; 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-ONLY,AVX512-SLOW,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-ONLY,AVX512-FAST,AVX512F-FAST,AVX512F-ONLY-FAST,FALLBACK6
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512DQ-ONLY,AVX512-SLOW,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,AVX512DQ-ONLY,AVX512-FAST,AVX512F-FAST,AVX512DQ-FAST,FALLBACK8
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-ONLY,AVX512-SLOW,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-ONLY,AVX512-FAST,AVX512BW-FAST,AVX512BW-ONLY-FAST,FALLBACK10
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512DQBW-ONLY,AVX512-SLOW,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,AVX512DQBW-ONLY,AVX512-FAST,AVX512BW-FAST,AVX512DQBW-FAST,FALLBACK12

; These patterns are produced by LoopVectorizer for interleaved stores.

define void @store_i8_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride8_vf2:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT:    movdqa (%rdi), %xmm0
; SSE-NEXT:    movdqa (%rdx), %xmm1
; SSE-NEXT:    movdqa (%r8), %xmm2
; SSE-NEXT:    movdqa (%r11), %xmm3
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; SSE-NEXT:    pxor %xmm1, %xmm1
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm2[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,1,0,2]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[3,1,2,3,4,5,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]
; SSE-NEXT:    packuswb %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,3,2,3]
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE-NEXT:    movdqa %xmm0, (%rax)
; SSE-NEXT:    retq
;
; AVX-LABEL: store_i8_stride8_vf2:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX-NEXT:    vmovdqa (%rdi), %xmm0
; AVX-NEXT:    vmovdqa (%rdx), %xmm1
; AVX-NEXT:    vmovdqa (%r8), %xmm2
; AVX-NEXT:    vmovdqa (%r11), %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; AVX-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
; AVX-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
; AVX-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15]
; AVX-NEXT:    vmovdqa %xmm0, (%rax)
; AVX-NEXT:    retq
  %in.vec0 = load <2 x i8>, ptr %in.vecptr0, align 64
  %in.vec1 = load <2 x i8>, ptr %in.vecptr1, align 64
  %in.vec2 = load <2 x i8>, ptr %in.vecptr2, align 64
  %in.vec3 = load <2 x i8>, ptr %in.vecptr3, align 64
  %in.vec4 = load <2 x i8>, ptr %in.vecptr4, align 64
  %in.vec5 = load <2 x i8>, ptr %in.vecptr5, align 64
  %in.vec6 = load <2 x i8>, ptr %in.vecptr6, align 64
  %in.vec7 = load <2 x i8>, ptr %in.vecptr7, align 64
  %1 = shufflevector <2 x i8> %in.vec0, <2 x i8> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %2 = shufflevector <2 x i8> %in.vec2, <2 x i8> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %3 = shufflevector <2 x i8> %in.vec4, <2 x i8> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = shufflevector <2 x i8> %in.vec6, <2 x i8> %in.vec7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %5 = shufflevector <4 x i8> %1, <4 x i8> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = shufflevector <4 x i8> %3, <4 x i8> %4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %7 = shufflevector <8 x i8> %5, <8 x i8> %6, <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>
  %interleaved.vec = shufflevector <16 x i8> %7, <16 x i8> poison, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
  store <16 x i8> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i8_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride8_vf4:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT:    movdqa (%rdi), %xmm0
; SSE-NEXT:    movdqa (%rdx), %xmm1
; SSE-NEXT:    movdqa (%r8), %xmm2
; SSE-NEXT:    movdqa (%r11), %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:    punpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1]
; SSE-NEXT:    pxor %xmm6, %xmm6
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,5,7,5]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,4,5,6,4]
; SSE-NEXT:    packuswb %xmm5, %xmm7
; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [65535,65535,65535,0,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm4, %xmm5
; SSE-NEXT:    pandn %xmm7, %xmm5
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm7[0,1,2,3,7,5,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,6,4,6,7]
; SSE-NEXT:    packuswb %xmm8, %xmm7
; SSE-NEXT:    pand %xmm4, %xmm7
; SSE-NEXT:    por %xmm5, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm5[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm9 = xmm5[0,1,0,2,4,5,6,7]
; SSE-NEXT:    packuswb %xmm8, %xmm9
; SSE-NEXT:    movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm5, %xmm8
; SSE-NEXT:    pandn %xmm9, %xmm8
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm9 = xmm6[1,3,2,3,4,5,6,7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm9, %xmm6
; SSE-NEXT:    pand %xmm5, %xmm6
; SSE-NEXT:    por %xmm8, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,1,1,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm3[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm7, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm2[0,1,2,3,5,7,6,7]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm2
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    por %xmm2, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm1[0,1,3,1,4,5,6,7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm3, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm0[3,1,2,3,4,5,6,7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm3, %xmm0
; SSE-NEXT:    pand %xmm5, %xmm0
; SSE-NEXT:    pandn %xmm1, %xmm5
; SSE-NEXT:    por %xmm0, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE-NEXT:    movdqa %xmm0, 16(%rax)
; SSE-NEXT:    movdqa %xmm6, (%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i8_stride8_vf4:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa (%r11), %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],mem[0],xmm2[1],mem[1]
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm2 = [2,6,10,14,3,7,11,15,2,6,10,14,3,7,11,15]
; AVX1-ONLY-NEXT:    # xmm2 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm2, %xmm1, %xmm3
; AVX1-ONLY-NEXT:    vpshufb %xmm2, %xmm0, %xmm2
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm3 = [0,4,8,12,1,5,9,13,0,4,8,12,1,5,9,13]
; AVX1-ONLY-NEXT:    # xmm3 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX1-ONLY-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX1-ONLY-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-LABEL: store_i8_stride8_vf4:
; AVX2:       # %bb.0:
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-NEXT:    vmovdqa (%rsi), %xmm1
; AVX2-NEXT:    vmovdqa (%rdx), %xmm2
; AVX2-NEXT:    vmovdqa (%rcx), %xmm3
; AVX2-NEXT:    vinserti128 $1, (%r11), %ymm3, %ymm3
; AVX2-NEXT:    vinserti128 $1, (%r10), %ymm2, %ymm2
; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]
; AVX2-NEXT:    vinserti128 $1, (%r9), %ymm1, %ymm1
; AVX2-NEXT:    vinserti128 $1, (%r8), %ymm0, %ymm0
; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
; AVX2-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31]
; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7]
; AVX2-NEXT:    vpermd %ymm0, %ymm1, %ymm0
; AVX2-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
  %in.vec0 = load <4 x i8>, ptr %in.vecptr0, align 64
  %in.vec1 = load <4 x i8>, ptr %in.vecptr1, align 64
  %in.vec2 = load <4 x i8>, ptr %in.vecptr2, align 64
  %in.vec3 = load <4 x i8>, ptr %in.vecptr3, align 64
  %in.vec4 = load <4 x i8>, ptr %in.vecptr4, align 64
  %in.vec5 = load <4 x i8>, ptr %in.vecptr5, align 64
  %in.vec6 = load <4 x i8>, ptr %in.vecptr6, align 64
  %in.vec7 = load <4 x i8>, ptr %in.vecptr7, align 64
  %1 = shufflevector <4 x i8> %in.vec0, <4 x i8> %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 i8> %in.vec2, <4 x i8> %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 i8> %in.vec4, <4 x i8> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = shufflevector <4 x i8> %in.vec6, <4 x i8> %in.vec7, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %5 = shufflevector <8 x i8> %1, <8 x i8> %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>
  %6 = shufflevector <8 x i8> %3, <8 x i8> %4, <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>
  %7 = shufflevector <16 x i8> %5, <16 x i8> %6, <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>
  %interleaved.vec = shufflevector <32 x i8> %7, <32 x i8> poison, <32 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 29, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 30, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27, i32 31>
  store <32 x i8> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i8_stride8_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride8_vf8:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSE-NEXT:    movq {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm3 = mem[0],zero
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; SSE-NEXT:    movq {{.*#+}} xmm5 = mem[0],zero
; SSE-NEXT:    movq {{.*#+}} xmm3 = mem[0],zero
; SSE-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3],xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm5[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,1]
; SSE-NEXT:    movdqa {{.*#+}} xmm3 = [65535,65535,65535,0,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm3, %xmm6
; SSE-NEXT:    pandn %xmm4, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm2[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,0,2,1]
; SSE-NEXT:    pand %xmm3, %xmm4
; SSE-NEXT:    por %xmm6, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,3,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm1[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,1,1,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [65535,0,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm4, %xmm8
; SSE-NEXT:    pandn %xmm6, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm4, %xmm6
; SSE-NEXT:    por %xmm8, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm5[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    movdqa %xmm3, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm2[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    pand %xmm3, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,3,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm1[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3]
; SSE-NEXT:    movdqa %xmm4, %xmm9
; SSE-NEXT:    pandn %xmm7, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm4, %xmm7
; SSE-NEXT:    por %xmm9, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm5[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3]
; SSE-NEXT:    movdqa %xmm3, %xmm9
; SSE-NEXT:    pandn %xmm8, %xmm9
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm2[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3]
; SSE-NEXT:    pand %xmm3, %xmm8
; SSE-NEXT:    por %xmm9, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[1,3,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm1[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[2,1,3,3]
; SSE-NEXT:    movdqa %xmm4, %xmm10
; SSE-NEXT:    pandn %xmm9, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm4, %xmm9
; SSE-NEXT:    por %xmm10, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    pand %xmm3, %xmm2
; SSE-NEXT:    pandn %xmm5, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,3,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    por %xmm0, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE-NEXT:    movdqa %xmm0, 48(%rax)
; SSE-NEXT:    movdqa %xmm9, 32(%rax)
; SSE-NEXT:    movdqa %xmm7, 16(%rax)
; SSE-NEXT:    movdqa %xmm6, (%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i8_stride8_vf8:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; 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 {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX1-ONLY-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX1-ONLY-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm4 = [0,0,2,10,0,0,3,11,0,0,2,10,0,0,3,11]
; AVX1-ONLY-NEXT:    # xmm4 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
; AVX1-ONLY-NEXT:    vpshufb %xmm4, %xmm2, %xmm4
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm5 = [2,10,0,0,3,11,0,0,2,10,0,0,3,11,0,0]
; AVX1-ONLY-NEXT:    # xmm5 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm5, %xmm1, %xmm6
; AVX1-ONLY-NEXT:    vpshufb %xmm5, %xmm0, %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3],xmm5[4,5],xmm4[6,7]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm5 = [0,0,0,8,0,0,1,9,0,0,0,8,0,0,1,9]
; AVX1-ONLY-NEXT:    # xmm5 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
; AVX1-ONLY-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm6 = [0,8,0,0,1,9,0,0,0,8,0,0,1,9,0,0]
; AVX1-ONLY-NEXT:    # xmm6 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm6, %xmm1, %xmm7
; AVX1-ONLY-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm5 = [0,0,6,14,0,0,7,15,0,0,6,14,0,0,7,15]
; AVX1-ONLY-NEXT:    # xmm5 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
; AVX1-ONLY-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm6 = [6,14,0,0,7,15,0,0,6,14,0,0,7,15,0,0]
; AVX1-ONLY-NEXT:    # xmm6 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm6, %xmm1, %xmm7
; AVX1-ONLY-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5],xmm5[6,7]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm6 = [0,0,4,12,0,0,5,13,0,0,4,12,0,0,5,13]
; AVX1-ONLY-NEXT:    # xmm6 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm6, %xmm3, %xmm3
; AVX1-ONLY-NEXT:    vpshufb %xmm6, %xmm2, %xmm2
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX1-ONLY-NEXT:    vmovddup {{.*#+}} xmm3 = [4,12,0,0,5,13,0,0,4,12,0,0,5,13,0,0]
; AVX1-ONLY-NEXT:    # xmm3 = mem[0,0]
; AVX1-ONLY-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX1-ONLY-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX1-ONLY-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm4, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i8_stride8_vf8:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; 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:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-SLOW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,ymm1[0,8],zero,zero,zero,zero,zero,zero,ymm1[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm1[18,26],zero,zero,zero,zero,zero,zero,ymm1[19,27]
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[2,3,0,1]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,zero,ymm3[0,8],zero,zero,zero,zero,zero,zero,ymm3[1,9],zero,zero,zero,zero,ymm3[18,26],zero,zero,zero,zero,zero,zero,ymm3[19,27],zero,zero
; AVX2-SLOW-NEXT:    vpor %ymm4, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm0[0,8],zero,zero,zero,zero,zero,zero,ymm0[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[18,26],zero,zero,zero,zero,zero,zero,ymm0[19,27],zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[2,3,0,1]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,ymm5[0,8],zero,zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,ymm5[18,26],zero,zero,zero,zero,zero,zero,ymm5[19,27],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpor %ymm6, %ymm4, %ymm4
; AVX2-SLOW-NEXT:    vpor %ymm2, %ymm4, %ymm2
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,ymm1[4,12],zero,zero,zero,zero,zero,zero,ymm1[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm1[22,30],zero,zero,zero,zero,zero,zero,ymm1[23,31]
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,ymm3[4,12],zero,zero,zero,zero,zero,zero,ymm3[5,13],zero,zero,zero,zero,ymm3[22,30],zero,zero,zero,zero,zero,zero,ymm3[23,31],zero,zero
; AVX2-SLOW-NEXT:    vpor %ymm3, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[4,12],zero,zero,zero,zero,zero,zero,ymm0[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[22,30],zero,zero,zero,zero,zero,zero,ymm0[23,31],zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = zero,zero,ymm5[4,12],zero,zero,zero,zero,zero,zero,ymm5[5,13],zero,zero,zero,zero,ymm5[22,30],zero,zero,zero,zero,zero,zero,ymm5[23,31],zero,zero,zero,zero,zero,zero
; AVX2-SLOW-NEXT:    vpor %ymm3, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpor %ymm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 32(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i8_stride8_vf8:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; 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:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-FAST-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [0,2,4,6,0,2,4,6]
; AVX2-FAST-NEXT:    # ymm2 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm2, %ymm3
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm5 = [0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7]
; AVX2-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [1,3,5,7,1,3,5,7]
; AVX2-FAST-NEXT:    # ymm3 = mem[0,1,0,1]
; AVX2-FAST-NEXT:    vpermd %ymm1, %ymm3, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpermd %ymm0, %ymm3, %ymm0
; AVX2-FAST-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 32(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i8_stride8_vf8:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; 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:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-FAST-PERLANE-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,ymm1[0,8],zero,zero,zero,zero,zero,zero,ymm1[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm1[18,26],zero,zero,zero,zero,zero,zero,ymm1[19,27]
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm3 = ymm1[2,3,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,zero,ymm3[0,8],zero,zero,zero,zero,zero,zero,ymm3[1,9],zero,zero,zero,zero,ymm3[18,26],zero,zero,zero,zero,zero,zero,ymm3[19,27],zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpor %ymm4, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm0[0,8],zero,zero,zero,zero,zero,zero,ymm0[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[18,26],zero,zero,zero,zero,zero,zero,ymm0[19,27],zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm5 = ymm0[2,3,0,1]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm6 = zero,zero,ymm5[0,8],zero,zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,ymm5[18,26],zero,zero,zero,zero,zero,zero,ymm5[19,27],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpor %ymm6, %ymm4, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpor %ymm2, %ymm4, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,ymm1[4,12],zero,zero,zero,zero,zero,zero,ymm1[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm1[22,30],zero,zero,zero,zero,zero,zero,ymm1[23,31]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,ymm3[4,12],zero,zero,zero,zero,zero,zero,ymm3[5,13],zero,zero,zero,zero,ymm3[22,30],zero,zero,zero,zero,zero,zero,ymm3[23,31],zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpor %ymm3, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[4,12],zero,zero,zero,zero,zero,zero,ymm0[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[22,30],zero,zero,zero,zero,zero,zero,ymm0[23,31],zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm3 = zero,zero,ymm5[4,12],zero,zero,zero,zero,zero,zero,ymm5[5,13],zero,zero,zero,zero,ymm5[22,30],zero,zero,zero,zero,zero,zero,ymm5[23,31],zero,zero,zero,zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vpor %ymm3, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpor %ymm1, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-SLOW-LABEL: store_i8_stride8_vf8:
; AVX512F-SLOW:       # %bb.0:
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; 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:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512F-SLOW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX512F-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm1[u,u,u,u,4,12],zero,zero,ymm1[u,u,u,u,5,13],zero,zero,ymm1[u,u,u,u],zero,zero,ymm1[22,30,u,u,u,u],zero,zero,ymm1[23,31]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[4,12],zero,zero,ymm0[u,u,u,u,5,13],zero,zero,ymm0[u,u,u,u],zero,zero,ymm0[22,30,u,u,u,u],zero,zero,ymm0[23,31,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,0,8],zero,zero,ymm1[u,u,u,u,1,9],zero,zero,ymm1[u,u,u,u],zero,zero,ymm1[18,26,u,u,u,u],zero,zero,ymm1[19,27]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm0[0,8],zero,zero,ymm0[u,u,u,u,1,9],zero,zero,ymm0[u,u,u,u],zero,zero,ymm0[18,26,u,u,u,u],zero,zero,ymm0[19,27,u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u],zero,zero,ymm1[4,12,u,u,u,u],zero,zero,ymm1[5,13,u,u,u,u,22,30],zero,zero,ymm1[u,u,u,u,23,31],zero,zero
; AVX512F-SLOW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = zero,zero,ymm0[4,12,u,u,u,u],zero,zero,ymm0[5,13,u,u,u,u,22,30],zero,zero,ymm0[u,u,u,u,23,31],zero,zero,ymm0[u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u],zero,zero,ymm1[0,8,u,u,u,u],zero,zero,ymm1[1,9,u,u,u,u,18,26],zero,zero,ymm1[u,u,u,u,19,27],zero,zero
; AVX512F-SLOW-NEXT:    vpshufb {{.*#+}} ymm0 = zero,zero,ymm0[0,8,u,u,u,u],zero,zero,ymm0[1,9,u,u,u,u,18,26],zero,zero,ymm0[u,u,u,u,19,27],zero,zero,ymm0[u,u,u,u]
; AVX512F-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm0
; AVX512F-SLOW-NEXT:    vpord %zmm0, %zmm2, %zmm0
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512F-SLOW-NEXT:    vzeroupper
; AVX512F-SLOW-NEXT:    retq
;
; AVX512F-FAST-LABEL: store_i8_stride8_vf8:
; AVX512F-FAST:       # %bb.0:
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; 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 {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512F-FAST-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX512F-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm2 = [1,3,5,7,1,3,5,7]
; AVX512F-FAST-NEXT:    # ymm2 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm1, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15]
; AVX512F-FAST-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
; AVX512F-FAST-NEXT:    vpermd %ymm0, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm5 = [0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15,u,u,u,u]
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7]
; AVX512F-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [0,2,4,6,0,2,4,6]
; AVX512F-FAST-NEXT:    # ymm3 = mem[0,1,0,1]
; AVX512F-FAST-NEXT:    vpermd %ymm1, %ymm3, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm4, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vpermd %ymm0, %ymm3, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512F-FAST-NEXT:    vzeroupper
; AVX512F-FAST-NEXT:    retq
;
; AVX512BW-SLOW-LABEL: store_i8_stride8_vf8:
; AVX512BW-SLOW:       # %bb.0:
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX512BW-SLOW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpshufb {{.*#+}} zmm1 = zmm0[0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,u,u,18,26,u,u,u,u,u,u,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u]
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[2,3,0,1,6,7,4,5]
; AVX512BW-SLOW-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,u,u,19,27,u,u,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u,u,u]
; AVX512BW-SLOW-NEXT:    movl $287445282, %ecx # imm = 0x11221122
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT:    vmovdqu16 %zmm0, %zmm1 {%k1}
; AVX512BW-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm2[2,3,0,1,2,3,0,1]
; AVX512BW-SLOW-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,u,u,19,27,u,u,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u]
; AVX512BW-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3,0,1,2,3]
; AVX512BW-SLOW-NEXT:    vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,u,u,18,26,u,u,u,u,u,u,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63]
; AVX512BW-SLOW-NEXT:    movl $1149781128, %ecx # imm = 0x44884488
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT:    vmovdqu16 %zmm0, %zmm2 {%k1}
; AVX512BW-SLOW-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm1, (%rax)
; AVX512BW-SLOW-NEXT:    vzeroupper
; AVX512BW-SLOW-NEXT:    retq
;
; AVX512BW-FAST-LABEL: store_i8_stride8_vf8:
; AVX512BW-FAST:       # %bb.0:
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512BW-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512BW-FAST-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX512BW-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [8,10,12,14,8,10,12,14,9,11,13,15,9,11,13,15]
; AVX512BW-FAST-NEXT:    vpermd %zmm1, %zmm2, %zmm1
; AVX512BW-FAST-NEXT:    vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,18,22,26,30,u,u,u,u,19,23,27,31,u,u,u,u,32,36,40,44,u,u,u,u,33,37,41,45,u,u,u,u,50,54,58,62,u,u,u,u,51,55,59,63]
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [0,2,4,6,0,2,4,6,1,3,5,7,1,3,5,7]
; AVX512BW-FAST-NEXT:    vpermd %zmm0, %zmm2, %zmm0
; AVX512BW-FAST-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,18,22,26,30,u,u,u,u,19,23,27,31,u,u,u,u,32,36,40,44,u,u,u,u,33,37,41,45,u,u,u,u,50,54,58,62,u,u,u,u,51,55,59,63,u,u,u,u]
; AVX512BW-FAST-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512BW-FAST-NEXT:    kmovd %ecx, %k1
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512BW-FAST-NEXT:    vzeroupper
; AVX512BW-FAST-NEXT:    retq
  %in.vec0 = load <8 x i8>, ptr %in.vecptr0, align 64
  %in.vec1 = load <8 x i8>, ptr %in.vecptr1, align 64
  %in.vec2 = load <8 x i8>, ptr %in.vecptr2, align 64
  %in.vec3 = load <8 x i8>, ptr %in.vecptr3, align 64
  %in.vec4 = load <8 x i8>, ptr %in.vecptr4, align 64
  %in.vec5 = load <8 x i8>, ptr %in.vecptr5, align 64
  %in.vec6 = load <8 x i8>, ptr %in.vecptr6, align 64
  %in.vec7 = load <8 x i8>, ptr %in.vecptr7, align 64
  %1 = shufflevector <8 x i8> %in.vec0, <8 x i8> %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 i8> %in.vec2, <8 x i8> %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 i8> %in.vec4, <8 x i8> %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 <8 x i8> %in.vec6, <8 x i8> %in.vec7, <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 <16 x i8> %1, <16 x i8> %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>
  %6 = shufflevector <16 x i8> %3, <16 x i8> %4, <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>
  %7 = shufflevector <32 x i8> %5, <32 x i8> %6, <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>
  %interleaved.vec = shufflevector <64 x i8> %7, <64 x i8> poison, <64 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>
  store <64 x i8> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i8_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride8_vf16:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movdqa (%rdi), %xmm10
; SSE-NEXT:    movdqa (%rsi), %xmm9
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%rdx), %xmm1
; SSE-NEXT:    movdqa (%rcx), %xmm6
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%r8), %xmm3
; SSE-NEXT:    movdqa (%r9), %xmm11
; SSE-NEXT:    movdqa (%r10), %xmm4
; SSE-NEXT:    movdqa (%rax), %xmm13
; SSE-NEXT:    movdqa %xmm4, %xmm12
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm12[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535,65535,0,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm2, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    movdqa %xmm3, %xmm14
; SSE-NEXT:    punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm11[0],xmm14[1],xmm11[1],xmm14[2],xmm11[2],xmm14[3],xmm11[3],xmm14[4],xmm11[4],xmm14[5],xmm11[5],xmm14[6],xmm11[6],xmm14[7],xmm11[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm14[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm2, %xmm7
; SSE-NEXT:    por %xmm5, %xmm7
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm6[0],xmm15[1],xmm6[1],xmm15[2],xmm6[2],xmm15[3],xmm6[3],xmm15[4],xmm6[4],xmm15[5],xmm6[5],xmm15[6],xmm6[6],xmm15[7],xmm6[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm15[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,1,1,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm0, %xmm8
; SSE-NEXT:    pandn %xmm5, %xmm8
; SSE-NEXT:    movdqa %xmm10, %xmm6
; SSE-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3],xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm6[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm9
; SSE-NEXT:    por %xmm8, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm9[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm12[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    movdqa %xmm2, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm14[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    pand %xmm2, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm15[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,1,1,3]
; SSE-NEXT:    movdqa %xmm0, %xmm9
; SSE-NEXT:    pandn %xmm8, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm8
; SSE-NEXT:    por %xmm9, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm7[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm12[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3]
; SSE-NEXT:    movdqa %xmm2, %xmm9
; SSE-NEXT:    pandn %xmm8, %xmm9
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm14[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3]
; SSE-NEXT:    pand %xmm2, %xmm8
; SSE-NEXT:    por %xmm9, %xmm8
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm15[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[2,1,3,3]
; SSE-NEXT:    movdqa %xmm0, %xmm5
; SSE-NEXT:    pandn %xmm9, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm6[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm9
; SSE-NEXT:    por %xmm5, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm8[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm12[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    movdqa %xmm2, %xmm9
; SSE-NEXT:    pandn %xmm5, %xmm9
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm14[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    pand %xmm2, %xmm5
; SSE-NEXT:    por %xmm9, %xmm5
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm15[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[2,1,3,3]
; SSE-NEXT:    movdqa %xmm0, %xmm12
; SSE-NEXT:    pandn %xmm9, %xmm12
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm6
; SSE-NEXT:    por %xmm12, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm6[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm12 = xmm12[0],xmm5[0],xmm12[1],xmm5[1]
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm13[8],xmm4[9],xmm13[9],xmm4[10],xmm13[10],xmm4[11],xmm13[11],xmm4[12],xmm13[12],xmm4[13],xmm13[13],xmm4[14],xmm13[14],xmm4[15],xmm13[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm4[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    movdqa %xmm2, %xmm6
; SSE-NEXT:    pandn %xmm5, %xmm6
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm11[8],xmm3[9],xmm11[9],xmm3[10],xmm11[10],xmm3[11],xmm11[11],xmm3[12],xmm11[12],xmm3[13],xmm11[13],xmm3[14],xmm11[14],xmm3[15],xmm11[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm3[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    pand %xmm2, %xmm5
; SSE-NEXT:    por %xmm6, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[8],mem[8],xmm1[9],mem[9],xmm1[10],mem[10],xmm1[11],mem[11],xmm1[12],mem[12],xmm1[13],mem[13],xmm1[14],mem[14],xmm1[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm1[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,1,1,3]
; SSE-NEXT:    movdqa %xmm0, %xmm9
; SSE-NEXT:    pandn %xmm6, %xmm9
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
; SSE-NEXT:    # xmm10 = xmm10[8],mem[8],xmm10[9],mem[9],xmm10[10],mem[10],xmm10[11],mem[11],xmm10[12],mem[12],xmm10[13],mem[13],xmm10[14],mem[14],xmm10[15],mem[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm10[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm6
; SSE-NEXT:    por %xmm9, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm6[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm4[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    movdqa %xmm2, %xmm6
; SSE-NEXT:    pandn %xmm5, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm3[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    pand %xmm2, %xmm5
; SSE-NEXT:    por %xmm6, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,3,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm1[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,1,1,3]
; SSE-NEXT:    movdqa %xmm0, %xmm11
; SSE-NEXT:    pandn %xmm5, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm10[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm5
; SSE-NEXT:    por %xmm11, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    movdqa %xmm2, %xmm11
; SSE-NEXT:    pandn %xmm6, %xmm11
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    pand %xmm2, %xmm6
; SSE-NEXT:    por %xmm11, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm11 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[2,1,3,3]
; SSE-NEXT:    movdqa %xmm0, %xmm13
; SSE-NEXT:    pandn %xmm11, %xmm13
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm11
; SSE-NEXT:    por %xmm13, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
; SSE-NEXT:    pand %xmm2, %xmm3
; SSE-NEXT:    pandn %xmm4, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm10[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm0, %xmm3
; SSE-NEXT:    pandn %xmm1, %xmm0
; SSE-NEXT:    por %xmm3, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa %xmm0, 96(%rax)
; SSE-NEXT:    movdqa %xmm11, 112(%rax)
; SSE-NEXT:    movdqa %xmm5, 80(%rax)
; SSE-NEXT:    movdqa %xmm9, 64(%rax)
; SSE-NEXT:    movdqa %xmm12, 32(%rax)
; SSE-NEXT:    movdqa %xmm8, 48(%rax)
; SSE-NEXT:    movdqa %xmm7, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i8_stride8_vf16:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm4
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm5
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa (%r11), %xmm8
; AVX1-ONLY-NEXT:    vmovdqa (%r10), %xmm9
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm7[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm10[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[0,0,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm0 = xmm11[0,1,2],xmm0[3],xmm11[4,5,6],xmm0[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm7[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[0,0,2,1]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm10[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[0,0,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3],xmm12[4,5,6],xmm11[7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm11, %ymm0
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm11 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm11[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm12 = xmm12[0],zero,xmm12[1],zero
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm13[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm14 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm14[0],xmm12[1],xmm14[2,3,4],xmm12[5],xmm14[6,7]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm14 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm11[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm14 = xmm14[0],xmm15[1],xmm14[2,3,4],xmm15[5],xmm14[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm12, %ymm14, %ymm12
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm12[0],ymm0[1],ymm12[2],ymm0[3],ymm12[4],ymm0[5],ymm12[6],ymm0[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm7[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[0,2,2,3]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm10[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm12 = xmm14[0,1,2],xmm12[3],xmm14[4,5,6],xmm12[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[0,2,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm7 = xmm10[0,1,2],xmm7[3],xmm10[4,5,6],xmm7[7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm12, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm11[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[2,1,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm13[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm10 = xmm12[0],xmm10[1],xmm12[2,3,4],xmm10[5],xmm12[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm11[2,1,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm12 = xmm13[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm11 = xmm12[0],xmm11[1],xmm12[2,3,4],xmm11[5],xmm12[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm11, %ymm10
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2],ymm7[3],ymm10[4],ymm7[5],ymm10[6],ymm7[7]
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm8[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[0,0,2,1]
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm5[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm6 = xmm6[0,1,2],xmm9[3],xmm6[4,5,6],xmm9[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm8[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm9[0,0,2,1]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm5[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[0,0,2,1]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3],xmm10[4,5,6],xmm9[7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm9, %ymm6
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3,4],xmm4[5],xmm2[6,7]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm3[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0],xmm9[1],xmm4[2,3,4],xmm9[5],xmm4[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm2
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2],ymm6[3],ymm2[4],ymm6[5],ymm2[6],ymm6[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm8[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3],xmm6[4,5,6],xmm4[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm6[3],xmm5[4,5,6],xmm6[7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[2,1,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm1[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3,4],xmm5[5],xmm6[6,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,1,3,3]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3,4],xmm3[5],xmm1[6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7]
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 64(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm7, 32(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-ONLY-LABEL: store_i8_stride8_vf16:
; AVX2-ONLY:       # %bb.0:
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-ONLY-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovdqa (%rdx), %xmm1
; AVX2-ONLY-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-ONLY-NEXT:    vmovdqa (%r11), %xmm3
; AVX2-ONLY-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX2-ONLY-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX2-ONLY-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm4
; AVX2-ONLY-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm3
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm5 = ymm3[0,2,0,2]
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11]
; AVX2-ONLY-NEXT:    vpshufb %ymm6, %ymm5, %ymm2
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm7 = ymm4[0,2,0,2]
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm8 = [u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u]
; AVX2-ONLY-NEXT:    vpshufb %ymm8, %ymm7, %ymm9
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm2 = ymm9[0,1,2],ymm2[3],ymm9[4,5,6],ymm2[7],ymm9[8,9,10],ymm2[11],ymm9[12,13,14],ymm2[15]
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm9 = ymm1[0,2,0,2]
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm10 = [u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u]
; AVX2-ONLY-NEXT:    vpshufb %ymm10, %ymm9, %ymm11
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm12 = ymm0[0,2,0,2]
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm13 = [0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u]
; AVX2-ONLY-NEXT:    vpshufb %ymm13, %ymm12, %ymm14
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15]
; AVX2-ONLY-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7]
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15]
; AVX2-ONLY-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm14 = [u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u]
; AVX2-ONLY-NEXT:    vpshufb %ymm14, %ymm7, %ymm7
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15]
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u]
; AVX2-ONLY-NEXT:    vpshufb %ymm7, %ymm9, %ymm9
; AVX2-ONLY-NEXT:    vmovdqa {{.*#+}} ymm15 = [4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u,u,u]
; AVX2-ONLY-NEXT:    vpshufb %ymm15, %ymm12, %ymm12
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15]
; AVX2-ONLY-NEXT:    vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7]
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3]
; AVX2-ONLY-NEXT:    vpshufb %ymm6, %ymm3, %ymm6
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3]
; AVX2-ONLY-NEXT:    vpshufb %ymm8, %ymm4, %ymm8
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15]
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX2-ONLY-NEXT:    vpshufb %ymm10, %ymm1, %ymm8
; AVX2-ONLY-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX2-ONLY-NEXT:    vpshufb %ymm13, %ymm0, %ymm9
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15]
; AVX2-ONLY-NEXT:    vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7]
; AVX2-ONLY-NEXT:    vpshufb %ymm11, %ymm3, %ymm3
; AVX2-ONLY-NEXT:    vpshufb %ymm14, %ymm4, %ymm4
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-ONLY-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
; AVX2-ONLY-NEXT:    vpshufb %ymm15, %ymm0, %ymm0
; AVX2-ONLY-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7,8],ymm1[9],ymm0[10,11,12],ymm1[13],ymm0[14,15]
; AVX2-ONLY-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4],ymm3[5],ymm0[6],ymm3[7]
; AVX2-ONLY-NEXT:    vmovdqa %ymm0, 96(%rax)
; AVX2-ONLY-NEXT:    vmovdqa %ymm6, 64(%rax)
; AVX2-ONLY-NEXT:    vmovdqa %ymm5, 32(%rax)
; AVX2-ONLY-NEXT:    vmovdqa %ymm2, (%rax)
; AVX2-ONLY-NEXT:    vzeroupper
; AVX2-ONLY-NEXT:    retq
;
; AVX512F-LABEL: store_i8_stride8_vf16:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512F-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512F-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512F-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-NEXT:    vmovdqa (%r11), %xmm3
; AVX512F-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512F-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512F-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX512F-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm3
; AVX512F-NEXT:    vpermq {{.*#+}} ymm4 = ymm3[0,2,0,2]
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm5 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15]
; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm6
; AVX512F-NEXT:    vpermq {{.*#+}} ymm7 = ymm2[0,2,0,2]
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm8 = [u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u]
; AVX512F-NEXT:    vpshufb %ymm8, %ymm7, %ymm9
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm6 = ymm9[0,1,2],ymm6[3],ymm9[4,5,6],ymm6[7],ymm9[8,9,10],ymm6[11],ymm9[12,13,14],ymm6[15]
; AVX512F-NEXT:    vpermq {{.*#+}} ymm9 = ymm1[0,2,0,2]
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm10 = [u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u]
; AVX512F-NEXT:    vpshufb %ymm10, %ymm9, %ymm11
; AVX512F-NEXT:    vpermq {{.*#+}} ymm12 = ymm0[0,2,0,2]
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm13 = [4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u,u,u]
; AVX512F-NEXT:    vpshufb %ymm13, %ymm12, %ymm14
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15]
; AVX512F-NEXT:    vpblendd {{.*#+}} ymm6 = ymm11[0],ymm6[1],ymm11[2],ymm6[3],ymm11[4],ymm6[5],ymm11[6],ymm6[7]
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11]
; AVX512F-NEXT:    vpshufb %ymm11, %ymm4, %ymm4
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm14 = [u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u]
; AVX512F-NEXT:    vpshufb %ymm14, %ymm7, %ymm7
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm4 = ymm7[0,1,2],ymm4[3],ymm7[4,5,6],ymm4[7],ymm7[8,9,10],ymm4[11],ymm7[12,13,14],ymm4[15]
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm7 = [u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u]
; AVX512F-NEXT:    vpshufb %ymm7, %ymm9, %ymm9
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u]
; AVX512F-NEXT:    vpshufb %ymm15, %ymm12, %ymm12
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15]
; AVX512F-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0],ymm4[1],ymm9[2],ymm4[3],ymm9[4],ymm4[5],ymm9[6],ymm4[7]
; AVX512F-NEXT:    vinserti64x4 $1, %ymm6, %zmm4, %zmm4
; AVX512F-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3]
; AVX512F-NEXT:    vpshufb %ymm5, %ymm3, %ymm5
; AVX512F-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[1,3,1,3]
; AVX512F-NEXT:    vpshufb %ymm8, %ymm2, %ymm6
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm5 = ymm6[0,1,2],ymm5[3],ymm6[4,5,6],ymm5[7],ymm6[8,9,10],ymm5[11],ymm6[12,13,14],ymm5[15]
; AVX512F-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX512F-NEXT:    vpshufb %ymm10, %ymm1, %ymm6
; AVX512F-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX512F-NEXT:    vpshufb %ymm13, %ymm0, %ymm8
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2,3,4],ymm6[5],ymm8[6,7,8],ymm6[9],ymm8[10,11,12],ymm6[13],ymm8[14,15]
; AVX512F-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4],ymm5[5],ymm6[6],ymm5[7]
; AVX512F-NEXT:    vpshufb %ymm11, %ymm3, %ymm3
; AVX512F-NEXT:    vpshufb %ymm14, %ymm2, %ymm2
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5,6],ymm3[7],ymm2[8,9,10],ymm3[11],ymm2[12,13,14],ymm3[15]
; AVX512F-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
; AVX512F-NEXT:    vpshufb %ymm15, %ymm0, %ymm0
; AVX512F-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7,8],ymm1[9],ymm0[10,11,12],ymm1[13],ymm0[14,15]
; AVX512F-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7]
; AVX512F-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm0
; AVX512F-NEXT:    vmovdqa64 %zmm0, 64(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512F-NEXT:    vzeroupper
; AVX512F-NEXT:    retq
;
; AVX512BW-SLOW-LABEL: store_i8_stride8_vf16:
; AVX512BW-SLOW:       # %bb.0:
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-SLOW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-SLOW-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512BW-SLOW-NEXT:    vmovdqa (%r8), %xmm2
; AVX512BW-SLOW-NEXT:    vmovdqa (%r11), %xmm3
; AVX512BW-SLOW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512BW-SLOW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512BW-SLOW-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm3
; AVX512BW-SLOW-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm3
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm2
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm4 = zmm2[0,2,0,2,4,6,4,6]
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm5, %zmm4, %zmm4
; AVX512BW-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm3 = zmm3[4,5,6,7,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm6 = zmm3[0,2,0,2,4,6,4,6]
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm7, %zmm6, %zmm6
; AVX512BW-SLOW-NEXT:    movl $8913032, %ecx # imm = 0x880088
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT:    vmovdqu16 %zmm6, %zmm4 {%k1}
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm6 = zmm0[0,2,0,2,4,6,4,6]
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm8 = [0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u,u,u]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm8, %zmm6, %zmm6
; AVX512BW-SLOW-NEXT:    vshufi64x2 {{.*#+}} zmm1 = zmm1[4,5,6,7,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm9 = zmm1[0,2,0,2,4,6,4,6]
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm10, %zmm9, %zmm9
; AVX512BW-SLOW-NEXT:    movl $2228258, %ecx # imm = 0x220022
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k2
; AVX512BW-SLOW-NEXT:    vmovdqu16 %zmm9, %zmm6 {%k2}
; AVX512BW-SLOW-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k3
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm4, %zmm6 {%k3}
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[1,3,1,3,5,7,5,7]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm5, %zmm2, %zmm2
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm3 = zmm3[1,3,1,3,5,7,5,7]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm7, %zmm3, %zmm3
; AVX512BW-SLOW-NEXT:    vmovdqu16 %zmm3, %zmm2 {%k1}
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm8, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} zmm1 = zmm1[1,3,1,3,5,7,5,7]
; AVX512BW-SLOW-NEXT:    vpshufb %zmm10, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vmovdqu16 %zmm1, %zmm0 {%k2}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm0, 64(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm6, (%rax)
; AVX512BW-SLOW-NEXT:    vzeroupper
; AVX512BW-SLOW-NEXT:    retq
;
; AVX512BW-FAST-LABEL: store_i8_stride8_vf16:
; AVX512BW-FAST:       # %bb.0:
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FAST-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-FAST-NEXT:    vmovdqa (%rdx), %xmm1
; AVX512BW-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX512BW-FAST-NEXT:    vmovdqa (%r11), %xmm3
; AVX512BW-FAST-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512BW-FAST-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm4
; AVX512BW-FAST-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm3
; AVX512BW-FAST-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm5
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm6 = [0,2,0,2,12,14,12,14]
; AVX512BW-FAST-NEXT:    vpermt2q %zmm5, %zmm6, %zmm3
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm7 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15]
; AVX512BW-FAST-NEXT:    vpshufb %zmm7, %zmm3, %zmm3
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm2
; AVX512BW-FAST-NEXT:    vpermq {{.*#+}} zmm8 = zmm2[0,2,0,2,4,6,4,6]
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u]
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm8, %zmm8
; AVX512BW-FAST-NEXT:    movl $8913032, %ecx # imm = 0x880088
; AVX512BW-FAST-NEXT:    kmovd %ecx, %k1
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm3, %zmm8 {%k1}
; AVX512BW-FAST-NEXT:    vpermt2q %zmm4, %zmm6, %zmm1
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u]
; AVX512BW-FAST-NEXT:    vpshufb %zmm3, %zmm1, %zmm1
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpermq {{.*#+}} zmm6 = zmm0[0,2,0,2,4,6,4,6]
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11,u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15,u,u,u,u,u,u]
; AVX512BW-FAST-NEXT:    vpshufb %zmm10, %zmm6, %zmm6
; AVX512BW-FAST-NEXT:    movl $2228258, %ecx # imm = 0x220022
; AVX512BW-FAST-NEXT:    kmovd %ecx, %k2
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm1, %zmm6 {%k2}
; AVX512BW-FAST-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512BW-FAST-NEXT:    kmovd %ecx, %k3
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm8, %zmm6 {%k3}
; AVX512BW-FAST-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [5,7,5,7,5,7,5,7]
; AVX512BW-FAST-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FAST-NEXT:    vpermq %zmm5, %zmm1, %zmm5
; AVX512BW-FAST-NEXT:    vpshufb %zmm7, %zmm5, %zmm5
; AVX512BW-FAST-NEXT:    vpermq {{.*#+}} zmm2 = zmm2[1,3,1,3,5,7,5,7]
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm2, %zmm2
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm5, %zmm2 {%k1}
; AVX512BW-FAST-NEXT:    vpermq %zmm4, %zmm1, %zmm1
; AVX512BW-FAST-NEXT:    vpshufb %zmm3, %zmm1, %zmm1
; AVX512BW-FAST-NEXT:    vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7]
; AVX512BW-FAST-NEXT:    vpshufb %zmm10, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm1, %zmm0 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm0, 64(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm6, (%rax)
; AVX512BW-FAST-NEXT:    vzeroupper
; AVX512BW-FAST-NEXT:    retq
  %in.vec0 = load <16 x i8>, ptr %in.vecptr0, align 64
  %in.vec1 = load <16 x i8>, ptr %in.vecptr1, align 64
  %in.vec2 = load <16 x i8>, ptr %in.vecptr2, align 64
  %in.vec3 = load <16 x i8>, ptr %in.vecptr3, align 64
  %in.vec4 = load <16 x i8>, ptr %in.vecptr4, align 64
  %in.vec5 = load <16 x i8>, ptr %in.vecptr5, align 64
  %in.vec6 = load <16 x i8>, ptr %in.vecptr6, align 64
  %in.vec7 = load <16 x i8>, ptr %in.vecptr7, align 64
  %1 = shufflevector <16 x i8> %in.vec0, <16 x i8> %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 i8> %in.vec2, <16 x i8> %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 i8> %in.vec4, <16 x i8> %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 <16 x i8> %in.vec6, <16 x i8> %in.vec7, <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 <32 x i8> %1, <32 x i8> %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>
  %6 = shufflevector <32 x i8> %3, <32 x i8> %4, <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>
  %7 = shufflevector <64 x i8> %5, <64 x i8> %6, <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>
  %interleaved.vec = shufflevector <128 x i8> %7, <128 x i8> poison, <128 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 96, i32 112, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 97, i32 113, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 98, i32 114, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 99, i32 115, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 100, i32 116, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 101, i32 117, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 102, i32 118, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 103, i32 119, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 104, i32 120, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 105, i32 121, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 106, i32 122, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 107, i32 123, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 108, i32 124, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 109, i32 125, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 110, i32 126, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95, i32 111, i32 127>
  store <128 x i8> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i8_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride8_vf32:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $232, %rsp
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movdqa (%rdi), %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%rsi), %xmm4
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%rdx), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%rcx), %xmm8
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%r8), %xmm13
; SSE-NEXT:    movdqa (%r9), %xmm12
; SSE-NEXT:    movdqa (%r10), %xmm14
; SSE-NEXT:    movdqa (%rax), %xmm11
; SSE-NEXT:    movdqa %xmm14, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3],xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm2[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm2, %xmm15
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa {{.*#+}} xmm9 = [65535,65535,65535,0,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm9, %xmm6
; SSE-NEXT:    pandn %xmm2, %xmm6
; SSE-NEXT:    movdqa %xmm13, %xmm3
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm7
; SSE-NEXT:    por %xmm6, %xmm7
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm1[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,1,1,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm8 = [65535,0,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm8, %xmm10
; SSE-NEXT:    pandn %xmm6, %xmm10
; SSE-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,1,1]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm6
; SSE-NEXT:    por %xmm10, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm15[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm6, %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm3[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm6
; SSE-NEXT:    por %xmm7, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm1[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm10
; SSE-NEXT:    pandn %xmm7, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm7
; SSE-NEXT:    por %xmm10, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhbw {{.*#+}} xmm14 = xmm14[8],xmm11[8],xmm14[9],xmm11[9],xmm14[10],xmm11[10],xmm14[11],xmm11[11],xmm14[12],xmm11[12],xmm14[13],xmm11[13],xmm14[14],xmm11[14],xmm14[15],xmm11[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm14[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm6
; SSE-NEXT:    pandn %xmm5, %xmm6
; SSE-NEXT:    punpckhbw {{.*#+}} xmm13 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm13[2,1,3,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm4
; SSE-NEXT:    por %xmm6, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,3,2,3]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; SSE-NEXT:    # xmm12 = xmm12[8],mem[8],xmm12[9],mem[9],xmm12[10],mem[10],xmm12[11],mem[11],xmm12[12],mem[12],xmm12[13],mem[13],xmm12[14],mem[14],xmm12[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm12[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm3, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
; SSE-NEXT:    # xmm11 = xmm11[8],mem[8],xmm11[9],mem[9],xmm11[10],mem[10],xmm11[11],mem[11],xmm11[12],mem[12],xmm11[13],mem[13],xmm11[14],mem[14],xmm11[15],mem[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm11[1,1,1,1]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm1
; SSE-NEXT:    por %xmm5, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm14[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm1, %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm13[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm3, %xmm1
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm12[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm4
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm11[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm3
; SSE-NEXT:    por %xmm4, %xmm3
; SSE-NEXT:    movdqa 16(%r10), %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 16(%rax), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm5
; SSE-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm5[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm1, %xmm3
; SSE-NEXT:    movdqa 16(%r8), %xmm12
; SSE-NEXT:    movdqa 16(%r9), %xmm11
; SSE-NEXT:    movdqa %xmm12, %xmm4
; SSE-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3],xmm4[4],xmm11[4],xmm4[5],xmm11[5],xmm4[6],xmm11[6],xmm4[7],xmm11[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm4[2,1,3,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm4, (%rsp) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm3, %xmm0
; SSE-NEXT:    movdqa 16(%rdx), %xmm13
; SSE-NEXT:    movdqa 16(%rcx), %xmm7
; SSE-NEXT:    movdqa %xmm13, %xmm3
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3],xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa 16(%rdi), %xmm14
; SSE-NEXT:    movdqa 16(%rsi), %xmm6
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm1[1,1,1,1]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm15
; SSE-NEXT:    por %xmm2, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm15[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm5[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm0, %xmm2
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm4[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm15
; SSE-NEXT:    pandn %xmm2, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm2
; SSE-NEXT:    por %xmm15, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
; SSE-NEXT:    # xmm10 = xmm10[8],mem[8],xmm10[9],mem[9],xmm10[10],mem[10],xmm10[11],mem[11],xmm10[12],mem[12],xmm10[13],mem[13],xmm10[14],mem[14],xmm10[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm10[0,2,2,3,4,5,6,7]
; SSE-NEXT:    movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm0, %xmm2
; SSE-NEXT:    movdqa %xmm12, %xmm15
; SSE-NEXT:    punpckhbw {{.*#+}} xmm15 = xmm15[8],xmm11[8],xmm15[9],xmm11[9],xmm15[10],xmm11[10],xmm15[11],xmm11[11],xmm15[12],xmm11[12],xmm15[13],xmm11[13],xmm15[14],xmm11[14],xmm15[15],xmm11[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm15[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm13 = xmm13[8],xmm7[8],xmm13[9],xmm7[9],xmm13[10],xmm7[10],xmm13[11],xmm7[11],xmm13[12],xmm7[12],xmm13[13],xmm7[13],xmm13[14],xmm7[14],xmm13[15],xmm7[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm13[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm14 = xmm14[8],xmm6[8],xmm14[9],xmm6[9],xmm14[10],xmm6[10],xmm14[11],xmm6[11],xmm14[12],xmm6[12],xmm14[13],xmm6[13],xmm14[14],xmm6[14],xmm14[15],xmm6[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm14[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm1
; SSE-NEXT:    por %xmm3, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm10[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm15[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm13[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm14[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm10[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm10[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm0, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm5[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm12[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm0, %xmm2
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm5[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm4
; SSE-NEXT:    pandn %xmm2, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm12[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm2
; SSE-NEXT:    por %xmm4, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm4
; SSE-NEXT:    pandn %xmm2, %xmm4
; SSE-NEXT:    movdqa (%rsp), %xmm6 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm4, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm5
; SSE-NEXT:    pandn %xmm4, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm11[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm4
; SSE-NEXT:    por %xmm5, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm10[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm5
; SSE-NEXT:    pandn %xmm4, %xmm5
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm6[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm4
; SSE-NEXT:    por %xmm5, %xmm4
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm7[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm10
; SSE-NEXT:    pandn %xmm5, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm11[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm5
; SSE-NEXT:    por %xmm10, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm4[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    movdqa %xmm9, %xmm10
; SSE-NEXT:    pandn %xmm5, %xmm10
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm15[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm5
; SSE-NEXT:    por %xmm10, %xmm5
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm13[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[2,1,3,3]
; SSE-NEXT:    movdqa %xmm8, %xmm12
; SSE-NEXT:    pandn %xmm10, %xmm12
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm14[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm10
; SSE-NEXT:    por %xmm12, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm10 = xmm10[0],xmm5[0],xmm10[1],xmm5[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm11 = xmm15[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[0,2,2,3]
; SSE-NEXT:    pand %xmm9, %xmm11
; SSE-NEXT:    pandn %xmm5, %xmm9
; SSE-NEXT:    por %xmm11, %xmm9
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm13[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[2,1,3,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm14[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm8, %xmm6
; SSE-NEXT:    pandn %xmm5, %xmm8
; SSE-NEXT:    por %xmm6, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm9[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa %xmm6, 224(%rax)
; SSE-NEXT:    movdqa %xmm10, 240(%rax)
; SSE-NEXT:    movdqa %xmm4, 160(%rax)
; SSE-NEXT:    movdqa %xmm2, 176(%rax)
; SSE-NEXT:    movdqa %xmm0, 96(%rax)
; SSE-NEXT:    movdqa %xmm3, 112(%rax)
; SSE-NEXT:    movdqa %xmm1, 32(%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, 192(%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, 128(%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, 64(%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, (%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    addq $232, %rsp
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i8_stride8_vf32:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $72, %rsp
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT:    vmovdqa (%r10), %xmm8
; AVX1-ONLY-NEXT:    vmovdqa (%rax), %xmm4
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0]
; AVX1-ONLY-NEXT:    vandnps %ymm2, %ymm0, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm7
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm9
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm12[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm12[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm2, %ymm3, %ymm5
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm14
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm14[0],xmm2[0],xmm14[1],xmm2[1],xmm14[2],xmm2[2],xmm14[3],xmm2[3],xmm14[4],xmm2[4],xmm14[5],xmm2[5],xmm14[6],xmm2[6],xmm14[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm2, %ymm6
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm10
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm11
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm15 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm15[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm15[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm13 = xmm13[0],zero,xmm13[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm13, %ymm2, %ymm13
; AVX1-ONLY-NEXT:    vbroadcastsd {{.*#+}} ymm2 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535]
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vandnps %ymm13, %ymm2, %ymm13
; AVX1-ONLY-NEXT:    vorps %ymm6, %ymm13, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4],ymm5[5],ymm6[6],ymm5[7]
; AVX1-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm1[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm5, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm12[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm12[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vmovdqa 16(%r10), %xmm12
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm0, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm5 = ymm5[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm5, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm15[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm15[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm3[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm6, %ymm3
; AVX1-ONLY-NEXT:    vmovdqa 16(%rax), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm5 = ymm5[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm5, %ymm2, %ymm5
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm5, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vmovdqa 16(%r9), %xmm13
; AVX1-ONLY-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4],ymm1[5],ymm3[6],ymm1[7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15]
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm4[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vmovdqa 16(%r8), %xmm8
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm0, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
; AVX1-ONLY-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm4[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm2, %ymm3
; AVX1-ONLY-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm5 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm5 = xmm14[8],mem[8],xmm14[9],mem[9],xmm14[10],mem[10],xmm14[11],mem[11],xmm14[12],mem[12],xmm14[13],mem[13],xmm14[14],mem[14],xmm14[15],mem[15]
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm5[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm5[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4],ymm1[5],ymm3[6],ymm1[7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm11 = xmm6[0],xmm12[0],xmm6[1],xmm12[1],xmm6[2],xmm12[2],xmm6[3],xmm12[3],xmm6[4],xmm12[4],xmm6[5],xmm12[5],xmm6[6],xmm12[6],xmm6[7],xmm12[7]
; AVX1-ONLY-NEXT:    vmovdqa %xmm12, %xmm6
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm11[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm8[0],xmm13[0],xmm8[1],xmm13[1],xmm8[2],xmm13[2],xmm8[3],xmm13[3],xmm8[4],xmm13[4],xmm8[5],xmm13[5],xmm8[6],xmm13[6],xmm8[7],xmm13[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm12[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm12[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm0, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm5
; AVX1-ONLY-NEXT:    vmovdqa 16(%rcx), %xmm7
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdx), %xmm4
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm10[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm10[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm13, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vmovdqa 16(%rsi), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm13[0],xmm3[0],xmm13[1],xmm3[1],xmm13[2],xmm3[2],xmm13[3],xmm3[3],xmm13[4],xmm3[4],xmm13[5],xmm3[5],xmm13[6],xmm3[6],xmm13[7],xmm3[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm14[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm15 = xmm14[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm15 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm15, %ymm9, %ymm9
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm9, %ymm9
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm9, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm1[0],ymm5[1],ymm1[2],ymm5[3],ymm1[4],ymm5[5],ymm1[6],ymm5[7]
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm6[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm5 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm5 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm5[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm9 = xmm5[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm0, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm7[8],xmm4[9],xmm7[9],xmm4[10],xmm7[10],xmm4[11],xmm7[11],xmm4[12],xmm7[12],xmm4[13],xmm7[13],xmm4[14],xmm7[14],xmm4[15],xmm7[15]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm1[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm13[8],xmm3[8],xmm13[9],xmm3[9],xmm13[10],xmm3[10],xmm13[11],xmm3[11],xmm13[12],xmm3[12],xmm13[13],xmm3[13],xmm13[14],xmm3[14],xmm13[15],xmm3[15]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm3[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm9 = xmm3[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm2, %ymm4
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm7, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm13 = ymm4[0],ymm8[1],ymm4[2],ymm8[3],ymm4[4],ymm8[5],ymm4[6],ymm8[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm6[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm5[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm6, %ymm5
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm0, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm1[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm5, %ymm1
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm11[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm11[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm12[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm12[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm14[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm10[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm10[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vandnps %ymm5, %ymm2, %ymm5
; AVX1-ONLY-NEXT:    vorps %ymm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm5[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm6[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm0, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm5, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm6[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vandnps %ymm5, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vorps %ymm2, %ymm4, %ymm2
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2],ymm0[3],ymm2[4],ymm0[5],ymm2[6],ymm0[7]
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 128(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm13, 224(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm15, 160(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX1-ONLY-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT:    addq $72, %rsp
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i8_stride8_vf32:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    subq $88, %rsp
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-SLOW-NEXT:    vmovdqa (%r10), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa (%rax), %xmm3
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %xmm4
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm5
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm12[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm12[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm6
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3],ymm6[4,5,6],ymm0[7],ymm6[8,9,10],ymm0[11],ymm6[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %xmm6
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm7
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm13[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm8, %ymm9
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %xmm10
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm11
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm14[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm14[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm15, %ymm8, %ymm15
; AVX2-SLOW-NEXT:    vmovaps 16(%r10), %xmm8
; AVX2-SLOW-NEXT:    vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm15 = ymm9[0],ymm15[1],ymm9[2,3,4],ymm15[5],ymm9[6,7,8],ymm15[9],ymm9[10,11,12],ymm15[13],ymm9[14,15]
; AVX2-SLOW-NEXT:    vmovdqa 16(%rax), %xmm9
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm15[0],ymm0[1],ymm15[2],ymm0[3],ymm15[4],ymm0[5],ymm15[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm12[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm12, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vmovdqa 16(%r9), %xmm8
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm13[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm13[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm13, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm14[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm14, %ymm13, %ymm13
; AVX2-SLOW-NEXT:    vmovdqa 16(%r8), %xmm15
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm13[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm13[1],ymm1[2,3,4],ymm13[5],ymm1[6,7,8],ymm13[9],ymm1[10,11,12],ymm13[13],ymm1[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm1[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm1[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm1[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm1[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm7, %ymm7
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm7[1],ymm2[2,3,4],ymm7[5],ymm2[6,7,8],ymm7[9],ymm2[10,11,12],ymm7[13],ymm2[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2],ymm0[3],ymm2[4],ymm0[5],ymm2[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-SLOW-NEXT:    vmovdqa %xmm9, %xmm5
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3],xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm10[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm10[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm11 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3],xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
; AVX2-SLOW-NEXT:    vmovdqa %xmm8, %xmm9
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm11[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm11[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX2-SLOW-NEXT:    vmovdqa 16(%rsi), %xmm8
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm13 = ymm13[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm7 = ymm13[0,1,2],ymm7[3],ymm13[4,5,6],ymm7[7],ymm13[8,9,10],ymm7[11],ymm13[12,13,14],ymm7[15]
; AVX2-SLOW-NEXT:    vmovdqa 16(%rdi), %xmm4
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm13 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm14 = xmm13[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm14 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm13[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm14, %ymm2
; AVX2-SLOW-NEXT:    vmovdqa 16(%rcx), %xmm3
; AVX2-SLOW-NEXT:    vmovdqa 16(%rdx), %xmm1
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm14[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm14[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm12, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4],ymm0[5],ymm2[6,7,8],ymm0[9],ymm2[10,11,12],ymm0[13],ymm2[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2],ymm7[3],ymm0[4],ymm7[5],ymm0[6],ymm7[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm2[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm12, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm12 = xmm15[8],xmm9[8],xmm15[9],xmm9[9],xmm15[10],xmm9[10],xmm15[11],xmm9[11],xmm15[12],xmm9[12],xmm15[13],xmm9[13],xmm15[14],xmm9[14],xmm15[15],xmm9[15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm15 = xmm12[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm12[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm15, %ymm7
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm7[0,1,2],ymm0[3],ymm7[4,5,6],ymm0[7],ymm7[8,9,10],ymm0[11],ymm7[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm4[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm7 = xmm4[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm3[8],xmm1[9],xmm3[9],xmm1[10],xmm3[10],xmm1[11],xmm3[11],xmm1[12],xmm3[12],xmm1[13],xmm3[13],xmm1[14],xmm3[14],xmm1[15],xmm3[15]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm1[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7,8],ymm3[9],ymm5[10,11,12],ymm3[13],ymm5[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2],ymm0[3],ymm3[4],ymm0[5],ymm3[6],ymm0[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm12[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm12[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm4, %ymm1
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3,4],ymm1[5],ymm3[6,7,8],ymm1[9],ymm3[10,11,12],ymm1[13],ymm3[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm10[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm10[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm11[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm11[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm13[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm14[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm14[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm4[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm5[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm5[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm6[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3,4],ymm5[5],ymm4[6,7,8],ymm5[9],ymm4[10,11,12],ymm5[13],ymm4[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa %ymm3, 64(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm2, 128(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm1, 192(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 224(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 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, 32(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-SLOW-NEXT:    addq $88, %rsp
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i8_stride8_vf32:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    subq $72, %rsp
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FAST-NEXT:    vmovdqa (%r10), %xmm5
; AVX2-FAST-NEXT:    vmovdqa (%rax), %xmm6
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm8
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm3 = ymm8[0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,16,17,20,21,16,17,20,21,24,25,26,27,20,21,22,23]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %xmm1
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm9
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm9[0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,20,21,18,19,20,21,18,19,24,25,26,27,22,23,22,23]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm11 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %xmm7
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %xmm3
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm10
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} xmm4 = xmm12[4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm13, %ymm13
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm15
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm15[0],xmm7[0],xmm15[1],xmm7[1],xmm15[2],xmm7[2],xmm15[3],xmm7[3],xmm15[4],xmm7[4],xmm15[5],xmm7[5],xmm15[6],xmm7[6],xmm15[7],xmm7[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm0
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm14 = ymm0[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm13 = ymm13[0],ymm14[1],ymm13[2,3,4],ymm14[5],ymm13[6,7,8],ymm14[9],ymm13[10,11,12],ymm14[13],ymm13[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0],ymm11[1],ymm13[2],ymm11[3],ymm13[4],ymm11[5],ymm13[6],ymm11[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm8, %ymm8
; AVX2-FAST-NEXT:    vmovdqa %ymm4, %ymm14
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm9, %ymm9
; AVX2-FAST-NEXT:    vmovdqa %ymm4, %ymm13
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5,6],ymm8[7],ymm9[8,9,10],ymm8[11],ymm9[12,13,14],ymm8[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm12, %ymm12, %ymm9
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm11 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm9, %ymm9
; AVX2-FAST-NEXT:    vmovdqa %ymm11, %ymm12
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2,3,4],ymm0[5],ymm9[6,7,8],ymm0[9],ymm9[10,11,12],ymm0[13],ymm9[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm0, (%rsp) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm14, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpshufb %ymm13, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm7[8],xmm15[9],xmm7[9],xmm15[10],xmm7[10],xmm15[11],xmm7[11],xmm15[12],xmm7[12],xmm15[13],xmm7[13],xmm15[14],xmm7[14],xmm15[15],xmm7[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm2
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm11 = xmm10[8],xmm3[8],xmm10[9],xmm3[9],xmm10[10],xmm3[10],xmm10[11],xmm3[11],xmm10[12],xmm3[12],xmm10[13],xmm3[13],xmm10[14],xmm3[14],xmm10[15],xmm3[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm11, %ymm11, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 16(%r10), %xmm8
; AVX2-FAST-NEXT:    vmovdqa 16(%rax), %xmm6
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3],xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm12
; AVX2-FAST-NEXT:    vmovdqa 16(%r9), %xmm7
; AVX2-FAST-NEXT:    vmovdqa 16(%r8), %xmm5
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm10
; AVX2-FAST-NEXT:    vpshufb %ymm14, %ymm12, %ymm3
; AVX2-FAST-NEXT:    vpshufb %ymm13, %ymm10, %ymm4
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm15 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-NEXT:    vmovdqa 16(%rcx), %xmm4
; AVX2-FAST-NEXT:    vmovdqa 16(%rdx), %xmm2
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm9
; AVX2-FAST-NEXT:    vmovdqa 16(%rsi), %xmm1
; AVX2-FAST-NEXT:    vmovdqa 16(%rdi), %xmm0
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm14
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm14 = ymm14[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm13 = ymm9[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3,4],ymm13[5],ymm14[6,7,8],ymm13[9],ymm14[10,11,12],ymm13[13],ymm14[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm15 = ymm13[0],ymm15[1],ymm13[2],ymm15[3],ymm13[4],ymm15[5],ymm13[6],ymm15[7]
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15]
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm7[8],xmm5[9],xmm7[9],xmm5[10],xmm7[10],xmm5[11],xmm7[11],xmm5[12],xmm7[12],xmm5[13],xmm7[13],xmm5[14],xmm7[14],xmm5[15],xmm7[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm6, %ymm6, %ymm6
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm7 = ymm6[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm8 = ymm5[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3],ymm8[4,5,6],ymm7[7],ymm8[8,9,10],ymm7[11],ymm8[12,13,14],ymm7[15]
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15]
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7,8],ymm4[9],ymm0[10,11,12],ymm4[13],ymm0[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2],ymm7[3],ymm0[4],ymm7[5],ymm0[6],ymm7[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm6, %ymm4
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm8 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %xmm8, %xmm1, %xmm5
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm5 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm12, %ymm2
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm10, %ymm4
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
; AVX2-FAST-NEXT:    vpshufb %xmm8, %xmm3, %xmm4
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vpshufb %ymm5, %ymm9, %ymm4
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm4, %ymm4
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-FAST-NEXT:    vpshufb %ymm5, %ymm4, %ymm4
; AVX2-FAST-NEXT:    vpshufb %xmm8, %xmm11, %xmm5
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm6, %ymm5
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7,8],ymm4[9],ymm5[10,11,12],ymm4[13],ymm5[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa %ymm3, 64(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm2, 128(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm1, 192(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 224(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm15, 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 (%rsp), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FAST-NEXT:    addq $72, %rsp
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i8_stride8_vf32:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    subq $72, %rsp
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r10), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rax), %xmm6
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm3 = ymm8[0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,16,17,20,21,16,17,20,21,24,25,26,27,20,21,22,23]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r9), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r8), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm9[0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,20,21,18,19,20,21,18,19,24,25,26,27,22,23,22,23]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm11 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %xmm7
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm10
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} xmm4 = xmm12[4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm13, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm15
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm15[0],xmm7[0],xmm15[1],xmm7[1],xmm15[2],xmm7[2],xmm15[3],xmm7[3],xmm15[4],xmm7[4],xmm15[5],xmm7[5],xmm15[6],xmm7[6],xmm15[7],xmm7[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm14 = ymm0[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm13 = ymm13[0],ymm14[1],ymm13[2,3,4],ymm14[5],ymm13[6,7,8],ymm14[9],ymm13[10,11,12],ymm14[13],ymm13[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0],ymm11[1],ymm13[2],ymm11[3],ymm13[4],ymm11[5],ymm13[6],ymm11[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm4, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm4, %ymm14
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm4, %ymm9, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm4, %ymm13
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5,6],ymm8[7],ymm9[8,9,10],ymm8[11],ymm9[12,13,14],ymm8[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm4 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm4, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm12, %ymm12, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm11 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm9, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm11, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2,3,4],ymm0[5],ymm9[6,7,8],ymm0[9],ymm9[10,11,12],ymm0[13],ymm9[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, (%rsp) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm14, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm13, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm7[8],xmm15[9],xmm7[9],xmm15[10],xmm7[10],xmm15[11],xmm7[11],xmm15[12],xmm7[12],xmm15[13],xmm7[13],xmm15[14],xmm7[14],xmm15[15],xmm7[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm11 = xmm10[8],xmm3[8],xmm10[9],xmm3[9],xmm10[10],xmm3[10],xmm10[11],xmm3[11],xmm10[12],xmm3[12],xmm10[13],xmm3[13],xmm10[14],xmm3[14],xmm10[15],xmm3[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm11, %ymm11, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm4, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%r10), %xmm8
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rax), %xmm6
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3],xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm12
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%r9), %xmm7
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%r8), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm14, %ymm12, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm13, %ymm10, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm15 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rcx), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rdx), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rsi), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rdi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm14
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm14 = ymm14[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm13 = ymm9[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3,4],ymm13[5],ymm14[6,7,8],ymm13[9],ymm14[10,11,12],ymm13[13],ymm14[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm15 = ymm13[0],ymm15[1],ymm13[2],ymm15[3],ymm13[4],ymm15[5],ymm13[6],ymm15[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm7[8],xmm5[9],xmm7[9],xmm5[10],xmm7[10],xmm5[11],xmm7[11],xmm5[12],xmm7[12],xmm5[13],xmm7[13],xmm5[14],xmm7[14],xmm5[15],xmm7[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm6, %ymm6, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm7 = ymm6[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm8 = ymm5[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3],ymm8[4,5,6],ymm7[7],ymm8[8,9,10],ymm7[11],ymm8[12,13,14],ymm7[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7,8],ymm4[9],ymm0[10,11,12],ymm4[13],ymm0[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2],ymm7[3],ymm0[4],ymm7[5],ymm0[6],ymm7[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm6, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm8 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm8, %xmm1, %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm5 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm12, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm10, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm8, %xmm3, %xmm4
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm9, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm4, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm5, %ymm4, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm8, %xmm11, %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm6, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7,8],ymm4[9],ymm5[10,11,12],ymm4[13],ymm5[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm3, 64(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm2, 128(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm1, 192(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm0, 224(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm15, 160(%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:    vmovups (%rsp), %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 $72, %rsp
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-ONLY-SLOW-LABEL: store_i8_stride8_vf32:
; AVX512F-ONLY-SLOW:       # %bb.0:
; AVX512F-ONLY-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r10), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%r10), %xmm11
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rax), %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%rax), %xmm12
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm21
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm22
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r9), %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%r9), %xmm13
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%r8), %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%r8), %xmm14
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm4, %xmm23
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm3, %xmm24
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm2[0,1,2,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm5[0,1,2],ymm1[3],ymm5[4,5,6],ymm1[7],ymm5[8,9,10],ymm1[11],ymm5[12,13,14],ymm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm5, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm2[0,1,1,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm16
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdi), %xmm7
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm7[8],xmm1[8],xmm7[9],xmm1[9],xmm7[10],xmm1[10],xmm7[11],xmm1[11],xmm7[12],xmm1[12],xmm7[13],xmm1[13],xmm7[14],xmm1[14],xmm7[15],xmm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm25
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rcx), %xmm8
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa (%rdx), %xmm9
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm15 = xmm10[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm10[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm15, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm3[2,1,3,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm10[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm10[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm18
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm3[0,1,2,3,4,5,5,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm3[0,1,2,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm2, %ymm10
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%rcx), %xmm5
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm15, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm10[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm10[0,1,2],ymm1[3],ymm10[4,5,6],ymm1[7],ymm10[8,9,10],ymm1[11],ymm10[12,13,14],ymm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm3[0,1,1,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm10, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3],ymm3[4,5,6],ymm0[7],ymm3[8,9,10],ymm0[11],ymm3[12,13,14],ymm0[15]
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%rdx), %xmm10
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm17
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3],xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm15
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%rsi), %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm19 = xmm0[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm19 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm20 = xmm0[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti32x4 $1, %xmm20, %ymm19, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm15 = ymm15[2,1,3,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm15[1],ymm1[2,3,4],ymm15[5],ymm1[6,7,8],ymm15[9],ymm1[10,11,12],ymm15[13],ymm1[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm15 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm15, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm6, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7,8],ymm3[9],ymm0[10,11,12],ymm3[13],ymm0[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm15
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,4,5,5,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm3[0,1,2,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm11, %ymm6, %ymm6
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm11, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm6 = ymm6[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3],ymm6[4,5,6],ymm1[7],ymm6[8,9,10],ymm1[11],ymm6[12,13,14],ymm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm3[0,1,1,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm6, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3],ymm3[4,5,6],ymm0[7],ymm3[8,9,10],ymm0[11],ymm3[12,13,14],ymm0[15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm11
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm10[8],xmm5[8],xmm10[9],xmm5[9],xmm10[10],xmm5[10],xmm10[11],xmm5[11],xmm10[12],xmm5[12],xmm10[13],xmm5[13],xmm10[14],xmm5[14],xmm10[15],xmm5[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm4[2,1,3,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm4, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm21, %xmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm22, %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm23, %xmm3
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm24, %xmm4
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm1[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm5, %ymm1
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,1,1,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %xmm25, %xmm2
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3],xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm2[2,3,2,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm2[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5]
; AVX512F-ONLY-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm5[0,1,2,3,4,6,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm6
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm6[2,1,3,3,6,5,7,7]
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm4, %ymm2
; AVX512F-ONLY-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm5[0,2,2,3,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512F-ONLY-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm6, %ymm4
; AVX512F-ONLY-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3,4],ymm4[5],ymm2[6,7,8],ymm4[9],ymm2[10,11,12],ymm4[13],ymm2[14,15]
; AVX512F-ONLY-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512F-ONLY-SLOW-NEXT:    kmovw %ecx, %k1
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm16, %zmm18 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm17, %zmm15 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm11, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa32 %zmm1, %zmm2 {%k1}
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm15, 128(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vmovdqa64 %zmm18, 64(%rax)
; AVX512F-ONLY-SLOW-NEXT:    vzeroupper
; AVX512F-ONLY-SLOW-NEXT:    retq
;
; AVX512F-ONLY-FAST-LABEL: store_i8_stride8_vf32:
; AVX512F-ONLY-FAST:       # %bb.0:
; AVX512F-ONLY-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r10), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rax), %xmm2
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm2, %xmm21
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm1, %xmm22
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r9), %xmm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%r8), %xmm4
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm4, %xmm24
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm3, %xmm25
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm4, %ymm20
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm3, %ymm2, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm3, %ymm27
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm23
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rcx), %xmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdx), %xmm10
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm10[8],xmm1[8],xmm10[9],xmm1[9],xmm10[10],xmm1[10],xmm10[11],xmm1[11],xmm10[12],xmm1[12],xmm10[13],xmm1[13],xmm10[14],xmm1[14],xmm10[15],xmm1[15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm1, %xmm30
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rsi), %xmm12
; AVX512F-ONLY-FAST-NEXT:    vmovdqa (%rdi), %xmm13
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm3, %ymm2, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa %ymm3, %ymm14
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm4, %ymm0, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm4, %ymm28
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3,4],ymm3[5],ymm2[6,7,8],ymm3[9],ymm2[10,11,12],ymm3[13],ymm2[14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm3, %ymm29
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} xmm9 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm3, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7,8],ymm0[9],ymm1[10,11,12],ymm0[13],ymm1[14,15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm19
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%r10), %xmm8
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%rax), %xmm11
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%r9), %xmm7
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%r8), %xmm6
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm0[u,u,u,u,u,u,8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31]
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm1[0,1,2,3,8,9,u,u,8,9,10,11,10,11,u,u,16,17,18,19,28,29,u,u,28,29,26,27,30,31,u,u]
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,0,1,u,u,u,u,u,u,2,3,u,u,u,u,u,u,20,21,u,u,u,u,u,u,22,23]
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,1,2,3,0,1,u,u,8,9,10,11,2,3,u,u,20,21,18,19,20,21,u,u,24,25,26,27,22,23,u,u]
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm17
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%rsi), %xmm5
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%rdi), %xmm4
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm14, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm14, %ymm16
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%rcx), %xmm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa 16(%rdx), %xmm2
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm15 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm14 = ymm15[u,u,8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4],ymm14[5],ymm1[6,7,8],ymm14[9],ymm1[10,11,12],ymm14[13],ymm1[14,15]
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm14 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm9, %xmm0, %xmm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm9, %xmm31
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm14, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vpshufb {{.*#+}} ymm14 = ymm15[u,u,0,1,u,u,u,u,u,u,2,3,u,u,u,u,u,u,20,21,u,u,u,u,u,u,22,23,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm14[1],ymm0[2,3,4],ymm14[5],ymm0[6,7,8],ymm14[9],ymm0[10,11,12],ymm14[13],ymm0[14,15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm18
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm8[8],xmm11[9],xmm8[9],xmm11[10],xmm8[10],xmm11[11],xmm8[11],xmm11[12],xmm8[12],xmm11[13],xmm8[13],xmm11[14],xmm8[14],xmm11[15],xmm8[15]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm8, %ymm0, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm20, %ymm11
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm11, %ymm1, %ymm7
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5,6],ymm6[7],ymm7[8,9,10],ymm6[11],ymm7[12,13,14],ymm6[15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm26, %ymm7
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm27, %ymm14
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm14, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm20
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
; AVX512F-ONLY-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm16, %ymm15
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm28, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm6, %ymm1, %ymm4
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %ymm29, %ymm9
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm9, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm31, %xmm0
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm4, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm21, %xmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm22, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm24, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm25, %xmm4
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm8, %ymm2, %ymm4
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm11, %ymm3, %ymm5
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm14, %ymm3, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %xmm30, %xmm3
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7]
; AVX512F-ONLY-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm5
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm15, %ymm5, %ymm5
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm6, %ymm3, %ymm6
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15]
; AVX512F-ONLY-FAST-NEXT:    vpshufb %ymm9, %ymm3, %ymm3
; AVX512F-ONLY-FAST-NEXT:    vpshufb %xmm0, %xmm4, %xmm6
; AVX512F-ONLY-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512F-ONLY-FAST-NEXT:    vinserti128 $1, %xmm6, %ymm4, %ymm4
; AVX512F-ONLY-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7,8],ymm3[9],ymm4[10,11,12],ymm3[13],ymm4[14,15]
; AVX512F-ONLY-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm3, %zmm3
; AVX512F-ONLY-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512F-ONLY-FAST-NEXT:    kmovw %ecx, %k1
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm23, %zmm19 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm17, %zmm18 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm20, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa32 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm1, 192(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm18, 128(%rax)
; AVX512F-ONLY-FAST-NEXT:    vmovdqa64 %zmm19, 64(%rax)
; AVX512F-ONLY-FAST-NEXT:    vzeroupper
; AVX512F-ONLY-FAST-NEXT:    retq
;
; AVX512DQ-SLOW-LABEL: store_i8_stride8_vf32:
; AVX512DQ-SLOW:       # %bb.0:
; AVX512DQ-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%rsi), %xmm10
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdi), %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%rdi), %xmm11
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm19
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm20
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm5
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rcx), %xmm1
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%rcx), %xmm12
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rdx), %xmm2
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%rdx), %xmm13
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm21
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm22
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm6[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm6[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm7, %ymm7
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[2,1,3,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2,3,4],ymm7[5],ymm5[6,7,8],ymm7[9],ymm5[10,11,12],ymm7[13],ymm5[14,15]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm7, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm6[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm6
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[2,3,4],ymm6[5],ymm0[6,7,8],ymm6[9],ymm0[10,11,12],ymm6[13],ymm0[14,15]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm0, %zmm16
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r10), %xmm5
; AVX512DQ-SLOW-NEXT:    vmovdqa (%rax), %xmm6
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm0[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm0[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm7, %ymm9
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r9), %xmm7
; AVX512DQ-SLOW-NEXT:    vmovdqa (%r8), %xmm8
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm15 = xmm14[0,1,2,3,4,5,5,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm14[0,1,2,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm15, %ymm1
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm9[3],ymm1[4,5,6],ymm9[7],ymm1[8,9,10],ymm9[11],ymm1[12,13,14],ymm9[15]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm0[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm9, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm14[0,1,1,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm14[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm14, %ymm9, %ymm9
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm9[0,1,2],ymm0[3],ymm9[4,5,6],ymm0[7],ymm9[8,9,10],ymm0[11],ymm9[12,13,14],ymm0[15]
; AVX512DQ-SLOW-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512DQ-SLOW-NEXT:    kmovw %ecx, %k1
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm1, %zmm0, %zmm16 {%k1}
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm0[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm15 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm15[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm15[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm14, %ymm2
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm1, %ymm1
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%r10), %xmm14
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm15[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm15[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm9, %ymm2
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%rax), %xmm15
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm17
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti32x4 $1, %xmm2, %ymm0, %ymm18
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%r9), %xmm3
; AVX512DQ-SLOW-NEXT:    vmovdqa 16(%r8), %xmm9
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm9[0],xmm3[0],xmm9[1],xmm3[1],xmm9[2],xmm3[2],xmm9[3],xmm3[3],xmm9[4],xmm3[4],xmm9[5],xmm3[5],xmm9[6],xmm3[6],xmm9[7],xmm3[7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,5,5,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm4, %ymm1
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,1,1,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm4, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm18[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5,6],ymm4[7],ymm2[8,9,10],ymm4[11],ymm2[12,13,14],ymm4[15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5,6],ymm1[7],ymm0[8,9,10],ymm1[11],ymm0[12,13,14],ymm1[15]
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm2, %zmm0, %zmm17 {%k1}
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm4[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm4[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm10[2,1,3,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm4[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm10 = xmm10[0],zero,xmm10[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm4[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm10, %ymm2
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm10
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512DQ-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm9[8],xmm3[8],xmm9[9],xmm3[9],xmm9[10],xmm3[10],xmm9[11],xmm3[11],xmm9[12],xmm3[12],xmm9[13],xmm3[13],xmm9[14],xmm3[14],xmm9[15],xmm3[15]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm4, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[0,1,1,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm1, %zmm0, %zmm10 {%k1}
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm19, %xmm0
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm20, %xmm1
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm21, %xmm3
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %xmm22, %xmm4
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm9 = xmm3[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm4, %ymm4
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm4[2,1,3,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm4, %ymm2
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX512DQ-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX512DQ-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7]
; AVX512DQ-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,6,5,7,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm1[0,0,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm5, %ymm1
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,1,1,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX512DQ-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX512DQ-SLOW-NEXT:    vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm10, 192(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm17, 128(%rax)
; AVX512DQ-SLOW-NEXT:    vmovdqa64 %zmm16, 64(%rax)
; AVX512DQ-SLOW-NEXT:    vzeroupper
; AVX512DQ-SLOW-NEXT:    retq
;
; AVX512DQ-FAST-LABEL: store_i8_stride8_vf32:
; AVX512DQ-FAST:       # %bb.0:
; AVX512DQ-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FAST-NEXT:    vmovdqa (%rcx), %xmm1
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdx), %xmm2
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm2, %xmm19
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm1, %xmm20
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm2, %ymm0, %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm2, %ymm18
; AVX512DQ-FAST-NEXT:    vmovdqa (%rsi), %xmm3
; AVX512DQ-FAST-NEXT:    vmovdqa (%rdi), %xmm4
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm4, %xmm21
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm3, %xmm22
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqa %ymm4, %ymm5
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3,4],ymm1[5],ymm3[6,7,8],ymm1[9],ymm3[10,11,12],ymm1[13],ymm3[14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm3, %ymm23
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} xmm4 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX512DQ-FAST-NEXT:    vpshufb %xmm4, %xmm2, %xmm3
; AVX512DQ-FAST-NEXT:    vmovdqa %xmm4, %xmm10
; AVX512DQ-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4],ymm0[5],ymm2[6,7,8],ymm0[9],ymm2[10,11,12],ymm0[13],ymm2[14,15]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm17
; AVX512DQ-FAST-NEXT:    vmovdqa (%r10), %xmm1
; AVX512DQ-FAST-NEXT:    vmovdqa (%rax), %xmm11
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm1[8],xmm11[9],xmm1[9],xmm11[10],xmm1[10],xmm11[11],xmm1[11],xmm11[12],xmm1[12],xmm11[13],xmm1[13],xmm11[14],xmm1[14],xmm11[15],xmm1[15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm1, %xmm28
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm2, %ymm0, %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm2, %ymm24
; AVX512DQ-FAST-NEXT:    vmovdqa (%r9), %xmm12
; AVX512DQ-FAST-NEXT:    vmovdqa (%r8), %xmm13
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm4, %ymm25
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512DQ-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm3, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm3, %ymm27
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX512DQ-FAST-NEXT:    movw $-21846, %r11w # imm = 0xAAAA
; AVX512DQ-FAST-NEXT:    kmovw %r11d, %k1
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm1, %zmm0, %zmm17 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%rsi), %xmm9
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%rdi), %xmm8
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm0
; AVX512DQ-FAST-NEXT:    vpshufb %ymm5, %ymm0, %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm5, %ymm30
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%rcx), %xmm7
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%rdx), %xmm6
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm4
; AVX512DQ-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm4[u,u,8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u]
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4],ymm5[5],ymm3[6,7,8],ymm5[9],ymm3[10,11,12],ymm5[13],ymm3[14,15]
; AVX512DQ-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512DQ-FAST-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm10, %xmm29
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
; AVX512DQ-FAST-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,0,1,u,u,u,u,u,u,2,3,u,u,u,u,u,u,20,21,u,u,u,u,u,u,22,23,u,u,u,u]
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3,4],ymm4[5],ymm2[6,7,8],ymm4[9],ymm2[10,11,12],ymm4[13],ymm2[14,15]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm16
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%r10), %xmm5
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%rax), %xmm4
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm0
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%r9), %xmm3
; AVX512DQ-FAST-NEXT:    vmovdqa 16(%r8), %xmm2
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm15 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX512DQ-FAST-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[u,u,u,u,u,u,8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31]
; AVX512DQ-FAST-NEXT:    vpshufb {{.*#+}} ymm14 = ymm15[0,1,2,3,8,9,u,u,8,9,10,11,10,11,u,u,16,17,18,19,28,29,u,u,28,29,26,27,30,31,u,u]
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm14[0,1,2],ymm1[3],ymm14[4,5,6],ymm1[7],ymm14[8,9,10],ymm1[11],ymm14[12,13,14],ymm1[15]
; AVX512DQ-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,0,1,u,u,u,u,u,u,2,3,u,u,u,u,u,u,20,21,u,u,u,u,u,u,22,23]
; AVX512DQ-FAST-NEXT:    vpshufb {{.*#+}} ymm14 = ymm15[0,1,2,3,0,1,u,u,8,9,10,11,2,3,u,u,20,21,18,19,20,21,u,u,24,25,26,27,22,23,u,u]
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm14[0,1,2],ymm0[3],ymm14[4,5,6],ymm0[7],ymm14[8,9,10],ymm0[11],ymm14[12,13,14],ymm0[15]
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm1, %zmm0, %zmm16 {%k1}
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm6
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm30, %ymm14
; AVX512DQ-FAST-NEXT:    vpshufb %ymm14, %ymm6, %ymm6
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm18, %ymm10
; AVX512DQ-FAST-NEXT:    vpshufb %ymm10, %ymm0, %ymm7
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3,4],ymm7[5],ymm6[6,7,8],ymm7[9],ymm6[10,11,12],ymm7[13],ymm6[14,15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm23, %ymm9
; AVX512DQ-FAST-NEXT:    vpshufb %ymm9, %ymm0, %ymm0
; AVX512DQ-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm29, %xmm15
; AVX512DQ-FAST-NEXT:    vpshufb %xmm15, %xmm1, %xmm1
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm7, %ymm1
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7,8],ymm0[9],ymm1[10,11,12],ymm0[13],ymm1[14,15]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm0, %zmm0
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
; AVX512DQ-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm24, %ymm5
; AVX512DQ-FAST-NEXT:    vpshufb %ymm5, %ymm1, %ymm3
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm25, %ymm6
; AVX512DQ-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm4
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm26, %ymm7
; AVX512DQ-FAST-NEXT:    vpshufb %ymm7, %ymm1, %ymm1
; AVX512DQ-FAST-NEXT:    vmovdqa64 %ymm27, %ymm8
; AVX512DQ-FAST-NEXT:    vpshufb %ymm8, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15]
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm3, %zmm1, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm19, %xmm1
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm20, %xmm2
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm21, %xmm2
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm22, %xmm3
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm3
; AVX512DQ-FAST-NEXT:    vpshufb %ymm14, %ymm3, %ymm3
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512DQ-FAST-NEXT:    vpshufb %ymm10, %ymm1, %ymm4
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm9, %ymm1, %ymm1
; AVX512DQ-FAST-NEXT:    vpshufb %xmm15, %xmm2, %xmm4
; AVX512DQ-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15]
; AVX512DQ-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512DQ-FAST-NEXT:    vmovdqa64 %xmm28, %xmm2
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3],xmm11[4],xmm2[4],xmm11[5],xmm2[5],xmm11[6],xmm2[6],xmm11[7],xmm2[7]
; AVX512DQ-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vpshufb %ymm5, %ymm2, %ymm4
; AVX512DQ-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX512DQ-FAST-NEXT:    vpshufb %ymm6, %ymm3, %ymm5
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX512DQ-FAST-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT:    vpshufb %ymm8, %ymm3, %ymm3
; AVX512DQ-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX512DQ-FAST-NEXT:    vinserti32x8 $1, %ymm4, %zmm2, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm1, (%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm0, 192(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm16, 128(%rax)
; AVX512DQ-FAST-NEXT:    vmovdqa64 %zmm17, 64(%rax)
; AVX512DQ-FAST-NEXT:    vzeroupper
; AVX512DQ-FAST-NEXT:    retq
;
; AVX512BW-SLOW-LABEL: store_i8_stride8_vf32:
; AVX512BW-SLOW:       # %bb.0:
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-SLOW-NEXT:    vmovdqa64 (%r11), %xmm25
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%r11), %xmm11
; AVX512BW-SLOW-NEXT:    vmovdqa (%r10), %xmm1
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%r10), %xmm12
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm25[8],xmm1[9],xmm25[9],xmm1[10],xmm25[10],xmm1[11],xmm25[11],xmm1[12],xmm25[12],xmm1[13],xmm25[13],xmm1[14],xmm25[14],xmm1[15],xmm25[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm5
; AVX512BW-SLOW-NEXT:    vmovdqa (%r9), %xmm2
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%r9), %xmm13
; AVX512BW-SLOW-NEXT:    vmovdqa (%r8), %xmm3
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%r8), %xmm14
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm4, %zmm4, %zmm19
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm23 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,16,17,20,52,20,21,21,53,16,17,22,54,22,21,23,55]
; AVX512BW-SLOW-NEXT:    vpermt2w %zmm5, %zmm23, %zmm19
; AVX512BW-SLOW-NEXT:    vmovdqa (%rsi), %xmm5
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%rsi), %xmm15
; AVX512BW-SLOW-NEXT:    vmovdqa (%rdi), %xmm6
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rdi), %xmm16
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm7[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm8, %ymm10
; AVX512BW-SLOW-NEXT:    vmovdqa (%rcx), %xmm8
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rcx), %xmm17
; AVX512BW-SLOW-NEXT:    vmovdqa (%rdx), %xmm9
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rdx), %xmm18
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm21 = xmm20[0,0,2,1,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm21 = xmm21[0],zero,xmm21[1],zero
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm22 = xmm20[0,2,2,3,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm22 = xmm22[0],zero,xmm22[1],zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm22, %ymm21, %ymm4
; AVX512BW-SLOW-NEXT:    vpblendw {{.*#+}} ymm4 = ymm10[0],ymm4[1],ymm10[2,3,4],ymm4[5],ymm10[6,7,8],ymm4[9],ymm10[10,11,12],ymm4[13],ymm10[14,15]
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} ymm22 = [4,20,u,u,5,21,u,u,6,22,u,u,7,23,u,u]
; AVX512BW-SLOW-NEXT:    vpermt2w %ymm20, %ymm22, %ymm7
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm4, %zmm7
; AVX512BW-SLOW-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512BW-SLOW-NEXT:    kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm19, %zmm7 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm4, %zmm4, %zmm4
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm19 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm19, %zmm19, %zmm20
; AVX512BW-SLOW-NEXT:    vpermt2w %zmm4, %zmm23, %zmm20
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm19 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm21 = xmm4[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm21, %ymm19, %ymm10
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm19 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm21 = xmm19[0,0,2,1,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm21 = xmm21[0],zero,xmm21[1],zero
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm24 = xmm19[0,2,2,3,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm24 = xmm24[0],zero,xmm24[1],zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm24, %ymm21, %ymm0
; AVX512BW-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm10[0],ymm0[1],ymm10[2,3,4],ymm0[5],ymm10[6,7,8],ymm0[9],ymm10[10,11,12],ymm0[13],ymm10[14,15]
; AVX512BW-SLOW-NEXT:    vpermt2w %ymm19, %ymm22, %ymm4
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm19
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm20, %zmm19 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm4, %zmm4, %zmm4
; AVX512BW-SLOW-NEXT:    vpermt2w %zmm0, %zmm23, %zmm4
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm16[8],xmm15[8],xmm16[9],xmm15[9],xmm16[10],xmm15[10],xmm16[11],xmm15[11],xmm16[12],xmm15[12],xmm16[13],xmm15[13],xmm16[14],xmm15[14],xmm16[15],xmm15[15]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm11 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm11 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm11 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm11[0,0,2,1,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm12 = xmm12[0],zero,xmm12[1],zero
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm11[0,2,2,3,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm13 = xmm13[0],zero,xmm13[1],zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm13, %ymm12, %ymm12
; AVX512BW-SLOW-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0],ymm12[1],ymm10[2,3,4],ymm12[5],ymm10[6,7,8],ymm12[9],ymm10[10,11,12],ymm12[13],ymm10[14,15]
; AVX512BW-SLOW-NEXT:    vpermt2w %ymm11, %ymm22, %ymm0
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm10, %zmm0
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm4, %zmm0 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm25[0],xmm1[1],xmm25[1],xmm1[2],xmm25[2],xmm1[3],xmm25[3],xmm1[4],xmm25[4],xmm1[5],xmm25[5],xmm1[6],xmm25[6],xmm1[7],xmm25[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm2
; AVX512BW-SLOW-NEXT:    vpermt2w %zmm1, %zmm23, %zmm2
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm1[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm4[0,0,2,1,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX512BW-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm4[0,2,2,3,4,5,6,7]
; AVX512BW-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX512BW-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4],ymm5[5],ymm3[6,7,8],ymm5[9],ymm3[10,11,12],ymm5[13],ymm3[14,15]
; AVX512BW-SLOW-NEXT:    vpermt2w %ymm4, %ymm22, %ymm1
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm3, %zmm1
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm2, %zmm1 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm1, (%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm0, 192(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm19, 128(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm7, 64(%rax)
; AVX512BW-SLOW-NEXT:    vzeroupper
; AVX512BW-SLOW-NEXT:    retq
;
; AVX512BW-FAST-LABEL: store_i8_stride8_vf32:
; AVX512BW-FAST:       # %bb.0:
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FAST-NEXT:    vmovdqa (%r11), %xmm0
; AVX512BW-FAST-NEXT:    vmovdqa 16(%r11), %xmm7
; AVX512BW-FAST-NEXT:    vmovdqa (%r10), %xmm1
; AVX512BW-FAST-NEXT:    vmovdqa 16(%r10), %xmm8
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm5
; AVX512BW-FAST-NEXT:    vmovdqa (%r9), %xmm2
; AVX512BW-FAST-NEXT:    vmovdqa 16(%r9), %xmm9
; AVX512BW-FAST-NEXT:    vmovdqa (%r8), %xmm3
; AVX512BW-FAST-NEXT:    vmovdqa 16(%r8), %xmm10
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm4, %zmm4, %zmm11
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm4 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,16,17,20,52,20,21,21,53,16,17,22,54,22,21,23,55]
; AVX512BW-FAST-NEXT:    vpermt2w %zmm5, %zmm4, %zmm11
; AVX512BW-FAST-NEXT:    vmovdqa (%rcx), %xmm5
; AVX512BW-FAST-NEXT:    vmovdqa 16(%rcx), %xmm12
; AVX512BW-FAST-NEXT:    vmovdqa (%rdx), %xmm13
; AVX512BW-FAST-NEXT:    vmovdqa 16(%rdx), %xmm14
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm13[8],xmm5[8],xmm13[9],xmm5[9],xmm13[10],xmm5[10],xmm13[11],xmm5[11],xmm13[12],xmm5[12],xmm13[13],xmm5[13],xmm13[14],xmm5[14],xmm13[15],xmm5[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm6, %zmm6, %zmm15
; AVX512BW-FAST-NEXT:    vmovdqa64 (%rsi), %xmm16
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%rsi), %xmm17
; AVX512BW-FAST-NEXT:    vmovdqa64 (%rdi), %xmm18
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%rdi), %xmm19
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm6, %zmm6, %zmm6
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm20 = [0,32,u,u,1,33,u,u,2,34,u,u,3,35,u,u,20,52,u,u,21,53,u,u,22,54,u,u,23,55,u,u]
; AVX512BW-FAST-NEXT:    vpermt2w %zmm15, %zmm20, %zmm6
; AVX512BW-FAST-NEXT:    movw $-21846, %cx # imm = 0xAAAA
; AVX512BW-FAST-NEXT:    kmovd %ecx, %k1
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm11, %zmm6 {%k1}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm11 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm11, %zmm11, %zmm11
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm15 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm15, %zmm15, %zmm15
; AVX512BW-FAST-NEXT:    vpermt2w %zmm11, %zmm4, %zmm15
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3],xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm11, %zmm11, %zmm11
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm21 = xmm19[0],xmm17[0],xmm19[1],xmm17[1],xmm19[2],xmm17[2],xmm19[3],xmm17[3],xmm19[4],xmm17[4],xmm19[5],xmm17[5],xmm19[6],xmm17[6],xmm19[7],xmm17[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm21, %zmm21, %zmm21
; AVX512BW-FAST-NEXT:    vpermt2w %zmm11, %zmm20, %zmm21
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm15, %zmm21 {%k1}
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm7, %zmm7, %zmm7
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm8, %zmm8, %zmm8
; AVX512BW-FAST-NEXT:    vpermt2w %zmm7, %zmm4, %zmm8
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm7, %zmm7, %zmm7
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm19[8],xmm17[8],xmm19[9],xmm17[9],xmm19[10],xmm17[10],xmm19[11],xmm17[11],xmm19[12],xmm17[12],xmm19[13],xmm17[13],xmm19[14],xmm17[14],xmm19[15],xmm17[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm9, %zmm9, %zmm9
; AVX512BW-FAST-NEXT:    vpermt2w %zmm7, %zmm20, %zmm9
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm8, %zmm9 {%k1}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
; AVX512BW-FAST-NEXT:    vpermt2w %zmm0, %zmm4, %zmm1
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3],xmm13[4],xmm5[4],xmm13[5],xmm5[5],xmm13[6],xmm5[6],xmm13[7],xmm5[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm2
; AVX512BW-FAST-NEXT:    vpermt2w %zmm0, %zmm20, %zmm2
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm1, %zmm2 {%k1}
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm21, 128(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm6, 64(%rax)
; AVX512BW-FAST-NEXT:    vzeroupper
; AVX512BW-FAST-NEXT:    retq
  %in.vec0 = load <32 x i8>, ptr %in.vecptr0, align 64
  %in.vec1 = load <32 x i8>, ptr %in.vecptr1, align 64
  %in.vec2 = load <32 x i8>, ptr %in.vecptr2, align 64
  %in.vec3 = load <32 x i8>, ptr %in.vecptr3, align 64
  %in.vec4 = load <32 x i8>, ptr %in.vecptr4, align 64
  %in.vec5 = load <32 x i8>, ptr %in.vecptr5, align 64
  %in.vec6 = load <32 x i8>, ptr %in.vecptr6, align 64
  %in.vec7 = load <32 x i8>, ptr %in.vecptr7, align 64
  %1 = shufflevector <32 x i8> %in.vec0, <32 x i8> %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 i8> %in.vec2, <32 x i8> %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 i8> %in.vec4, <32 x i8> %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 <32 x i8> %in.vec6, <32 x i8> %in.vec7, <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 <64 x i8> %1, <64 x i8> %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>
  %6 = shufflevector <64 x i8> %3, <64 x i8> %4, <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>
  %7 = shufflevector <128 x i8> %5, <128 x i8> %6, <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>
  %interleaved.vec = shufflevector <256 x i8> %7, <256 x i8> poison, <256 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 192, i32 224, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 193, i32 225, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 194, i32 226, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 195, i32 227, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 196, i32 228, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 197, i32 229, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 198, i32 230, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 199, i32 231, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 200, i32 232, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 201, i32 233, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 202, i32 234, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 203, i32 235, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 204, i32 236, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 205, i32 237, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 206, i32 238, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 207, i32 239, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 208, i32 240, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 209, i32 241, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 210, i32 242, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 211, i32 243, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 212, i32 244, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 213, i32 245, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 214, i32 246, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 215, i32 247, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 216, i32 248, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 217, i32 249, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 218, i32 250, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 219, i32 251, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 220, i32 252, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 221, i32 253, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 222, i32 254, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191, i32 223, i32 255>
  store <256 x i8> %interleaved.vec, ptr %out.vec, align 64
  ret void
}

define void @store_i8_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride8_vf64:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $312, %rsp # imm = 0x138
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movdqa (%rdi), %xmm3
; SSE-NEXT:    movdqa (%rsi), %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%rdx), %xmm4
; SSE-NEXT:    movdqa (%rcx), %xmm8
; SSE-NEXT:    movdqa (%r8), %xmm6
; SSE-NEXT:    movdqa (%r9), %xmm9
; SSE-NEXT:    movdqa (%r10), %xmm7
; SSE-NEXT:    movdqa (%rax), %xmm10
; SSE-NEXT:    movdqa %xmm7, %xmm0
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,1]
; SSE-NEXT:    movdqa {{.*#+}} xmm13 = [65535,65535,65535,0,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm13, %xmm12
; SSE-NEXT:    pandn %xmm2, %xmm12
; SSE-NEXT:    movdqa %xmm6, %xmm11
; SSE-NEXT:    punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3],xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm11[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm14
; SSE-NEXT:    por %xmm12, %xmm14
; SSE-NEXT:    movdqa %xmm4, %xmm12
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3],xmm12[4],xmm8[4],xmm12[5],xmm8[5],xmm12[6],xmm8[6],xmm12[7],xmm8[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm12[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm2, %xmm15
; SSE-NEXT:    movdqa %xmm3, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm5
; SSE-NEXT:    por %xmm15, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm14[0],xmm5[1],xmm14[1]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm5, %xmm14
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm11[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm14, %xmm5
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm12[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm14 = xmm14[0],xmm5[0],xmm14[1],xmm5[1]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm0[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm5, %xmm14
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm11[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm14, %xmm5
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm14 = xmm14[0],xmm5[0],xmm14[1],xmm5[1]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm11[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm11
; SSE-NEXT:    por %xmm5, %xmm11
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm12[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm12 = xmm0[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm12
; SSE-NEXT:    por %xmm5, %xmm12
; SSE-NEXT:    movdqa 16(%r8), %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm11[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm12[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhbw {{.*#+}} xmm7 = xmm7[8],xmm10[8],xmm7[9],xmm10[9],xmm7[10],xmm10[10],xmm7[11],xmm10[11],xmm7[12],xmm10[12],xmm7[13],xmm10[13],xmm7[14],xmm10[14],xmm7[15],xmm10[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm7[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm10
; SSE-NEXT:    pandn %xmm5, %xmm10
; SSE-NEXT:    punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm9[8],xmm6[9],xmm9[9],xmm6[10],xmm9[10],xmm6[11],xmm9[11],xmm6[12],xmm9[12],xmm6[13],xmm9[13],xmm6[14],xmm9[14],xmm6[15],xmm9[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm6[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm10, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm4[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm9
; SSE-NEXT:    pandn %xmm8, %xmm9
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; SSE-NEXT:    # xmm3 = xmm3[8],mem[8],xmm3[9],mem[9],xmm3[10],mem[10],xmm3[11],mem[11],xmm3[12],mem[12],xmm3[13],mem[13],xmm3[14],mem[14],xmm3[15],mem[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm3[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm8
; SSE-NEXT:    por %xmm9, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm7[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm8
; SSE-NEXT:    pandn %xmm5, %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm8, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm4[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm9
; SSE-NEXT:    pandn %xmm8, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm3[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm8
; SSE-NEXT:    por %xmm9, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm7[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm8
; SSE-NEXT:    pandn %xmm5, %xmm8
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm5[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm9
; SSE-NEXT:    por %xmm8, %xmm9
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm5, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm5[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm10
; SSE-NEXT:    por %xmm8, %xmm10
; SSE-NEXT:    movdqa 16(%r10), %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    movdqa 16(%rax), %xmm7
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm8, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm4, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm3
; SSE-NEXT:    por %xmm8, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm5, %xmm10
; SSE-NEXT:    punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3],xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm10[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm4
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    movdqa 16(%r9), %xmm6
; SSE-NEXT:    movdqa %xmm0, %xmm11
; SSE-NEXT:    punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm11[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm3[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm14
; SSE-NEXT:    por %xmm4, %xmm14
; SSE-NEXT:    movdqa 16(%rdx), %xmm3
; SSE-NEXT:    movdqa 16(%rcx), %xmm8
; SSE-NEXT:    movdqa %xmm3, %xmm12
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3],xmm12[4],xmm8[4],xmm12[5],xmm8[5],xmm12[6],xmm8[6],xmm12[7],xmm8[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm12[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm4, %xmm15
; SSE-NEXT:    movdqa 16(%rdi), %xmm4
; SSE-NEXT:    movdqa 16(%rsi), %xmm9
; SSE-NEXT:    movdqa %xmm4, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3],xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm15, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm10[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm7, %xmm14
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm11[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm7
; SSE-NEXT:    por %xmm14, %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm12[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm14 = xmm14[0],xmm7[0],xmm14[1],xmm7[1]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm10[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm7, %xmm14
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm11[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm7
; SSE-NEXT:    por %xmm14, %xmm7
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm14 = xmm14[0],xmm7[0],xmm14[1],xmm7[1]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm10[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm10
; SSE-NEXT:    pandn %xmm7, %xmm10
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm11[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm7
; SSE-NEXT:    por %xmm10, %xmm7
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm12[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm11
; SSE-NEXT:    pandn %xmm10, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm2[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm10
; SSE-NEXT:    por %xmm11, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm5[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm10
; SSE-NEXT:    pandn %xmm7, %xmm10
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm6[8],xmm0[9],xmm6[9],xmm0[10],xmm6[10],xmm0[11],xmm6[11],xmm0[12],xmm6[12],xmm0[13],xmm6[13],xmm0[14],xmm6[14],xmm0[15],xmm6[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm10, %xmm6
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm8[8],xmm3[9],xmm8[9],xmm3[10],xmm8[10],xmm3[11],xmm8[11],xmm3[12],xmm8[12],xmm3[13],xmm8[13],xmm3[14],xmm8[14],xmm3[15],xmm8[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm3[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm9[8],xmm4[9],xmm9[9],xmm4[10],xmm9[10],xmm4[11],xmm9[11],xmm4[12],xmm9[12],xmm4[13],xmm9[13],xmm4[14],xmm9[14],xmm4[15],xmm9[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm5[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm6, %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm0[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm7, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm6, %xmm7
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm0[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm7, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm3[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm6
; SSE-NEXT:    pandn %xmm5, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm6, %xmm0
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm5
; SSE-NEXT:    pandn %xmm3, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm3
; SSE-NEXT:    por %xmm5, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 32(%r10), %xmm0
; SSE-NEXT:    movdqa 32(%rax), %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm0, %xmm10
; SSE-NEXT:    punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3],xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm3 = xmm10[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm4
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    movdqa 32(%r8), %xmm3
; SSE-NEXT:    movdqa 32(%r9), %xmm7
; SSE-NEXT:    movdqa %xmm3, %xmm11
; SSE-NEXT:    punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm7[0],xmm11[1],xmm7[1],xmm11[2],xmm7[2],xmm11[3],xmm7[3],xmm11[4],xmm7[4],xmm11[5],xmm7[5],xmm11[6],xmm7[6],xmm11[7],xmm7[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm11[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm5[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm14
; SSE-NEXT:    por %xmm4, %xmm14
; SSE-NEXT:    movdqa 32(%rdx), %xmm4
; SSE-NEXT:    movdqa 32(%rcx), %xmm8
; SSE-NEXT:    movdqa %xmm4, %xmm12
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3],xmm12[4],xmm8[4],xmm12[5],xmm8[5],xmm12[6],xmm8[6],xmm12[7],xmm8[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm12[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm5, %xmm15
; SSE-NEXT:    movdqa 32(%rdi), %xmm5
; SSE-NEXT:    movdqa 32(%rsi), %xmm9
; SSE-NEXT:    movdqa %xmm5, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3],xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm6
; SSE-NEXT:    por %xmm15, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm14[0],xmm6[1],xmm14[1]
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm10[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm6, %xmm14
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm11[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm14, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm12[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm10[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm6, %xmm14
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm11[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm14, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm10[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm10
; SSE-NEXT:    pandn %xmm6, %xmm10
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm11[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm10, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm12[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm11
; SSE-NEXT:    pandn %xmm10, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm2[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm10
; SSE-NEXT:    por %xmm11, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm0[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm10
; SSE-NEXT:    pandn %xmm6, %xmm10
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm7[8],xmm3[9],xmm7[9],xmm3[10],xmm7[10],xmm3[11],xmm7[11],xmm3[12],xmm7[12],xmm3[13],xmm7[13],xmm3[14],xmm7[14],xmm3[15],xmm7[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm3[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm10, %xmm6
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm4[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm9[8],xmm5[9],xmm9[9],xmm5[10],xmm9[10],xmm5[11],xmm9[11],xmm5[12],xmm9[12],xmm5[13],xmm9[13],xmm5[14],xmm9[14],xmm5[15],xmm9[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm0[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm6, %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm3[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm7, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm4[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, (%rsp) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm0[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm6, %xmm7
; SSE-NEXT:    pshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm7, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm6
; SSE-NEXT:    pandn %xmm0, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm6, %xmm0
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm4[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm4
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm3
; SSE-NEXT:    por %xmm4, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 48(%r10), %xmm9
; SSE-NEXT:    movdqa 48(%rax), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm6
; SSE-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm6[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm3
; SSE-NEXT:    pandn %xmm0, %xmm3
; SSE-NEXT:    movdqa 48(%r8), %xmm8
; SSE-NEXT:    movdqa 48(%r9), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm8, %xmm4
; SSE-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm4[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm10
; SSE-NEXT:    por %xmm3, %xmm10
; SSE-NEXT:    movdqa 48(%rdx), %xmm7
; SSE-NEXT:    movdqa 48(%rcx), %xmm12
; SSE-NEXT:    movdqa %xmm7, %xmm3
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm3[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm0, %xmm15
; SSE-NEXT:    movdqa 48(%rdi), %xmm5
; SSE-NEXT:    movdqa 48(%rsi), %xmm11
; SSE-NEXT:    movdqa %xmm5, %xmm0
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3],xmm0[4],xmm11[4],xmm0[5],xmm11[5],xmm0[6],xmm11[6],xmm0[7],xmm11[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm10 = xmm6[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm10, %xmm14
; SSE-NEXT:    pshuflw {{.*#+}} xmm10 = xmm4[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm10
; SSE-NEXT:    por %xmm14, %xmm10
; SSE-NEXT:    pshuflw {{.*#+}} xmm14 = xmm3[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm15
; SSE-NEXT:    pandn %xmm14, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm15, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm15 = xmm15[0],xmm10[0],xmm15[1],xmm10[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm6[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm14
; SSE-NEXT:    pandn %xmm10, %xmm14
; SSE-NEXT:    pshufhw {{.*#+}} xmm10 = xmm4[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm10
; SSE-NEXT:    por %xmm14, %xmm10
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm3[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    pandn %xmm14, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    por %xmm2, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm10[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm14[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm10 = xmm10[0],xmm2[0],xmm10[1],xmm2[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm6
; SSE-NEXT:    pandn %xmm2, %xmm6
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm4[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm2
; SSE-NEXT:    por %xmm6, %xmm2
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm4
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm0
; SSE-NEXT:    por %xmm4, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
; SSE-NEXT:    # xmm9 = xmm9[8],mem[8],xmm9[9],mem[9],xmm9[10],mem[10],xmm9[11],mem[11],xmm9[12],mem[12],xmm9[13],mem[13],xmm9[14],mem[14],xmm9[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm9[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm0, %xmm2
; SSE-NEXT:    punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
; SSE-NEXT:    # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm8[0,1,1,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm7 = xmm7[8],xmm12[8],xmm7[9],xmm12[9],xmm7[10],xmm12[10],xmm7[11],xmm12[11],xmm7[12],xmm12[12],xmm7[13],xmm12[13],xmm7[14],xmm12[14],xmm7[15],xmm12[15]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm7[0,0,2,1,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm11[8],xmm5[9],xmm11[9],xmm5[10],xmm11[10],xmm5[11],xmm11[11],xmm5[12],xmm11[12],xmm5[13],xmm11[13],xmm5[14],xmm11[14],xmm5[15],xmm11[15]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,0,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm9[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm0, %xmm2
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm8[2,1,3,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1]
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm2, %xmm0
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm7[0,2,2,3,4,5,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,1,1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm9[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm8[0,1,2,3,4,5,5,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm7[0,1,2,3,4,4,6,5]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3]
; SSE-NEXT:    movdqa %xmm1, %xmm11
; SSE-NEXT:    pandn %xmm3, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[2,2,2,2]
; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm3
; SSE-NEXT:    por %xmm11, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm9[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,6,5,7,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3]
; SSE-NEXT:    pand %xmm13, %xmm8
; SSE-NEXT:    pandn %xmm2, %xmm13
; SSE-NEXT:    por %xmm8, %xmm13
; SSE-NEXT:    pshufhw {{.*#+}} xmm2 = xmm7[0,1,2,3,4,6,6,7]
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,1,3,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[3,3,3,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5]
; SSE-NEXT:    pand %xmm1, %xmm5
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    por %xmm5, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm13[1,3,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movdqa %xmm1, 496(%rax)
; SSE-NEXT:    movdqa %xmm3, 480(%rax)
; SSE-NEXT:    movdqa %xmm0, 464(%rax)
; SSE-NEXT:    movdqa %xmm4, 448(%rax)
; SSE-NEXT:    movdqa %xmm6, 432(%rax)
; SSE-NEXT:    movdqa %xmm10, 416(%rax)
; SSE-NEXT:    movdqa %xmm15, 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, 368(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%rax)
; SSE-NEXT:    movaps (%rsp), %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, 320(%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, 272(%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, 224(%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 {{[-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, 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, 128(%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, 80(%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, 32(%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:    addq $312, %rsp # imm = 0x138
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: store_i8_stride8_vf64:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $360, %rsp # imm = 0x168
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT:    vmovdqa (%r10), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa (%rax), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vbroadcastsd {{.*#+}} ymm13 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0]
; AVX1-ONLY-NEXT:    vandnps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa (%r9), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa (%r8), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm4[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm4[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm13, %ymm2
; AVX1-ONLY-NEXT:    vorps %ymm0, %ymm2, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa (%rcx), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa (%rdx), %xmm3
; AVX1-ONLY-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm3[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa (%rsi), %xmm5
; AVX1-ONLY-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa (%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm6[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm8 = xmm6[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm8, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm8 = ymm2[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vbroadcastsd {{.*#+}} ymm5 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535]
; AVX1-ONLY-NEXT:    vandnps %ymm8, %ymm5, %ymm8
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vorps %ymm7, %ymm8, %ymm7
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm7[0],ymm0[1],ymm7[2],ymm0[3],ymm7[4],ymm0[5],ymm7[6],ymm0[7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm4[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm4[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm1, %ymm4
; AVX1-ONLY-NEXT:    vmovdqa 48(%r10), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm4, %ymm13, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm6[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm3[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vmovdqa 48(%rax), %xmm2
; AVX1-ONLY-NEXT:    vmovdqa %xmm2, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vandnps %ymm6, %ymm5, %ymm6
; AVX1-ONLY-NEXT:    vorps %ymm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0],ymm0[1],ymm4[2],ymm0[3],ymm4[4],ymm0[5],ymm4[6],ymm0[7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm11 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm11[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm11[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm6
; AVX1-ONLY-NEXT:    vmovdqa 48(%r9), %xmm4
; AVX1-ONLY-NEXT:    vmovdqa 48(%r8), %xmm0
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm12 = xmm0[8],xmm4[8],xmm0[9],xmm4[9],xmm0[10],xmm4[10],xmm0[11],xmm4[11],xmm0[12],xmm4[12],xmm0[13],xmm4[13],xmm0[14],xmm4[14],xmm0[15],xmm4[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm12[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm12[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm8, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm6, %ymm13, %ymm6
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm7 = ymm7[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm7, %ymm13, %ymm7
; AVX1-ONLY-NEXT:    vorps %ymm6, %ymm7, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa 48(%rsi), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa 48(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm10[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm9, %ymm14
; AVX1-ONLY-NEXT:    vmovdqa 48(%rcx), %xmm7
; AVX1-ONLY-NEXT:    vmovdqa 48(%rdx), %xmm9
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm15[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm15[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm14, %ymm3
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm5, %ymm1
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm11[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm12[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm12[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm3, %ymm13, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm15[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm15[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm10[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm11 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm11, %ymm10
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm10, %ymm10
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm10, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4],ymm1[5],ymm3[6],ymm1[7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm3, %ymm10
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm3[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm10[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm13, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm0, %ymm4
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm0[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm8, %ymm6
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm7[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm7[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vandnps %ymm8, %ymm5, %ymm8
; AVX1-ONLY-NEXT:    vorps %ymm6, %ymm8, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm6[0],ymm4[1],ymm6[2],ymm4[3],ymm6[4],ymm4[5],ymm6[6],ymm4[7]
; AVX1-ONLY-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm3, %ymm13, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm7[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 32(%r10), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 32(%rax), %xmm4
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm11 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm11[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm11[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm6
; AVX1-ONLY-NEXT:    vmovdqa 32(%r9), %xmm1
; AVX1-ONLY-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 32(%r8), %xmm0
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm12 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm12[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm12[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm8, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm6, %ymm13, %ymm6
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm7 = ymm7[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm7, %ymm13, %ymm7
; AVX1-ONLY-NEXT:    vorps %ymm6, %ymm7, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa 32(%rsi), %xmm6
; AVX1-ONLY-NEXT:    vmovdqa 32(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm10[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm9, %ymm14
; AVX1-ONLY-NEXT:    vmovdqa 32(%rcx), %xmm7
; AVX1-ONLY-NEXT:    vmovdqa 32(%rdx), %xmm9
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm15[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm15[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm14, %ymm3
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm5, %ymm1
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm11[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm12[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm12[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm3, %ymm13, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm15[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm15[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm11 = xmm10[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm11 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm11, %ymm10
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm10, %ymm10
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm10, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4],ymm1[5],ymm3[6],ymm1[7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm4[0],mem[0],xmm4[1],mem[1],xmm4[2],mem[2],xmm4[3],mem[3],xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm4
; AVX1-ONLY-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm3 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm3[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm3[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm13, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm0, %ymm4
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm0[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm8, %ymm6
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm7[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm7[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm6, %ymm6
; AVX1-ONLY-NEXT:    vandnps %ymm8, %ymm5, %ymm8
; AVX1-ONLY-NEXT:    vorps %ymm6, %ymm8, %ymm6
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm6[0],ymm4[1],ymm6[2],ymm4[3],ymm6[4],ymm4[5],ymm6[6],ymm4[7]
; AVX1-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm4, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm3, %ymm13, %ymm3
; AVX1-ONLY-NEXT:    vorps %ymm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm7[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 16(%r10), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 16(%rax), %xmm8
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm8[8],xmm0[8],xmm8[9],xmm0[9],xmm8[10],xmm0[10],xmm8[11],xmm0[11],xmm8[12],xmm0[12],xmm8[13],xmm0[13],xmm8[14],xmm0[14],xmm8[15],xmm0[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm2[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vmovdqa 16(%r9), %xmm0
; AVX1-ONLY-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovdqa 16(%r8), %xmm9
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm0[8],xmm9[9],xmm0[9],xmm9[10],xmm0[10],xmm9[11],xmm0[11],xmm9[12],xmm0[12],xmm9[13],xmm0[13],xmm9[14],xmm0[14],xmm9[15],xmm0[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm1[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm1[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm13, %ymm3
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm4, %ymm13, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm4, %ymm15
; AVX1-ONLY-NEXT:    vmovdqa 16(%rsi), %xmm11
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm6[8],xmm11[8],xmm6[9],xmm11[9],xmm6[10],xmm11[10],xmm6[11],xmm11[11],xmm6[12],xmm11[12],xmm6[13],xmm11[13],xmm6[14],xmm11[14],xmm6[15],xmm11[15]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm10[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vmovdqa 16(%rcx), %xmm7
; AVX1-ONLY-NEXT:    vmovdqa 16(%rdx), %xmm4
; AVX1-ONLY-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm7[8],xmm4[9],xmm7[9],xmm4[10],xmm7[10],xmm4[11],xmm7[11],xmm4[12],xmm7[12],xmm4[13],xmm7[13],xmm4[14],xmm7[14],xmm4[15],xmm7[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm0[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm12 = xmm12[0],zero,xmm12[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm0[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm14, %ymm12, %ymm12
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm3, %ymm3
; AVX1-ONLY-NEXT:    vandnps %ymm12, %ymm5, %ymm12
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm12, %ymm3
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm3[0],ymm15[1],ymm3[2],ymm15[3],ymm3[4],ymm15[5],ymm3[6],ymm15[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm2, %ymm13, %ymm2
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT:    vorps %ymm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm10[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm3 = xmm10[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm0, %ymm5, %ymm0
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vorps %ymm0, %ymm2, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm10 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX1-ONLY-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm8 = xmm8[0],mem[0],xmm8[1],mem[1],xmm8[2],mem[2],xmm8[3],mem[3],xmm8[4],mem[4],xmm8[5],mem[5],xmm8[6],mem[6],xmm8[7],mem[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm8[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm8[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm9 = xmm9[0],mem[0],xmm9[1],mem[1],xmm9[2],mem[2],xmm9[3],mem[3],xmm9[4],mem[4],xmm9[5],mem[5],xmm9[6],mem[6],xmm9[7],mem[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm9[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm9[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT:    vorps %ymm0, %ymm1, %ymm1
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3],xmm6[4],xmm11[4],xmm6[5],xmm11[5],xmm6[6],xmm11[6],xmm6[7],xmm11[7]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
; AVX1-ONLY-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm3[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm8[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm8[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm9[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm9[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm2, %ymm13, %ymm2
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm4, %ymm13, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm2, %ymm4, %ymm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm3, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7]
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm4 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm4[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm4[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm6 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm6 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm6[0,1,1,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm6[2,1,3,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandnps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
; AVX1-ONLY-NEXT:    vandps %ymm2, %ymm13, %ymm2
; AVX1-ONLY-NEXT:    vorps %ymm0, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm7 = xmm0[1,1,1,1]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm7, %ymm8, %ymm7
; AVX1-ONLY-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15]
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm8[0,0,2,1,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero
; AVX1-ONLY-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm8[0,2,2,3,4,5,6,7]
; AVX1-ONLY-NEXT:    vpmovzxdq {{.*#+}} xmm11 = xmm11[0],zero,xmm11[1],zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm11, %ymm9, %ymm9
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm7, %ymm7
; AVX1-ONLY-NEXT:    vandnps %ymm9, %ymm5, %ymm9
; AVX1-ONLY-NEXT:    vorps %ymm7, %ymm9, %ymm7
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0],ymm2[1],ymm7[2],ymm2[3],ymm7[4],ymm2[5],ymm7[6],ymm2[7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm4[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm6[0,1,2,3,4,5,5,7]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandnps %ymm2, %ymm13, %ymm2
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vandps %ymm4, %ymm13, %ymm4
; AVX1-ONLY-NEXT:    vorps %ymm2, %ymm4, %ymm2
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm8[0,1,2,3,4,4,6,5]
; AVX1-ONLY-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,4,6,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm6 = xmm0[2,3,2,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX1-ONLY-NEXT:    vinsertf128 $1, %xmm0, %ymm6, %ymm0
; AVX1-ONLY-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7]
; AVX1-ONLY-NEXT:    vandnps %ymm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT:    vandps %ymm5, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vorps %ymm4, %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7]
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm7, 64(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 160(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm10, 224(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm15, 192(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 352(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 416(%rax)
; AVX1-ONLY-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 384(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 448(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX1-ONLY-NEXT:    addq $360, %rsp # imm = 0x168
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-SLOW-LABEL: store_i8_stride8_vf64:
; AVX2-SLOW:       # %bb.0:
; AVX2-SLOW-NEXT:    subq $328, %rsp # imm = 0x148
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-SLOW-NEXT:    vmovdqa (%r10), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%rax), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vmovdqa (%r9), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%r8), %xmm3
; AVX2-SLOW-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm5 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vmovdqa (%rsi), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-SLOW-NEXT:    vmovdqa (%rcx), %xmm0
; AVX2-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa (%rdx), %xmm6
; AVX2-SLOW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm6[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm6[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX2-SLOW-NEXT:    vmovdqa 48(%r10), %xmm0
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm7[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm7 = ymm1[0],ymm7[1],ymm1[2,3,4],ymm7[5],ymm1[6,7,8],ymm7[9],ymm1[10,11,12],ymm7[13],ymm1[14,15]
; AVX2-SLOW-NEXT:    vmovdqa 48(%rax), %xmm1
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2],ymm5[3],ymm7[4],ymm5[5],ymm7[6],ymm5[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm3
; AVX2-SLOW-NEXT:    vmovdqa 48(%r9), %xmm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm5[3],ymm3[4,5,6],ymm5[7],ymm3[8,9,10],ymm5[11],ymm3[12,13,14],ymm5[15]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm5, %ymm4
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm6[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3,4],ymm5[5],ymm4[6,7,8],ymm5[9],ymm4[10,11,12],ymm5[13],ymm4[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm8[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm8[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm4
; AVX2-SLOW-NEXT:    vmovdqa 48(%r8), %xmm3
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm9[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm9[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm11 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-SLOW-NEXT:    vmovdqa 48(%rsi), %xmm4
; AVX2-SLOW-NEXT:    vmovdqa 48(%rdi), %xmm5
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm10[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm12
; AVX2-SLOW-NEXT:    vmovdqa 48(%rcx), %xmm6
; AVX2-SLOW-NEXT:    vmovdqa 48(%rdx), %xmm7
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm13 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm13[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm13[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm15, %ymm14, %ymm14
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3,4],ymm14[5],ymm12[6,7,8],ymm14[9],ymm12[10,11,12],ymm14[13],ymm12[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2],ymm11[3],ymm12[4],ymm11[5],ymm12[6],ymm11[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm8[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm9[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm11, %ymm9
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5,6],ymm8[7],ymm9[8,9,10],ymm8[11],ymm9[12,13,14],ymm8[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm10[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm13[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm13[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm10[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3,4],ymm10[5],ymm9[6,7,8],ymm10[9],ymm9[10,11,12],ymm10[13],ymm9[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2],ymm8[3],ymm9[4],ymm8[5],ymm9[6],ymm8[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm0[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm3[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm5, %ymm4
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm5[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm6
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2,3,4],ymm6[5],ymm4[6,7,8],ymm6[9],ymm4[10,11,12],ymm6[13],ymm4[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2],ymm1[3],ymm4[4],ymm1[5],ymm4[6],ymm1[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm3[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm5[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm5[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX2-SLOW-NEXT:    vmovdqa 32(%r10), %xmm0
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 32(%rax), %xmm1
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm8[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm8[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm4
; AVX2-SLOW-NEXT:    vmovdqa 32(%r9), %xmm2
; AVX2-SLOW-NEXT:    vmovdqa 32(%r8), %xmm3
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm9[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm9[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm11 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm4
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm10[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm12
; AVX2-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm6
; AVX2-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm7
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm13 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm13[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm13[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm15, %ymm14, %ymm14
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3,4],ymm14[5],ymm12[6,7,8],ymm14[9],ymm12[10,11,12],ymm14[13],ymm12[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2],ymm11[3],ymm12[4],ymm11[5],ymm12[6],ymm11[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm8[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm9[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm11, %ymm9
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm9 = ymm9[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5,6],ymm8[7],ymm9[8,9,10],ymm8[11],ymm9[12,13,14],ymm8[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm10[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm13[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm11 = xmm13[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm11, %ymm10, %ymm10
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm10[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3,4],ymm10[5],ymm9[6,7,8],ymm10[9],ymm9[10,11,12],ymm10[13],ymm9[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2],ymm8[3],ymm9[4],ymm8[5],ymm9[6],ymm8[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm0[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15]
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm3[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm5, %ymm4
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm5[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm6
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2,3,4],ymm6[5],ymm4[6,7,8],ymm6[9],ymm4[10,11,12],ymm6[13],ymm4[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2],ymm1[3],ymm4[4],ymm1[5],ymm4[6],ymm1[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm3[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm5[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm5[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-SLOW-NEXT:    vmovdqu %ymm0, (%rsp) # 32-byte Spill
; AVX2-SLOW-NEXT:    vmovdqa 16(%r10), %xmm13
; AVX2-SLOW-NEXT:    vmovdqa 16(%rax), %xmm12
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm12[8],xmm13[8],xmm12[9],xmm13[9],xmm12[10],xmm13[10],xmm12[11],xmm13[11],xmm12[12],xmm13[12],xmm12[13],xmm13[13],xmm12[14],xmm13[14],xmm12[15],xmm13[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm8[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm8[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vmovdqa 16(%r9), %xmm10
; AVX2-SLOW-NEXT:    vmovdqa 16(%r8), %xmm9
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm4 = xmm9[8],xmm10[8],xmm9[9],xmm10[9],xmm9[10],xmm10[10],xmm9[11],xmm10[11],xmm9[12],xmm10[12],xmm9[13],xmm10[13],xmm9[14],xmm10[14],xmm9[15],xmm10[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm4[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm11 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vmovdqa 16(%rsi), %xmm7
; AVX2-SLOW-NEXT:    vmovdqa 16(%rdi), %xmm6
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
; AVX2-SLOW-NEXT:    vmovdqa 16(%rcx), %xmm5
; AVX2-SLOW-NEXT:    vmovdqa 16(%rdx), %xmm3
; AVX2-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm15 = xmm0[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm0[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm14, %ymm15, %ymm14
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm14[1],ymm2[2,3,4],ymm14[5],ymm2[6,7,8],ymm14[9],ymm2[10,11,12],ymm14[13],ymm2[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm2[0],ymm11[1],ymm2[2],ymm11[3],ymm2[4],ymm11[5],ymm2[6],ymm11[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm8[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm8 = xmm4[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm1[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm4, %ymm1
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm4, %ymm0
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7,8],ymm0[9],ymm1[10,11,12],ymm0[13],ymm1[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm8 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7]
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm2[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm2[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm4, %ymm4
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6],ymm0[7],ymm4[8,9,10],ymm0[11],ymm4[12,13,14],ymm0[15]
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm4[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm6
; AVX2-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm3[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm3[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3,4],ymm5[5],ymm6[6,7,8],ymm5[9],ymm6[10,11,12],ymm5[13],ymm6[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm1[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm5, %ymm1
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm5, %ymm2
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm4[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3,4],ymm3[5],ymm2[6,7,8],ymm3[9],ymm2[10,11,12],ymm3[13],ymm2[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm2[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm4 = xmm4[8],mem[8],xmm4[9],mem[9],xmm4[10],mem[10],xmm4[11],mem[11],xmm4[12],mem[12],xmm4[13],mem[13],xmm4[14],mem[14],xmm4[15],mem[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm4[0,1,1,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm4[2,1,3,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0,1,2],ymm3[3],ymm5[4,5,6],ymm3[7],ymm5[8,9,10],ymm3[11],ymm5[12,13,14],ymm3[15]
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm5[1,1,1,1]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm7, %ymm6
; AVX2-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-SLOW-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
; AVX2-SLOW-NEXT:    # xmm7 = xmm7[8],mem[8],xmm7[9],mem[9],xmm7[10],mem[10],xmm7[11],mem[11],xmm7[12],mem[12],xmm7[13],mem[13],xmm7[14],mem[14],xmm7[15],mem[15]
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm7[0,0,2,1,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero
; AVX2-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm7[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT:    vpmovzxdq {{.*#+}} xmm10 = xmm10[0],zero,xmm10[1],zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm9, %ymm9
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm6 = ymm6[0],ymm9[1],ymm6[2,3,4],ymm9[5],ymm6[6,7,8],ymm9[9],ymm6[10,11,12],ymm9[13],ymm6[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2],ymm3[3],ymm6[4],ymm3[5],ymm6[6],ymm3[7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm2[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm6, %ymm2
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,4,5,5,7]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm6, %ymm4
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm5[2,3,2,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[3,3,3,3]
; AVX2-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm4, %ymm4
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm7[0,1,2,3,4,4,6,5]
; AVX2-SLOW-NEXT:    vpshufhw {{.*#+}} xmm6 = xmm7[0,1,2,3,4,6,6,7]
; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm5, %ymm5
; AVX2-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm5[2,1,3,3,6,5,7,7]
; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3,4],ymm5[5],ymm4[6,7,8],ymm5[9],ymm4[10,11,12],ymm5[13],ymm4[14,15]
; AVX2-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2],ymm2[3],ymm4[4],ymm2[5],ymm4[6],ymm2[7]
; AVX2-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT:    vmovdqa %ymm2, 96(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm3, 64(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm1, 160(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm0, 128(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm8, 224(%rax)
; AVX2-SLOW-NEXT:    vmovdqa %ymm11, 192(%rax)
; AVX2-SLOW-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 288(%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, 352(%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, 416(%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, 480(%rax)
; AVX2-SLOW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT:    vmovaps %ymm0, 448(%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, 32(%rax)
; AVX2-SLOW-NEXT:    addq $328, %rsp # imm = 0x148
; AVX2-SLOW-NEXT:    vzeroupper
; AVX2-SLOW-NEXT:    retq
;
; AVX2-FAST-LABEL: store_i8_stride8_vf64:
; AVX2-FAST:       # %bb.0:
; AVX2-FAST-NEXT:    subq $392, %rsp # imm = 0x188
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FAST-NEXT:    vmovdqa (%r10), %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rax), %xmm0
; AVX2-FAST-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31]
; AVX2-FAST-NEXT:    vmovdqa (%r9), %xmm3
; AVX2-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%r8), %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX2-FAST-NEXT:    vmovdqa (%rcx), %xmm4
; AVX2-FAST-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rdx), %xmm2
; AVX2-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vmovdqa (%rsi), %xmm5
; AVX2-FAST-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vmovdqa (%rdi), %xmm4
; AVX2-FAST-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm5
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vmovdqa %ymm3, %ymm9
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqa %ymm3, %ymm10
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm3 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %xmm3, %xmm4, %xmm1
; AVX2-FAST-NEXT:    vmovdqa %xmm3, %xmm13
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm3, %ymm1
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 48(%r10), %xmm8
; AVX2-FAST-NEXT:    vmovdqa 48(%rax), %xmm3
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm8[8],xmm3[9],xmm8[9],xmm3[10],xmm8[10],xmm3[11],xmm8[11],xmm3[12],xmm8[12],xmm3[13],xmm8[13],xmm3[14],xmm8[14],xmm3[15],xmm8[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm6
; AVX2-FAST-NEXT:    vmovdqa 48(%r9), %xmm4
; AVX2-FAST-NEXT:    vmovdqa 48(%r8), %xmm5
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm7
; AVX2-FAST-NEXT:    vpshufb %ymm9, %ymm6, %ymm0
; AVX2-FAST-NEXT:    vpshufb %ymm10, %ymm7, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm11 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX2-FAST-NEXT:    vmovdqa 48(%rsi), %xmm10
; AVX2-FAST-NEXT:    vmovdqa 48(%rdi), %xmm9
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm9[8],xmm10[8],xmm9[9],xmm10[9],xmm9[10],xmm10[10],xmm9[11],xmm10[11],xmm9[12],xmm10[12],xmm9[13],xmm10[13],xmm9[14],xmm10[14],xmm9[15],xmm10[15]
; AVX2-FAST-NEXT:    vpshufb %xmm13, %xmm0, %xmm1
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm13, %ymm1
; AVX2-FAST-NEXT:    vmovdqa 48(%rcx), %xmm13
; AVX2-FAST-NEXT:    vmovdqa 48(%rdx), %xmm2
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm2[8],xmm13[8],xmm2[9],xmm13[9],xmm2[10],xmm13[10],xmm2[11],xmm13[11],xmm2[12],xmm13[12],xmm2[13],xmm13[13],xmm2[14],xmm13[14],xmm2[15],xmm13[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm15, %ymm14
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4],ymm14[5],ymm1[6,7,8],ymm14[9],ymm1[10,11,12],ymm14[13],ymm1[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2],ymm11[3],ymm1[4],ymm11[5],ymm1[6],ymm11[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm6, %ymm1
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm7, %ymm6
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3],ymm6[4,5,6],ymm1[7],ymm6[8,9,10],ymm1[11],ymm6[12,13,14],ymm1[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm15, %ymm6
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm14 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %ymm14, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[2,3,4],ymm6[5],ymm0[6,7,8],ymm6[9],ymm0[10,11,12],ymm6[13],ymm0[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm0, %ymm3
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm8, %ymm1, %ymm4
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1],xmm2[2],xmm13[2],xmm2[3],xmm13[3],xmm2[4],xmm13[4],xmm2[5],xmm13[5],xmm2[6],xmm13[6],xmm2[7],xmm13[7]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm10 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %xmm10, %xmm4, %xmm5
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm6, %ymm5
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm2
; AVX2-FAST-NEXT:    vpshufb %ymm14, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 32(%r10), %xmm1
; AVX2-FAST-NEXT:    vmovdqa 32(%rax), %xmm3
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm6
; AVX2-FAST-NEXT:    vmovdqa 32(%r9), %xmm4
; AVX2-FAST-NEXT:    vmovdqa 32(%r8), %xmm5
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm7
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm6, %ymm0
; AVX2-FAST-NEXT:    vpshufb %ymm8, %ymm7, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm8 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX2-FAST-NEXT:    vmovdqa 32(%rsi), %xmm0
; AVX2-FAST-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
; AVX2-FAST-NEXT:    vpshufb %xmm10, %xmm9, %xmm10
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm11 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm10, %ymm11, %ymm10
; AVX2-FAST-NEXT:    vmovdqa 32(%rcx), %xmm11
; AVX2-FAST-NEXT:    vmovdqa 32(%rdx), %xmm13
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm14, %ymm14, %ymm14
; AVX2-FAST-NEXT:    vpshufb {{.*#+}} ymm15 = ymm14[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31]
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0],ymm15[1],ymm10[2,3,4],ymm15[5],ymm10[6,7,8],ymm15[9],ymm10[10,11,12],ymm15[13],ymm10[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm8 = ymm10[0],ymm8[1],ymm10[2],ymm8[3],ymm10[4],ymm8[5],ymm10[6],ymm8[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa %ymm12, %ymm10
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm6, %ymm6
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm7, %ymm7
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5,6],ymm6[7],ymm7[8,9,10],ymm6[11],ymm7[12,13,14],ymm6[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm14, %ymm7
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm9, %ymm9, %ymm8
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm14 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %ymm14, %ymm8, %ymm8
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2,3,4],ymm7[5],ymm8[6,7,8],ymm7[9],ymm8[10,11,12],ymm7[13],ymm8[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm9 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm9, %ymm1, %ymm4
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm3, %ymm5
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm11 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %xmm11, %xmm0, %xmm2
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm6, %ymm2
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm5, %ymm6
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7,8],ymm6[9],ymm2[10,11,12],ymm6[13],ymm2[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[3],ymm2[4],ymm4[5],ymm2[6],ymm4[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vpshufb %ymm10, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm5, %ymm2
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpshufb %ymm14, %ymm0, %ymm0
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa 16(%r10), %xmm4
; AVX2-FAST-NEXT:    vmovdqa 16(%rax), %xmm2
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm6
; AVX2-FAST-NEXT:    vmovdqa 16(%r9), %xmm1
; AVX2-FAST-NEXT:    vmovdqa 16(%r8), %xmm0
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm8
; AVX2-FAST-NEXT:    vpshufb %ymm9, %ymm6, %ymm3
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm8, %ymm5
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm7 = ymm5[0,1,2],ymm3[3],ymm5[4,5,6],ymm3[7],ymm5[8,9,10],ymm3[11],ymm5[12,13,14],ymm3[15]
; AVX2-FAST-NEXT:    vmovdqa 16(%rsi), %xmm5
; AVX2-FAST-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
; AVX2-FAST-NEXT:    vpshufb %xmm11, %xmm9, %xmm10
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm10, %ymm13, %ymm10
; AVX2-FAST-NEXT:    vmovdqa 16(%rcx), %xmm13
; AVX2-FAST-NEXT:    vmovdqa 16(%rdx), %xmm14
; AVX2-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm11 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm15, %ymm11
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2,3,4],ymm11[5],ymm10[6,7,8],ymm11[9],ymm10[10,11,12],ymm11[13],ymm10[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2],ymm7[3],ymm10[4],ymm7[5],ymm10[6],ymm7[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm6, %ymm6
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm8, %ymm8
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm11 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm15, %ymm8
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm9, %ymm9, %ymm9
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm10 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7]
; AVX2-FAST-NEXT:    vmovdqu %ymm6, (%rsp) # 32-byte Spill
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm1
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm2
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm1, %ymm0
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm9 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-NEXT:    vpshufb %ymm9, %ymm2, %ymm4
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6],ymm0[7],ymm4[8,9,10],ymm0[11],ymm4[12,13,14],ymm0[15]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX2-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} xmm14 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-NEXT:    vpshufb %xmm14, %xmm3, %xmm5
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm8, %ymm5
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm4
; AVX2-FAST-NEXT:    vmovdqa {{.*#+}} ymm13 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-NEXT:    vpshufb %ymm13, %ymm4, %ymm8
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm8[1],ymm5[2,3,4],ymm8[5],ymm5[6,7,8],ymm8[9],ymm5[10,11,12],ymm8[13],ymm5[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm1, %ymm1
; AVX2-FAST-NEXT:    vmovdqa %ymm7, %ymm15
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15]
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm4, %ymm2
; AVX2-FAST-NEXT:    vmovdqa %ymm11, %ymm7
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vmovdqa %ymm10, %ymm11
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7,8],ymm2[9],ymm3[10,11,12],ymm2[13],ymm3[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm3 = xmm3[8],mem[8],xmm3[9],mem[9],xmm3[10],mem[10],xmm3[11],mem[11],xmm3[12],mem[12],xmm3[13],mem[13],xmm3[14],mem[14],xmm3[15],mem[15]
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm4
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vpshufb %ymm9, %ymm3, %ymm5
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15]
; AVX2-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX2-FAST-NEXT:    # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15]
; AVX2-FAST-NEXT:    vpshufb %xmm14, %xmm8, %xmm9
; AVX2-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm9, %ymm10, %ymm9
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX2-FAST-NEXT:    vpshufb %ymm13, %ymm5, %ymm10
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3,4],ymm10[5],ymm9[6,7,8],ymm10[9],ymm9[10,11,12],ymm10[13],ymm9[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0],ymm4[1],ymm9[2],ymm4[3],ymm9[4],ymm4[5],ymm9[6],ymm4[7]
; AVX2-FAST-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
; AVX2-FAST-NEXT:    vpshufb %ymm15, %ymm3, %ymm3
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX2-FAST-NEXT:    vpshufb %ymm7, %ymm5, %ymm3
; AVX2-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm8, %ymm5
; AVX2-FAST-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
; AVX2-FAST-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7,8],ymm3[9],ymm5[10,11,12],ymm3[13],ymm5[14,15]
; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7]
; AVX2-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT:    vmovdqa %ymm2, 96(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm4, 64(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm1, 160(%rax)
; AVX2-FAST-NEXT:    vmovdqa %ymm0, 128(%rax)
; AVX2-FAST-NEXT:    vmovups (%rsp), %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, 192(%rax)
; AVX2-FAST-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT:    vmovaps %ymm0, 288(%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, 352(%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, 416(%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, 480(%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, (%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 $392, %rsp # imm = 0x188
; AVX2-FAST-NEXT:    vzeroupper
; AVX2-FAST-NEXT:    retq
;
; AVX2-FAST-PERLANE-LABEL: store_i8_stride8_vf64:
; AVX2-FAST-PERLANE:       # %bb.0:
; AVX2-FAST-PERLANE-NEXT:    subq $392, %rsp # imm = 0x188
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%r10), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rax), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31]
; 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 (%r8), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rcx), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdx), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rsi), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm5 = ymm5[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm3, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm3, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm3 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm3, %xmm4, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %xmm3, %xmm13
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm3, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm12 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%r10), %xmm8
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%rax), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm8[8],xmm3[9],xmm8[9],xmm3[10],xmm8[10],xmm3[11],xmm8[11],xmm3[12],xmm8[12],xmm3[13],xmm8[13],xmm3[14],xmm8[14],xmm3[15],xmm8[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%r9), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%r8), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm9, %ymm6, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm10, %ymm7, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm11 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%rsi), %xmm10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%rdi), %xmm9
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm9[8],xmm10[8],xmm9[9],xmm10[9],xmm9[10],xmm10[10],xmm9[11],xmm10[11],xmm9[12],xmm10[12],xmm9[13],xmm10[13],xmm9[14],xmm10[14],xmm9[15],xmm10[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm13, %xmm0, %xmm1
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm13, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%rcx), %xmm13
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 48(%rdx), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm2[8],xmm13[8],xmm2[9],xmm13[9],xmm2[10],xmm13[10],xmm2[11],xmm13[11],xmm2[12],xmm13[12],xmm2[13],xmm13[13],xmm2[14],xmm13[14],xmm2[15],xmm13[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm15, %ymm14
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4],ymm14[5],ymm1[6,7,8],ymm14[9],ymm1[10,11,12],ymm14[13],ymm1[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2],ymm11[3],ymm1[4],ymm11[5],ymm1[6],ymm11[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm6, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm7, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3],ymm6[4,5,6],ymm1[7],ymm6[8,9,10],ymm1[11],ymm6[12,13,14],ymm1[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm7 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm15, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm14 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm14, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[2,3,4],ymm6[5],ymm0[6,7,8],ymm6[9],ymm0[10,11,12],ymm6[13],ymm0[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm15, %ymm0, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm8, %ymm1, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1],xmm2[2],xmm13[2],xmm2[3],xmm13[3],xmm2[4],xmm13[4],xmm2[5],xmm13[5],xmm2[6],xmm13[6],xmm2[7],xmm13[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm10 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm10, %xmm4, %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm6, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm6 = ymm2[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm14, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r10), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rax), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r9), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%r8), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm15, %ymm6, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm8, %ymm7, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm8 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rsi), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm10, %xmm9, %xmm10
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm11 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm10, %ymm11, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rcx), %xmm11
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 32(%rdx), %xmm13
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm14, %ymm14, %ymm14
; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm15 = ymm14[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31]
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0],ymm15[1],ymm10[2,3,4],ymm15[5],ymm10[6,7,8],ymm15[9],ymm10[10,11,12],ymm15[13],ymm10[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm8 = ymm10[0],ymm8[1],ymm10[2],ymm8[3],ymm10[4],ymm8[5],ymm10[6],ymm8[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm12, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm6, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm15 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm15, %ymm7, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5,6],ymm6[7],ymm7[8,9,10],ymm6[11],ymm7[12,13,14],ymm6[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm12 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm14, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm9, %ymm9, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm14 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm14, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2,3,4],ymm7[5],ymm8[6,7,8],ymm7[9],ymm8[10,11,12],ymm7[13],ymm8[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm9 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm9, %ymm1, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm3, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm11 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm11, %xmm0, %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm2, %ymm6, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm5, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7,8],ymm6[9],ymm2[10,11,12],ymm6[13],ymm2[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[3],ymm2[4],ymm4[5],ymm2[6],ymm4[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm10, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm15, %ymm3, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm5, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm14, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%r10), %xmm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rax), %xmm2
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%r9), %xmm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%r8), %xmm0
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm9, %ymm6, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm8, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm7 = ymm5[0,1,2],ymm3[3],ymm5[4,5,6],ymm3[7],ymm5[8,9,10],ymm3[11],ymm5[12,13,14],ymm3[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rsi), %xmm5
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm9 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm11, %xmm9, %xmm10
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm10, %ymm13, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rcx), %xmm13
; AVX2-FAST-PERLANE-NEXT:    vmovdqa 16(%rdx), %xmm14
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm11 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm15, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2,3,4],ymm11[5],ymm10[6,7,8],ymm11[9],ymm10[10,11,12],ymm11[13],ymm10[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2],ymm7[3],ymm10[4],ymm7[5],ymm10[6],ymm7[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm6, %ymm6
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm7 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm8, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm11 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm15, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm9, %ymm9, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm10 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm10, %ymm9, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqu %ymm6, (%rsp) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm9 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm9, %ymm2, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6],ymm0[7],ymm4[8,9,10],ymm0[11],ymm4[12,13,14],ymm0[15]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX2-FAST-PERLANE-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} xmm14 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm14, %xmm3, %xmm5
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm8, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm13 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm13, %ymm4, %ymm8
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm5 = ymm5[0],ymm8[1],ymm5[2,3,4],ymm8[5],ymm5[6,7,8],ymm8[9],ymm5[10,11,12],ymm8[13],ymm5[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm7, %ymm15
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm4, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm11, %ymm7
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm10, %ymm11
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7,8],ymm2[9],ymm3[10,11,12],ymm2[13],ymm3[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm3 = xmm3[8],mem[8],xmm3[9],mem[9],xmm3[10],mem[10],xmm3[11],mem[11],xmm3[12],mem[12],xmm3[13],mem[13],xmm3[14],mem[14],xmm3[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm6, %ymm2, %ymm4
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm9, %ymm3, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT:    # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %xmm14, %xmm8, %xmm9
; AVX2-FAST-PERLANE-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm9, %ymm10, %ymm9
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm13, %ymm5, %ymm10
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3,4],ymm10[5],ymm9[6,7,8],ymm10[9],ymm9[10,11,12],ymm10[13],ymm9[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0],ymm4[1],ymm9[2],ymm4[3],ymm9[4],ymm4[5],ymm9[6],ymm4[7]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm15, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15]
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm7, %ymm5, %ymm3
; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm8, %ymm8, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7,8],ymm3[9],ymm5[10,11,12],ymm3[13],ymm5[14,15]
; AVX2-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7]
; AVX2-FAST-PERLANE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm2, 96(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm4, 64(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm1, 160(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovdqa %ymm0, 128(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups (%rsp), %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, 288(%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, 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, 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, 480(%rax)
; AVX2-FAST-PERLANE-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 448(%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:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT:    addq $392, %rsp # imm = 0x188
; AVX2-FAST-PERLANE-NEXT:    vzeroupper
; AVX2-FAST-PERLANE-NEXT:    retq
;
; AVX512F-SLOW-LABEL: store_i8_stride8_vf64:
; AVX512F-SLOW:       # %bb.0:
; AVX512F-SLOW-NEXT:    subq $648, %rsp # imm = 0x288
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-SLOW-NEXT:    vmovdqa (%rcx), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 32(%rcx), %xmm10
; AVX512F-SLOW-NEXT:    vmovdqa 48(%rcx), %xmm2
; AVX512F-SLOW-NEXT:    vmovdqa (%rdx), %xmm0
; AVX512F-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 48(%rdx), %xmm3
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa (%r10), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 48(%r10), %xmm4
; AVX512F-SLOW-NEXT:    vmovdqa (%rax), %xmm0
; AVX512F-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 48(%rax), %xmm5
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa (%r9), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 48(%r9), %xmm12
; AVX512F-SLOW-NEXT:    vmovdqa (%r8), %xmm0
; AVX512F-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 48(%r8), %xmm13
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm8 = xmm0[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm8
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm9 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm11 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm11, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm0[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm14, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, (%rsp) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm1
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm0[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm0, %ymm14, %ymm28
; AVX512F-SLOW-NEXT:    vmovdqa 32(%r10), %xmm0
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm3[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm14, %ymm2, %ymm11
; AVX512F-SLOW-NEXT:    vmovdqa 32(%rax), %xmm2
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm14 = xmm3[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm3, %ymm14, %ymm31
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm3[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm5, %ymm4, %ymm24
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm3, %ymm4, %ymm22
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm3 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm4 = xmm3[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm3[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm5, %ymm4, %ymm20
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm3, %ymm4, %ymm18
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3],xmm1[4],xmm10[4],xmm1[5],xmm10[5],xmm1[6],xmm10[6],xmm1[7],xmm10[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm3[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm12, %ymm5, %ymm30
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm3, %ymm5, %ymm29
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm3[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm3[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm12, %ymm5, %ymm23
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm3, %ymm5, %ymm21
; AVX512F-SLOW-NEXT:    vmovdqa 32(%r9), %xmm3
; AVX512F-SLOW-NEXT:    vmovdqa 32(%r8), %xmm5
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3],xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm12[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm14 = xmm12[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm14, %ymm13, %ymm19
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm13 = xmm12[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm12, %ymm13, %ymm17
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm10[8],xmm1[9],xmm10[9],xmm1[10],xmm10[10],xmm1[11],xmm10[11],xmm1[12],xmm10[12],xmm1[13],xmm10[13],xmm1[14],xmm10[14],xmm1[15],xmm10[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm1[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm12 = xmm1[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm12, %ymm10, %ymm4
; AVX512F-SLOW-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm1[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm10, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm3[8],xmm5[9],xmm3[9],xmm5[10],xmm3[10],xmm5[11],xmm3[11],xmm5[12],xmm3[12],xmm5[13],xmm3[13],xmm5[14],xmm3[14],xmm5[15],xmm3[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm0[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 16(%rcx), %xmm13
; AVX512F-SLOW-NEXT:    vmovdqa 16(%rdx), %xmm12
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 16(%r10), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa 16(%rax), %xmm2
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm26
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm16
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm1, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa 16(%r9), %xmm15
; AVX512F-SLOW-NEXT:    vmovdqa 16(%r8), %xmm14
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3],xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm7[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm6 = xmm7[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vmovdqa (%rsi), %xmm1
; AVX512F-SLOW-NEXT:    vmovdqa (%rdi), %xmm2
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm2, %xmm25
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm1, %xmm27
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm0[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm0[1,1,1,1]
; AVX512F-SLOW-NEXT:    vmovdqa 48(%rsi), %xmm2
; AVX512F-SLOW-NEXT:    vmovdqa 48(%rdi), %xmm1
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm0[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm3, %ymm6, %ymm3
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm5, %ymm4
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm4, %zmm4
; AVX512F-SLOW-NEXT:    vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm3 = mem[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm5 = mem[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm3, %zmm6
; AVX512F-SLOW-NEXT:    vpbroadcastq {{.*#+}} zmm3 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535]
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm4, %zmm3, %zmm6
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm4 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm5 = mem[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm5 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm8, %zmm5, %zmm5
; AVX512F-SLOW-NEXT:    vpbroadcastq {{.*#+}} zmm8 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0]
; AVX512F-SLOW-NEXT:    vpandnq %zmm4, %zmm8, %zmm4
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm5, %zmm5
; AVX512F-SLOW-NEXT:    movw $-21846, %ax # imm = 0xAAAA
; AVX512F-SLOW-NEXT:    kmovw %eax, %k1
; AVX512F-SLOW-NEXT:    vpord %zmm4, %zmm5, %zmm6 {%k1}
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm10, %ymm0
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512F-SLOW-NEXT:    vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm2 = mem[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm4 = mem[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm9
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm0, %zmm3, %zmm9
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm0 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm2 = mem[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512F-SLOW-NEXT:    vpshufd $96, (%rsp), %ymm2 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm2 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm28[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm4, %zmm2, %zmm2
; AVX512F-SLOW-NEXT:    vpandnq %zmm0, %zmm8, %zmm0
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm2, %zmm2
; AVX512F-SLOW-NEXT:    vpord %zmm0, %zmm2, %zmm9 {%k1}
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm11[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm31[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm11
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm0, %zmm3, %zmm11
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm24[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm22[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm20[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm18[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-SLOW-NEXT:    vpandnq %zmm0, %zmm8, %zmm0
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm1, %zmm1
; AVX512F-SLOW-NEXT:    vpord %zmm0, %zmm1, %zmm11 {%k1}
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm7[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm7[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm12[8],xmm13[8],xmm12[9],xmm13[9],xmm12[10],xmm13[10],xmm12[11],xmm13[11],xmm12[12],xmm13[12],xmm12[13],xmm13[13],xmm12[14],xmm13[14],xmm12[15],xmm13[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm2[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm1, %ymm0, %ymm28
; AVX512F-SLOW-NEXT:    vmovdqa 32(%rsi), %xmm0
; AVX512F-SLOW-NEXT:    vmovdqa 32(%rdi), %xmm1
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm4[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm7 = xmm4[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,1,1,1]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm4, %ymm7, %ymm4
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm5
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm4 = ymm30[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm29[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm4, %zmm4
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm5, %zmm3, %zmm4
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm5 = ymm23[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm21[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm5, %zmm5
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm19[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm17[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm10, %zmm7, %zmm7
; AVX512F-SLOW-NEXT:    vpandnq %zmm5, %zmm8, %zmm5
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm7, %zmm7
; AVX512F-SLOW-NEXT:    vpord %zmm5, %zmm7, %zmm4 {%k1}
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm2, %ymm5, %ymm18
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm26, %xmm2
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm16, %xmm5
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm2[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm7, %ymm5, %ymm21
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm2, %ymm5, %ymm22
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm14[8],xmm15[8],xmm14[9],xmm15[9],xmm14[10],xmm15[10],xmm14[11],xmm15[11],xmm14[12],xmm15[12],xmm14[13],xmm15[13],xmm14[14],xmm15[14],xmm14[15],xmm15[15]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm5[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm7 = xmm5[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm7, %ymm2, %ymm26
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm7 = xmm5[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm5, %ymm7, %ymm19
; AVX512F-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512F-SLOW-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm5 # 16-byte Folded Reload
; AVX512F-SLOW-NEXT:    # xmm5 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm5[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm5[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm13, %ymm10, %ymm14
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm5[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm5, %ymm10, %ymm15
; AVX512F-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512F-SLOW-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm5 # 16-byte Folded Reload
; AVX512F-SLOW-NEXT:    # xmm5 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm5[0,0,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm13 = xmm5[0,2,2,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm13, %ymm10, %ymm17
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm5[0,1,2,3,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm5, %ymm10, %ymm20
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512F-SLOW-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload
; AVX512F-SLOW-NEXT:    # xmm1 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm1[0,1,1,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vpshuflw {{.*#+}} xmm10 = xmm1[2,1,3,3,4,5,6,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm0
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm10 = xmm1[0,1,2,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm10, %ymm1
; AVX512F-SLOW-NEXT:    vmovdqa 16(%rsi), %xmm10
; AVX512F-SLOW-NEXT:    vmovdqa 16(%rdi), %xmm13
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm12 = xmm13[0],xmm10[0],xmm13[1],xmm10[1],xmm13[2],xmm10[2],xmm13[3],xmm10[3],xmm13[4],xmm10[4],xmm13[5],xmm10[5],xmm13[6],xmm10[6],xmm13[7],xmm10[7]
; AVX512F-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm10 = xmm13[8],xmm10[8],xmm13[9],xmm10[9],xmm13[10],xmm10[10],xmm13[11],xmm10[11],xmm13[12],xmm10[12],xmm13[13],xmm10[13],xmm13[14],xmm10[14],xmm13[15],xmm10[15]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm5[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm23 = xmm5[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm23, %ymm13, %ymm13
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm5, %ymm23, %ymm5
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm13, %zmm5, %zmm13
; AVX512F-SLOW-NEXT:    vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm5 = mem[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm23 = mem[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm23, %zmm5, %zmm5
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm13, %zmm3, %zmm5
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm13 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm23 = mem[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm23, %zmm13, %zmm13
; AVX512F-SLOW-NEXT:    vpandnq %zmm13, %zmm8, %zmm13
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm23 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm24 = mem[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm24, %zmm23, %zmm23
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm23, %zmm23
; AVX512F-SLOW-NEXT:    vpord %zmm13, %zmm23, %zmm5 {%k1}
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm13 = xmm12[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm23 = xmm12[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm24 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm12[1,1,1,1]
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm25, %xmm2
; AVX512F-SLOW-NEXT:    vmovdqa64 %xmm27, %xmm7
; AVX512F-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3],xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm23, %ymm13, %ymm13
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm12, %ymm24, %ymm12
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm23 = xmm10[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm24 = xmm10[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm29 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm10 = xmm10[1,1,1,1]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm13, %zmm12, %zmm12
; AVX512F-SLOW-NEXT:    vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm13 = mem[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm30 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm30 = mem[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm30, %zmm13, %zmm13
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm12, %zmm3, %zmm13
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm12 = xmm2[2,3,2,3]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm30 = xmm2[3,3,3,3]
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm31 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm25 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm27 = mem[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm27, %zmm25, %zmm25
; AVX512F-SLOW-NEXT:    vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm27 = mem[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
; AVX512F-SLOW-NEXT:    # ymm16 = mem[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm16, %zmm27, %zmm16
; AVX512F-SLOW-NEXT:    vpandnq %zmm25, %zmm8, %zmm25
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm16, %zmm16
; AVX512F-SLOW-NEXT:    vpord %zmm25, %zmm16, %zmm13 {%k1}
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm16 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm23, %ymm16, %ymm16
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm10, %ymm29, %ymm10
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm16, %zmm10, %zmm10
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm16 = ymm28[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm18 = ymm18[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm18, %zmm16, %zmm16
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm10, %zmm3, %zmm16
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm21[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm18 = ymm22[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm18, %zmm10, %zmm10
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm18 = ymm26[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm19[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm18, %zmm7
; AVX512F-SLOW-NEXT:    vpandnq %zmm10, %zmm8, %zmm10
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm7, %zmm7
; AVX512F-SLOW-NEXT:    vpord %zmm10, %zmm7, %zmm16 {%k1}
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm30[0],zero,zero,zero,xmm30[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti128 $1, %xmm10, %ymm7, %ymm7
; AVX512F-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-SLOW-NEXT:    vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm2, %zmm2
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm7 = ymm14[0,1,1,3,4,5,5,7]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm10 = ymm15[2,1,3,3,6,5,7,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm10, %zmm7, %zmm7
; AVX512F-SLOW-NEXT:    vpternlogq $184, %zmm2, %zmm3, %zmm7
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm2 = ymm17[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm3 = ymm20[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512F-SLOW-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
; AVX512F-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-SLOW-NEXT:    vpandnq %zmm2, %zmm8, %zmm1
; AVX512F-SLOW-NEXT:    vpandq %zmm8, %zmm0, %zmm0
; AVX512F-SLOW-NEXT:    vpord %zmm1, %zmm0, %zmm7 {%k1}
; AVX512F-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm7, (%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm16, 192(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm13, 128(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm5, 320(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm4, 256(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm11, 448(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm9, 384(%rax)
; AVX512F-SLOW-NEXT:    vmovdqa64 %zmm6, 64(%rax)
; AVX512F-SLOW-NEXT:    addq $648, %rsp # imm = 0x288
; AVX512F-SLOW-NEXT:    vzeroupper
; AVX512F-SLOW-NEXT:    retq
;
; AVX512F-FAST-LABEL: store_i8_stride8_vf64:
; AVX512F-FAST:       # %bb.0:
; AVX512F-FAST-NEXT:    subq $392, %rsp # imm = 0x188
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512F-FAST-NEXT:    vmovdqa (%rcx), %xmm2
; AVX512F-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa 48(%rcx), %xmm0
; AVX512F-FAST-NEXT:    vmovdqa (%rdx), %xmm15
; AVX512F-FAST-NEXT:    vmovdqa 48(%rdx), %xmm1
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm15[8],xmm2[8],xmm15[9],xmm2[9],xmm15[10],xmm2[10],xmm15[11],xmm2[11],xmm15[12],xmm2[12],xmm15[13],xmm2[13],xmm15[14],xmm2[14],xmm15[15],xmm2[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15]
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa (%r10), %xmm2
; AVX512F-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa (%rax), %xmm14
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm14[8],xmm2[8],xmm14[9],xmm2[9],xmm14[10],xmm2[10],xmm14[11],xmm2[11],xmm14[12],xmm2[12],xmm14[13],xmm2[13],xmm14[14],xmm2[14],xmm14[15],xmm2[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm10 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7]
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm19
; AVX512F-FAST-NEXT:    vmovdqa (%r9), %xmm3
; AVX512F-FAST-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa (%r8), %xmm2
; AVX512F-FAST-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX512F-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vmovdqa %ymm4, %ymm9
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7]
; AVX512F-FAST-NEXT:    vpshufb %ymm4, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vmovdqa %ymm4, %ymm12
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm21
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa 48(%r10), %xmm2
; AVX512F-FAST-NEXT:    vmovdqa 48(%rax), %xmm3
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm4, %ymm5
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa 48(%r9), %xmm4
; AVX512F-FAST-NEXT:    vmovdqa 48(%r8), %xmm5
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm7, %ymm7, %ymm7
; AVX512F-FAST-NEXT:    vpshufb %ymm9, %ymm7, %ymm13
; AVX512F-FAST-NEXT:    vpshufb %ymm12, %ymm7, %ymm7
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm13, %zmm7, %zmm7
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512F-FAST-NEXT:    vmovdqa 32(%rcx), %xmm0
; AVX512F-FAST-NEXT:    vmovdqa 32(%rdx), %xmm1
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm9, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512F-FAST-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm20
; AVX512F-FAST-NEXT:    vmovdqa 32(%r10), %xmm2
; AVX512F-FAST-NEXT:    vmovdqa 32(%rax), %xmm3
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm4, %ymm5
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm22
; AVX512F-FAST-NEXT:    vmovdqa 32(%r9), %xmm4
; AVX512F-FAST-NEXT:    vmovdqa 32(%r8), %xmm5
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm7, %ymm7, %ymm7
; AVX512F-FAST-NEXT:    vpshufb %ymm9, %ymm7, %ymm13
; AVX512F-FAST-NEXT:    vpshufb %ymm12, %ymm7, %ymm7
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm13, %zmm7, %zmm23
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm24
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm25
; AVX512F-FAST-NEXT:    vmovdqa 16(%rcx), %xmm0
; AVX512F-FAST-NEXT:    vmovdqa 16(%rdx), %xmm1
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm9, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vpshufb %ymm12, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm26
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm2, %ymm3
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm27
; AVX512F-FAST-NEXT:    vmovdqa 16(%r10), %xmm2
; AVX512F-FAST-NEXT:    vmovdqa 16(%rax), %xmm3
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm4, %ymm5
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm4, %ymm4
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm4, %zmm28
; AVX512F-FAST-NEXT:    vmovdqa 16(%r9), %xmm4
; AVX512F-FAST-NEXT:    vmovdqa 16(%r8), %xmm5
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm7 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm7, %ymm7, %ymm7
; AVX512F-FAST-NEXT:    vpshufb %ymm9, %ymm7, %ymm13
; AVX512F-FAST-NEXT:    vpshufb %ymm12, %ymm7, %ymm7
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm13, %zmm7, %zmm29
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm30
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm31
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vpshufb %ymm9, %ymm0, %ymm1
; AVX512F-FAST-NEXT:    vmovdqa64 %ymm9, %ymm17
; AVX512F-FAST-NEXT:    vpshufb %ymm12, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vmovdqa64 %ymm12, %ymm18
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm16
; AVX512F-FAST-NEXT:    vmovdqa (%rsi), %xmm12
; AVX512F-FAST-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm9[8],xmm12[8],xmm9[9],xmm12[9],xmm9[10],xmm12[10],xmm9[11],xmm12[11],xmm9[12],xmm12[12],xmm9[13],xmm12[13],xmm9[14],xmm12[14],xmm9[15],xmm12[15]
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} xmm7 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm0, %xmm2
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vmovdqa {{.*#+}} ymm5 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm0, %ymm0
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm4
; AVX512F-FAST-NEXT:    vpbroadcastq {{.*#+}} zmm3 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535]
; AVX512F-FAST-NEXT:    vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm4 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpbroadcastq {{.*#+}} zmm0 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0]
; AVX512F-FAST-NEXT:    vpandnq %zmm19, %zmm0, %zmm19
; AVX512F-FAST-NEXT:    vpandq %zmm0, %zmm21, %zmm21
; AVX512F-FAST-NEXT:    movw $-21846, %ax # imm = 0xAAAA
; AVX512F-FAST-NEXT:    kmovw %eax, %k1
; AVX512F-FAST-NEXT:    vpord %zmm19, %zmm21, %zmm4 {%k1}
; AVX512F-FAST-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX512F-FAST-NEXT:    # xmm15 = xmm15[0],mem[0],xmm15[1],mem[1],xmm15[2],mem[2],xmm15[3],mem[3],xmm15[4],mem[4],xmm15[5],mem[5],xmm15[6],mem[6],xmm15[7],mem[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX512F-FAST-NEXT:    vpshufb %ymm6, %ymm15, %ymm6
; AVX512F-FAST-NEXT:    vpshufb %ymm8, %ymm15, %ymm8
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm8, %zmm19
; AVX512F-FAST-NEXT:    vmovdqa 48(%rsi), %xmm15
; AVX512F-FAST-NEXT:    vmovdqa 48(%rdi), %xmm8
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm15[0],xmm8[1],xmm15[1],xmm8[2],xmm15[2],xmm8[3],xmm15[3],xmm8[4],xmm15[4],xmm8[5],xmm15[5],xmm8[6],xmm15[6],xmm8[7],xmm15[7]
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm6, %xmm1
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm21 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti32x4 $1, %xmm1, %ymm21, %ymm1
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm6, %ymm6, %ymm6
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm6, %ymm6
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm6, %zmm1, %zmm6
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm15[8],xmm8[9],xmm15[9],xmm8[10],xmm15[10],xmm8[11],xmm15[11],xmm8[12],xmm15[12],xmm8[13],xmm15[13],xmm8[14],xmm15[14],xmm8[15],xmm15[15]
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm15 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm15, %ymm8
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm8, %zmm15
; AVX512F-FAST-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm1 # 16-byte Folded Reload
; AVX512F-FAST-NEXT:    # xmm1 = xmm14[0],mem[0],xmm14[1],mem[1],xmm14[2],mem[2],xmm14[3],mem[3],xmm14[4],mem[4],xmm14[5],mem[5],xmm14[6],mem[6],xmm14[7],mem[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm11, %ymm1, %ymm8
; AVX512F-FAST-NEXT:    vpshufb %ymm10, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm1, %zmm13
; AVX512F-FAST-NEXT:    vmovdqa 32(%rsi), %xmm1
; AVX512F-FAST-NEXT:    vmovdqa 32(%rdi), %xmm8
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm11 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3],xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm11, %xmm14
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm21 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti32x4 $1, %xmm14, %ymm21, %ymm14
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm11, %ymm11, %ymm11
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm11, %ymm11
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm11, %zmm14, %zmm11
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm1[8],xmm8[9],xmm1[9],xmm8[10],xmm1[10],xmm8[11],xmm1[11],xmm8[12],xmm1[12],xmm8[13],xmm1[13],xmm8[14],xmm1[14],xmm8[15],xmm1[15]
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm14 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm14, %ymm8
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm8, %zmm14
; AVX512F-FAST-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512F-FAST-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX512F-FAST-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vmovdqa64 %ymm17, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm2, %ymm1, %ymm8
; AVX512F-FAST-NEXT:    vmovdqa64 %ymm18, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm1, %zmm21
; AVX512F-FAST-NEXT:    vmovdqa 16(%rsi), %xmm8
; AVX512F-FAST-NEXT:    vmovdqa 16(%rdi), %xmm10
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm1, %xmm2
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm17 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti32x4 $1, %xmm2, %ymm17, %ymm2
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm1, %ymm1
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512F-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm10 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm10, %ymm8
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm2, %ymm2
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm8, %zmm2
; AVX512F-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3],xmm9[4],xmm12[4],xmm9[5],xmm12[5],xmm9[6],xmm12[6],xmm9[7],xmm12[7]
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm8, %ymm9
; AVX512F-FAST-NEXT:    vpshufb %ymm5, %ymm9, %ymm5
; AVX512F-FAST-NEXT:    vpshufb %xmm7, %xmm8, %xmm7
; AVX512F-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
; AVX512F-FAST-NEXT:    vinserti128 $1, %xmm7, %ymm8, %ymm7
; AVX512F-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512F-FAST-NEXT:    vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpandnq {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm7 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpandq {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm8 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpord %zmm7, %zmm8, %zmm6 {%k1}
; AVX512F-FAST-NEXT:    vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm15 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpandnq (%rsp), %zmm0, %zmm7 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpandq {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm8 # 64-byte Folded Reload
; AVX512F-FAST-NEXT:    vpord %zmm7, %zmm8, %zmm15 {%k1}
; AVX512F-FAST-NEXT:    vpternlogq $226, %zmm20, %zmm3, %zmm11
; AVX512F-FAST-NEXT:    vpandnq %zmm22, %zmm0, %zmm7
; AVX512F-FAST-NEXT:    vpandq %zmm0, %zmm23, %zmm8
; AVX512F-FAST-NEXT:    vpord %zmm7, %zmm8, %zmm11 {%k1}
; AVX512F-FAST-NEXT:    vpternlogq $226, %zmm24, %zmm3, %zmm14
; AVX512F-FAST-NEXT:    vpandnq %zmm25, %zmm0, %zmm7
; AVX512F-FAST-NEXT:    vpandq %zmm0, %zmm26, %zmm8
; AVX512F-FAST-NEXT:    vpord %zmm7, %zmm8, %zmm14 {%k1}
; AVX512F-FAST-NEXT:    vpternlogq $226, %zmm27, %zmm3, %zmm1
; AVX512F-FAST-NEXT:    vpandnq %zmm28, %zmm0, %zmm7
; AVX512F-FAST-NEXT:    vpandq %zmm0, %zmm29, %zmm8
; AVX512F-FAST-NEXT:    vpord %zmm7, %zmm8, %zmm1 {%k1}
; AVX512F-FAST-NEXT:    vpternlogq $226, %zmm30, %zmm3, %zmm2
; AVX512F-FAST-NEXT:    vpandnq %zmm31, %zmm0, %zmm7
; AVX512F-FAST-NEXT:    vpandq %zmm0, %zmm16, %zmm8
; AVX512F-FAST-NEXT:    vpord %zmm7, %zmm8, %zmm2 {%k1}
; AVX512F-FAST-NEXT:    vpternlogq $226, %zmm19, %zmm3, %zmm5
; AVX512F-FAST-NEXT:    vpandnq %zmm13, %zmm0, %zmm3
; AVX512F-FAST-NEXT:    vpandq %zmm0, %zmm21, %zmm0
; AVX512F-FAST-NEXT:    vpord %zmm3, %zmm0, %zmm5 {%k1}
; AVX512F-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm5, (%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm2, 192(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm1, 128(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm14, 320(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm11, 256(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm15, 448(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm6, 384(%rax)
; AVX512F-FAST-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512F-FAST-NEXT:    addq $392, %rsp # imm = 0x188
; AVX512F-FAST-NEXT:    vzeroupper
; AVX512F-FAST-NEXT:    retq
;
; AVX512BW-SLOW-LABEL: store_i8_stride8_vf64:
; AVX512BW-SLOW:       # %bb.0:
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-SLOW-NEXT:    vmovdqa (%r10), %xmm1
; AVX512BW-SLOW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%r10), %xmm13
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%r10), %xmm22
; AVX512BW-SLOW-NEXT:    vmovdqa64 48(%r10), %xmm19
; AVX512BW-SLOW-NEXT:    vmovdqa (%rax), %xmm0
; AVX512BW-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rax), %xmm16
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%rax), %xmm23
; AVX512BW-SLOW-NEXT:    vmovdqa64 48(%rax), %xmm20
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vmovdqa (%r9), %xmm5
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%r9), %xmm17
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%r9), %xmm25
; AVX512BW-SLOW-NEXT:    vmovdqa64 48(%r9), %xmm21
; AVX512BW-SLOW-NEXT:    vmovdqa (%r8), %xmm7
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%r8), %xmm18
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%r8), %xmm26
; AVX512BW-SLOW-NEXT:    vmovdqa64 48(%r8), %xmm24
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm10 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3,16,17,20,21,20,21,21,23,16,17,22,21,22,21,23,23]
; AVX512BW-SLOW-NEXT:    vpermw %zmm1, %zmm10, %zmm4
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3,16,17,20,20,20,20,22,21,16,17,20,22,20,22,22,23]
; AVX512BW-SLOW-NEXT:    movl $-2004318072, %eax # imm = 0x88888888
; AVX512BW-SLOW-NEXT:    kmovd %eax, %k1
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm9, %zmm4 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa (%rsi), %xmm0
; AVX512BW-SLOW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-SLOW-NEXT:    vmovdqa64 48(%rsi), %xmm28
; AVX512BW-SLOW-NEXT:    vmovdqa (%rdi), %xmm6
; AVX512BW-SLOW-NEXT:    vmovdqa64 48(%rdi), %xmm30
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm0[8],xmm6[9],xmm0[9],xmm6[10],xmm0[10],xmm6[11],xmm0[11],xmm6[12],xmm0[12],xmm6[13],xmm0[13],xmm6[14],xmm0[14],xmm6[15],xmm0[15]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm8 = xmm0[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm8, %ymm1, %ymm1
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm8, %ymm0
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm3
; AVX512BW-SLOW-NEXT:    vmovdqa (%rcx), %xmm8
; AVX512BW-SLOW-NEXT:    vmovdqa 48(%rcx), %xmm1
; AVX512BW-SLOW-NEXT:    vmovdqa (%rdx), %xmm11
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm12 = xmm11[8],xmm8[8],xmm11[9],xmm8[9],xmm11[10],xmm8[10],xmm11[11],xmm8[11],xmm11[12],xmm8[12],xmm11[13],xmm8[13],xmm11[14],xmm8[14],xmm11[15],xmm8[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm12, %zmm12, %zmm14
; AVX512BW-SLOW-NEXT:    vmovdqa64 {{.*#+}} zmm12 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7,20,20,18,19,22,21,22,21,20,22,18,19,22,23,22,23]
; AVX512BW-SLOW-NEXT:    movl $572662306, %eax # imm = 0x22222222
; AVX512BW-SLOW-NEXT:    kmovd %eax, %k2
; AVX512BW-SLOW-NEXT:    vpermw %zmm14, %zmm12, %zmm3 {%k2}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm15
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm14 = xmm24[0],xmm21[0],xmm24[1],xmm21[1],xmm24[2],xmm21[2],xmm24[3],xmm21[3],xmm24[4],xmm21[4],xmm24[5],xmm21[5],xmm24[6],xmm21[6],xmm24[7],xmm21[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm14, %zmm14, %zmm14
; AVX512BW-SLOW-NEXT:    vpermw %zmm14, %zmm10, %zmm14
; AVX512BW-SLOW-NEXT:    vpermw %zmm15, %zmm9, %zmm14 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm15 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm27 = xmm15[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm27 = xmm27[0],zero,zero,zero,xmm27[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm29 = xmm15[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm29 = xmm29[0],zero,zero,zero,xmm29[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm29, %ymm27, %ymm27
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm29 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm15 = xmm15[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm15 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm15, %ymm29, %ymm15
; AVX512BW-SLOW-NEXT:    vmovdqa 48(%rdx), %xmm0
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm27, %zmm15, %zmm15
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm27 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm27, %zmm27, %zmm27
; AVX512BW-SLOW-NEXT:    vpermw %zmm27, %zmm12, %zmm15 {%k2}
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%rsi), %xmm29
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15]
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%rdi), %xmm31
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm24[8],xmm21[8],xmm24[9],xmm21[9],xmm24[10],xmm21[10],xmm24[11],xmm21[11],xmm24[12],xmm21[12],xmm24[13],xmm21[13],xmm24[14],xmm21[14],xmm24[15],xmm21[15]
; AVX512BW-SLOW-NEXT:    vmovdqa64 32(%rcx), %xmm27
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm19, %zmm19, %zmm21
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm20, %zmm20, %zmm19
; AVX512BW-SLOW-NEXT:    vpermw %zmm19, %zmm10, %zmm19
; AVX512BW-SLOW-NEXT:    vpermw %zmm21, %zmm9, %zmm19 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm21 = xmm20[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm24 = xmm20[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm24, %ymm21, %ymm21
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm24 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm20 = xmm20[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm20, %ymm24, %ymm20
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm21, %zmm20, %zmm20
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm12, %zmm20 {%k2}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm26[0],xmm25[0],xmm26[1],xmm25[1],xmm26[2],xmm25[2],xmm26[3],xmm25[3],xmm26[4],xmm25[4],xmm26[5],xmm25[5],xmm26[6],xmm25[6],xmm26[7],xmm25[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vpermw %zmm1, %zmm10, %zmm21
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm9, %zmm21 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm31[0],xmm29[0],xmm31[1],xmm29[1],xmm31[2],xmm29[2],xmm31[3],xmm29[3],xmm31[4],xmm29[4],xmm31[5],xmm29[5],xmm31[6],xmm29[6],xmm31[7],xmm29[7]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm24 = xmm0[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm24, %ymm1, %ymm1
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm24 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm0, %ymm24, %ymm0
; AVX512BW-SLOW-NEXT:    vmovdqa 32(%rdx), %xmm2
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm24
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm27[0],xmm2[1],xmm27[1],xmm2[2],xmm27[2],xmm2[3],xmm27[3],xmm2[4],xmm27[4],xmm2[5],xmm27[5],xmm2[6],xmm27[6],xmm2[7],xmm27[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm12, %zmm24 {%k2}
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rsi), %xmm28
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rdi), %xmm30
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm26[8],xmm25[8],xmm26[9],xmm25[9],xmm26[10],xmm25[10],xmm26[11],xmm25[11],xmm26[12],xmm25[12],xmm26[13],xmm25[13],xmm26[14],xmm25[14],xmm26[15],xmm25[15]
; AVX512BW-SLOW-NEXT:    vmovdqa64 16(%rcx), %xmm25
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vpermw %zmm1, %zmm10, %zmm22
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm9, %zmm22 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm31[8],xmm29[8],xmm31[9],xmm29[9],xmm31[10],xmm29[10],xmm31[11],xmm29[11],xmm31[12],xmm29[12],xmm31[13],xmm29[13],xmm31[14],xmm29[14],xmm31[15],xmm29[15]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm23 = xmm0[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm23, %ymm1, %ymm1
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti32x4 $1, %xmm0, %ymm23, %ymm0
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm23
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm27[8],xmm2[9],xmm27[9],xmm2[10],xmm27[10],xmm2[11],xmm27[11],xmm2[12],xmm27[12],xmm2[13],xmm27[13],xmm2[14],xmm27[14],xmm2[15],xmm27[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm12, %zmm23 {%k2}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm16[0],xmm13[0],xmm16[1],xmm13[1],xmm16[2],xmm13[2],xmm16[3],xmm13[3],xmm16[4],xmm13[4],xmm16[5],xmm13[5],xmm16[6],xmm13[6],xmm16[7],xmm13[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vpermw %zmm1, %zmm10, %zmm26
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm9, %zmm26 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512BW-SLOW-NEXT:    vmovdqa 16(%rdx), %xmm2
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm27
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm25[0],xmm2[1],xmm25[1],xmm2[2],xmm25[2],xmm2[3],xmm25[3],xmm2[4],xmm25[4],xmm2[5],xmm25[5],xmm2[6],xmm25[6],xmm2[7],xmm25[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm12, %zmm27 {%k2}
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm16[8],xmm13[8],xmm16[9],xmm13[9],xmm16[10],xmm13[10],xmm16[11],xmm13[11],xmm16[12],xmm13[12],xmm16[13],xmm13[13],xmm16[14],xmm13[14],xmm16[15],xmm13[15]
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm13
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm1, %zmm1, %zmm0
; AVX512BW-SLOW-NEXT:    vpermw %zmm0, %zmm10, %zmm0
; AVX512BW-SLOW-NEXT:    vpermw %zmm13, %zmm9, %zmm0 {%k1}
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm1 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15]
; AVX512BW-SLOW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
; AVX512BW-SLOW-NEXT:    # xmm13 = xmm13[0],mem[0],xmm13[1],mem[1],xmm13[2],mem[2],xmm13[3],mem[3],xmm13[4],mem[4],xmm13[5],mem[5],xmm13[6],mem[6],xmm13[7],mem[7]
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm5, %zmm5, %zmm5
; AVX512BW-SLOW-NEXT:    vpermw %zmm5, %zmm10, %zmm5
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm13, %zmm13, %zmm7
; AVX512BW-SLOW-NEXT:    vpermw %zmm7, %zmm9, %zmm5 {%k1}
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm7 = xmm1[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm9 = xmm1[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm9, %ymm7, %ymm7
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm9 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm9, %ymm1
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm7, %zmm1, %zmm1
; AVX512BW-SLOW-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm25[8],xmm2[9],xmm25[9],xmm2[10],xmm25[10],xmm2[11],xmm25[11],xmm2[12],xmm25[12],xmm2[13],xmm25[13],xmm2[14],xmm25[14],xmm2[15],xmm25[15]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm2, %zmm2, %zmm2
; AVX512BW-SLOW-NEXT:    vpermw %zmm2, %zmm12, %zmm1 {%k2}
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm2 # 16-byte Folded Reload
; AVX512BW-SLOW-NEXT:    # xmm2 = xmm6[0],mem[0],xmm6[1],mem[1],xmm6[2],mem[2],xmm6[3],mem[3],xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm6 = xmm2[2,3,2,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm7 = xmm2[3,3,3,3]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm7, %ymm6, %ymm6
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm7 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
; AVX512BW-SLOW-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512BW-SLOW-NEXT:    vinserti128 $1, %xmm2, %ymm7, %ymm2
; AVX512BW-SLOW-NEXT:    vinserti64x4 $1, %ymm6, %zmm2, %zmm2
; AVX512BW-SLOW-NEXT:    vpunpcklbw {{.*#+}} xmm6 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
; AVX512BW-SLOW-NEXT:    vinserti32x4 $2, %xmm6, %zmm6, %zmm6
; AVX512BW-SLOW-NEXT:    vpermw %zmm6, %zmm12, %zmm2 {%k2}
; AVX512BW-SLOW-NEXT:    movw $-21846, %ax # imm = 0xAAAA
; AVX512BW-SLOW-NEXT:    kmovd %eax, %k1
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm4, %zmm3 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm14, %zmm15 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm19, %zmm20 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm21, %zmm24 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm22, %zmm23 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm26, %zmm27 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512BW-SLOW-NEXT:    vmovdqa32 %zmm5, %zmm2 {%k1}
; AVX512BW-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm1, 192(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm27, 128(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm23, 320(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm24, 256(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm20, 448(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm15, 384(%rax)
; AVX512BW-SLOW-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512BW-SLOW-NEXT:    vzeroupper
; AVX512BW-SLOW-NEXT:    retq
;
; AVX512BW-FAST-LABEL: store_i8_stride8_vf64:
; AVX512BW-FAST:       # %bb.0:
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FAST-NEXT:    vmovdqa (%r10), %xmm0
; AVX512BW-FAST-NEXT:    vmovdqa 16(%r10), %xmm14
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%r10), %xmm18
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%r10), %xmm17
; AVX512BW-FAST-NEXT:    vmovdqa (%rax), %xmm1
; AVX512BW-FAST-NEXT:    vmovdqa 16(%rax), %xmm15
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%rax), %xmm19
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%rax), %xmm20
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm2
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm2, %zmm2, %zmm3
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm2 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7,0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15]
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm3, %zmm6
; AVX512BW-FAST-NEXT:    vmovdqa (%r9), %xmm3
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%r9), %xmm21
; AVX512BW-FAST-NEXT:    vmovdqa (%r8), %xmm4
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%r8), %xmm22
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm5, %ymm5, %ymm5
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm5, %zmm5, %zmm7
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm5 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7,0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15]
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm7, %zmm16
; AVX512BW-FAST-NEXT:    movl $-2004318072, %eax # imm = 0x88888888
; AVX512BW-FAST-NEXT:    kmovd %eax, %k1
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm6, %zmm16 {%k1}
; AVX512BW-FAST-NEXT:    vmovdqa (%rcx), %xmm6
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%rcx), %xmm23
; AVX512BW-FAST-NEXT:    vmovdqa (%rdx), %xmm7
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm8, %ymm8
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm8, %zmm8, %zmm8
; AVX512BW-FAST-NEXT:    vmovdqa64 {{.*#+}} zmm9 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15,8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15]
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm8, %zmm24
; AVX512BW-FAST-NEXT:    vmovdqa (%rsi), %xmm10
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%rsi), %xmm25
; AVX512BW-FAST-NEXT:    vmovdqa (%rdi), %xmm11
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%rdi), %xmm28
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm8 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm8, %ymm8, %ymm13
; AVX512BW-FAST-NEXT:    vmovdqa {{.*#+}} ymm12 = [8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,12,13,u,u,u,u,u,u,14,15,u,u,u,u,u,u]
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm13, %ymm26
; AVX512BW-FAST-NEXT:    vmovdqa {{.*#+}} xmm13 = [4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u]
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm8, %xmm27
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm27, %ymm8, %ymm8
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm26, %zmm8, %zmm8
; AVX512BW-FAST-NEXT:    movl $572662306, %eax # imm = 0x22222222
; AVX512BW-FAST-NEXT:    kmovd %eax, %k2
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm24, %zmm8 {%k2}
; AVX512BW-FAST-NEXT:    movw $-21846, %ax # imm = 0xAAAA
; AVX512BW-FAST-NEXT:    kmovd %eax, %k3
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm16, %zmm8 {%k3}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm16 = xmm20[0],xmm17[0],xmm20[1],xmm17[1],xmm20[2],xmm17[2],xmm20[3],xmm17[3],xmm20[4],xmm17[4],xmm20[5],xmm17[5],xmm20[6],xmm17[6],xmm20[7],xmm17[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm16, %ymm16, %ymm16
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm16, %zmm16, %zmm16
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm16, %zmm16
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm24 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3],xmm22[4],xmm21[4],xmm22[5],xmm21[5],xmm22[6],xmm21[6],xmm22[7],xmm21[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm24, %ymm24, %ymm24
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm24, %zmm24, %zmm24
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm24, %zmm24
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm16, %zmm24 {%k1}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm16 = xmm28[0],xmm25[0],xmm28[1],xmm25[1],xmm28[2],xmm25[2],xmm28[3],xmm25[3],xmm28[4],xmm25[4],xmm28[5],xmm25[5],xmm28[6],xmm25[6],xmm28[7],xmm25[7]
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm16, %xmm26
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm27 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm26, %ymm27, %ymm26
; AVX512BW-FAST-NEXT:    vmovdqa64 48(%rdx), %xmm30
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm16, %ymm16, %ymm16
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm16, %ymm16
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm16, %zmm26, %zmm16
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm26 = xmm30[0],xmm23[0],xmm30[1],xmm23[1],xmm30[2],xmm23[2],xmm30[3],xmm23[3],xmm30[4],xmm23[4],xmm30[5],xmm23[5],xmm30[6],xmm23[6],xmm30[7],xmm23[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm26, %ymm26, %ymm26
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm26, %zmm26, %zmm26
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm26, %zmm26
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm26, %zmm16 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%r9), %xmm26
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm24, %zmm16 {%k3}
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%r8), %xmm27
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm20[8],xmm17[8],xmm20[9],xmm17[9],xmm20[10],xmm17[10],xmm20[11],xmm17[11],xmm20[12],xmm17[12],xmm20[13],xmm17[13],xmm20[14],xmm17[14],xmm20[15],xmm17[15]
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%rcx), %xmm24
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm20 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%rsi), %xmm21
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm17, %ymm17, %ymm17
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm17, %zmm17, %zmm17
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm17, %zmm17
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm20, %ymm20, %ymm20
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm20, %zmm20, %zmm20
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm20, %zmm20
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm17, %zmm20 {%k1}
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%rdi), %xmm29
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm30[8],xmm23[8],xmm30[9],xmm23[9],xmm30[10],xmm23[10],xmm30[11],xmm23[11],xmm30[12],xmm23[12],xmm30[13],xmm23[13],xmm30[14],xmm23[14],xmm30[15],xmm23[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm17, %ymm17, %ymm17
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm17, %zmm17, %zmm17
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm17, %zmm22
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm17 = xmm28[8],xmm25[8],xmm28[9],xmm25[9],xmm28[10],xmm25[10],xmm28[11],xmm25[11],xmm28[12],xmm25[12],xmm28[13],xmm25[13],xmm28[14],xmm25[14],xmm28[15],xmm25[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm17, %ymm17, %ymm23
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm23, %ymm23
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm17, %xmm25
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm17 = xmm17[0],zero,zero,zero,xmm17[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm25, %ymm17, %ymm17
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm23, %zmm17, %zmm17
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm22, %zmm17 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm20, %zmm17 {%k3}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm20 = xmm19[0],xmm18[0],xmm19[1],xmm18[1],xmm19[2],xmm18[2],xmm19[3],xmm18[3],xmm19[4],xmm18[4],xmm19[5],xmm18[5],xmm19[6],xmm18[6],xmm19[7],xmm18[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm20, %ymm20, %ymm20
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm20, %zmm20, %zmm20
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm20, %zmm20
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm22 = xmm27[0],xmm26[0],xmm27[1],xmm26[1],xmm27[2],xmm26[2],xmm27[3],xmm26[3],xmm27[4],xmm26[4],xmm27[5],xmm26[5],xmm27[6],xmm26[6],xmm27[7],xmm26[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm22, %ymm22, %ymm22
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm22, %zmm22, %zmm22
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm22, %zmm22
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm20, %zmm22 {%k1}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm20 = xmm29[0],xmm21[0],xmm29[1],xmm21[1],xmm29[2],xmm21[2],xmm29[3],xmm21[3],xmm29[4],xmm21[4],xmm29[5],xmm21[5],xmm29[6],xmm21[6],xmm29[7],xmm21[7]
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm20, %xmm23
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm25 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm23, %ymm25, %ymm23
; AVX512BW-FAST-NEXT:    vmovdqa64 32(%rdx), %xmm28
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm20, %ymm20, %ymm20
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm20, %ymm20
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm20, %zmm23, %zmm20
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm23 = xmm28[0],xmm24[0],xmm28[1],xmm24[1],xmm28[2],xmm24[2],xmm28[3],xmm24[3],xmm28[4],xmm24[4],xmm28[5],xmm24[5],xmm28[6],xmm24[6],xmm28[7],xmm24[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm23, %ymm23, %ymm23
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm23, %zmm23, %zmm23
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm23, %zmm23
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm23, %zmm20 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%r9), %xmm23
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm22, %zmm20 {%k3}
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%r8), %xmm25
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm18 = xmm19[8],xmm18[8],xmm19[9],xmm18[9],xmm19[10],xmm18[10],xmm19[11],xmm18[11],xmm19[12],xmm18[12],xmm19[13],xmm18[13],xmm19[14],xmm18[14],xmm19[15],xmm18[15]
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%rcx), %xmm19
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm26 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15]
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%rsi), %xmm22
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm18, %ymm18, %ymm18
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm18, %zmm18, %zmm18
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm18, %zmm18
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm26, %ymm26, %ymm26
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm26, %zmm26, %zmm26
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm26, %zmm27
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm18, %zmm27 {%k1}
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%rdi), %xmm26
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm18 = xmm28[8],xmm24[8],xmm28[9],xmm24[9],xmm28[10],xmm24[10],xmm28[11],xmm24[11],xmm28[12],xmm24[12],xmm28[13],xmm24[13],xmm28[14],xmm24[14],xmm28[15],xmm24[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm18, %ymm18, %ymm18
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm18, %zmm18, %zmm18
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm18, %zmm24
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm18 = xmm29[8],xmm21[8],xmm29[9],xmm21[9],xmm29[10],xmm21[10],xmm29[11],xmm21[11],xmm29[12],xmm21[12],xmm29[13],xmm21[13],xmm29[14],xmm21[14],xmm29[15],xmm21[15]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm18, %ymm18, %ymm21
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm21, %ymm21
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm18, %xmm28
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm18 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm28, %ymm18, %ymm18
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm21, %zmm18, %zmm18
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm24, %zmm18 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm27, %zmm18 {%k3}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm21 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm21, %ymm21, %ymm21
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm21, %zmm21, %zmm21
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm21, %zmm21
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm24 = xmm25[0],xmm23[0],xmm25[1],xmm23[1],xmm25[2],xmm23[2],xmm25[3],xmm23[3],xmm25[4],xmm23[4],xmm25[5],xmm23[5],xmm25[6],xmm23[6],xmm25[7],xmm23[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm24, %ymm24, %ymm24
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm24, %zmm24, %zmm24
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm24, %zmm24
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm21, %zmm24 {%k1}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm21 = xmm26[0],xmm22[0],xmm26[1],xmm22[1],xmm26[2],xmm22[2],xmm26[3],xmm22[3],xmm26[4],xmm22[4],xmm26[5],xmm22[5],xmm26[6],xmm22[6],xmm26[7],xmm22[7]
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm21, %xmm27
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm28 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm27, %ymm28, %ymm27
; AVX512BW-FAST-NEXT:    vmovdqa64 16(%rdx), %xmm28
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm21, %ymm21, %ymm21
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm21, %ymm21
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm21, %zmm27, %zmm21
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm27 = xmm28[0],xmm19[0],xmm28[1],xmm19[1],xmm28[2],xmm19[2],xmm28[3],xmm19[3],xmm28[4],xmm19[4],xmm28[5],xmm19[5],xmm28[6],xmm19[6],xmm28[7],xmm19[7]
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm27, %ymm27, %ymm27
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm27, %zmm27, %zmm27
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm27, %zmm27
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm27, %zmm21 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm24, %zmm21 {%k3}
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15]
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm15 = xmm25[8],xmm23[8],xmm25[9],xmm23[9],xmm25[10],xmm23[10],xmm25[11],xmm23[11],xmm25[12],xmm23[12],xmm25[13],xmm23[13],xmm25[14],xmm23[14],xmm25[15],xmm23[15]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm14, %ymm14, %ymm14
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm14, %zmm14, %zmm14
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm14, %zmm14
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm15, %ymm15, %ymm15
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm15, %zmm15, %zmm15
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm15, %zmm15
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm14, %zmm15 {%k1}
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm14 = xmm28[8],xmm19[8],xmm28[9],xmm19[9],xmm28[10],xmm19[10],xmm28[11],xmm19[11],xmm28[12],xmm19[12],xmm28[13],xmm19[13],xmm28[14],xmm19[14],xmm28[15],xmm19[15]
; AVX512BW-FAST-NEXT:    vpunpckhbw {{.*#+}} xmm19 = xmm26[8],xmm22[8],xmm26[9],xmm22[9],xmm26[10],xmm22[10],xmm26[11],xmm22[11],xmm26[12],xmm22[12],xmm26[13],xmm22[13],xmm26[14],xmm22[14],xmm26[15],xmm22[15]
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm19, %xmm22
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm23 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm22, %ymm23, %ymm22
; AVX512BW-FAST-NEXT:    vinserti32x4 $1, %xmm19, %ymm19, %ymm19
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm19, %ymm19
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm19, %zmm22, %zmm19
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm14, %ymm14, %ymm14
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm14, %zmm14, %zmm14
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm14, %zmm14
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm14, %zmm19 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm15, %zmm19 {%k3}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpshufb %zmm2, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm1, %ymm1
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
; AVX512BW-FAST-NEXT:    vpshufb %zmm5, %zmm1, %zmm1
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm0, %zmm1 {%k1}
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpshufb %zmm9, %zmm0, %zmm0
; AVX512BW-FAST-NEXT:    vpunpcklbw {{.*#+}} xmm2 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm3
; AVX512BW-FAST-NEXT:    vpshufb %ymm12, %ymm3, %ymm3
; AVX512BW-FAST-NEXT:    vpshufb %xmm13, %xmm2, %xmm4
; AVX512BW-FAST-NEXT:    vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
; AVX512BW-FAST-NEXT:    vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512BW-FAST-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512BW-FAST-NEXT:    vmovdqu16 %zmm0, %zmm2 {%k2}
; AVX512BW-FAST-NEXT:    vmovdqa32 %zmm1, %zmm2 {%k3}
; AVX512BW-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm2, (%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm19, 192(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm21, 128(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm18, 320(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm20, 256(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm17, 448(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm16, 384(%rax)
; AVX512BW-FAST-NEXT:    vmovdqa64 %zmm8, 64(%rax)
; AVX512BW-FAST-NEXT:    vzeroupper
; AVX512BW-FAST-NEXT:    retq
  %in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
  %in.vec1 = load <64 x i8>, ptr %in.vecptr1, align 64
  %in.vec2 = load <64 x i8>, ptr %in.vecptr2, align 64
  %in.vec3 = load <64 x i8>, ptr %in.vecptr3, align 64
  %in.vec4 = load <64 x i8>, ptr %in.vecptr4, align 64
  %in.vec5 = load <64 x i8>, ptr %in.vecptr5, align 64
  %in.vec6 = load <64 x i8>, ptr %in.vecptr6, align 64
  %in.vec7 = load <64 x i8>, ptr %in.vecptr7, align 64
  %1 = shufflevector <64 x i8> %in.vec0, <64 x i8> %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 i8> %in.vec2, <64 x i8> %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 i8> %in.vec4, <64 x i8> %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 <64 x i8> %in.vec6, <64 x i8> %in.vec7, <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 <128 x i8> %1, <128 x i8> %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>
  %6 = shufflevector <128 x i8> %3, <128 x i8> %4, <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>
  %7 = shufflevector <256 x i8> %5, <256 x i8> %6, <512 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, i32 384, i32 385, i32 386, i32 387, i32 388, i32 389, i32 390, i32 391, i32 392, i32 393, i32 394, i32 395, i32 396, i32 397, i32 398, i32 399, i32 400, i32 401, i32 402, i32 403, i32 404, i32 405, i32 406, i32 407, i32 408, i32 409, i32 410, i32 411, i32 412, i32 413, i32 414, i32 415, i32 416, i32 417, i32 418, i32 419, i32 420, i32 421, i32 422, i32 423, i32 424, i32 425, i32 426, i32 427, i32 428, i32 429, i32 430, i32 431, i32 432, i32 433, i32 434, i32 435, i32 436, i32 437, i32 438, i32 439, i32 440, i32 441, i32 442, i32 443, i32 444, i32 445, i32 446, i32 447, i32 448, i32 449, i32 450, i32 451, i32 452, i32 453, i32 454, i32 455, i32 456, i32 457, i32 458, i32 459, i32 460, i32 461, i32 462, i32 463, i32 464, i32 465, i32 466, i32 467, i32 468, i32 469, i32 470, i32 471, i32 472, i32 473, i32 474, i32 475, i32 476, i32 477, i32 478, i32 479, i32 480, i32 481, i32 482, i32 483, i32 484, i32 485, i32 486, i32 487, i32 488, i32 489, i32 490, i32 491, i32 492, i32 493, i32 494, i32 495, i32 496, i32 497, i32 498, i32 499, i32 500, i32 501, i32 502, i32 503, i32 504, i32 505, i32 506, i32 507, i32 508, i32 509, i32 510, i32 511>
  %interleaved.vec = shufflevector <512 x i8> %7, <512 x i8> poison, <512 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 320, i32 384, i32 448, i32 1, i32 65, i32 129, i32 193, i32 257, i32 321, i32 385, i32 449, i32 2, i32 66, i32 130, i32 194, i32 258, i32 322, i32 386, i32 450, i32 3, i32 67, i32 131, i32 195, i32 259, i32 323, i32 387, i32 451, i32 4, i32 68, i32 132, i32 196, i32 260, i32 324, i32 388, i32 452, i32 5, i32 69, i32 133, i32 197, i32 261, i32 325, i32 389, i32 453, i32 6, i32 70, i32 134, i32 198, i32 262, i32 326, i32 390, i32 454, i32 7, i32 71, i32 135, i32 199, i32 263, i32 327, i32 391, i32 455, i32 8, i32 72, i32 136, i32 200, i32 264, i32 328, i32 392, i32 456, i32 9, i32 73, i32 137, i32 201, i32 265, i32 329, i32 393, i32 457, i32 10, i32 74, i32 138, i32 202, i32 266, i32 330, i32 394, i32 458, i32 11, i32 75, i32 139, i32 203, i32 267, i32 331, i32 395, i32 459, i32 12, i32 76, i32 140, i32 204, i32 268, i32 332, i32 396, i32 460, i32 13, i32 77, i32 141, i32 205, i32 269, i32 333, i32 397, i32 461, i32 14, i32 78, i32 142, i32 206, i32 270, i32 334, i32 398, i32 462, i32 15, i32 79, i32 143, i32 207, i32 271, i32 335, i32 399, i32 463, i32 16, i32 80, i32 144, i32 208, i32 272, i32 336, i32 400, i32 464, i32 17, i32 81, i32 145, i32 209, i32 273, i32 337, i32 401, i32 465, i32 18, i32 82, i32 146, i32 210, i32 274, i32 338, i32 402, i32 466, i32 19, i32 83, i32 147, i32 211, i32 275, i32 339, i32 403, i32 467, i32 20, i32 84, i32 148, i32 212, i32 276, i32 340, i32 404, i32 468, i32 21, i32 85, i32 149, i32 213, i32 277, i32 341, i32 405, i32 469, i32 22, i32 86, i32 150, i32 214, i32 278, i32 342, i32 406, i32 470, i32 23, i32 87, i32 151, i32 215, i32 279, i32 343, i32 407, i32 471, i32 24, i32 88, i32 152, i32 216, i32 280, i32 344, i32 408, i32 472, i32 25, i32 89, i32 153, i32 217, i32 281, i32 345, i32 409, i32 473, i32 26, i32 90, i32 154, i32 218, i32 282, i32 346, i32 410, i32 474, i32 27, i32 91, i32 155, i32 219, i32 283, i32 347, i32 411, i32 475, i32 28, i32 92, i32 156, i32 220, i32 284, i32 348, i32 412, i32 476, i32 29, i32 93, i32 157, i32 221, i32 285, i32 349, i32 413, i32 477, i32 30, i32 94, i32 158, i32 222, i32 286, i32 350, i32 414, i32 478, i32 31, i32 95, i32 159, i32 223, i32 287, i32 351, i32 415, i32 479, i32 32, i32 96, i32 160, i32 224, i32 288, i32 352, i32 416, i32 480, i32 33, i32 97, i32 161, i32 225, i32 289, i32 353, i32 417, i32 481, i32 34, i32 98, i32 162, i32 226, i32 290, i32 354, i32 418, i32 482, i32 35, i32 99, i32 163, i32 227, i32 291, i32 355, i32 419, i32 483, i32 36, i32 100, i32 164, i32 228, i32 292, i32 356, i32 420, i32 484, i32 37, i32 101, i32 165, i32 229, i32 293, i32 357, i32 421, i32 485, i32 38, i32 102, i32 166, i32 230, i32 294, i32 358, i32 422, i32 486, i32 39, i32 103, i32 167, i32 231, i32 295, i32 359, i32 423, i32 487, i32 40, i32 104, i32 168, i32 232, i32 296, i32 360, i32 424, i32 488, i32 41, i32 105, i32 169, i32 233, i32 297, i32 361, i32 425, i32 489, i32 42, i32 106, i32 170, i32 234, i32 298, i32 362, i32 426, i32 490, i32 43, i32 107, i32 171, i32 235, i32 299, i32 363, i32 427, i32 491, i32 44, i32 108, i32 172, i32 236, i32 300, i32 364, i32 428, i32 492, i32 45, i32 109, i32 173, i32 237, i32 301, i32 365, i32 429, i32 493, i32 46, i32 110, i32 174, i32 238, i32 302, i32 366, i32 430, i32 494, i32 47, i32 111, i32 175, i32 239, i32 303, i32 367, i32 431, i32 495, i32 48, i32 112, i32 176, i32 240, i32 304, i32 368, i32 432, i32 496, i32 49, i32 113, i32 177, i32 241, i32 305, i32 369, i32 433, i32 497, i32 50, i32 114, i32 178, i32 242, i32 306, i32 370, i32 434, i32 498, i32 51, i32 115, i32 179, i32 243, i32 307, i32 371, i32 435, i32 499, i32 52, i32 116, i32 180, i32 244, i32 308, i32 372, i32 436, i32 500, i32 53, i32 117, i32 181, i32 245, i32 309, i32 373, i32 437, i32 501, i32 54, i32 118, i32 182, i32 246, i32 310, i32 374, i32 438, i32 502, i32 55, i32 119, i32 183, i32 247, i32 311, i32 375, i32 439, i32 503, i32 56, i32 120, i32 184, i32 248, i32 312, i32 376, i32 440, i32 504, i32 57, i32 121, i32 185, i32 249, i32 313, i32 377, i32 441, i32 505, i32 58, i32 122, i32 186, i32 250, i32 314, i32 378, i32 442, i32 506, i32 59, i32 123, i32 187, i32 251, i32 315, i32 379, i32 443, i32 507, i32 60, i32 124, i32 188, i32 252, i32 316, i32 380, i32 444, i32 508, i32 61, i32 125, i32 189, i32 253, i32 317, i32 381, i32 445, i32 509, i32 62, i32 126, i32 190, i32 254, i32 318, i32 382, i32 446, i32 510, i32 63, i32 127, i32 191, i32 255, i32 319, i32 383, i32 447, i32 511>
  store <512 x i8> %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:
; AVX1: {{.*}}
; AVX512: {{.*}}
; AVX512-FAST: {{.*}}
; AVX512-SLOW: {{.*}}
; AVX512BW: {{.*}}
; AVX512BW-ONLY: {{.*}}
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512BW-ONLY-SLOW: {{.*}}
; AVX512DQ-ONLY: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512DQBW-ONLY: {{.*}}
; AVX512DQBW-SLOW: {{.*}}
; AVX512F-ONLY: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
; FALLBACK11: {{.*}}
; FALLBACK12: {{.*}}
; FALLBACK2: {{.*}}
; FALLBACK3: {{.*}}
; FALLBACK4: {{.*}}
; FALLBACK5: {{.*}}
; FALLBACK6: {{.*}}
; FALLBACK7: {{.*}}
; FALLBACK8: {{.*}}
; FALLBACK9: {{.*}}