = <58,u,2,3,4,5,59,u,8,9,10,11,60,u,14,15,16,17,61,u,20,21,22,23,62,u,26,27,28,29,63,u>
; AVX512BW-NEXT:    vpermi2w %zmm1, %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [0,58,2,3,4,5,6,59,8,9,10,11,12,60,14,15,16,17,18,61,20,21,22,23,24,62,26,27,28,29,30,63]
; AVX512BW-NEXT:    vpermi2w %zmm0, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 320(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm11, 256(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm10, 192(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm9, 128(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm8, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm6, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %in.vec0 = load <32 x i16>, ptr %in.vecptr0, align 64
  %in.vec1 = load <32 x i16>, ptr %in.vecptr1, align 64
  %in.vec2 = load <32 x i16>, ptr %in.vecptr2, align 64
  %in.vec3 = load <32 x i16>, ptr %in.vecptr3, align 64
  %in.vec4 = load <32 x i16>, ptr %in.vecptr4, align 64
  %in.vec5 = load <32 x i16>, ptr %in.vecptr5, align 64
  %1 = shufflevector <32 x i16> %in.vec0, <32 x i16> %in.vec1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %2 = shufflevector <32 x i16> %in.vec2, <32 x i16> %in.vec3, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %3 = shufflevector <32 x i16> %in.vec4, <32 x i16> %in.vec5, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = shufflevector <64 x i16> %1, <64 x i16> %2, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
  %5 = shufflevector <64 x i16> %3, <64 x i16> poison, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %6 = shufflevector <128 x i16> %4, <128 x i16> %5, <192 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191>
  %interleaved.vec = shufflevector <192 x i16> %6, <192 x i16> poison, <192 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191>
  store <192 x i16> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i16_stride6_vf64:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $792, %rsp # imm = 0x318
; SSE-NEXT:    movdqa (%rdi), %xmm10
; SSE-NEXT:    movdqa 16(%rdi), %xmm2
; SSE-NEXT:    movdqa (%rsi), %xmm3
; SSE-NEXT:    movdqa 16(%rsi), %xmm1
; SSE-NEXT:    movdqa (%rdx), %xmm11
; SSE-NEXT:    movdqa 16(%rdx), %xmm8
; SSE-NEXT:    movdqa (%rcx), %xmm6
; SSE-NEXT:    movdqa 16(%rcx), %xmm4
; SSE-NEXT:    movdqa (%r8), %xmm9
; SSE-NEXT:    movdqa (%r9), %xmm12
; SSE-NEXT:    movdqa %xmm11, %xmm5
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm5[3,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm9[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,2],xmm5[0,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0,1,3]
; SSE-NEXT:    movaps {{.*#+}} xmm14 = [65535,0,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    andps %xmm14, %xmm0
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm12[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm5
; SSE-NEXT:    andnps %xmm7, %xmm5
; SSE-NEXT:    orps %xmm0, %xmm5
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm11 = xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm11[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm9[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,2],xmm3[2,3]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm0
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm12[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm5
; SSE-NEXT:    andnps %xmm3, %xmm5
; SSE-NEXT:    orps %xmm0, %xmm5
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm8, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm0
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 16(%r8), %xmm5
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm5[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,2],xmm6[0,1]
; SSE-NEXT:    movdqa 16(%r9), %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm7
; SSE-NEXT:    andnps %xmm6, %xmm7
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm0
; SSE-NEXT:    orps %xmm0, %xmm7
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm8 = xmm8[4],xmm4[4],xmm8[5],xmm4[5],xmm8[6],xmm4[6],xmm8[7],xmm4[7]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm0
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movdqa %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm8[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,2],xmm1[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm4
; SSE-NEXT:    andnps %xmm1, %xmm4
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm0
; SSE-NEXT:    orps %xmm0, %xmm4
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 32(%rdx), %xmm0
; SSE-NEXT:    movdqa 32(%rcx), %xmm1
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 32(%rdi), %xmm2
; SSE-NEXT:    movdqa 32(%rsi), %xmm6
; SSE-NEXT:    movdqa %xmm2, %xmm7
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 32(%r8), %xmm4
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm4[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[1,2],xmm8[0,1]
; SSE-NEXT:    movdqa 32(%r9), %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm8, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm7
; SSE-NEXT:    orps %xmm7, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm0[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm6[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm7
; SSE-NEXT:    andnps %xmm6, %xmm7
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm1
; SSE-NEXT:    orps %xmm1, %xmm7
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 48(%rdx), %xmm0
; SSE-NEXT:    movdqa 48(%rcx), %xmm1
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 48(%rdi), %xmm2
; SSE-NEXT:    movdqa 48(%rsi), %xmm7
; SSE-NEXT:    movdqa %xmm2, %xmm8
; SSE-NEXT:    punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 48(%r8), %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm11 = xmm6[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,2],xmm11[0,1]
; SSE-NEXT:    movdqa 48(%r9), %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm11 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm11, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm8
; SSE-NEXT:    orps %xmm8, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm0[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm6[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm7[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm8
; SSE-NEXT:    andnps %xmm7, %xmm8
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm1
; SSE-NEXT:    orps %xmm1, %xmm8
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 64(%rdx), %xmm0
; SSE-NEXT:    movdqa 64(%rcx), %xmm1
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 64(%rdi), %xmm2
; SSE-NEXT:    movdqa 64(%rsi), %xmm8
; SSE-NEXT:    movdqa %xmm2, %xmm11
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 64(%r8), %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm12 = xmm7[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[1,2],xmm12[0,1]
; SSE-NEXT:    movdqa 64(%r9), %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm12 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm12[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm12, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm11
; SSE-NEXT:    orps %xmm11, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm0[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm7[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm8[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm8, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm1
; SSE-NEXT:    orps %xmm1, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 80(%rdx), %xmm0
; SSE-NEXT:    movdqa 80(%rcx), %xmm1
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 80(%rdi), %xmm2
; SSE-NEXT:    movdqa 80(%rsi), %xmm11
; SSE-NEXT:    movdqa %xmm2, %xmm12
; SSE-NEXT:    punpcklwd {{.*#+}} xmm12 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 80(%r8), %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm3[2,1,3,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, %xmm8
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[1,2],xmm15[0,1]
; SSE-NEXT:    movdqa 80(%r9), %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm15[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm15, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm12
; SSE-NEXT:    orps %xmm12, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm11[4],xmm1[5],xmm11[5],xmm1[6],xmm11[6],xmm1[7],xmm11[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm0[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm11 = xmm8[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm11[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm11 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm11, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm1
; SSE-NEXT:    orps %xmm1, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 96(%rdx), %xmm0
; SSE-NEXT:    movdqa 96(%rcx), %xmm1
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 96(%rdi), %xmm2
; SSE-NEXT:    movdqa 96(%rsi), %xmm12
; SSE-NEXT:    movdqa %xmm2, %xmm15
; SSE-NEXT:    punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm3[3,3]
; SSE-NEXT:    movdqa 96(%r8), %xmm11
; SSE-NEXT:    pshuflw {{.*#+}} xmm13 = xmm11[2,1,3,3,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[1,2],xmm13[0,1]
; SSE-NEXT:    movdqa 96(%r9), %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm13 = xmm8[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm13 = xmm13[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm13, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm15
; SSE-NEXT:    orps %xmm15, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm12[4],xmm1[5],xmm12[5],xmm1[6],xmm12[6],xmm1[7],xmm12[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm0[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm12 = xmm11[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm12[2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm12 = xmm8[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm12[2,1,2,3]
; SSE-NEXT:    movaps %xmm14, %xmm10
; SSE-NEXT:    andnps %xmm12, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm1
; SSE-NEXT:    orps %xmm1, %xmm10
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 112(%rdx), %xmm0
; SSE-NEXT:    movdqa 112(%rcx), %xmm3
; SSE-NEXT:    movdqa %xmm0, %xmm8
; SSE-NEXT:    punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 112(%rdi), %xmm1
; SSE-NEXT:    movdqa 112(%rsi), %xmm13
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm8[3,3]
; SSE-NEXT:    movdqa 112(%r8), %xmm12
; SSE-NEXT:    pshuflw {{.*#+}} xmm10 = xmm12[2,1,3,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm12, %xmm8
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[1,2],xmm10[0,1]
; SSE-NEXT:    movdqa 112(%r9), %xmm2
; SSE-NEXT:    pshuflw {{.*#+}} xmm10 = xmm2[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,1,2,1]
; SSE-NEXT:    movaps %xmm14, %xmm12
; SSE-NEXT:    andnps %xmm10, %xmm12
; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm15
; SSE-NEXT:    orps %xmm15, %xmm12
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm0[3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm8[0,1,2,3,6,5,7,7]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,2],xmm10[2,3]
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0,1,3]
; SSE-NEXT:    andps %xmm14, %xmm1
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm2[0,1,2,3,4,6,6,7]
; SSE-NEXT:    movdqa %xmm2, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[2,1,2,3]
; SSE-NEXT:    andnps %xmm10, %xmm14
; SSE-NEXT:    orps %xmm1, %xmm14
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movaps %xmm3, %xmm10
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[1,3]
; SSE-NEXT:    movaps %xmm0, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,0],xmm1[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm13
; SSE-NEXT:    pslldq {{.*#+}} xmm13 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm13[0,1,2,3,4,5]
; SSE-NEXT:    movdqa {{.*#+}} xmm8 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pandn %xmm13, %xmm0
; SSE-NEXT:    andps %xmm8, %xmm10
; SSE-NEXT:    por %xmm10, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm10 = xmm9[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[1,1],xmm3[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa {{.*#+}} xmm12 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    pshufd {{.*#+}} xmm13 = xmm1[0,0,1,1]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm13, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm10
; SSE-NEXT:    por %xmm10, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movaps %xmm3, %xmm10
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm9, %xmm13
; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,1],xmm0[1,3]
; SSE-NEXT:    movaps %xmm0, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,0],xmm13[0,2]
; SSE-NEXT:    movdqa %xmm1, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm13 = xmm1[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm0
; SSE-NEXT:    movdqa %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pandn %xmm1, %xmm0
; SSE-NEXT:    andps %xmm8, %xmm10
; SSE-NEXT:    por %xmm10, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[1,1],xmm3[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm13, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm9
; SSE-NEXT:    por %xmm9, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; SSE-NEXT:    movaps %xmm9, %xmm2
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm5, %xmm10
; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,1],xmm0[1,3]
; SSE-NEXT:    movaps %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm10[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm10
; SSE-NEXT:    pslldq {{.*#+}} xmm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm10[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pandn %xmm10, %xmm0
; SSE-NEXT:    andps %xmm8, %xmm2
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm9[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm5[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1],xmm9[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm3[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm1[0,0,1,1]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm9, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm2
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; SSE-NEXT:    movaps %xmm10, %xmm2
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm5, %xmm9
; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,1],xmm0[1,3]
; SSE-NEXT:    movaps %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm9[0,2]
; SSE-NEXT:    movdqa %xmm1, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm1[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm0
; SSE-NEXT:    movdqa %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pandn %xmm1, %xmm0
; SSE-NEXT:    andps %xmm8, %xmm2
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm10[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm10[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm9, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm5
; SSE-NEXT:    por %xmm5, %xmm0
; SSE-NEXT:    movdqa %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; SSE-NEXT:    movaps %xmm9, %xmm2
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movdqa %xmm4, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[1,3]
; SSE-NEXT:    movaps %xmm0, %xmm1
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm0, %xmm3
; SSE-NEXT:    pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm3, %xmm5
; SSE-NEXT:    andps %xmm8, %xmm2
; SSE-NEXT:    por %xmm2, %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm9[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm4[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1],xmm9[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm1[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn %xmm3, %xmm1
; SSE-NEXT:    andps %xmm12, %xmm2
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    movaps %xmm5, %xmm2
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movdqa %xmm4, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,1],xmm1[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm3[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm0
; SSE-NEXT:    movdqa %xmm8, %xmm9
; SSE-NEXT:    pandn %xmm0, %xmm9
; SSE-NEXT:    andps %xmm8, %xmm2
; SSE-NEXT:    por %xmm2, %xmm9
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm5[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm5[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,0],xmm1[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm3, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm4
; SSE-NEXT:    por %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT:    movdqa %xmm6, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[1,3]
; SSE-NEXT:    movaps %xmm1, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    pslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm2, %xmm5
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm6[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm3[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
; SSE-NEXT:    movdqa %xmm12, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    andps %xmm12, %xmm0
; SSE-NEXT:    por %xmm0, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
; SSE-NEXT:    movdqa %xmm6, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,1],xmm3[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm1, %xmm5
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[1,1],xmm4[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm2, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm6
; SSE-NEXT:    por %xmm6, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT:    movdqa %xmm7, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[1,3]
; SSE-NEXT:    movaps %xmm1, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    pslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm2, %xmm5
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm7[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm3[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
; SSE-NEXT:    movdqa %xmm12, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    andps %xmm12, %xmm0
; SSE-NEXT:    por %xmm0, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
; SSE-NEXT:    movdqa %xmm7, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,1],xmm3[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm1, %xmm5
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[1,1],xmm4[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pandn %xmm2, %xmm0
; SSE-NEXT:    andps %xmm12, %xmm7
; SSE-NEXT:    por %xmm7, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    movaps %xmm5, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm3[1,3]
; SSE-NEXT:    movaps %xmm3, %xmm4
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm2
; SSE-NEXT:    pslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm6
; SSE-NEXT:    pandn %xmm2, %xmm6
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm6
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm5[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm1[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm5[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm4[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,1,1]
; SSE-NEXT:    movdqa %xmm12, %xmm4
; SSE-NEXT:    pandn %xmm2, %xmm4
; SSE-NEXT:    andps %xmm12, %xmm0
; SSE-NEXT:    por %xmm0, %xmm4
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    movaps %xmm5, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm4[0]
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,1],xmm4[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm3
; SSE-NEXT:    movdqa %xmm8, %xmm6
; SSE-NEXT:    pandn %xmm3, %xmm6
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm6
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm5[1]
; SSE-NEXT:    movdqa %xmm1, %xmm0
; SSE-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm5[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm4[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    andps %xmm12, %xmm0
; SSE-NEXT:    por %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT:    movdqa %xmm11, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[1,3]
; SSE-NEXT:    movaps %xmm1, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    pslldq {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm13
; SSE-NEXT:    pandn %xmm2, %xmm13
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm13
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm11[1,1,1,1,4,5,6,7]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm3[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
; SSE-NEXT:    movdqa %xmm1, %xmm4
; SSE-NEXT:    movdqa %xmm12, %xmm10
; SSE-NEXT:    pandn %xmm2, %xmm10
; SSE-NEXT:    andps %xmm12, %xmm0
; SSE-NEXT:    por %xmm0, %xmm10
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; SSE-NEXT:    movaps %xmm7, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT:    movdqa %xmm11, %xmm2
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,1],xmm1[1,3]
; SSE-NEXT:    movaps %xmm1, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[0,2]
; SSE-NEXT:    movdqa %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm6
; SSE-NEXT:    pandn %xmm1, %xmm6
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm6
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm7[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm11 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[1,1],xmm7[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,0],xmm3[0,2]
; SSE-NEXT:    movdqa %xmm12, %xmm4
; SSE-NEXT:    pandn %xmm5, %xmm4
; SSE-NEXT:    andps %xmm12, %xmm11
; SSE-NEXT:    por %xmm11, %xmm4
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; SSE-NEXT:    movaps %xmm7, %xmm0
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movaps %xmm2, %xmm5
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm1[1,3]
; SSE-NEXT:    movaps %xmm1, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm5[0,2]
; SSE-NEXT:    movdqa %xmm15, %xmm5
; SSE-NEXT:    pslldq {{.*#+}} xmm5 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm5[0,1,2,3,4,5]
; SSE-NEXT:    movdqa %xmm8, %xmm11
; SSE-NEXT:    pandn %xmm5, %xmm11
; SSE-NEXT:    andps %xmm8, %xmm0
; SSE-NEXT:    por %xmm0, %xmm11
; SSE-NEXT:    movaps %xmm7, %xmm0
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm7[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm2[1,1,1,1,4,5,6,7]
; SSE-NEXT:    movaps %xmm2, %xmm7
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,0],xmm3[0,2]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm15[0,0,1,1]
; SSE-NEXT:    movdqa %xmm15, %xmm9
; SSE-NEXT:    movdqa %xmm12, %xmm15
; SSE-NEXT:    pandn %xmm2, %xmm15
; SSE-NEXT:    andps %xmm12, %xmm5
; SSE-NEXT:    por %xmm5, %xmm15
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movaps %xmm3, %xmm2
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm7, %xmm5
; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,1],xmm1[1,3]
; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm5[0,2]
; SSE-NEXT:    andps %xmm8, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm9[2,2,3,3]
; SSE-NEXT:    pslld $16, %xmm9
; SSE-NEXT:    pandn %xmm9, %xmm8
; SSE-NEXT:    por %xmm2, %xmm8
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]
; SSE-NEXT:    psrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[1,1],xmm3[1,1]
; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,0],xmm1[0,2]
; SSE-NEXT:    andps %xmm12, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm12
; SSE-NEXT:    por %xmm7, %xmm12
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa %xmm12, 736(%rax)
; SSE-NEXT:    movdqa %xmm8, 720(%rax)
; SSE-NEXT:    movdqa %xmm15, 688(%rax)
; SSE-NEXT:    movdqa %xmm11, 672(%rax)
; SSE-NEXT:    movdqa %xmm4, 640(%rax)
; SSE-NEXT:    movdqa %xmm6, 624(%rax)
; SSE-NEXT:    movdqa %xmm10, 592(%rax)
; SSE-NEXT:    movdqa %xmm13, 576(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 544(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 528(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 496(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 480(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 448(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 432(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 400(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 384(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rax)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movaps %xmm14, 752(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 704(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 656(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 608(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 560(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 512(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 464(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 416(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    addq $792, %rsp # imm = 0x318
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i16_stride6_vf64:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $504, %rsp # imm = 0x1F8
; AVX1-ONLY-NEXT:    vmovdqa 80(%rcx), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa 80(%rdx), %xmm2
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa 80(%rsi), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa 80(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm1[2,3,2,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm5[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm6, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 80(%r8), %xmm2
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0],xmm6[1,2],xmm3[3]
; AVX1-ONLY-NEXT:    vmovdqa 80(%r9), %xmm3
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0],xmm8[1],xmm7[2,3,4,5,6],xmm8[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm6, %xmm6
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3],xmm2[4,5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm3, %xmm7
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3,4],xmm7[5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm6, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm5 = xmm4[0],xmm5[1],xmm4[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm6[3],xmm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm4, %xmm4
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3,4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3,4,5,6],xmm5[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 64(%rsi), %xmm8
; AVX1-ONLY-NEXT:    vmovdqa 64(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX1-ONLY-NEXT:    vmovdqa 64(%rcx), %xmm10
; AVX1-ONLY-NEXT:    vmovdqa 64(%rdx), %xmm11
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm7[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm7[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm4[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm12 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 64(%r8), %xmm5
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm13 = xmm12[0],xmm6[1],xmm12[2,3]
; AVX1-ONLY-NEXT:    vmovdqa 64(%r9), %xmm6
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm6[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1,2],xmm14[3],xmm13[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm12, %xmm12
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm5[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3,4,5],xmm13[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm6[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0],xmm13[1],xmm12[2,3,4,5,6],xmm13[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm12, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm1[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm12, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6],ymm0[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm1 = xmm0[0,1],xmm2[0],xmm0[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm12 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm12[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6],ymm1[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm3 = xmm1[0,1],xmm5[0],xmm1[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm6[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm8[5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm1
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm6[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm3[3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm7[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm2[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm5[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm6[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3,4,5,6],xmm2[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm5[4,5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm6, %xmm1
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm1[5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 48(%rcx), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa 48(%rdx), %xmm2
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa 48(%rsi), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa 48(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 48(%r8), %xmm2
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0],xmm6[1,2],xmm3[3]
; AVX1-ONLY-NEXT:    vmovdqa 48(%r9), %xmm3
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0],xmm8[1],xmm7[2,3,4,5,6],xmm8[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm6, %xmm6
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3],xmm2[4,5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm3, %xmm7
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3,4],xmm7[5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm6, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm5 = xmm4[0],xmm5[1],xmm4[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm6[3],xmm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm4, %xmm4
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3,4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3,4,5,6],xmm5[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 32(%rcx), %xmm8
; AVX1-ONLY-NEXT:    vmovdqa 32(%rdx), %xmm9
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm6[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm6[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vmovdqa 32(%rsi), %xmm10
; AVX1-ONLY-NEXT:    vmovdqa 32(%rdi), %xmm11
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm7[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm7, %ymm5
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm12 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 32(%r8), %xmm4
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm13 = xmm12[0],xmm5[1],xmm12[2,3]
; AVX1-ONLY-NEXT:    vmovdqa 32(%r9), %xmm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm5[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1,2],xmm14[3],xmm13[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm12, %xmm12
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm4[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3,4,5],xmm13[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm5[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0],xmm13[1],xmm12[2,3,4,5,6],xmm13[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm12, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm1[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm12, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6],ymm0[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm1 = xmm0[0,1],xmm2[0],xmm0[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm12 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm12[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6],ymm1[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm3 = xmm1[0,1],xmm4[0],xmm1[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm5[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm8[5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm1
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm5[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm3[3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm6[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm2[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm7[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm4[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm5[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3,4,5,6],xmm2[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm4[4,5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm5, %xmm1
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm1[5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 112(%rcx), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa 112(%rdx), %xmm2
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm4[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa 112(%rsi), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa 112(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 112(%r8), %xmm2
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0],xmm6[1,2],xmm3[3]
; AVX1-ONLY-NEXT:    vmovdqa 112(%r9), %xmm3
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm8[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm7[0],xmm8[1],xmm7[2,3,4,5,6],xmm8[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm6, %xmm6
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3],xmm2[4,5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm3, %xmm7
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2,3,4],xmm7[5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm6, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm5 = xmm4[0],xmm5[1],xmm4[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm6[3],xmm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm4, %xmm4
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3,4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3,4,5,6],xmm5[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 96(%rcx), %xmm8
; AVX1-ONLY-NEXT:    vmovdqa 96(%rdx), %xmm9
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm6[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm6[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vmovdqa 96(%rsi), %xmm10
; AVX1-ONLY-NEXT:    vmovdqa 96(%rdi), %xmm11
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm7[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm7, %ymm5
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm12 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 96(%r8), %xmm4
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm13 = xmm12[0],xmm5[1],xmm12[2,3]
; AVX1-ONLY-NEXT:    vmovdqa 96(%r9), %xmm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm5[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1,2],xmm14[3],xmm13[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm12, %xmm12
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm4[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3,4,5],xmm13[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm5[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0],xmm13[1],xmm12[2,3,4,5,6],xmm13[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm12, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm1[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm12, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6],ymm0[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm1 = xmm0[0,1],xmm2[0],xmm0[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm12 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm12[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3],xmm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6],ymm1[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm3 = xmm1[0,1],xmm4[0],xmm1[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm5[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm8[5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm1
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm5[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm3[3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm6[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm2[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm7[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm4[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm5[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3,4,5,6],xmm2[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm4[4,5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm5, %xmm1
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm1[5],xmm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 16(%rcx), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdx), %xmm1
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm11[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vmovdqa 16(%rsi), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm10[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
; AVX1-ONLY-NEXT:    vmovdqa 16(%r8), %xmm3
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm5 = xmm2[0],xmm1[1,2],xmm2[3]
; AVX1-ONLY-NEXT:    vmovdqa 16(%r9), %xmm2
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm2[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0],xmm6[1],xmm5[2,3,4,5,6],xmm6[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm1, %xmm1
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm3[4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm2, %xmm5
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm5[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm2[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3],xmm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3,4,5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm14 = xmm0[0],xmm1[1],xmm0[2,3,4,5,6],xmm1[7]
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm9
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm8
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm7[1,1,2,2]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm7[2,2,3,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm5
; AVX1-ONLY-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[2,3,2,3]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm12 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm1
; AVX1-ONLY-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm13 = xmm12[0],xmm0[1],xmm12[2,3]
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm0[2,2,3,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1,2],xmm15[3],xmm13[4,5,6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm12, %xmm12
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm15 = xmm1[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3,4,5],xmm15[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm15 = xmm0[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[2,1,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3,4,5,6],xmm15[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm11[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm15, %ymm11
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm10[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm15, %ymm10
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm11 = ymm10[0],ymm11[1],ymm10[2,3],ymm11[4],ymm10[5,6],ymm11[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm10 = xmm11[0,1],xmm3[0],xmm11[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm2[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm10[0,1,2,3,4],xmm15[5],xmm10[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm11, %xmm11
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm3[2,3],xmm11[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3],xmm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm3[1,1,2,2]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0],ymm8[1],ymm6[2,3],ymm8[4],ymm6[5,6],ymm8[7]
; AVX1-ONLY-NEXT:    vinsertps {{.*#+}} xmm8 = xmm6[0,1],xmm1[0],xmm6[3]
; AVX1-ONLY-NEXT:    vpslldq {{.*#+}} xmm9 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm8 = xmm8[0,1,2,3,4],xmm9[5],xmm8[6,7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm6, %xmm6
; AVX1-ONLY-NEXT:    vpmovzxwd {{.*#+}} xmm9 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm9[2,3],xmm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2],xmm9[3],xmm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[0,0,1,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,3,2,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,2,1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0],xmm3[1,2],xmm4[3]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3,4,5,6],xmm5[7]
; AVX1-ONLY-NEXT:    vextractf128 $1, %ymm3, %xmm3
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3],xmm1[4,5],xmm3[6,7]
; AVX1-ONLY-NEXT:    vpslld $16, %xmm0, %xmm0
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4],xmm0[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, 48(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, 32(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, 16(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm8, (%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, 112(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm10, 96(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm12, 80(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, 64(%rax)
; AVX1-ONLY-NEXT:    vmovdqa %xmm14, 176(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 160(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 144(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 128(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 624(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 608(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 592(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 576(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 688(%rax)
; AVX1-ONLY-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 672(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 656(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 640(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 752(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 736(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 720(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 704(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 240(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 224(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 208(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 192(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 304(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 288(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 272(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 256(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 368(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 352(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 336(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 320(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 432(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 416(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 400(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 384(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 496(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 480(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 464(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 448(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 560(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 544(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 528(%rax)
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %xmm0, 512(%rax)
; AVX1-ONLY-NEXT:    addq $504, %rsp # imm = 0x1F8
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i16_stride6_vf64:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    subq $1544, %rsp # imm = 0x608
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm6
; AVX2-SLOW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, %xmm3
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm4
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm7
; AVX2-SLOW-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm15
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm14
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%r8), %xmm8
; AVX2-SLOW-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm13 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm15[0,1,2,1]
; AVX2-SLOW-NEXT:    vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm14[0,1,2,1]
; AVX2-SLOW-NEXT:    vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r9), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm8[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 64(%rcx), %xmm12
; AVX2-SLOW-NEXT:    vmovdqa 64(%rdx), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-SLOW-NEXT:    vmovdqa 64(%rsi), %xmm10
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm10[0,1,2,1]
; AVX2-SLOW-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa 64(%r8), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa 64(%r9), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 96(%rcx), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vmovdqa 96(%rdx), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-SLOW-NEXT:    vmovdqa 96(%rsi), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vmovdqa 96(%rdi), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa 96(%r8), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa 96(%r9), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %ymm2
; AVX2-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdx), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rcx), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-SLOW-NEXT:    vmovdqa 32(%rsi), %ymm9
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm9[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdi), %ymm11
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm11[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r8), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r9), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 64(%rdx), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 64(%rcx), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-SLOW-NEXT:    vmovdqa 64(%rsi), %ymm8
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm8[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqa 64(%rdi), %ymm7
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm7[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa 64(%r8), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa 64(%r9), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 96(%rdx), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 96(%rcx), %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-SLOW-NEXT:    vmovdqa 96(%rsi), %ymm6
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm6[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vmovdqa 96(%rdi), %ymm2
; AVX2-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,1,2,3,6,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa 96(%r8), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqa 96(%r9), %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa %xmm3, %xmm5
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-SLOW-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm3[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm2
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm14[0],zero,xmm14[1],zero,xmm14[2],zero,xmm14[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = mem[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = mem[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,0,2,2]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vpmovzxwd (%rsp), %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = mem[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[2],mem[2],ymm2[3],mem[3],ymm2[8],mem[8],ymm2[9],mem[9],ymm2[10],mem[10],ymm2[11],mem[11]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm11[0],ymm9[0],ymm11[1],ymm9[1],ymm11[2],ymm9[2],ymm11[3],ymm9[3],ymm11[8],ymm9[8],ymm11[9],ymm9[9],ymm11[10],ymm9[10],ymm11[11],ymm9[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[2],mem[2],ymm2[3],mem[3],ymm2[8],mem[8],ymm2[9],mem[9],ymm2[10],mem[10],ymm2[11],mem[11]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm2 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm2 = ymm13[0],mem[0],ymm13[1],mem[1],ymm13[2],mem[2],ymm13[3],mem[3],ymm13[8],mem[8],ymm13[9],mem[9],ymm13[10],mem[10],ymm13[11],mem[11]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[8],ymm6[8],ymm10[9],ymm6[9],ymm10[10],ymm6[10],ymm10[11],ymm6[11]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[8],ymm9[8],ymm8[9],ymm9[9],ymm8[10],ymm9[10],ymm8[11],ymm9[11]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-SLOW-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %xmm5, %xmm1, %xmm2
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm0[0,1,0,1]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm4 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm2, %ymm3, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm3 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
; AVX2-SLOW-NEXT:    vpshufb %xmm5, %xmm14, %xmm3
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6],ymm3[7]
; AVX2-SLOW-NEXT:    vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm3 = mem[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm2, %ymm3, %ymm0
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm0 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm0 = xmm15[4],mem[4],xmm15[5],mem[5],xmm15[6],mem[6],xmm15[7],mem[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %xmm5, %xmm1, %xmm2
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX2-SLOW-NEXT:    vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = mem[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm0, %ymm2, %ymm14
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %xmm5, %xmm1, %xmm1
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-SLOW-NEXT:    vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm1 = mem[2,3,2,3]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm0, %ymm1, %ymm3
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm1 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm5 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1],ymm1[2],ymm5[3,4],ymm1[5],ymm5[6,7]
; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm5 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %ymm5, %ymm0, %ymm15
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm15[1],ymm1[2,3],ymm15[4],ymm1[5,6],ymm15[7]
; AVX2-SLOW-NEXT:    vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm15 = mem[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm15[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm1, %ymm15, %ymm2
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm15 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm0 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4],ymm15[5],ymm0[6,7]
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpshufb %ymm5, %ymm1, %ymm15
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3],ymm15[4],ymm0[5,6],ymm15[7]
; AVX2-SLOW-NEXT:    vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm15 = mem[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm15[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm0, %ymm15, %ymm1
; AVX2-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm15 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm0 = ymm13[4],mem[4],ymm13[5],mem[5],ymm13[6],mem[6],ymm13[7],mem[7],ymm13[12],mem[12],ymm13[13],mem[13],ymm13[14],mem[14],ymm13[15],mem[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4],ymm15[5],ymm0[6,7]
; AVX2-SLOW-NEXT:    vpshufb %ymm5, %ymm12, %ymm15
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3],ymm15[4],ymm0[5,6],ymm15[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm11[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm15[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm0, %ymm15, %ymm0
; AVX2-SLOW-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm12 # 32-byte Folded Reload
; AVX2-SLOW-NEXT:    # ymm12 = ymm10[4],mem[4],ymm10[5],mem[5],ymm10[6],mem[6],ymm10[7],mem[7],ymm10[12],mem[12],ymm10[13],mem[13],ymm10[14],mem[14],ymm10[15],mem[15]
; AVX2-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm13 = ymm8[4],ymm9[4],ymm8[5],ymm9[5],ymm8[6],ymm9[6],ymm8[7],ymm9[7],ymm8[12],ymm9[12],ymm8[13],ymm9[13],ymm8[14],ymm9[14],ymm8[15],ymm9[15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm13[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1],ymm12[2],ymm13[3,4],ymm12[5],ymm13[6,7]
; AVX2-SLOW-NEXT:    vpshufb %ymm5, %ymm6, %ymm5
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm12[0],ymm5[1],ymm12[2,3],ymm5[4],ymm12[5,6],ymm5[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm7[2,3,2,3,6,7,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm10[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX2-SLOW-NEXT:    vpblendvb %ymm4, %ymm5, %ymm10, %ymm4
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa %ymm4, 736(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm4, 672(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 544(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm1, 352(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm2, 160(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 704(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm3, 640(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 576(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 512(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm14, 448(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 384(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 608(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 416(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-SLOW-NEXT:    addq $1544, %rsp # imm = 0x608
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i16_stride6_vf64:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    subq $1592, %rsp # imm = 0x638
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rsi), %xmm7
; AVX2-FAST-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX2-FAST-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rcx), %xmm15
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rdx), %xmm14
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX2-FAST-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm5 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm7, %xmm2
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm6, %xmm3
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r8), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r9), %xmm13
; AVX2-FAST-NEXT:    vpshufb %xmm1, %xmm13, %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 64(%rsi), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX2-FAST-NEXT:    vmovdqa 64(%rcx), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 64(%rdx), %xmm4
; AVX2-FAST-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vmovdqa 64(%r8), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7]
; AVX2-FAST-NEXT:    vmovdqa 64(%r9), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 96(%rsi), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FAST-NEXT:    vmovdqa 96(%rdi), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; AVX2-FAST-NEXT:    vmovdqa 96(%rcx), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqa 96(%rdx), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqa 96(%r8), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqa 96(%r9), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, (%rsp) # 16-byte Spill
; AVX2-FAST-NEXT:    vpshufb %xmm1, %xmm2, %xmm1
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %ymm3
; AVX2-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %ymm4
; AVX2-FAST-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %ymm12
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqa (%r8), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX2-FAST-NEXT:    # ymm6 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rdi), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-NEXT:    vmovdqa 32(%rdx), %ymm10
; AVX2-FAST-NEXT:    vmovdqa 32(%rcx), %ymm11
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r8), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa 32(%r9), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 64(%rdi), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 64(%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-NEXT:    vmovdqa 64(%rdx), %ymm8
; AVX2-FAST-NEXT:    vmovdqa 64(%rcx), %ymm9
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqa 64(%r8), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa 64(%r9), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 96(%rdi), %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 96(%rsi), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX2-FAST-NEXT:    vmovdqa 96(%rdx), %ymm3
; AVX2-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 96(%rcx), %ymm7
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqa 96(%r8), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqa 96(%r9), %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm5, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,0,2,2,1,0,2,2]
; AVX2-FAST-NEXT:    # ymm1 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
; AVX2-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm0 = mem[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm0, %ymm3
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm2
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6],ymm2[7]
; AVX2-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm13[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm2
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6],ymm2[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = mem[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
; AVX2-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX2-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, (%rsp), %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = mem[0,0,2,1,4,5,6,7]
; AVX2-FAST-NEXT:    vpbroadcastq %xmm2, %ymm2
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm12[0],ymm1[1],ymm12[1],ymm1[2],ymm12[2],ymm1[3],ymm12[3],ymm1[8],ymm12[8],ymm1[9],ymm12[9],ymm1[10],ymm12[10],ymm1[11],ymm12[11]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = [5,4,2,2,5,4,6,6]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm2
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm14[0],ymm3[1],ymm14[1],ymm3[2],ymm14[2],ymm3[3],ymm14[3],ymm3[8],ymm14[8],ymm3[9],ymm14[9],ymm3[10],ymm14[10],ymm3[11],ymm14[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6],ymm2[7]
; AVX2-FAST-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm3 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm3 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[8],ymm11[8],ymm10[9],ymm11[9],ymm10[10],ymm11[10],ymm10[11],ymm11[11]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm2
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm3 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[2],mem[2],ymm3[3],mem[3],ymm3[8],mem[8],ymm3[9],mem[9],ymm3[10],mem[10],ymm3[11],mem[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6],ymm2[7]
; AVX2-FAST-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm3 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm3 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[8],ymm9[8],ymm8[9],ymm9[9],ymm8[10],ymm9[10],ymm8[11],ymm9[11]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm2
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm3 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[2],mem[2],ymm3[3],mem[3],ymm3[8],mem[8],ymm3[9],mem[9],ymm3[10],mem[10],ymm3[11],mem[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6],ymm2[7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm13[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[2],ymm7[2],ymm11[3],ymm7[3],ymm11[8],ymm7[8],ymm11[9],ymm7[9],ymm11[10],ymm7[10],ymm11[11],ymm7[11]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[8],ymm10[8],ymm9[9],ymm10[9],ymm9[10],ymm10[10],ymm9[11],ymm10[11]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX2-FAST-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm5[4],xmm1[5],xmm5[5],xmm1[6],xmm5[6],xmm1[7],xmm5[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = <1,2,1,2,u,u,3,3>
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm6, %ymm0
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm4 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm4, %xmm1, %xmm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5,6],ymm3[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm2 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[0,1,0,1]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm3, %ymm5, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm5 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT:    vpermd %ymm3, %ymm6, %ymm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm4, %xmm0, %xmm5
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3],ymm5[4],ymm3[5,6],ymm5[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm3, %ymm5, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT:    vpermd %ymm3, %ymm6, %ymm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7]
; AVX2-FAST-NEXT:    vpshufb %xmm4, %xmm15, %xmm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5,6],ymm3[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm0, %ymm3, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm6, %ymm0
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm6 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm6 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[1,1,1,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3],ymm4[4],ymm0[5,6],ymm4[7]
; AVX2-FAST-NEXT:    vmovdqa (%rsp), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm0, %ymm2, %ymm5
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm2 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm0[4],ymm14[4],ymm0[5],ymm14[5],ymm0[6],ymm14[6],ymm0[7],ymm14[7],ymm0[12],ymm14[12],ymm0[13],ymm14[13],ymm0[14],ymm14[14],ymm0[15],ymm14[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [5,6,5,6,5,6,7,7]
; AVX2-FAST-NEXT:    vpermd %ymm2, %ymm6, %ymm2
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm0, %ymm15
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm15[1],ymm2[2,3],ymm15[4],ymm2[5,6],ymm15[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm15 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm0, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm2, %ymm14, %ymm3
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm14 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm0 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-NEXT:    vpermd %ymm14, %ymm6, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1],ymm0[2],ymm14[3,4],ymm0[5],ymm14[6,7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm14[1],ymm0[2,3],ymm14[4],ymm0[5,6],ymm14[7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm2, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm0, %ymm14, %ymm2
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm14 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm0 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-NEXT:    vpermd %ymm14, %ymm6, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1],ymm0[2],ymm14[3,4],ymm0[5],ymm14[6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm13, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm14[1],ymm0[2,3],ymm14[4],ymm0[5,6],ymm14[7]
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm12, %ymm14
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm0, %ymm14, %ymm0
; AVX2-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm13 # 32-byte Folded Reload
; AVX2-FAST-NEXT:    # ymm13 = ymm11[4],mem[4],ymm11[5],mem[5],ymm11[6],mem[6],ymm11[7],mem[7],ymm11[12],mem[12],ymm11[13],mem[13],ymm11[14],mem[14],ymm11[15],mem[15]
; AVX2-FAST-NEXT:    vpermd %ymm13, %ymm6, %ymm6
; AVX2-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm13 = ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[12],ymm10[12],ymm9[13],ymm10[13],ymm9[14],ymm10[14],ymm9[15],ymm10[15]
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[3,3,3,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1],ymm13[2],ymm6[3,4],ymm13[5],ymm6[6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm7, %ymm4
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3],ymm4[4],ymm6[5,6],ymm4[7]
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm8, %ymm6
; AVX2-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
; AVX2-FAST-NEXT:    vpblendvb %ymm1, %ymm4, %ymm6, %ymm1
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa %ymm1, 736(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm1, 672(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 544(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm2, 352(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm3, 160(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 704(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm5, 640(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 576(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 512(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 448(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 384(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 608(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 416(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-NEXT:    addq $1592, %rsp # imm = 0x638
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i16_stride6_vf64:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    subq $1464, %rsp # imm = 0x5B8
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rsi), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdi), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm4, %xmm14
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %xmm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rcx), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdx), %xmm7
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1],xmm1[2],xmm9[2],xmm1[3],xmm9[3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm9, %xmm15
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %xmm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r8), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r9), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r9), %xmm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm8 = [255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm6[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rsi), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rcx), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rdx), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%r8), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%r9), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rsi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rcx), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rdx), %xmm13
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm2[0],xmm13[1],xmm2[1],xmm13[2],xmm2[2],xmm13[3],xmm2[3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, %xmm11
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%r8), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%r9), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %ymm10
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[1],ymm10[1],ymm0[2],ymm10[2],ymm0[3],ymm10[3],ymm0[8],ymm10[8],ymm0[9],ymm10[9],ymm0[10],ymm10[10],ymm0[11],ymm10[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r9), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdi), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rsi), %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdx), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rcx), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r8), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r9), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rdi), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rsi), %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rdx), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%rcx), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%r8), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 64(%r9), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rdi), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rsi), %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rdx), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%rcx), %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%r8), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 96(%r9), %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm14, %xmm6
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm14, %xmm4
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm15, %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm9, %xmm3
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm9[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1,2],ymm4[3],ymm1[4,5],ymm4[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm7, %xmm4, %xmm4
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm4[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm4 = [255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm1, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm8, %xmm8
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm8[2],ymm1[3,4],ymm8[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $246, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm8 = mem[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm7, %xmm8, %xmm8
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm1, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm8, %xmm8
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm8[2],ymm1[3,4],ymm8[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $246, (%rsp), %xmm8 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm8 = mem[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm7, %xmm8, %xmm8
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm1, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm0, %xmm8, %xmm0
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm11, %xmm15
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm14[2,1,3,3,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm7, %xmm11, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm10, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm7, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[1],ymm1[1],ymm7[2],ymm1[2],ymm7[3],ymm1[3],ymm7[8],ymm1[8],ymm7[9],ymm1[9],ymm7[10],ymm1[10],ymm7[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm7 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm7[2],ymm1[3,4],ymm7[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm7 = mem[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0],ymm1[1,2],ymm7[3],ymm1[4,5],ymm7[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vbroadcasti128 {{.*#+}} ymm7 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX2-FAST-PERLANE-NEXT:    # ymm7 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm1, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm12[0],ymm8[0],ymm12[1],ymm8[1],ymm12[2],ymm8[2],ymm12[3],ymm8[3],ymm12[8],ymm8[8],ymm12[9],ymm8[9],ymm12[10],ymm8[10],ymm12[11],ymm8[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm8[2],ymm1[3,4],ymm8[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm8 = mem[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm1, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm12[0],ymm8[0],ymm12[1],ymm8[1],ymm12[2],ymm8[2],ymm12[3],ymm8[3],ymm12[8],ymm8[8],ymm12[9],ymm8[9],ymm12[10],ymm8[10],ymm12[11],ymm8[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm8[2],ymm1[3,4],ymm8[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshuflw $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm8 = mem[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm1, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm0, %ymm8, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm8[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm4, %ymm0, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm0 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm6 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm6, %xmm3, %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm5 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm5, %xmm0, %xmm0
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm0[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm3 = [255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0,255,255,255,255,255,255,255,255,255,255,0,0]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm2, %ymm4, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm4 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm6, %xmm0, %xmm4
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3],ymm4[4],ymm2[5,6],ymm4[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm5, %xmm0, %xmm4
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm2, %ymm4, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm6, %xmm1, %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm5, %xmm1, %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm0, %ymm2, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm6, %xmm14, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm5, %xmm11, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm0, %ymm1, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm10[4],ymm0[5],ymm10[5],ymm0[6],ymm10[6],ymm0[7],ymm10[7],ymm0[12],ymm10[12],ymm0[13],ymm10[13],ymm0[14],ymm10[14],ymm0[15],ymm10[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm5 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1],ymm1[2],ymm5[3,4],ymm1[5],ymm5[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm5 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm0, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm6[1],ymm1[2,3],ymm6[4],ymm1[5,6],ymm6[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm6 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm0, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm1, %ymm12, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm12 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm0 = ymm9[4],mem[4],ymm9[5],mem[5],ymm9[6],mem[6],ymm9[7],mem[7],ymm9[12],mem[12],ymm9[13],mem[13],ymm9[14],mem[14],ymm9[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm12[2],ymm0[3,4],ymm12[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm1, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3],ymm12[4],ymm0[5,6],ymm12[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm1, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm0, %ymm12, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm12 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm0 = ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm12[2],ymm0[3,4],ymm12[5],ymm0[6,7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm9, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3],ymm12[4],ymm0[5,6],ymm12[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm9, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm0, %ymm12, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm12 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm12 = ymm9[4],mem[4],ymm9[5],mem[5],ymm9[6],mem[6],ymm9[7],mem[7],ymm9[12],mem[12],ymm9[13],mem[13],ymm9[14],mem[14],ymm9[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm11 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # ymm11 = ymm9[4],mem[4],ymm9[5],mem[5],ymm9[6],mem[6],ymm9[7],mem[7],ymm9[12],mem[12],ymm9[13],mem[13],ymm9[14],mem[14],ymm9[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
; AVX2-FAST-PERLANE-NEXT:    vpshufd {{.*#+}} ymm11 = ymm11[1,2,3,3,5,6,7,7]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm12[2],ymm11[3,4],ymm12[5],ymm11[6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm8, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm5 = ymm11[0],ymm5[1],ymm11[2,3],ymm5[4],ymm11[5,6],ymm5[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm8, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
; AVX2-FAST-PERLANE-NEXT:    vpblendvb %ymm3, %ymm5, %ymm6, %ymm3
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm3, 736(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm3, 704(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm3, 672(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm0, 544(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 512(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm1, 352(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm2, 160(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm4, 640(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 608(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 576(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm7, 448(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 416(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 384(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FAST-PERLANE-NEXT:    addq $1464, %rsp # imm = 0x5B8
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-ONLY-SLOW-LABEL: store_i16_stride6_vf64:
; AVX512F-ONLY-SLOW:       # %bb.0:
; AVX512F-ONLY-SLOW-NEXT:    subq $408, %rsp # imm = 0x198
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rsi), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdi), %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rcx), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdx), %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rcx), %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdx), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %ymm3, %ymm31
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-SLOW-NEXT:    kmovw %eax, %k1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm1, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r8), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{.*#+}} xmm13 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %xmm13, %xmm2, %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 (%r9), %xmm19
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm19[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r9), %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm25 = ymm0[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rsi), %xmm9
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rdi), %xmm15
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rsi), %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rdi), %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm3[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rcx), %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rdx), %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,2,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rcx), %ymm7
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%rdx), %ymm8
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm10[1,0,2,2,5,4,6,6]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm6[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm10[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm5, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm5 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%r8), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm6 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm5, %ymm10
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm6 = ymm10[0,1],ymm6[2],ymm10[3,4],ymm6[5],ymm10[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%r8), %xmm6
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %xmm13, %xmm6, %xmm12
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0],ymm12[1],ymm5[2,3],ymm12[4],ymm5[5,6],ymm12[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm21 = zmm5[0,1,2,3],zmm10[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%r9), %xmm10
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm10[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %xmm10, %xmm11
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm23 = ymm5[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 96(%r9), %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm24 = ymm10[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[8],ymm10[8],ymm12[9],ymm10[9],ymm12[10],ymm10[10],ymm12[11],ymm10[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm7 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm0[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm12[0],ymm8[0],ymm12[1],ymm8[1],ymm12[2],ymm8[2],ymm12[3],ymm8[3],ymm12[8],ymm8[8],ymm12[9],ymm8[9],ymm12[10],ymm8[10],ymm12[11],ymm8[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm7[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm10, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm8, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{.*#+}} ymm7 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa %ymm7, %ymm14
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm7
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0],ymm1[1],ymm7[2,3],ymm1[4],ymm7[5,6],ymm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm5[2,3,2,3,6,7,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rcx), %ymm8
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rdx), %ymm7
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm0[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm7[4],ymm8[4],ymm7[5],ymm8[5],ymm7[6],ymm8[6],ymm7[7],ymm8[7],ymm7[12],ymm8[12],ymm7[13],ymm8[13],ymm7[14],ymm8[14],ymm7[15],ymm8[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm0[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rsi), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm0[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[8],ymm10[8],ymm12[9],ymm10[9],ymm12[10],ymm10[10],ymm12[11],ymm10[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm12 = ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[12],ymm2[12],ymm0[13],ymm2[13],ymm0[14],ymm2[14],ymm0[15],ymm2[15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm10, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm1, %zmm5 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%r8), %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %ymm14, %ymm1, %ymm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %ymm14, %ymm17
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm5, %ymm12
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6],ymm10[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm12[0],ymm5[1,2],ymm12[3],ymm5[4,5],ymm12[6],ymm5[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm28 = zmm5[0,1,2,3],zmm10[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%r9), %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm22 = ymm10[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm5[2,3,2,3,6,7,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm10[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm20 = ymm10[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[1,0,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm15[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm12[4],xmm9[4],xmm12[5],xmm9[5],xmm12[6],xmm9[6],xmm12[7],xmm9[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm10, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512F-ONLY-SLOW-NEXT:    kmovw %eax, %k2
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm9, %zmm4, %zmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm3, %zmm4 {%k2}
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm6[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm4, %ymm9
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0],ymm9[1,2],ymm3[3],ymm9[4,5],ymm3[6],ymm9[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm26 = zmm4[0,1,2,3],zmm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm11[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm11, %xmm16
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm27
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rcx), %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rdx), %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,0,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm9 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm10 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm9, %ymm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rsi), %xmm9
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%rdi), %xmm15
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm14 = xmm9[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm11 = xmm15[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm14[4],xmm11[5],xmm14[5],xmm11[6],xmm14[6],xmm11[7],xmm14[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm10, %zmm6, %zmm6
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm12, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm6, %zmm10 {%k2}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%r8), %xmm6
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm6[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm10, %ymm12
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0],ymm12[1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm0, %zmm11
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm12 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm10[0,1],ymm12[2],ymm10[3,4],ymm12[5],ymm10[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm29 = zmm10[0,1,2,3],zmm11[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,2,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[1,0,2,2,5,4,6,6]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm9, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm0, %zmm2 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %xmm13, %xmm6, %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm30 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 64(%r9), %xmm15
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm15[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm0[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm7
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm10
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rsi), %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdi), %ymm8
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm8[0],ymm2[0],ymm8[1],ymm2[1],ymm8[2],ymm2[2],ymm8[3],ymm2[3],ymm8[8],ymm2[8],ymm8[9],ymm2[9],ymm8[10],ymm2[10],ymm8[11],ymm2[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm0[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm11 = xmm0[1,2,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rcx), %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%rdx), %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm14 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm14 = ymm14[1,0,2,2,5,4,6,6]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm9, %zmm3, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm11[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm14[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm9, %zmm11
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm3, %zmm11 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r8), %ymm9
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm11, %ymm14
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm14[0,1],ymm3[2],ymm14[3,4],ymm3[5],ymm14[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm14
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r8), %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %xmm13, %xmm3, %xmm13
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0],ymm13[1],ymm11[2,3],ymm13[4],ymm11[5,6],ymm13[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm18 = zmm11[0,1,2,3],zmm14[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm13 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm11 = ymm13[0],ymm11[0],ymm13[1],ymm11[1],ymm13[2],ymm11[2],ymm13[3],ymm11[3],ymm13[8],ymm11[8],ymm13[9],ymm11[9],ymm13[10],ymm11[10],ymm13[11],ymm11[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm2[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm8[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm13 = ymm13[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm13[0],ymm1[0],ymm13[1],ymm1[1],ymm13[2],ymm1[2],ymm13[3],ymm1[3],ymm13[8],ymm1[8],ymm13[9],ymm1[9],ymm13[10],ymm1[10],ymm13[11],ymm1[11]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm8[4],ymm2[4],ymm8[5],ymm2[5],ymm8[6],ymm2[6],ymm8[7],ymm2[7],ymm8[12],ymm2[12],ymm8[13],ymm2[13],ymm8[14],ymm2[14],ymm8[15],ymm2[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm11, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %ymm17, %ymm13
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %ymm13, %ymm9, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm17 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %ymm31, %ymm8
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm8[4],ymm2[5],ymm8[5],ymm2[6],ymm8[6],ymm2[7],ymm8[7],ymm2[12],ymm8[12],ymm2[13],ymm8[13],ymm2[14],ymm8[14],ymm2[15],ymm8[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm11[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm2[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm9[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm2[4],ymm11[4],ymm2[5],ymm11[5],ymm2[6],ymm11[6],ymm2[7],ymm11[7],ymm2[12],ymm11[12],ymm2[13],ymm11[13],ymm2[14],ymm11[14],ymm2[15],ymm11[15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm9, %zmm8, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpshufb %ymm13, %ymm2, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm1, %ymm8
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm8[0],ymm0[1],ymm8[2,3],ymm0[4],ymm8[5,6],ymm0[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm25, %zmm2, %zmm8
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm8 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm24, %zmm23, %zmm9
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm21, %zmm25, %zmm9
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm12, %zmm6
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm16, %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm2[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,4,4,4]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm16 = ymm12[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm21 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm1 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm15[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm24
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,0,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm1, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm7[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm10[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r9), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm15[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm12[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm10, %zmm7
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm5, %zmm7 {%k2}
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm7, %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm3[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm10[0],ymm5[1,2],ymm10[3],ymm5[4,5],ymm10[6],ymm5[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm1[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm10[0,2,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm7[0,1],ymm3[2],ymm7[3,4],ymm3[5],ymm7[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 32(%r9), %ymm7
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm5[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,0,2,2]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm5, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm14 = xmm2[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm4[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm14[4],xmm2[5],xmm14[5],xmm2[6],xmm14[6],xmm2[7],xmm14[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm14 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} ymm14 = ymm14[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[2,3,2,3,6,7,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm7[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm12, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm5, %zmm2 {%k2}
; AVX512F-ONLY-SLOW-NEXT:    vextracti64x4 $1, %zmm2, %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm4[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm12[0],ymm5[1,2],ymm12[3],ymm5[4,5],ymm12[6],ymm5[7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm11[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} ymm12 = ymm12[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm11[2,3,2,3,6,7,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} ymm13 = ymm13[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm23 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm23, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
; AVX512F-ONLY-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm5[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm19, %xmm11
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm11[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpbroadcastq %xmm5, %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm11[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,4,4]
; AVX512F-ONLY-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm30, %zmm25, %zmm6
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm15, %zmm10, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm18, %zmm25, %zmm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm15 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm15 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm20, %zmm22, %zmm19
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm28, %zmm18, %zmm19
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm14, %zmm7
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm17, %zmm18, %zmm7
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm13, %zmm12, %zmm12
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm21, %zmm18, %zmm12
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm16, %zmm27, %zmm13
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm26, %zmm14, %zmm13
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm24, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm29, %zmm14, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm4, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm3, %zmm14, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm5, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    vpternlogq $184, %zmm2, %zmm14, %zmm3
; AVX512F-ONLY-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm1, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm12, 128(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm7, 320(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm10, 256(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm6, 448(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm0, 384(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm13, 576(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm19, 512(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm15, 704(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm9, 640(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm8, 64(%rax)
; AVX512F-ONLY-SLOW-NEXT:    addq $408, %rsp # imm = 0x198
; AVX512F-ONLY-SLOW-NEXT:    vzeroupper
; AVX512F-ONLY-SLOW-NEXT:    retq
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride6_vf64:
; AVX512F-ONLY-FAST:       # %bb.0:
; AVX512F-ONLY-FAST-NEXT:    subq $1064, %rsp # imm = 0x428
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rcx), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdx), %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rcx), %xmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rcx), %xmm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rcx), %xmm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdx), %xmm9
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm9, (%rsp) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rdx), %xmm8
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rcx), %ymm11
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rdx), %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm11[0],ymm2[1],ymm11[1],ymm2[2],ymm11[2],ymm2[3],ymm11[3],ymm2[8],ymm11[8],ymm2[9],ymm11[9],ymm2[10],ymm11[10],ymm2[11],ymm11[11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rcx), %xmm5
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rdx), %xmm6
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm6, %xmm23
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm4, %zmm16
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rcx), %ymm14
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rdx), %ymm6
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm6[0],ymm14[0],ymm6[1],ymm14[1],ymm6[2],ymm14[2],ymm6[3],ymm14[3],ymm6[8],ymm14[8],ymm6[9],ymm14[9],ymm6[10],ymm14[10],ymm6[11],ymm14[11]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm8, %xmm31
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm7, %xmm21
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm4, %zmm27
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rcx), %ymm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdx), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[8],ymm4[8],ymm0[9],ymm4[9],ymm0[10],ymm4[10],ymm0[11],ymm4[11]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm3[4],xmm9[5],xmm3[5],xmm9[6],xmm3[6],xmm9[7],xmm3[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm7, %zmm8, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rsi), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdi), %xmm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[1,1,1,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rsi), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdi), %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm22 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
; AVX512F-ONLY-FAST-NEXT:    vpermd %zmm1, %zmm22, %zmm18
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm12, %zmm7, %zmm1
; AVX512F-ONLY-FAST-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-FAST-NEXT:    kmovw %eax, %k2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm1, %zmm18 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm17 = [8,9,20,11,12,21,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm18, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r8), %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm1, %zmm17, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,9,2,3,8,5,6,11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r8), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm3, %xmm4
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm1, %ymm8, %ymm18
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r9), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm3, %xmm9
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm1[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r9), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm1[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rsi), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rdi), %xmm7
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm12 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[1,1,1,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rsi), %ymm13
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%rdi), %ymm15
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm12, %zmm3
; AVX512F-ONLY-FAST-NEXT:    vpermd %zmm16, %zmm22, %zmm16
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm3, %zmm16 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm16, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%r8), %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm12, %zmm17, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%r8), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm4, %xmm0, %xmm12
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, %xmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm12, %ymm8, %ymm16
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%r9), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm0, %xmm12
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm12[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 96(%r9), %ymm12
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm9[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm24 = [5,6,5,6,5,6,7,7]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm2[4],ymm11[4],ymm2[5],ymm11[5],ymm2[6],ymm11[6],ymm2[7],ymm11[7],ymm2[12],ymm11[12],ymm2[13],ymm11[13],ymm2[14],ymm11[14],ymm2[15],ymm11[15]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm9, %ymm24, %ymm9
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm11[0],ymm2[1],ymm11[1],ymm2[2],ymm11[2],ymm2[3],ymm11[3],ymm2[8],ymm11[8],ymm2[9],ymm11[9],ymm2[10],ymm11[10],ymm2[11],ymm11[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm2[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm2, %ymm13, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm2, %ymm15, %ymm10
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm2, %ymm19
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[1],ymm0[1],ymm10[2],ymm0[2],ymm10[3],ymm0[3],ymm10[8],ymm0[8],ymm10[9],ymm0[9],ymm10[10],ymm0[10],ymm10[11],ymm0[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm10 = ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[12],ymm13[12],ymm15[13],ymm13[13],ymm15[14],ymm13[14],ymm15[15],ymm13[15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm9, %zmm11, %zmm9
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm25
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm9, %zmm25 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [8,21,10,11,20,13,14,23]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm2, %ymm3, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %ymm2, %ymm13
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm0, %zmm20, %zmm9
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [12,1,2,13,4,5,14,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm0, %ymm2, %ymm25
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %ymm2, %ymm10
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm14[4],ymm6[5],ymm14[5],ymm6[6],ymm14[6],ymm6[7],ymm14[7],ymm6[12],ymm14[12],ymm6[13],ymm14[13],ymm6[14],ymm14[14],ymm6[15],ymm14[15]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm0, %ymm24, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm3 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm6[0],ymm3[0],ymm6[1],ymm3[1],ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[8],ymm3[8],ymm6[9],ymm3[9],ymm6[10],ymm3[10],ymm6[11],ymm3[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rsi), %ymm11
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm19, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm2, %ymm11, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rdi), %ymm15
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm2, %ymm15, %ymm9
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[8],ymm6[8],ymm9[9],ymm6[9],ymm9[10],ymm6[10],ymm9[11],ymm6[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm15[4],ymm11[4],ymm15[5],ymm11[5],ymm15[6],ymm11[6],ymm15[7],ymm11[7],ymm15[12],ymm11[12],ymm15[13],ymm11[13],ymm15[14],ymm11[14],ymm15[15],ymm11[15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm9, %zmm6, %zmm14
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm0, %zmm14 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm14, %zmm29
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%r8), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm13, %ymm0, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm3, %zmm20, %zmm29
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm3, %ymm10, %ymm14
; AVX512F-ONLY-FAST-NEXT:    vbroadcasti32x4 {{.*#+}} ymm19 = [1,0,2,2,1,0,2,2]
; AVX512F-ONLY-FAST-NEXT:    # ymm19 = mem[0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm23, %xmm2
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm3, %ymm19, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm9 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm1, %xmm9
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm7, %xmm10
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm3, %zmm3
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm9, %zmm1, %zmm30
; AVX512F-ONLY-FAST-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512F-ONLY-FAST-NEXT:    kmovw %eax, %k1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm3, %zmm30 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm23 = [16,9,10,17,12,13,18,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm4[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm1, %zmm23, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm21, %xmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm31, %xmm5
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm1, %ymm19, %ymm31
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm3 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rsi), %xmm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%rdi), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm7, %xmm6
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm1, %xmm5
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm31, %zmm3
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm6, %zmm31
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm3, %zmm31 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm31, %zmm28
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%r8), %xmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm3, %zmm23, %zmm28
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm15[0],ymm11[0],ymm15[1],ymm11[1],ymm15[2],ymm11[2],ymm15[3],ymm11[3],ymm15[8],ymm11[8],ymm15[9],ymm11[9],ymm15[10],ymm11[10],ymm15[11],ymm11[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512F-ONLY-FAST-NEXT:    vpermd %zmm27, %zmm22, %zmm27
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm1, %zmm27 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm27, %zmm21
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm0, %zmm17, %zmm21
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm4 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm4, %xmm2, %xmm0
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm0, %ymm8, %ymm27
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rsi), %xmm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm0[1,1,1,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rsi), %ymm11
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%rdi), %ymm15
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm15[0],ymm11[0],ymm15[1],ymm11[1],ymm15[2],ymm11[2],ymm15[3],ymm11[3],ymm15[8],ymm11[8],ymm15[9],ymm11[9],ymm15[10],ymm11[10],ymm15[11],ymm11[11]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm1[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm0, %zmm3 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r8), %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermi2d %zmm0, %zmm3, %zmm17
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r8), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm0, %ymm8, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm4[4],ymm1[5],ymm4[5],ymm1[6],ymm4[6],ymm1[7],ymm4[7],ymm1[12],ymm4[12],ymm1[13],ymm4[13],ymm1[14],ymm4[14],ymm1[15],ymm4[15]
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm8[0],ymm4[1],ymm8[1],ymm4[2],ymm8[2],ymm4[3],ymm8[3],ymm4[8],ymm8[8],ymm4[9],ymm8[9],ymm4[10],ymm8[10],ymm4[11],ymm8[11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm1, %ymm11, %ymm8
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm1, %ymm15, %ymm13
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm1, %ymm22
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm13[0],ymm8[0],ymm13[1],ymm8[1],ymm13[2],ymm8[2],ymm13[3],ymm8[3],ymm13[8],ymm8[8],ymm13[9],ymm8[9],ymm13[10],ymm8[10],ymm13[11],ymm8[11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%r9), %ymm13
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm12 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm12[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm15[4],ymm11[4],ymm15[5],ymm11[5],ymm15[6],ymm11[6],ymm15[7],ymm11[7],ymm15[12],ymm11[12],ymm15[13],ymm11[13],ymm15[14],ymm11[14],ymm15[15],ymm11[15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm12 = ymm13[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm12[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm0, %ymm24, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[3,3,3,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm4, %zmm4
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm11, %zmm8, %zmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm4, %zmm1 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm1, %zmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm10 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm10, %ymm2, %ymm8
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm8, %zmm20, %zmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 64(%r9), %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm9 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm0, %xmm8
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm8[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm13[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm8[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [12,1,2,13,4,5,14,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm2, %ymm6, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm2 = ymm0[4],ymm8[4],ymm0[5],ymm8[5],ymm0[6],ymm8[6],ymm0[7],ymm8[7],ymm0[12],ymm8[12],ymm0[13],ymm8[13],ymm0[14],ymm8[14],ymm0[15],ymm8[15]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm2, %ymm24, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r9), %xmm11
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm11, %xmm15
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm26 = ymm15[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm15 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} ymm12 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm12[0],ymm15[0],ymm12[1],ymm15[1],ymm12[2],ymm15[2],ymm12[3],ymm15[3],ymm12[8],ymm15[8],ymm12[9],ymm15[9],ymm12[10],ymm15[10],ymm12[11],ymm15[11]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm22, %ymm9
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm9, %ymm0, %ymm15
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm9, %ymm8, %ymm13
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm13 = ymm13[0],ymm15[0],ymm13[1],ymm15[1],ymm13[2],ymm15[2],ymm13[3],ymm15[3],ymm13[8],ymm15[8],ymm13[9],ymm15[9],ymm13[10],ymm15[10],ymm13[11],ymm15[11]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm15 = ymm8[4],ymm0[4],ymm8[5],ymm0[5],ymm8[6],ymm0[6],ymm8[7],ymm0[7],ymm8[12],ymm0[12],ymm8[13],ymm0[13],ymm8[14],ymm0[14],ymm8[15],ymm0[15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 32(%r9), %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm12, %zmm12
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm15, %zmm13, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm12, %zmm2 {%k2}
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm10, %ymm8, %ymm12
; AVX512F-ONLY-FAST-NEXT:    vpermi2d %zmm12, %zmm2, %zmm20
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm8[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm12, %ymm6, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512F-ONLY-FAST-NEXT:    # ymm6 = mem[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm6, %ymm0, %ymm8
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm8, %ymm24
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm8, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512F-ONLY-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm6, %ymm0, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm8, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rsp), %xmm10 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm0, %ymm19, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm13 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm7, %xmm13
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm5, %xmm15
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm12, %zmm0, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm13, %zmm5, %zmm5
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm0, %zmm5 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm5, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm6[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %zmm7, %zmm23, %zmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
; AVX512F-ONLY-FAST-NEXT:    vpermd %ymm7, %ymm19, %ymm7
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpsrldq {{.*#+}} xmm13 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm13 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # zmm13 = zmm18[0,1,2,3],mem[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm15 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm13, %zmm18, %zmm15
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm13 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # zmm13 = zmm16[0,1,2,3],mem[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm16 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm13, %zmm18, %zmm16
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm13 = zmm27[0,1,2,3],zmm21[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm19 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm13, %zmm18, %zmm19
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm11, %zmm26, %zmm11
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm3, %zmm18, %zmm11
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # zmm3 = zmm25[0,1,2,3],mem[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm13 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm17 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm3, %zmm17, %zmm13
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm14[0,1,2,3],zmm29[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm14 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm3, %zmm17, %zmm14
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm4[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm24[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm22[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm1, %zmm17, %zmm3
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm18 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # xmm18 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm12, %ymm4
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm7, %zmm4
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm1, %xmm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm8, %xmm12
; AVX512F-ONLY-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm7[4],xmm12[5],xmm7[5],xmm12[6],xmm7[6],xmm12[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm9[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm12, %ymm21
; AVX512F-ONLY-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm12 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm9, %xmm9
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm29 = ymm9[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm22 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # xmm22 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm7, %zmm1, %zmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm8[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm7, %ymm7
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm8, %xmm8
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # ymm24 = mem[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpermq $230, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # ymm25 = mem[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm26 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm4, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm9[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermi2d %zmm4, %zmm1, %zmm23
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm6[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm4, %ymm4
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm6, %xmm10
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm27 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm12, %xmm9, %xmm12
; AVX512F-ONLY-FAST-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm9[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpbroadcastq %xmm9, %ymm9
; AVX512F-ONLY-FAST-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm20[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm25, %zmm24, %zmm20
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm2, %zmm17, %zmm20
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,8,3,4,9,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm18, %ymm2, %ymm30
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm22, %ymm2, %ymm31
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm26, %ymm2, %ymm5
; AVX512F-ONLY-FAST-NEXT:    vpermt2d %ymm27, %ymm2, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT:    # zmm2 = zmm30[0,1,2,3],mem[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm29, %zmm21, %zmm17
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm2, %zmm18, %zmm17
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm31[0,1,2,3],zmm28[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm7, %zmm6
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm2, %zmm18, %zmm6
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm0[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm10, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm0, %zmm18, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm23[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm12, %zmm9, %zmm1
; AVX512F-ONLY-FAST-NEXT:    vpternlogq $184, %zmm0, %zmm18, %zmm1
; AVX512F-ONLY-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm1, (%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm2, 192(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm20, 128(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm3, 320(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm11, 256(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm19, 448(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm6, 384(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm17, 576(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm14, 512(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm13, 704(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm16, 640(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm15, 64(%rax)
; AVX512F-ONLY-FAST-NEXT:    addq $1064, %rsp # imm = 0x428
; AVX512F-ONLY-FAST-NEXT:    vzeroupper
; AVX512F-ONLY-FAST-NEXT:    retq
;
; AVX512DQ-SLOW-LABEL: store_i16_stride6_vf64:
; AVX512DQ-SLOW:       # %bb.0:
; AVX512DQ-SLOW-NEXT:    subq $408, %rsp # imm = 0x198
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rcx), %xmm0
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rcx), %ymm2
; AVX512DQ-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdx), %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdi), %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rsi), %ymm3
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512DQ-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %ymm3, %ymm31
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512DQ-SLOW-NEXT:    kmovw %eax, %k1
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r8), %ymm2
; AVX512DQ-SLOW-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm3, %xmm12
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa64 (%r9), %xmm19
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm19[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r9), %ymm0
; AVX512DQ-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm25 = ymm0[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rcx), %xmm15
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rdx), %xmm9
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm15[4],xmm9[5],xmm15[5],xmm9[6],xmm15[6],xmm9[7],xmm15[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[1,2,3,3]
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rcx), %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rdx), %ymm3
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm1[0],ymm3[1],ymm1[1],ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[8],ymm1[8],ymm3[9],ymm1[9],ymm3[10],ymm1[10],ymm3[11],ymm1[11]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm0[1,0,2,2,5,4,6,6]
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rsi), %xmm0
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rdi), %xmm2
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rsi), %ymm7
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%rdi), %ymm8
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm5
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm10, %zmm6, %zmm5 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%r8), %ymm4
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm6 = ymm4[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm5, %ymm10
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm6 = ymm10[0,1],ymm6[2],ymm10[3,4],ymm6[5],ymm10[6,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm10
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%r8), %xmm6
; AVX512DQ-SLOW-NEXT:    vpshufb %xmm12, %xmm6, %xmm11
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm12, %xmm18
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0],ymm11[1],ymm5[2,3],ymm11[4],ymm5[5,6],ymm11[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm21 = zmm5[0,1,2,3],zmm10[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%r9), %xmm10
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm10[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vmovdqa %xmm10, %xmm13
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm23 = ymm5[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa 96(%r9), %ymm5
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm24 = ymm10[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm7[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm10[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm11 = ymm8[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm11[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm7 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm11[0],ymm8[0],ymm11[1],ymm8[1],ymm11[2],ymm8[2],ymm11[3],ymm8[3],ymm11[8],ymm8[8],ymm11[9],ymm8[9],ymm11[10],ymm8[10],ymm11[11],ymm8[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm1[4],ymm3[5],ymm1[5],ymm3[6],ymm1[6],ymm3[7],ymm1[7],ymm3[12],ymm1[12],ymm3[13],ymm1[13],ymm3[14],ymm1[14],ymm3[15],ymm1[15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm10, %zmm3
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm1, %zmm8, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa {{.*#+}} ymm7 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512DQ-SLOW-NEXT:    vpshufb %ymm7, %ymm4, %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqa %ymm7, %ymm14
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm3, %ymm7
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0],ymm1[1],ymm7[2,3],ymm1[4],ymm7[5,6],ymm1[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm4 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm5[2,3,2,3,6,7,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rsi), %ymm12
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm12[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rdi), %ymm8
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm8[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm1[0],ymm3[1],ymm1[1],ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[8],ymm1[8],ymm3[9],ymm1[9],ymm3[10],ymm1[10],ymm3[11],ymm1[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm12[4],ymm8[5],ymm12[5],ymm8[6],ymm12[6],ymm8[7],ymm12[7],ymm8[12],ymm12[12],ymm8[13],ymm12[13],ymm8[14],ymm12[14],ymm8[15],ymm12[15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm1[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rcx), %ymm7
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rdx), %ymm4
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm11 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm11 = ymm4[4],ymm7[4],ymm4[5],ymm7[5],ymm4[6],ymm7[6],ymm4[7],ymm7[7],ymm4[12],ymm7[12],ymm4[13],ymm7[13],ymm4[14],ymm7[14],ymm4[15],ymm7[15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm11 = ymm11[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm3, %zmm5
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm11, %zmm10, %zmm5 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%r8), %ymm3
; AVX512DQ-SLOW-NEXT:    vpshufb %ymm14, %ymm3, %ymm10
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %ymm14, %ymm16
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm5, %ymm11
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0],ymm10[1],ymm11[2,3],ymm10[4],ymm11[5,6],ymm10[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm10, %zmm0, %zmm10
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm11[0],ymm5[1,2],ymm11[3],ymm5[4,5],ymm11[6],ymm5[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm28 = zmm5[0,1,2,3],zmm10[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%r9), %ymm5
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm22 = ymm10[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm5[2,3,2,3,6,7,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm10[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm20 = ymm10[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm15[0],xmm9[1],xmm15[1],xmm9[2],xmm15[2],xmm9[3],xmm15[3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,0,2,2]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm11 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm9, %ymm9
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512DQ-SLOW-NEXT:    kmovw %eax, %k2
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm10, %zmm0
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm9, %zmm2, %zmm0 {%k2}
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm6[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm9
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm9[1,2],ymm2[3],ymm9[4,5],ymm2[6],ymm9[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm6
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm26 = zmm0[0,1,2,3],zmm2[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm13[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm13, %xmm17
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm27
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rsi), %xmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm2[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm9[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rcx), %xmm9
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%rdx), %xmm10
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[1,0,2,2]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm14 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm15[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm6, %zmm11
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm0, %zmm14, %zmm11 {%k2}
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%r8), %xmm6
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm6[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm11, %ymm14
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm14[1,2],ymm0[3],ymm14[4,5],ymm0[6],ymm14[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm14 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm14, %ymm14
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm14[2],ymm11[3,4],ymm14[5],ymm11[6,7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm29 = zmm11[0,1,2,3],zmm0[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[1],ymm7[1],ymm4[2],ymm7[2],ymm4[3],ymm7[3],ymm4[8],ymm7[8],ymm4[9],ymm7[9],ymm4[10],ymm7[10],ymm4[11],ymm7[11]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[1,0,2,2,5,4,6,6]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm12[0],ymm8[1],ymm12[1],ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[8],ymm12[8],ymm8[9],ymm12[9],ymm8[10],ymm12[10],ymm8[11],ymm12[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm18, %xmm13
; AVX512DQ-SLOW-NEXT:    vpshufb %xmm13, %xmm6, %xmm2
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm30 = zmm0[0,1,2,3],zmm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa 64(%r9), %xmm2
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm0[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm6 = ymm0[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm7
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm10
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[1,2,3,3]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rcx), %ymm8
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdx), %ymm9
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm0[1,0,2,2,5,4,6,6]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm5
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm4
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm15 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm3, %zmm12
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm15, %zmm14, %zmm12 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r8), %ymm15
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm15[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm12, %ymm14
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm14[0,1],ymm3[2],ymm14[3,4],ymm3[5],ymm14[6,7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm14
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r8), %xmm3
; AVX512DQ-SLOW-NEXT:    vpshufb %xmm13, %xmm3, %xmm13
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0],ymm13[1],ymm12[2,3],ymm13[4],ymm12[5,6],ymm13[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm18 = zmm12[0,1,2,3],zmm14[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm12 = ymm1[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm0[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm13 = ymm13[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm12 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm1 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm13 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm13[0],ymm1[0],ymm13[1],ymm1[1],ymm13[2],ymm1[2],ymm13[3],ymm1[3],ymm13[8],ymm1[8],ymm13[9],ymm1[9],ymm13[10],ymm1[10],ymm13[11],ymm1[11]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm8[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm12, %zmm0
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm8, %zmm1, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %ymm16, %ymm14
; AVX512DQ-SLOW-NEXT:    vpshufb %ymm14, %ymm15, %ymm1
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm8
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1],ymm8[2,3],ymm1[4],ymm8[5,6],ymm1[7]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm8 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm8[0],ymm0[1,2],ymm8[3],ymm0[4,5],ymm8[6],ymm0[7]
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm16 = zmm0[0,1,2,3],zmm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %ymm31, %ymm9
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm31[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm8[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm9[4],ymm8[5],ymm9[5],ymm8[6],ymm9[6],ymm8[7],ymm9[7],ymm8[12],ymm9[12],ymm8[13],ymm9[13],ymm8[14],ymm9[14],ymm8[15],ymm9[15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} ymm9 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[12],ymm13[12],ymm12[13],ymm13[13],ymm12[14],ymm13[14],ymm12[15],ymm13[15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm9, %zmm8, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX512DQ-SLOW-NEXT:    vpshufb %ymm14, %ymm9, %ymm1
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm0, %ymm8
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm13 = ymm8[0],ymm1[1],ymm8[2,3],ymm1[4],ymm8[5,6],ymm1[7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7]
; AVX512DQ-SLOW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm25, %zmm1, %zmm8
; AVX512DQ-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm8 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm24, %zmm23, %zmm9
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm21, %zmm1, %zmm9
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm11, %zmm6
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm17, %xmm11
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm11[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,4,4]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm17 = ymm11[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm13, %zmm0, %zmm11
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm21 = zmm0[0,1,2,3],zmm11[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm5 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm2[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm4, %ymm24
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm25 = ymm2[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm5
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r9), %xmm0
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm13[1,0,2,2]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm7, %ymm7
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm7, %zmm10, %zmm5 {%k2}
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm5, %ymm7
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm3[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm10[0],ymm7[1,2],ymm10[3],ymm7[4,5],ymm10[6],ymm7[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm0[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm10[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm7
; AVX512DQ-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm3, %ymm3
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa 32(%r9), %ymm3
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm7[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm2[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm14 = xmm4[0,1,2,1]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT:    vpunpckhwd {{.*#+}} xmm13 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm14 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm13, %zmm7, %zmm7
; AVX512DQ-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm12 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpsrldq {{.*#+}} xmm15 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm15 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm3[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[1,0,2,2]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm12, %zmm13, %zmm7 {%k2}
; AVX512DQ-SLOW-NEXT:    vextracti64x4 $1, %zmm7, %ymm12
; AVX512DQ-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm2[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0],ymm12[1,2],ymm13[3],ymm12[4,5],ymm13[6],ymm12[7]
; AVX512DQ-SLOW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm13 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm11 = ymm4[2,3,2,3,6,7,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm11[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,1,2,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwd {{.*#+}} xmm23 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm23, %ymm4
; AVX512DQ-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm7[0,1],ymm4[2],ymm7[3,4],ymm4[5],ymm7[6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm0[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm7, %ymm7
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm12, %zmm0, %zmm12
; AVX512DQ-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm12[4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm19, %xmm2
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm2[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpbroadcastq %xmm12, %ymm12
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; AVX512DQ-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm30, %zmm1, %zmm6
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm14, %zmm10, %zmm10
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm18, %zmm1, %zmm10
; AVX512DQ-SLOW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm14 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm1 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm20, %zmm22, %zmm18
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm28, %zmm14, %zmm18
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm15, %zmm3
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm16, %zmm14, %zmm3
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm11, %zmm13, %zmm11
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm21, %zmm14, %zmm11
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm17, %zmm27, %zmm13
; AVX512DQ-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm26, %zmm14, %zmm13
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm25, %zmm24, %zmm15
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm29, %zmm14, %zmm15
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm7, %zmm0
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm5, %zmm14, %zmm0
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm12, %zmm2
; AVX512DQ-SLOW-NEXT:    vpternlogq $184, %zmm4, %zmm14, %zmm2
; AVX512DQ-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm3, 320(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm10, 256(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm6, 448(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm15, 384(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm13, 576(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm18, 512(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm1, 704(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm9, 640(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm8, 64(%rax)
; AVX512DQ-SLOW-NEXT:    addq $408, %rsp # imm = 0x198
; AVX512DQ-SLOW-NEXT:    vzeroupper
; AVX512DQ-SLOW-NEXT:    retq
;
; AVX512DQ-FAST-LABEL: store_i16_stride6_vf64:
; AVX512DQ-FAST:       # %bb.0:
; AVX512DQ-FAST-NEXT:    subq $920, %rsp # imm = 0x398
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rcx), %ymm8
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rdx), %ymm6
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rcx), %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rdx), %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rcx), %xmm7
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rdx), %xmm14
; AVX512DQ-FAST-NEXT:    vmovdqa (%rcx), %xmm4
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rcx), %xmm10
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rcx), %xmm11
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdx), %xmm5
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdx), %xmm12
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm12, (%rsp) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rdx), %xmm15
; AVX512DQ-FAST-NEXT:    vmovdqa (%rcx), %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdx), %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm6[0],ymm8[0],ymm6[1],ymm8[1],ymm6[2],ymm8[2],ymm6[3],ymm8[3],ymm6[8],ymm8[8],ymm6[9],ymm8[9],ymm6[10],ymm8[10],ymm6[11],ymm8[11]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm14[4],xmm7[4],xmm14[5],xmm7[5],xmm14[6],xmm7[6],xmm14[7],xmm7[7]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm3 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm15, %xmm28
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rcx), %ymm13
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdx), %ymm2
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm13[0],ymm2[1],ymm13[1],ymm2[2],ymm13[2],ymm2[3],ymm13[3],ymm2[8],ymm13[8],ymm2[9],ymm13[9],ymm2[10],ymm13[10],ymm2[11],ymm13[11]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm2, %ymm24
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm9 = xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm10, %xmm18
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm9, %zmm2
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rsi), %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm15 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm2, %ymm9
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rdi), %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm10
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm9 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm9[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm10 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm17 = [5,6,5,6,5,6,7,7]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm12 = ymm6[4],ymm8[4],ymm6[5],ymm8[5],ymm6[6],ymm8[6],ymm6[7],ymm8[7],ymm6[12],ymm8[12],ymm6[13],ymm8[13],ymm6[14],ymm8[14],ymm6[15],ymm8[15]
; AVX512DQ-FAST-NEXT:    vpermd %ymm12, %ymm17, %ymm12
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm8 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm8[0],ymm6[1],ymm8[1],ymm6[2],ymm8[2],ymm6[3],ymm8[3],ymm6[8],ymm8[8],ymm6[9],ymm8[9],ymm6[10],ymm8[10],ymm6[11],ymm8[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm10, %zmm9, %zmm21
; AVX512DQ-FAST-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512DQ-FAST-NEXT:    kmovw %eax, %k1
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm12, %zmm6, %zmm21 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [8,21,10,11,20,13,14,23]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm21, %zmm23
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%r8), %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
; AVX512DQ-FAST-NEXT:    vpshufb %ymm12, %ymm2, %ymm6
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm6, %zmm20, %zmm23
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm19 = [12,1,2,13,4,5,14,7]
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm6 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm6, %ymm19, %ymm21
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%r9), %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm5 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512DQ-FAST-NEXT:    # ymm5 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm5, %ymm2, %ymm6
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm6[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
; AVX512DQ-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm6
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm6[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rsi), %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm2, %ymm6
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rdi), %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm8
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm6 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[8],ymm6[8],ymm8[9],ymm6[9],ymm8[10],ymm6[10],ymm8[11],ymm6[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm8 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[3,3,3,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm9 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
; AVX512DQ-FAST-NEXT:    vpermd %ymm9, %ymm17, %ymm9
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm10 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[1],ymm10[1],ymm0[2],ymm10[2],ymm0[3],ymm10[3],ymm0[8],ymm10[8],ymm0[9],ymm10[9],ymm0[10],ymm10[10],ymm0[11],ymm10[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm6, %zmm27
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm9, %zmm0, %zmm27 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm27, %zmm29
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%r8), %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm12, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm0, %zmm20, %zmm29
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm0, %ymm19, %ymm27
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%r9), %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm5, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm5, %ymm31
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm4, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm4, %ymm22
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rsi), %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm2, %xmm0, %xmm6
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %xmm2, %xmm1, %xmm8
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm10 = [1,0,2,2,1,0,2,2]
; AVX512DQ-FAST-NEXT:    # ymm10 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm14, %xmm0
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm14[0],xmm7[0],xmm14[1],xmm7[1],xmm14[2],xmm7[2],xmm14[3],xmm7[3]
; AVX512DQ-FAST-NEXT:    vpermd %ymm9, %ymm10, %ymm9
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm14 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1],xmm7[2],xmm14[2],xmm7[3],xmm14[3]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm7, %ymm7
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm8, %zmm25
; AVX512DQ-FAST-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512DQ-FAST-NEXT:    kmovw %eax, %k2
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm7, %zmm9, %zmm25 {%k2}
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm16 = [16,9,10,17,12,13,18,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm25, %zmm26
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%r8), %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm0[2,1,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm6, %zmm16, %zmm26
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rsi), %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %xmm2, %xmm0, %xmm6
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%rdi), %xmm1
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %xmm2, %xmm1, %xmm14
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm6 = xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm28, %xmm2
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3]
; AVX512DQ-FAST-NEXT:    vpermd %ymm1, %ymm10, %ymm1
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm14, %zmm11
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm0, %zmm1, %zmm11 {%k2}
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm11, %zmm28
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%r8), %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm0, %zmm16, %zmm28
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rsi), %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm2, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm1
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm24, %ymm2
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm2[4],ymm13[4],ymm2[5],ymm13[5],ymm2[6],ymm13[6],ymm2[7],ymm13[7],ymm2[12],ymm13[12],ymm2[13],ymm13[13],ymm2[14],ymm13[14],ymm2[15],ymm13[15]
; AVX512DQ-FAST-NEXT:    vpermd %ymm5, %ymm17, %ymm5
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm4 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm6 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm4 = ymm6[0],ymm4[0],ymm6[1],ymm4[1],ymm6[2],ymm4[2],ymm6[3],ymm4[3],ymm6[8],ymm4[8],ymm6[9],ymm4[9],ymm6[10],ymm4[10],ymm6[11],ymm4[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm13
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm5, %zmm4, %zmm13 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r8), %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm12, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm0, %zmm20, %zmm4
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm0, %ymm19, %ymm13
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r9), %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm31, %ymm14
; AVX512DQ-FAST-NEXT:    vpshufb %ymm14, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm31 = ymm0[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm22, %ymm9
; AVX512DQ-FAST-NEXT:    vpshufb %ymm9, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm24 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa (%rsi), %ymm3
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512DQ-FAST-NEXT:    vpshufb %ymm15, %ymm2, %ymm5
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[2],ymm0[2],ymm5[3],ymm0[3],ymm5[8],ymm0[8],ymm5[9],ymm0[9],ymm5[10],ymm0[10],ymm5[11],ymm0[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
; AVX512DQ-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX512DQ-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} ymm6 = ymm8[4],ymm1[4],ymm8[5],ymm1[5],ymm8[6],ymm1[6],ymm8[7],ymm1[7],ymm8[12],ymm1[12],ymm8[13],ymm1[13],ymm8[14],ymm1[14],ymm8[15],ymm1[15]
; AVX512DQ-FAST-NEXT:    vpermd %ymm6, %ymm17, %ymm6
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm7 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} ymm15 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm7 = ymm15[0],ymm7[0],ymm15[1],ymm7[1],ymm15[2],ymm7[2],ymm15[3],ymm7[3],ymm15[8],ymm7[8],ymm15[9],ymm7[9],ymm15[10],ymm7[10],ymm15[11],ymm7[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm0
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm6, %zmm7, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa (%r8), %ymm15
; AVX512DQ-FAST-NEXT:    vpshufb %ymm12, %ymm15, %ymm5
; AVX512DQ-FAST-NEXT:    vpermi2d %zmm5, %zmm0, %zmm20
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm5 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm5, %ymm19, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa (%r9), %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT:    vpshufb %ymm14, %ymm1, %ymm5
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm22 = ymm5[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm9, %ymm1, %ymm5
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm19 = ymm5[2,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rsi), %xmm6
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm1, %xmm6, %xmm7
; AVX512DQ-FAST-NEXT:    vpshufb %xmm1, %xmm5, %xmm14
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm7[4],xmm14[5],xmm7[5],xmm14[6],xmm7[6],xmm14[7],xmm7[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm17 = ymm7[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm18, %xmm1
; AVX512DQ-FAST-NEXT:    vmovdqa (%rsp), %xmm8 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; AVX512DQ-FAST-NEXT:    vpermd %ymm12, %ymm10, %ymm12
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm9 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm7, %ymm7
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm17, %zmm14, %zmm17
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm7, %zmm12, %zmm17 {%k2}
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm17, %zmm18
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r8), %xmm1
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm1[2,1,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm1, %xmm30
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm9, %zmm16, %zmm18
; AVX512DQ-FAST-NEXT:    vmovdqa (%rsi), %xmm9
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdi), %xmm14
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm7, %xmm9, %xmm1
; AVX512DQ-FAST-NEXT:    vpshufb %xmm7, %xmm14, %xmm12
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm12[4],xmm1[4],xmm12[5],xmm1[5],xmm12[6],xmm1[6],xmm12[7],xmm1[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX512DQ-FAST-NEXT:    vpermd %ymm12, %ymm10, %ymm12
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm10 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm9[0],xmm14[1],xmm9[1],xmm14[2],xmm9[2],xmm14[3],xmm9[3]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm8, %ymm8
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm10, %zmm10
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm8, %zmm12, %zmm10 {%k2}
; AVX512DQ-FAST-NEXT:    vmovdqa (%r8), %xmm12
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm12[2,1,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpermi2d %zmm1, %zmm10, %zmm16
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm9[4],xmm14[5],xmm9[5],xmm14[6],xmm9[6],xmm14[7],xmm9[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm9 = ymm1[1,1,1,1]
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm1[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm8 = ymm1[1,1,1,1]
; AVX512DQ-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm2 = ymm1[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm7 = ymm1[1,1,1,1]
; AVX512DQ-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm3 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm23 = zmm21[0,1,2,3],zmm23[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm21 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm23, %zmm1, %zmm21
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm27 = zmm27[0,1,2,3],zmm29[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm23, %zmm23 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm27, %zmm1, %zmm23
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm13 = zmm13[0,1,2,3],zmm4[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm24, %zmm31, %zmm24
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm13, %zmm1, %zmm24
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm13 = zmm0[0,1,2,3],zmm20[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm19, %zmm22, %zmm19
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm13, %zmm1, %zmm19
; AVX512DQ-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm13 = [0,1,8,3,4,9,6,7]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm1, %ymm13, %ymm25
; AVX512DQ-FAST-NEXT:    vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm1, %ymm13, %ymm11
; AVX512DQ-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm30[0],zero,xmm30[1],zero,xmm30[2],zero,xmm30[3],zero
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm1, %ymm13, %ymm17
; AVX512DQ-FAST-NEXT:    vpmovzxwd {{.*#+}} xmm1 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm1, %ymm13, %ymm10
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm1 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
; AVX512DQ-FAST-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm13 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm14, %zmm9, %zmm13 {%k1}
; AVX512DQ-FAST-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm9 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm2, %zmm8, %zmm9 {%k1}
; AVX512DQ-FAST-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm3, %zmm7, %zmm2 {%k1}
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm15[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [8,9,20,11,12,21,14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm3, %zmm7, %zmm8
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} ymm20 = [0,9,2,3,8,5,6,11]
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm12, %ymm20, %ymm13
; AVX512DQ-FAST-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm12 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm9, %zmm15
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm12, %zmm7, %zmm15
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm12 = zmm25[0,1,2,3],zmm26[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm4
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm3, %xmm26
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm4, %ymm20, %ymm9
; AVX512DQ-FAST-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX512DQ-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX512DQ-FAST-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm5 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
; AVX512DQ-FAST-NEXT:    vmovdqa 96(%r9), %xmm6
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm6[0,0,2,1,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm14 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm14, %xmm6, %xmm3
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm27
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm12, %zmm25, %zmm27
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm22 = zmm11[0,1,2,3],zmm28[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm5, %zmm4, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa 64(%r9), %xmm4
; AVX512DQ-FAST-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm5 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm2, %zmm12
; AVX512DQ-FAST-NEXT:    vpermt2d %zmm5, %zmm7, %zmm12
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm4[0,0,2,1,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm5, %ymm5
; AVX512DQ-FAST-NEXT:    vpshufb %xmm14, %xmm4, %xmm11
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm11, %zmm5, %zmm5
; AVX512DQ-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm26, %xmm28
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm26, %xmm3
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm11
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm11, %ymm20, %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqa 32(%r9), %xmm11
; AVX512DQ-FAST-NEXT:    vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm3 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermi2d %zmm3, %zmm1, %zmm7
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm11[0,0,2,1,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm3, %ymm26
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm30, %xmm0
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm28, %xmm3
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX512DQ-FAST-NEXT:    vmovdqa (%r9), %xmm0
; AVX512DQ-FAST-NEXT:    vpermt2d %ymm3, %ymm20, %ymm1
; AVX512DQ-FAST-NEXT:    vpshufb %xmm14, %xmm11, %xmm3
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm20 = ymm3[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm14, %xmm0, %xmm14
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm11, %xmm11
; AVX512DQ-FAST-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX512DQ-FAST-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpbroadcastq %xmm0, %ymm0
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm28 = ymm3[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm3 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm29 = ymm6[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm30 = ymm4[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm4 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm31 = ymm11[0,1,0,1]
; AVX512DQ-FAST-NEXT:    vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT:    # ymm11 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm22, %zmm25, %zmm5
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm17 = zmm17[0,1,2,3],zmm18[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm20, %zmm26, %zmm18
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm17, %zmm25, %zmm18
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm10 = zmm10[0,1,2,3],zmm16[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm14, %zmm0, %zmm0
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm10, %zmm25, %zmm0
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm28, %zmm3
; AVX512DQ-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm8, %zmm10, %zmm3
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm8 = zmm9[0,1,2,3],zmm15[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm29, %zmm6
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm8, %zmm10, %zmm6
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm12[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm30, %zmm4
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm2, %zmm10, %zmm4
; AVX512DQ-FAST-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm7[0,1,2,3]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm11, %zmm31, %zmm2
; AVX512DQ-FAST-NEXT:    vpternlogq $184, %zmm1, %zmm10, %zmm2
; AVX512DQ-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm2, 256(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm4, 448(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm6, 640(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm18, 192(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm19, 128(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm24, 320(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm5, 384(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm27, 576(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm23, 512(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm21, 704(%rax)
; AVX512DQ-FAST-NEXT:    addq $920, %rsp # imm = 0x398
; AVX512DQ-FAST-NEXT:    vzeroupper
; AVX512DQ-FAST-NEXT:    retq
;
; AVX512BW-LABEL: store_i16_stride6_vf64:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm8
; AVX512BW-NEXT:    vmovdqa64 64(%rsi), %zmm24
; AVX512BW-NEXT:    vmovdqa64 (%rdx), %zmm1
; AVX512BW-NEXT:    vmovdqa64 64(%rdx), %zmm4
; AVX512BW-NEXT:    vmovdqa64 (%rcx), %zmm7
; AVX512BW-NEXT:    vmovdqa64 64(%rcx), %zmm9
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm13 = [0,0,0,32,3,35,0,0,1,33,4,36,0,0,2,34,0,0,0,32,3,35,0,0,1,33,4,36,0,0,2,34]
; AVX512BW-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm10
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm13, %zmm10
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm25 = [0,32,3,35,0,0,1,33,4,36,0,0,2,34,5,37,0,32,3,35,0,0,1,33,4,36,0,0,2,34,5,37]
; AVX512BW-NEXT:    # zmm25 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm25, %zmm3
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [29,61,0,0,27,59,30,62,0,0,28,60,31,63,0,0,29,61,0,0,27,59,30,62,0,0,28,60,31,63,0,0]
; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm12, %zmm15
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,27,59,30,62,0,0,28,60,31,63,0,0,29,61,0,0,27,59,30,62,0,0,28,60,31,63,0,0,29,61]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5
; AVX512BW-NEXT:    vpermt2w %zmm24, %zmm6, %zmm5
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm17 = [24,56,0,0,22,54,25,57,0,0,23,55,26,58,0,0,24,56,0,0,22,54,25,57,0,0,23,55,26,58,0,0]
; AVX512BW-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19
; AVX512BW-NEXT:    vpermt2w %zmm24, %zmm17, %zmm19
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [21,53,24,56,0,0,22,54,25,57,0,0,23,55,26,58,21,53,24,56,0,0,22,54,25,57,0,0,23,55,26,58]
; AVX512BW-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm14, %zmm11
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,16,48,19,51,0,0,17,49,20,52,0,0,18,50,0,0,16,48,19,51,0,0,17,49,20,52,0,0,18,50]
; AVX512BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm23
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm21, %zmm23
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [16,48,19,51,0,0,17,49,20,52,0,0,18,50,21,53,16,48,19,51,0,0,17,49,20,52,0,0,18,50,21,53]
; AVX512BW-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm16
; AVX512BW-NEXT:    vpermt2w %zmm24, %zmm18, %zmm16
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,0,11,43,14,46,0,0,12,44,15,47,0,0,13,45,0,0,11,43,14,46,0,0,12,44,15,47,0,0,13,45]
; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm20
; AVX512BW-NEXT:    vpermt2w %zmm24, %zmm22, %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm26
; AVX512BW-NEXT:    vpermt2w %zmm24, %zmm25, %zmm0
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm25 = [8,40,0,0,6,38,9,41,0,0,7,39,10,42,0,0,8,40,0,0,6,38,9,41,0,0,7,39,10,42,0,0]
; AVX512BW-NEXT:    # zmm25 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2w %zmm24, %zmm25, %zmm26
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [13,45,0,0,11,43,14,46,0,0,12,44,15,47,0,0,13,45,0,0,11,43,14,46,0,0,12,44,15,47,0,0]
; AVX512BW-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2w %zmm8, %zmm2, %zmm6
; AVX512BW-NEXT:    vpermi2w %zmm8, %zmm2, %zmm17
; AVX512BW-NEXT:    vpermi2w %zmm8, %zmm2, %zmm18
; AVX512BW-NEXT:    vpermi2w %zmm8, %zmm2, %zmm22
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm25, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm8
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm24, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm25
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm13, %zmm4
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm13 = [5,37,8,40,0,0,6,38,9,41,0,0,7,39,10,42,5,37,8,40,0,0,6,38,9,41,0,0,7,39,10,42]
; AVX512BW-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm13, %zmm25
; AVX512BW-NEXT:    vpermi2w %zmm7, %zmm1, %zmm12
; AVX512BW-NEXT:    vpermi2w %zmm7, %zmm1, %zmm14
; AVX512BW-NEXT:    vpermi2w %zmm7, %zmm1, %zmm21
; AVX512BW-NEXT:    vpermi2w %zmm7, %zmm1, %zmm24
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm13, %zmm1
; AVX512BW-NEXT:    movw $9362, %ax # imm = 0x2492
; AVX512BW-NEXT:    kmovd %eax, %k2
; AVX512BW-NEXT:    vmovdqa32 %zmm10, %zmm3 {%k2}
; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm7
; AVX512BW-NEXT:    movw $18724, %ax # imm = 0x4924
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vmovdqa32 %zmm15, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqa64 64(%r8), %zmm9
; AVX512BW-NEXT:    vmovdqa32 %zmm19, %zmm11 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm10 = <0,1,2,3,32,u,6,7,8,9,33,u,12,13,14,15,34,u,18,19,20,21,35,u,24,25,26,27,36,u,30,31>
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqa32 %zmm23, %zmm16 {%k2}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm13 = <58,u,2,3,4,5,59,u,8,9,10,11,60,u,14,15,16,17,61,u,20,21,22,23,62,u,26,27,28,29,63,u>
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm13, %zmm5
; AVX512BW-NEXT:    vmovdqa32 %zmm8, %zmm20 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm8 = <0,1,53,u,4,5,6,7,54,u,10,11,12,13,55,u,16,17,18,19,56,u,22,23,24,25,57,u,28,29,30,31>
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm8, %zmm11
; AVX512BW-NEXT:    vmovdqa32 %zmm26, %zmm25 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm15 = <0,1,2,3,48,u,6,7,8,9,49,u,12,13,14,15,50,u,18,19,20,21,51,u,24,25,26,27,52,u,30,31>
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm15, %zmm16
; AVX512BW-NEXT:    vmovdqa32 %zmm4, %zmm0 {%k2}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm4 = <42,u,2,3,4,5,43,u,8,9,10,11,44,u,14,15,16,17,45,u,20,21,22,23,46,u,26,27,28,29,47,u>
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm4, %zmm20
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm10, %zmm0
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm10 = <0,1,37,u,4,5,6,7,38,u,10,11,12,13,39,u,16,17,18,19,40,u,22,23,24,25,41,u,28,29,30,31>
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm10, %zmm25
; AVX512BW-NEXT:    vmovdqa32 %zmm12, %zmm6 {%k1}
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm13, %zmm6
; AVX512BW-NEXT:    vmovdqa32 %zmm17, %zmm14 {%k1}
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm8, %zmm14
; AVX512BW-NEXT:    vmovdqa32 %zmm21, %zmm18 {%k2}
; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm8
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm15, %zmm18
; AVX512BW-NEXT:    vmovdqa64 64(%r9), %zmm9
; AVX512BW-NEXT:    vmovdqa32 %zmm24, %zmm22 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [0,1,2,3,4,32,6,7,8,9,10,33,12,13,14,15,16,34,18,19,20,21,22,35,24,25,26,27,28,36,30,31]
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm12, %zmm3
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm4, %zmm22
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [0,58,2,3,4,5,6,59,8,9,10,11,12,60,14,15,16,17,18,61,20,21,22,23,24,62,26,27,28,29,30,63]
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm4, %zmm5
; AVX512BW-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [0,1,2,53,4,5,6,7,8,54,10,11,12,13,14,55,16,17,18,19,20,56,22,23,24,25,26,57,28,29,30,31]
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm2, %zmm11
; AVX512BW-NEXT:    vpermt2w %zmm7, %zmm10, %zmm1
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [0,1,2,3,4,48,6,7,8,9,10,49,12,13,14,15,16,50,18,19,20,21,22,51,24,25,26,27,28,52,30,31]
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm7, %zmm16
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [0,42,2,3,4,5,6,43,8,9,10,11,12,44,14,15,16,17,18,45,20,21,22,23,24,46,26,27,28,29,30,47]
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm10, %zmm20
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm12, %zmm0
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [0,1,2,37,4,5,6,7,8,38,10,11,12,13,14,39,16,17,18,19,20,40,22,23,24,25,26,41,28,29,30,31]
; AVX512BW-NEXT:    vpermt2w %zmm9, %zmm12, %zmm25
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm4, %zmm6
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm2, %zmm14
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm7, %zmm18
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm10, %zmm22
; AVX512BW-NEXT:    vpermt2w %zmm8, %zmm12, %zmm1
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm22, 128(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm18, 192(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm14, 256(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm6, 320(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, 384(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm20, 512(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm16, 576(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm11, 640(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, 704(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %in.vec0 = load <64 x i16>, ptr %in.vecptr0, align 64
  %in.vec1 = load <64 x i16>, ptr %in.vecptr1, align 64
  %in.vec2 = load <64 x i16>, ptr %in.vecptr2, align 64
  %in.vec3 = load <64 x i16>, ptr %in.vecptr3, align 64
  %in.vec4 = load <64 x i16>, ptr %in.vecptr4, align 64
  %in.vec5 = load <64 x i16>, ptr %in.vecptr5, align 64
  %1 = shufflevector <64 x i16> %in.vec0, <64 x i16> %in.vec1, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
  %2 = shufflevector <64 x i16> %in.vec2, <64 x i16> %in.vec3, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
  %3 = shufflevector <64 x i16> %in.vec4, <64 x i16> %in.vec5, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
  %4 = shufflevector <128 x i16> %1, <128 x i16> %2, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>
  %5 = shufflevector <128 x i16> %3, <128 x i16> poison, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %6 = shufflevector <256 x i16> %4, <256 x i16> %5, <384 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255, i32 256, i32 257, i32 258, i32 259, i32 260, i32 261, i32 262, i32 263, i32 264, i32 265, i32 266, i32 267, i32 268, i32 269, i32 270, i32 271, i32 272, i32 273, i32 274, i32 275, i32 276, i32 277, i32 278, i32 279, i32 280, i32 281, i32 282, i32 283, i32 284, i32 285, i32 286, i32 287, i32 288, i32 289, i32 290, i32 291, i32 292, i32 293, i32 294, i32 295, i32 296, i32 297, i32 298, i32 299, i32 300, i32 301, i32 302, i32 303, i32 304, i32 305, i32 306, i32 307, i32 308, i32 309, i32 310, i32 311, i32 312, i32 313, i32 314, i32 315, i32 316, i32 317, i32 318, i32 319, i32 320, i32 321, i32 322, i32 323, i32 324, i32 325, i32 326, i32 327, i32 328, i32 329, i32 330, i32 331, i32 332, i32 333, i32 334, i32 335, i32 336, i32 337, i32 338, i32 339, i32 340, i32 341, i32 342, i32 343, i32 344, i32 345, i32 346, i32 347, i32 348, i32 349, i32 350, i32 351, i32 352, i32 353, i32 354, i32 355, i32 356, i32 357, i32 358, i32 359, i32 360, i32 361, i32 362, i32 363, i32 364, i32 365, i32 366, i32 367, i32 368, i32 369, i32 370, i32 371, i32 372, i32 373, i32 374, i32 375, i32 376, i32 377, i32 378, i32 379, i32 380, i32 381, i32 382, i32 383>
  %interleaved.vec = shufflevector <384 x i16> %6, <384 x i16> poison, <384 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 320, i32 1, i32 65, i32 129, i32 193, i32 257, i32 321, i32 2, i32 66, i32 130, i32 194, i32 258, i32 322, i32 3, i32 67, i32 131, i32 195, i32 259, i32 323, i32 4, i32 68, i32 132, i32 196, i32 260, i32 324, i32 5, i32 69, i32 133, i32 197, i32 261, i32 325, i32 6, i32 70, i32 134, i32 198, i32 262, i32 326, i32 7, i32 71, i32 135, i32 199, i32 263, i32 327, i32 8, i32 72, i32 136, i32 200, i32 264, i32 328, i32 9, i32 73, i32 137, i32 201, i32 265, i32 329, i32 10, i32 74, i32 138, i32 202, i32 266, i32 330, i32 11, i32 75, i32 139, i32 203, i32 267, i32 331, i32 12, i32 76, i32 140, i32 204, i32 268, i32 332, i32 13, i32 77, i32 141, i32 205, i32 269, i32 333, i32 14, i32 78, i32 142, i32 206, i32 270, i32 334, i32 15, i32 79, i32 143, i32 207, i32 271, i32 335, i32 16, i32 80, i32 144, i32 208, i32 272, i32 336, i32 17, i32 81, i32 145, i32 209, i32 273, i32 337, i32 18, i32 82, i32 146, i32 210, i32 274, i32 338, i32 19, i32 83, i32 147, i32 211, i32 275, i32 339, i32 20, i32 84, i32 148, i32 212, i32 276, i32 340, i32 21, i32 85, i32 149, i32 213, i32 277, i32 341, i32 22, i32 86, i32 150, i32 214, i32 278, i32 342, i32 23, i32 87, i32 151, i32 215, i32 279, i32 343, i32 24, i32 88, i32 152, i32 216, i32 280, i32 344, i32 25, i32 89, i32 153, i32 217, i32 281, i32 345, i32 26, i32 90, i32 154, i32 218, i32 282, i32 346, i32 27, i32 91, i32 155, i32 219, i32 283, i32 347, i32 28, i32 92, i32 156, i32 220, i32 284, i32 348, i32 29, i32 93, i32 157, i32 221, i32 285, i32 349, i32 30, i32 94, i32 158, i32 222, i32 286, i32 350, i32 31, i32 95, i32 159, i32 223, i32 287, i32 351, i32 32, i32 96, i32 160, i32 224, i32 288, i32 352, i32 33, i32 97, i32 161, i32 225, i32 289, i32 353, i32 34, i32 98, i32 162, i32 226, i32 290, i32 354, i32 35, i32 99, i32 163, i32 227, i32 291, i32 355, i32 36, i32 100, i32 164, i32 228, i32 292, i32 356, i32 37, i32 101, i32 165, i32 229, i32 293, i32 357, i32 38, i32 102, i32 166, i32 230, i32 294, i32 358, i32 39, i32 103, i32 167, i32 231, i32 295, i32 359, i32 40, i32 104, i32 168, i32 232, i32 296, i32 360, i32 41, i32 105, i32 169, i32 233, i32 297, i32 361, i32 42, i32 106, i32 170, i32 234, i32 298, i32 362, i32 43, i32 107, i32 171, i32 235, i32 299, i32 363, i32 44, i32 108, i32 172, i32 236, i32 300, i32 364, i32 45, i32 109, i32 173, i32 237, i32 301, i32 365, i32 46, i32 110, i32 174, i32 238, i32 302, i32 366, i32 47, i32 111, i32 175, i32 239, i32 303, i32 367, i32 48, i32 112, i32 176, i32 240, i32 304, i32 368, i32 49, i32 113, i32 177, i32 241, i32 305, i32 369, i32 50, i32 114, i32 178, i32 242, i32 306, i32 370, i32 51, i32 115, i32 179, i32 243, i32 307, i32 371, i32 52, i32 116, i32 180, i32 244, i32 308, i32 372, i32 53, i32 117, i32 181, i32 245, i32 309, i32 373, i32 54, i32 118, i32 182, i32 246, i32 310, i32 374, i32 55, i32 119, i32 183, i32 247, i32 311, i32 375, i32 56, i32 120, i32 184, i32 248, i32 312, i32 376, i32 57, i32 121, i32 185, i32 249, i32 313, i32 377, i32 58, i32 122, i32 186, i32 250, i32 314, i32 378, i32 59, i32 123, i32 187, i32 251, i32 315, i32 379, i32 60, i32 124, i32 188, i32 252, i32 316, i32 380, i32 61, i32 125, i32 189, i32 253, i32 317, i32 381, i32 62, i32 126, i32 190, i32 254, i32 318, i32 382, i32 63, i32 127, i32 191, i32 255, i32 319, i32 383>
  store <384 x i16> %interleaved.vec, ptr %out.vec, align 64
  ret void
}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; AVX: {{.*}}
; AVX1: {{.*}}
; AVX2: {{.*}}
; AVX2-ONLY: {{.*}}
; AVX512: {{.*}}
; AVX512BW-FAST: {{.*}}
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512BW-ONLY-SLOW: {{.*}}
; AVX512BW-SLOW: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512DQBW-SLOW: {{.*}}
; AVX512F: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
; FALLBACK11: {{.*}}
; FALLBACK12: {{.*}}
; FALLBACK2: {{.*}}
; FALLBACK3: {{.*}}
; FALLBACK4: {{.*}}
; FALLBACK5: {{.*}}
; FALLBACK6: {{.*}}
; FALLBACK7: {{.*}}
; FALLBACK8: {{.*}}
; FALLBACK9: {{.*}}