Run DCE after a LoopFlatten test to reduce spurious output [nfc]

[llvm-project.git] / llvm / test / CodeGen / X86 / vector-interleaved-load-i64-stride-6.ll

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

; These patterns are produced by LoopVectorizer for interleaved loads.

define void @load_i64_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {
; SSE-LABEL: load_i64_stride6_vf2:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 64(%rdi), %xmm1
; SSE-NEXT:    movaps (%rdi), %xmm2
; SSE-NEXT:    movaps 16(%rdi), %xmm3
; SSE-NEXT:    movaps 32(%rdi), %xmm4
; SSE-NEXT:    movaps 48(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm2, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]
; SSE-NEXT:    movaps %xmm3, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm1[1]
; SSE-NEXT:    movaps %xmm4, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]
; SSE-NEXT:    movaps %xmm6, (%rsi)
; SSE-NEXT:    movaps %xmm2, (%rdx)
; SSE-NEXT:    movaps %xmm5, (%rcx)
; SSE-NEXT:    movaps %xmm3, (%r8)
; SSE-NEXT:    movaps %xmm1, (%r9)
; SSE-NEXT:    movaps %xmm4, (%rax)
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride6_vf2:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps (%rdi), %xmm0
; AVX-NEXT:    vmovaps 16(%rdi), %xmm1
; AVX-NEXT:    vmovaps 32(%rdi), %xmm2
; AVX-NEXT:    vmovaps 48(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]
; AVX-NEXT:    vmovaps 64(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; AVX-NEXT:    vmovaps %xmm4, (%rsi)
; AVX-NEXT:    vmovaps %xmm0, (%rdx)
; AVX-NEXT:    vmovaps %xmm5, (%rcx)
; AVX-NEXT:    vmovaps %xmm1, (%r8)
; AVX-NEXT:    vmovaps %xmm6, (%r9)
; AVX-NEXT:    vmovaps %xmm2, (%rax)
; AVX-NEXT:    retq
  %wide.vec = load <12 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 0, i32 6>
  %strided.vec1 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 1, i32 7>
  %strided.vec2 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 2, i32 8>
  %strided.vec3 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 3, i32 9>
  %strided.vec4 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 4, i32 10>
  %strided.vec5 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 5, i32 11>
  store <2 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <2 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <2 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <2 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <2 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <2 x i64> %strided.vec5, ptr %out.vec5, align 64
  ret void
}

define void @load_i64_stride6_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {
; SSE-LABEL: load_i64_stride6_vf4:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps 80(%rdi), %xmm3
; SSE-NEXT:    movaps 176(%rdi), %xmm5
; SSE-NEXT:    movaps 128(%rdi), %xmm0
; SSE-NEXT:    movaps 64(%rdi), %xmm6
; SSE-NEXT:    movaps 160(%rdi), %xmm7
; SSE-NEXT:    movaps 112(%rdi), %xmm1
; SSE-NEXT:    movaps (%rdi), %xmm8
; SSE-NEXT:    movaps 16(%rdi), %xmm4
; SSE-NEXT:    movaps 32(%rdi), %xmm2
; SSE-NEXT:    movaps 48(%rdi), %xmm9
; SSE-NEXT:    movaps 144(%rdi), %xmm10
; SSE-NEXT:    movaps 96(%rdi), %xmm11
; SSE-NEXT:    movaps %xmm11, %xmm12
; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm10[0]
; SSE-NEXT:    movaps %xmm8, %xmm13
; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm9[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm10[1]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm9[1]
; SSE-NEXT:    movaps %xmm1, %xmm9
; SSE-NEXT:    movlhps {{.*#+}} xmm9 = xmm9[0],xmm7[0]
; SSE-NEXT:    movaps %xmm4, %xmm10
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm6[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm7[1]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; SSE-NEXT:    movaps %xmm0, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    movaps %xmm2, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm3[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm5[1]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
; SSE-NEXT:    movaps %xmm13, (%rsi)
; SSE-NEXT:    movaps %xmm12, 16(%rsi)
; SSE-NEXT:    movaps %xmm8, (%rdx)
; SSE-NEXT:    movaps %xmm11, 16(%rdx)
; SSE-NEXT:    movaps %xmm10, (%rcx)
; SSE-NEXT:    movaps %xmm9, 16(%rcx)
; SSE-NEXT:    movaps %xmm4, (%r8)
; SSE-NEXT:    movaps %xmm1, 16(%r8)
; SSE-NEXT:    movaps %xmm7, (%r9)
; SSE-NEXT:    movaps %xmm6, 16(%r9)
; SSE-NEXT:    movaps %xmm2, (%rax)
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: load_i64_stride6_vf4:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX1-ONLY-NEXT:    vmovaps 96(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX1-ONLY-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX1-ONLY-NEXT:    vmovaps (%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovaps 16(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps 32(%rdi), %xmm7
; AVX1-ONLY-NEXT:    vmovaps 48(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm5[0],xmm8[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm8[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm3[0],ymm1[2],ymm3[2]
; AVX1-ONLY-NEXT:    vmovaps 64(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm6[0],xmm8[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm6[1],xmm8[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 80(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm7[0],xmm8[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm0[1],ymm3[3],ymm0[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm7[1],xmm8[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovaps %ymm4, (%rsi)
; AVX1-ONLY-NEXT:    vmovaps %ymm2, (%rdx)
; AVX1-ONLY-NEXT:    vmovaps %ymm5, (%rcx)
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%r8)
; AVX1-ONLY-NEXT:    vmovaps %ymm6, (%r9)
; AVX1-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-ONLY-LABEL: load_i64_stride6_vf4:
; AVX2-ONLY:       # %bb.0:
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovaps 128(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovaps (%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps 16(%rdi), %xmm4
; AVX2-ONLY-NEXT:    vmovaps 32(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovaps 48(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm6[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 104(%rdi), %ymm8
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm1[1],ymm8[3],ymm1[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 160(%rdi), %ymm6
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm6[0],ymm2[2],ymm6[2]
; AVX2-ONLY-NEXT:    vmovaps 64(%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm4[0],xmm8[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm8[1]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm0[1],ymm2[3],ymm0[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovaps 80(%rdi), %xmm4
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm4[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 136(%rdi), %ymm8
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm0[1],ymm8[3],ymm0[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovaps %ymm7, (%rsi)
; AVX2-ONLY-NEXT:    vmovaps %ymm3, (%rdx)
; AVX2-ONLY-NEXT:    vmovaps %ymm6, (%rcx)
; AVX2-ONLY-NEXT:    vmovaps %ymm2, (%r8)
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-ONLY-NEXT:    vzeroupper
; AVX2-ONLY-NEXT:    retq
;
; AVX512-SLOW-LABEL: load_i64_stride6_vf4:
; AVX512-SLOW:       # %bb.0:
; AVX512-SLOW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-SLOW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512-SLOW-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512-SLOW-NEXT:    vmovdqa {{.*#+}} ymm0 = <0,6,12,u>
; AVX512-SLOW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0
; AVX512-SLOW-NEXT:    vpbroadcastq 144(%rdi), %ymm1
; AVX512-SLOW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-SLOW-NEXT:    vmovdqa {{.*#+}} ymm1 = <1,7,13,u>
; AVX512-SLOW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1
; AVX512-SLOW-NEXT:    vmovdqa 128(%rdi), %ymm4
; AVX512-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm4[6,7]
; AVX512-SLOW-NEXT:    vmovdqa {{.*#+}} ymm5 = <10,0,6,u>
; AVX512-SLOW-NEXT:    vpermi2q %zmm2, %zmm3, %zmm5
; AVX512-SLOW-NEXT:    vmovdqa 160(%rdi), %xmm6
; AVX512-SLOW-NEXT:    vpbroadcastq %xmm6, %ymm7
; AVX512-SLOW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]
; AVX512-SLOW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512-SLOW-NEXT:    vmovdqa {{.*#+}} ymm7 = <11,1,7,u>
; AVX512-SLOW-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7
; AVX512-SLOW-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
; AVX512-SLOW-NEXT:    vmovdqa 160(%rdi), %ymm7
; AVX512-SLOW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]
; AVX512-SLOW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,3]
; AVX512-SLOW-NEXT:    vmovdqa {{.*#+}} xmm8 = [4,10]
; AVX512-SLOW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8
; AVX512-SLOW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]
; AVX512-SLOW-NEXT:    vpbroadcastq 136(%rdi), %ymm8
; AVX512-SLOW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm8[1],ymm7[1],ymm8[3],ymm7[3]
; AVX512-SLOW-NEXT:    vmovdqa {{.*#+}} xmm8 = [5,11]
; AVX512-SLOW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8
; AVX512-SLOW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm7[4,5,6,7]
; AVX512-SLOW-NEXT:    vmovdqa %ymm0, (%rsi)
; AVX512-SLOW-NEXT:    vmovdqa %ymm1, (%rdx)
; AVX512-SLOW-NEXT:    vmovdqa %ymm5, (%rcx)
; AVX512-SLOW-NEXT:    vmovdqa %ymm6, (%r8)
; AVX512-SLOW-NEXT:    vmovdqa %ymm4, (%r9)
; AVX512-SLOW-NEXT:    vmovdqa %ymm2, (%rax)
; AVX512-SLOW-NEXT:    vzeroupper
; AVX512-SLOW-NEXT:    retq
;
; AVX512-FAST-LABEL: load_i64_stride6_vf4:
; AVX512-FAST:       # %bb.0:
; AVX512-FAST-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FAST-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512-FAST-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} ymm0 = <0,6,12,u>
; AVX512-FAST-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0
; AVX512-FAST-NEXT:    vpbroadcastq 144(%rdi), %ymm1
; AVX512-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} ymm1 = <1,7,13,u>
; AVX512-FAST-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1
; AVX512-FAST-NEXT:    vmovdqa 128(%rdi), %ymm4
; AVX512-FAST-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm4[6,7]
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} ymm5 = <10,0,6,u>
; AVX512-FAST-NEXT:    vpermi2q %zmm2, %zmm3, %zmm5
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,2,4]
; AVX512-FAST-NEXT:    vmovdqa 160(%rdi), %ymm7
; AVX512-FAST-NEXT:    vpermi2q %ymm7, %ymm5, %ymm6
; AVX512-FAST-NEXT:    vinserti128 $1, 160(%rdi), %ymm0, %ymm5
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} ymm8 = <11,1,7,u>
; AVX512-FAST-NEXT:    vpermi2q %zmm2, %zmm3, %zmm8
; AVX512-FAST-NEXT:    vpblendd {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]
; AVX512-FAST-NEXT:    vbroadcasti128 {{.*#+}} ymm8 = [0,6,0,6]
; AVX512-FAST-NEXT:    # ymm8 = mem[0,1,0,1]
; AVX512-FAST-NEXT:    vpermi2q %ymm7, %ymm4, %ymm8
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} xmm4 = [4,10]
; AVX512-FAST-NEXT:    vpermi2q %zmm3, %zmm2, %zmm4
; AVX512-FAST-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]
; AVX512-FAST-NEXT:    vpbroadcastq 136(%rdi), %ymm8
; AVX512-FAST-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm8[1],ymm7[1],ymm8[3],ymm7[3]
; AVX512-FAST-NEXT:    vmovdqa {{.*#+}} xmm8 = [5,11]
; AVX512-FAST-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8
; AVX512-FAST-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm7[4,5,6,7]
; AVX512-FAST-NEXT:    vmovdqa %ymm0, (%rsi)
; AVX512-FAST-NEXT:    vmovdqa %ymm1, (%rdx)
; AVX512-FAST-NEXT:    vmovdqa %ymm6, (%rcx)
; AVX512-FAST-NEXT:    vmovdqa %ymm5, (%r8)
; AVX512-FAST-NEXT:    vmovdqa %ymm4, (%r9)
; AVX512-FAST-NEXT:    vmovdqa %ymm2, (%rax)
; AVX512-FAST-NEXT:    vzeroupper
; AVX512-FAST-NEXT:    retq
  %wide.vec = load <24 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 0, i32 6, i32 12, i32 18>
  %strided.vec1 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 1, i32 7, i32 13, i32 19>
  %strided.vec2 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 2, i32 8, i32 14, i32 20>
  %strided.vec3 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 3, i32 9, i32 15, i32 21>
  %strided.vec4 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 4, i32 10, i32 16, i32 22>
  %strided.vec5 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 5, i32 11, i32 17, i32 23>
  store <4 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <4 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <4 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <4 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <4 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <4 x i64> %strided.vec5, ptr %out.vec5, align 64
  ret void
}

define void @load_i64_stride6_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {
; SSE-LABEL: load_i64_stride6_vf8:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $24, %rsp
; SSE-NEXT:    movaps 160(%rdi), %xmm10
; SSE-NEXT:    movaps 256(%rdi), %xmm13
; SSE-NEXT:    movaps 208(%rdi), %xmm5
; SSE-NEXT:    movaps 352(%rdi), %xmm15
; SSE-NEXT:    movaps 304(%rdi), %xmm6
; SSE-NEXT:    movaps 64(%rdi), %xmm0
; SSE-NEXT:    movaps (%rdi), %xmm8
; SSE-NEXT:    movaps 16(%rdi), %xmm7
; SSE-NEXT:    movaps 48(%rdi), %xmm1
; SSE-NEXT:    movaps 144(%rdi), %xmm2
; SSE-NEXT:    movaps 96(%rdi), %xmm11
; SSE-NEXT:    movaps 240(%rdi), %xmm3
; SSE-NEXT:    movaps 192(%rdi), %xmm12
; SSE-NEXT:    movaps 336(%rdi), %xmm4
; SSE-NEXT:    movaps 288(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm14
; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm4[0]
; SSE-NEXT:    movaps %xmm14, (%rsp) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm4[1]
; SSE-NEXT:    movaps %xmm12, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm3[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm2[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm2[1]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm12
; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm1[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm7, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm6, %xmm14
; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm15[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm15[1]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm5, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm13[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm13[1]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 112(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm10[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm10[1]
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm10
; SSE-NEXT:    movaps %xmm10, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 320(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 272(%rdi), %xmm0
; SSE-NEXT:    movaps 224(%rdi), %xmm3
; SSE-NEXT:    movaps %xmm3, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
; SSE-NEXT:    movaps 176(%rdi), %xmm0
; SSE-NEXT:    movaps 128(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm11, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps %xmm12, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps %xmm9, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps %xmm7, 16(%rcx)
; SSE-NEXT:    movaps %xmm15, 32(%rcx)
; SSE-NEXT:    movaps %xmm14, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps %xmm13, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps %xmm2, 16(%r9)
; SSE-NEXT:    movaps %xmm4, 32(%r9)
; SSE-NEXT:    movaps %xmm6, 48(%r9)
; SSE-NEXT:    movaps %xmm8, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm1, 16(%rax)
; SSE-NEXT:    movaps %xmm3, 32(%rax)
; SSE-NEXT:    movaps %xmm5, 48(%rax)
; SSE-NEXT:    movaps %xmm10, (%rax)
; SSE-NEXT:    addq $24, %rsp
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: load_i64_stride6_vf8:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX1-ONLY-NEXT:    vmovaps 288(%rdi), %ymm4
; AVX1-ONLY-NEXT:    vmovaps 96(%rdi), %ymm7
; AVX1-ONLY-NEXT:    vmovaps 320(%rdi), %ymm5
; AVX1-ONLY-NEXT:    vmovaps 128(%rdi), %ymm3
; AVX1-ONLY-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm6
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm6[0],ymm3[0],ymm6[2],ymm3[2]
; AVX1-ONLY-NEXT:    vmovaps (%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovaps 16(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovaps 48(%rdi), %xmm10
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm9[0],xmm10[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm11
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX1-ONLY-NEXT:    vmovaps 240(%rdi), %xmm12
; AVX1-ONLY-NEXT:    vmovaps 192(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm13[0],xmm12[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm6[1],ymm3[1],ymm6[3],ymm3[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm6 = xmm9[1],xmm10[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm6 = xmm13[1],xmm12[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm10
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm7[0],ymm10[0],ymm7[2],ymm10[2]
; AVX1-ONLY-NEXT:    vmovaps 64(%rdi), %xmm11
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm11[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm12
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm4[0],ymm12[0],ymm4[2],ymm12[2]
; AVX1-ONLY-NEXT:    vmovaps 256(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovaps 208(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm13[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3],ymm9[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovaps 160(%rdi), %ymm15
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm10[1],ymm7[3],ymm10[3]
; AVX1-ONLY-NEXT:    vmovaps 32(%rdi), %xmm10
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm8[1],xmm11[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm4[1],ymm12[1],ymm4[3],ymm12[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm14[1],xmm13[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm8
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm8[0],ymm15[0],ymm8[2],ymm15[2]
; AVX1-ONLY-NEXT:    vmovaps 80(%rdi), %xmm12
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm10[0],xmm12[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm11 = ymm13[0,1,2,3],ymm11[4,5,6,7]
; AVX1-ONLY-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm13
; AVX1-ONLY-NEXT:    vmovaps %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm13[0],ymm0[0],ymm13[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 272(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps 224(%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm15[1],ymm8[3],ymm15[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm10 = xmm10[1],xmm12[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm10 = ymm13[1],ymm3[1],ymm13[3],ymm3[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm10[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rsi)
; AVX1-ONLY-NEXT:    vmovaps %ymm5, 32(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rdx)
; AVX1-ONLY-NEXT:    vmovaps %ymm9, 32(%rcx)
; AVX1-ONLY-NEXT:    vmovaps %ymm6, (%rcx)
; AVX1-ONLY-NEXT:    vmovaps %ymm4, 32(%r8)
; AVX1-ONLY-NEXT:    vmovaps %ymm7, (%r8)
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%r9)
; AVX1-ONLY-NEXT:    vmovaps %ymm11, (%r9)
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm8, (%rax)
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-ONLY-LABEL: load_i64_stride6_vf8:
; AVX2-ONLY:       # %bb.0:
; AVX2-ONLY-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovaps 288(%rdi), %ymm7
; AVX2-ONLY-NEXT:    vmovaps 128(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovaps 96(%rdi), %ymm9
; AVX2-ONLY-NEXT:    vmovaps (%rdi), %xmm4
; AVX2-ONLY-NEXT:    vmovaps 16(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovaps 48(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm6[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm9[0],ymm5[0],ymm9[2],ymm5[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 240(%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovaps 192(%rdi), %xmm10
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm10[0],xmm8[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm7[0],ymm2[0],ymm7[2],ymm2[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm12 = ymm12[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm12[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 104(%rdi), %ymm12
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm12 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm12[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 296(%rdi), %ymm6
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm2[1],ymm6[3],ymm2[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm10[1],xmm8[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 160(%rdi), %ymm8
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-ONLY-NEXT:    vmovaps 64(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm11[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 352(%rdi), %ymm10
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm10[0],ymm7[2],ymm10[2]
; AVX2-ONLY-NEXT:    vmovaps 256(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovaps 208(%rdi), %xmm14
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm13[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm10 = ymm15[0,1,2,3],ymm10[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovaps 160(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX2-ONLY-NEXT:    vmovaps 32(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm15[1],ymm9[3],ymm15[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm9 = ymm9[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm0[1],ymm7[3],ymm0[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm7 = ymm7[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovaps 80(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm15[0],ymm5[2],ymm15[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovaps 272(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovaps 224(%rdi), %xmm14
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm14[0],xmm13[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 136(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm12[1],xmm11[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 328(%rdi), %ymm11
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm0[1],ymm11[3],ymm0[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovaps %ymm3, 32(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm3, (%rsi)
; AVX2-ONLY-NEXT:    vmovaps %ymm6, 32(%rdx)
; AVX2-ONLY-NEXT:    vmovaps %ymm4, (%rdx)
; AVX2-ONLY-NEXT:    vmovaps %ymm10, 32(%rcx)
; AVX2-ONLY-NEXT:    vmovaps %ymm8, (%rcx)
; AVX2-ONLY-NEXT:    vmovaps %ymm7, 32(%r8)
; AVX2-ONLY-NEXT:    vmovaps %ymm9, (%r8)
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-ONLY-NEXT:    vmovaps %ymm5, (%r9)
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm2, (%rax)
; AVX2-ONLY-NEXT:    vzeroupper
; AVX2-ONLY-NEXT:    retq
;
; AVX512F-LABEL: load_i64_stride6_vf8:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT:    vmovdqa64 320(%rdi), %zmm6
; AVX512F-NEXT:    vmovdqa64 256(%rdi), %zmm7
; AVX512F-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 64(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 128(%rdi), %zmm3
; AVX512F-NEXT:    vmovdqa64 192(%rdi), %zmm4
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]
; AVX512F-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = <0,6,12,u>
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT:    movb $56, %dil
; AVX512F-NEXT:    kmovw %edi, %k1
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]
; AVX512F-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm8
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]
; AVX512F-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm6, %zmm7, %zmm9
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]
; AVX512F-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm6, %zmm7, %zmm10
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]
; AVX512F-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm7, %zmm6, %zmm11
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]
; AVX512F-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm7, %zmm6, %zmm12
; AVX512F-NEXT:    vpermt2q %zmm6, %zmm5, %zmm7
; AVX512F-NEXT:    movb $-64, %dil
; AVX512F-NEXT:    kmovw %edi, %k2
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]
; AVX512F-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm3, %zmm4, %zmm7
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm13 = <1,7,13,u>
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm0, %zmm13
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]
; AVX512F-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm6, %zmm7, %zmm8
; AVX512F-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]
; AVX512F-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm8 = <10,0,6,u>
; AVX512F-NEXT:    vpermi2q %zmm0, %zmm1, %zmm8
; AVX512F-NEXT:    movb $24, %dil
; AVX512F-NEXT:    kmovw %edi, %k2
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}
; AVX512F-NEXT:    movb $-32, %dil
; AVX512F-NEXT:    kmovw %edi, %k1
; AVX512F-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]
; AVX512F-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm9 = <11,1,7,u>
; AVX512F-NEXT:    vpermi2q %zmm0, %zmm1, %zmm9
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]
; AVX512F-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm0, %zmm5
; AVX512F-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm5
; AVX512F-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]
; AVX512F-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm7, %zmm0
; AVX512F-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm0
; AVX512F-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm13, (%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm8, (%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm9, (%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm5, (%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512F-NEXT:    vzeroupper
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride6_vf8:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm7
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm4
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]
; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm2 = <0,6,12,u>
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    movb $56, %dil
; AVX512BW-NEXT:    kmovd %edi, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]
; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm8
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]
; AVX512BW-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm7, %zmm9
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]
; AVX512BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm7, %zmm10
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]
; AVX512BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm7, %zmm6, %zmm11
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]
; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm7, %zmm6, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm5, %zmm7
; AVX512BW-NEXT:    movb $-64, %dil
; AVX512BW-NEXT:    kmovd %edi, %k2
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm7
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm13 = <1,7,13,u>
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm8 = <10,0,6,u>
; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm1, %zmm8
; AVX512BW-NEXT:    movb $24, %dil
; AVX512BW-NEXT:    kmovd %edi, %k2
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}
; AVX512BW-NEXT:    movb $-32, %dil
; AVX512BW-NEXT:    kmovd %edi, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm9 = <11,1,7,u>
; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm1, %zmm9
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm5
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm0
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm13, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm8, (%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm9, (%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, (%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %wide.vec = load <48 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42>
  %strided.vec1 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43>
  %strided.vec2 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44>
  %strided.vec3 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45>
  %strided.vec4 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46>
  %strided.vec5 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47>
  store <8 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <8 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <8 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <8 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <8 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <8 x i64> %strided.vec5, ptr %out.vec5, align 64
  ret void
}

define void @load_i64_stride6_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {
; SSE-LABEL: load_i64_stride6_vf16:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $408, %rsp # imm = 0x198
; SSE-NEXT:    movaps (%rdi), %xmm7
; SSE-NEXT:    movaps 624(%rdi), %xmm0
; SSE-NEXT:    movaps 576(%rdi), %xmm8
; SSE-NEXT:    movaps 240(%rdi), %xmm1
; SSE-NEXT:    movaps 192(%rdi), %xmm9
; SSE-NEXT:    movaps 720(%rdi), %xmm2
; SSE-NEXT:    movaps 672(%rdi), %xmm10
; SSE-NEXT:    movaps 336(%rdi), %xmm3
; SSE-NEXT:    movaps 288(%rdi), %xmm11
; SSE-NEXT:    movaps 432(%rdi), %xmm4
; SSE-NEXT:    movaps 384(%rdi), %xmm13
; SSE-NEXT:    movaps 528(%rdi), %xmm5
; SSE-NEXT:    movaps 480(%rdi), %xmm12
; SSE-NEXT:    movaps 144(%rdi), %xmm6
; SSE-NEXT:    movaps 96(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm6[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm6[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm13, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm4[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm3[1]
; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm2[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm1[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 48(%rdi), %xmm0
; SSE-NEXT:    movaps %xmm7, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 64(%rdi), %xmm0
; SSE-NEXT:    movaps 16(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 160(%rdi), %xmm0
; SSE-NEXT:    movaps 112(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 256(%rdi), %xmm0
; SSE-NEXT:    movaps 208(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 448(%rdi), %xmm0
; SSE-NEXT:    movaps 400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
; SSE-NEXT:    movaps 544(%rdi), %xmm0
; SSE-NEXT:    movaps 496(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 640(%rdi), %xmm0
; SSE-NEXT:    movaps 592(%rdi), %xmm15
; SSE-NEXT:    movaps %xmm15, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
; SSE-NEXT:    movaps 736(%rdi), %xmm0
; SSE-NEXT:    movaps 688(%rdi), %xmm11
; SSE-NEXT:    movaps %xmm11, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 176(%rdi), %xmm0
; SSE-NEXT:    movaps 128(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
; SSE-NEXT:    movaps 272(%rdi), %xmm0
; SSE-NEXT:    movaps 224(%rdi), %xmm12
; SSE-NEXT:    movaps %xmm12, %xmm14
; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 320(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm10
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps 464(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 560(%rdi), %xmm0
; SSE-NEXT:    movaps 512(%rdi), %xmm6
; SSE-NEXT:    movaps %xmm6, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]
; SSE-NEXT:    movaps 656(%rdi), %xmm0
; SSE-NEXT:    movaps 608(%rdi), %xmm3
; SSE-NEXT:    movaps %xmm3, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
; SSE-NEXT:    movaps 752(%rdi), %xmm0
; SSE-NEXT:    movaps 704(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rcx)
; SSE-NEXT:    movaps %xmm11, 112(%r8)
; SSE-NEXT:    movaps %xmm15, 96(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r8)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps %xmm2, 112(%r9)
; SSE-NEXT:    movaps %xmm4, 96(%r9)
; SSE-NEXT:    movaps %xmm7, 80(%r9)
; SSE-NEXT:    movaps %xmm8, 64(%r9)
; SSE-NEXT:    movaps %xmm10, 48(%r9)
; SSE-NEXT:    movaps %xmm14, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm1, 112(%rax)
; SSE-NEXT:    movaps %xmm3, 96(%rax)
; SSE-NEXT:    movaps %xmm6, 80(%rax)
; SSE-NEXT:    movaps %xmm5, 64(%rax)
; SSE-NEXT:    movaps %xmm9, 48(%rax)
; SSE-NEXT:    movaps %xmm12, 32(%rax)
; SSE-NEXT:    movaps %xmm13, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    addq $408, %rsp # imm = 0x198
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: load_i64_stride6_vf16:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $552, %rsp # imm = 0x228
; AVX1-ONLY-NEXT:    vmovaps 320(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 512(%rdi), %ymm8
; AVX1-ONLY-NEXT:    vmovups %ymm8, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX1-ONLY-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX1-ONLY-NEXT:    vmovaps (%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovaps 48(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm5[0],xmm6[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm4
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm4[0],ymm8[0],ymm4[2],ymm8[2]
; AVX1-ONLY-NEXT:    vmovaps 432(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovaps 384(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm7[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm7
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm1[0],ymm7[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 240(%rdi), %xmm11
; AVX1-ONLY-NEXT:    vmovaps 192(%rdi), %xmm12
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm10[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 704(%rdi), %ymm10
; AVX1-ONLY-NEXT:    vinsertf128 $1, 672(%rdi), %ymm0, %ymm13
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm13[0],ymm10[0],ymm13[2],ymm10[2]
; AVX1-ONLY-NEXT:    vmovaps 624(%rdi), %xmm15
; AVX1-ONLY-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm15[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm6[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd (%rsp), %ymm4, %ymm1 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm1 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm9[1],xmm8[1]
; AVX1-ONLY-NEXT:    vmovaps 480(%rdi), %ymm3
; AVX1-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX1-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm4 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm4 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps 16(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm12[1],xmm11[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm13[1],ymm10[1],ymm13[3],ymm10[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm14
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm14[0],ymm2[2],ymm14[2]
; AVX1-ONLY-NEXT:    vmovaps 64(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm6[0],xmm13[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 544(%rdi), %ymm0, %ymm11
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm11[0],ymm3[2],ymm11[2]
; AVX1-ONLY-NEXT:    vmovaps 448(%rdi), %xmm7
; AVX1-ONLY-NEXT:    vmovaps 400(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm7[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm6[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 288(%rdi), %ymm6
; AVX1-ONLY-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm15
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm6[0],ymm15[0],ymm6[2],ymm15[2]
; AVX1-ONLY-NEXT:    vmovaps 256(%rdi), %xmm4
; AVX1-ONLY-NEXT:    vmovaps 208(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm3[0],xmm4[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm10[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 672(%rdi), %ymm5
; AVX1-ONLY-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm2
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
; AVX1-ONLY-NEXT:    vmovaps 640(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps 592(%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm12[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm14[1],ymm9[3],ymm14[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm10[1],xmm13[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm9[1],ymm11[1],ymm9[3],ymm11[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm8[1],xmm7[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm11[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm15[1],ymm6[3],ymm15[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm4[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 544(%rdi), %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 512(%rdi), %ymm0, %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 464(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovaps 416(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm13[0],xmm14[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 160(%rdi), %ymm11
; AVX1-ONLY-NEXT:    vmovaps 32(%rdi), %xmm12
; AVX1-ONLY-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm10
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX1-ONLY-NEXT:    vmovaps 80(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm12[0],xmm9[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 352(%rdi), %ymm8
; AVX1-ONLY-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm7
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX1-ONLY-NEXT:    vmovaps 272(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps 224(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm3[0],xmm6[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 736(%rdi), %ymm5
; AVX1-ONLY-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm2
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm5[0],ymm2[2],ymm5[2]
; AVX1-ONLY-NEXT:    vmovaps 656(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm15[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm13[1],xmm14[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm12[1],xmm9[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 96(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 32(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 64(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, (%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 96(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 32(%rdx)
; AVX1-ONLY-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 64(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, (%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 96(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 32(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 64(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, (%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 32(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 96(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 64(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, (%r8)
; AVX1-ONLY-NEXT:    vmovaps %ymm4, 96(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 32(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, (%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm3, 64(%r9)
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm5, 96(%rax)
; AVX1-ONLY-NEXT:    addq $552, %rsp # imm = 0x228
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-ONLY-LABEL: load_i64_stride6_vf16:
; AVX2-ONLY:       # %bb.0:
; AVX2-ONLY-NEXT:    subq $488, %rsp # imm = 0x1E8
; AVX2-ONLY-NEXT:    vmovaps 320(%rdi), %ymm10
; AVX2-ONLY-NEXT:    vmovaps 288(%rdi), %ymm12
; AVX2-ONLY-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 512(%rdi), %ymm7
; AVX2-ONLY-NEXT:    vmovaps 480(%rdi), %ymm4
; AVX2-ONLY-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 128(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovaps 96(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps (%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps 48(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm5[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps %ymm0, %ymm15
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 432(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps 384(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm3[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]
; AVX2-ONLY-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 240(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovaps 192(%rdi), %xmm9
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm9[0],xmm11[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
; AVX2-ONLY-NEXT:    vmovaps %ymm10, %ymm4
; AVX2-ONLY-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 704(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovaps 672(%rdi), %ymm8
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm8[0],ymm0[0],ymm8[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps %ymm0, %ymm10
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm13 = ymm13[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 624(%rdi), %xmm14
; AVX2-ONLY-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm0[0],xmm14[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm12[0,1,2,3],ymm13[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm5[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 104(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm15[1],ymm5[3],ymm15[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm3[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 488(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm9[1],xmm11[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 296(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm14[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 680(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm10, %ymm14
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 160(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 16(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovaps 64(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm7[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 544(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 448(%rdi), %xmm4
; AVX2-ONLY-NEXT:    vmovaps 400(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 352(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 256(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps 208(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 736(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm0[0],ymm8[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 640(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovaps 592(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm1[0],xmm13[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm3[1]
; AVX2-ONLY-NEXT:    vmovaps 160(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm15[1],ymm10[3],ymm15[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
; AVX2-ONLY-NEXT:    vmovaps 544(%rdi), %ymm10
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm13[1]
; AVX2-ONLY-NEXT:    vmovaps 736(%rdi), %ymm4
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm4[1],ymm8[3],ymm4[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm12[1]
; AVX2-ONLY-NEXT:    vmovaps 352(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm9[1],ymm5[1],ymm9[3],ymm5[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 464(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovaps 416(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm7[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm8 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps 80(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm3[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 272(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps 224(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 656(%rdi), %xmm9
; AVX2-ONLY-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm9[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 712(%rdi), %ymm14
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 520(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm7[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 136(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm15[1],ymm0[3],ymm15[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 328(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm12[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 96(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 32(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 64(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, (%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 96(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 32(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 64(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, (%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 96(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 32(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 64(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, (%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 32(%r8)
; AVX2-ONLY-NEXT:    vmovups (%rsp), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 96(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 64(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, (%r8)
; AVX2-ONLY-NEXT:    vmovaps %ymm8, 96(%r9)
; AVX2-ONLY-NEXT:    vmovaps %ymm11, 32(%r9)
; AVX2-ONLY-NEXT:    vmovaps %ymm13, (%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm2, 64(%r9)
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm4, 64(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm9, 96(%rax)
; AVX2-ONLY-NEXT:    addq $488, %rsp # imm = 0x1E8
; AVX2-ONLY-NEXT:    vzeroupper
; AVX2-ONLY-NEXT:    retq
;
; AVX512F-LABEL: load_i64_stride6_vf16:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT:    vmovdqa64 448(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 384(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 512(%rdi), %zmm3
; AVX512F-NEXT:    vmovdqa64 576(%rdi), %zmm5
; AVX512F-NEXT:    vmovdqa64 704(%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512F-NEXT:    vmovdqa64 320(%rdi), %zmm8
; AVX512F-NEXT:    vmovdqa64 256(%rdi), %zmm12
; AVX512F-NEXT:    vmovdqa64 (%rdi), %zmm9
; AVX512F-NEXT:    vmovdqa64 64(%rdi), %zmm11
; AVX512F-NEXT:    vmovdqa64 128(%rdi), %zmm10
; AVX512F-NEXT:    vmovdqa64 192(%rdi), %zmm13
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]
; AVX512F-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512F-NEXT:    vpermt2q %zmm10, %zmm14, %zmm15
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm6 = <0,6,12,u>
; AVX512F-NEXT:    vmovdqa64 %zmm9, %zmm7
; AVX512F-NEXT:    vpermt2q %zmm11, %zmm6, %zmm7
; AVX512F-NEXT:    movb $56, %dil
; AVX512F-NEXT:    kmovw %edi, %k1
; AVX512F-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]
; AVX512F-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm12, %zmm15
; AVX512F-NEXT:    vpermt2q %zmm8, %zmm16, %zmm15
; AVX512F-NEXT:    movb $-64, %dil
; AVX512F-NEXT:    kmovw %edi, %k2
; AVX512F-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm16, %zmm15
; AVX512F-NEXT:    vpermi2q %zmm3, %zmm5, %zmm14
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm1, %zmm6
; AVX512F-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]
; AVX512F-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm17
; AVX512F-NEXT:    vpermt2q %zmm10, %zmm18, %zmm17
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm14 = <1,7,13,u>
; AVX512F-NEXT:    vmovdqa64 %zmm9, %zmm15
; AVX512F-NEXT:    vpermt2q %zmm11, %zmm14, %zmm15
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]
; AVX512F-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm12, %zmm19
; AVX512F-NEXT:    vpermt2q %zmm8, %zmm17, %zmm19
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm19
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm17, %zmm19
; AVX512F-NEXT:    vpermi2q %zmm3, %zmm5, %zmm18
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm1, %zmm14
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]
; AVX512F-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm10, %zmm21
; AVX512F-NEXT:    vpermt2q %zmm13, %zmm20, %zmm21
; AVX512F-NEXT:    vmovdqa64 {{.*#+}} ymm18 = <10,0,6,u>
; AVX512F-NEXT:    vmovdqa64 %zmm11, %zmm19
; AVX512F-NEXT:    vpermt2q %zmm9, %zmm18, %zmm19
; AVX512F-NEXT:    movb $24, %dil
; AVX512F-NEXT:    kmovw %edi, %k2
; AVX512F-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]
; AVX512F-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm12, %zmm22
; AVX512F-NEXT:    vpermt2q %zmm8, %zmm21, %zmm22
; AVX512F-NEXT:    movb $-32, %dil
; AVX512F-NEXT:    kmovw %edi, %k1
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}
; AVX512F-NEXT:    vpermi2q %zmm0, %zmm4, %zmm21
; AVX512F-NEXT:    vpermi2q %zmm5, %zmm3, %zmm20
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm2, %zmm18
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]
; AVX512F-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm10, %zmm23
; AVX512F-NEXT:    vpermt2q %zmm13, %zmm22, %zmm23
; AVX512F-NEXT:    vmovdqa64 {{.*#+}} ymm20 = <11,1,7,u>
; AVX512F-NEXT:    vmovdqa64 %zmm11, %zmm21
; AVX512F-NEXT:    vpermt2q %zmm9, %zmm20, %zmm21
; AVX512F-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]
; AVX512F-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm12, %zmm24
; AVX512F-NEXT:    vpermt2q %zmm8, %zmm23, %zmm24
; AVX512F-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}
; AVX512F-NEXT:    vpermi2q %zmm0, %zmm4, %zmm23
; AVX512F-NEXT:    vpermi2q %zmm5, %zmm3, %zmm22
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm2, %zmm20
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]
; AVX512F-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm8, %zmm23
; AVX512F-NEXT:    vpermt2q %zmm12, %zmm22, %zmm23
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]
; AVX512F-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm10, %zmm25
; AVX512F-NEXT:    vpermt2q %zmm13, %zmm24, %zmm25
; AVX512F-NEXT:    vmovdqa64 %zmm9, %zmm26
; AVX512F-NEXT:    vpermt2q %zmm11, %zmm16, %zmm26
; AVX512F-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm25
; AVX512F-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}
; AVX512F-NEXT:    vpermi2q %zmm5, %zmm3, %zmm24
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm1, %zmm16
; AVX512F-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm16
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm0, %zmm22
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]
; AVX512F-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm12, %zmm22, %zmm8
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]
; AVX512F-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm13, %zmm12, %zmm10
; AVX512F-NEXT:    vpermt2q %zmm11, %zmm17, %zmm9
; AVX512F-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm9
; AVX512F-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm12, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm17, %zmm1
; AVX512F-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm4, %zmm22, %zmm0
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm6, 64(%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm7, (%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm14, 64(%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm15, (%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm18, 64(%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm19, (%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm20, 64(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm21, (%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm16, 64(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm25, (%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm1, 64(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm9, (%rax)
; AVX512F-NEXT:    vzeroupper
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride6_vf16:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm8
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm12
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm9
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm11
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm10
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm13
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]
; AVX512BW-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm14, %zmm15
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm6 = <0,6,12,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm7
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm6, %zmm7
; AVX512BW-NEXT:    movb $56, %dil
; AVX512BW-NEXT:    kmovd %edi, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]
; AVX512BW-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm16, %zmm15
; AVX512BW-NEXT:    movb $-64, %dil
; AVX512BW-NEXT:    kmovd %edi, %k2
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm16, %zmm15
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm5, %zmm14
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]
; AVX512BW-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm17
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm18, %zmm17
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm14 = <1,7,13,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm14, %zmm15
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]
; AVX512BW-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm17, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm17, %zmm19
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm5, %zmm18
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm14
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]
; AVX512BW-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm21
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm20, %zmm21
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm18 = <10,0,6,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm18, %zmm19
; AVX512BW-NEXT:    movb $24, %dil
; AVX512BW-NEXT:    kmovd %edi, %k2
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]
; AVX512BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm22
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm21, %zmm22
; AVX512BW-NEXT:    movb $-32, %dil
; AVX512BW-NEXT:    kmovd %edi, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}
; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm4, %zmm21
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm20
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm2, %zmm18
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]
; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm23
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm22, %zmm23
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm20 = <11,1,7,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm21
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm20, %zmm21
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]
; AVX512BW-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm24
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm23, %zmm24
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}
; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm4, %zmm23
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm22
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm2, %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]
; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm23
; AVX512BW-NEXT:    vpermt2q %zmm12, %zmm22, %zmm23
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]
; AVX512BW-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm25
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm24, %zmm25
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm26
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm16, %zmm26
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm25
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm24
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm16
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm16
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm0, %zmm22
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]
; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm12, %zmm22, %zmm8
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]
; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm12, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm17, %zmm9
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm9
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm12, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm17, %zmm1
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm22, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm6, 64(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm7, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm14, 64(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm15, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm18, 64(%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm19, (%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm20, 64(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm21, (%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm16, 64(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm25, (%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm9, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %wide.vec = load <96 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90>
  %strided.vec1 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91>
  %strided.vec2 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92>
  %strided.vec3 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93>
  %strided.vec4 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94>
  %strided.vec5 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95>
  store <16 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <16 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <16 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <16 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <16 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <16 x i64> %strided.vec5, ptr %out.vec5, align 64
  ret void
}

define void @load_i64_stride6_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {
; SSE-LABEL: load_i64_stride6_vf32:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $1176, %rsp # imm = 0x498
; SSE-NEXT:    movaps 624(%rdi), %xmm0
; SSE-NEXT:    movaps 576(%rdi), %xmm9
; SSE-NEXT:    movaps 240(%rdi), %xmm3
; SSE-NEXT:    movaps 192(%rdi), %xmm8
; SSE-NEXT:    movaps 720(%rdi), %xmm1
; SSE-NEXT:    movaps 672(%rdi), %xmm11
; SSE-NEXT:    movaps 336(%rdi), %xmm5
; SSE-NEXT:    movaps 288(%rdi), %xmm10
; SSE-NEXT:    movaps 432(%rdi), %xmm4
; SSE-NEXT:    movaps 384(%rdi), %xmm12
; SSE-NEXT:    movaps 912(%rdi), %xmm2
; SSE-NEXT:    movaps 528(%rdi), %xmm6
; SSE-NEXT:    movaps 480(%rdi), %xmm14
; SSE-NEXT:    movaps 144(%rdi), %xmm7
; SSE-NEXT:    movaps 96(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm7[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm5[0]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm5[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm3[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm14, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm6[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm6[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm4[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm4[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm1[1]
; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 864(%rdi), %xmm0
; SSE-NEXT:    movaps %xmm0, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 816(%rdi), %xmm0
; SSE-NEXT:    movaps 768(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1104(%rdi), %xmm0
; SSE-NEXT:    movaps 1056(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1008(%rdi), %xmm0
; SSE-NEXT:    movaps 960(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1296(%rdi), %xmm0
; SSE-NEXT:    movaps 1248(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1200(%rdi), %xmm0
; SSE-NEXT:    movaps 1152(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1488(%rdi), %xmm0
; SSE-NEXT:    movaps 1440(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1392(%rdi), %xmm0
; SSE-NEXT:    movaps 1344(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps (%rdi), %xmm1
; SSE-NEXT:    movaps 48(%rdi), %xmm0
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 64(%rdi), %xmm0
; SSE-NEXT:    movaps 16(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 160(%rdi), %xmm0
; SSE-NEXT:    movaps 112(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 256(%rdi), %xmm0
; SSE-NEXT:    movaps 208(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 448(%rdi), %xmm0
; SSE-NEXT:    movaps 400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 544(%rdi), %xmm0
; SSE-NEXT:    movaps 496(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 640(%rdi), %xmm0
; SSE-NEXT:    movaps 592(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 736(%rdi), %xmm0
; SSE-NEXT:    movaps 688(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 832(%rdi), %xmm0
; SSE-NEXT:    movaps 784(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 928(%rdi), %xmm0
; SSE-NEXT:    movaps 880(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1024(%rdi), %xmm0
; SSE-NEXT:    movaps 976(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1120(%rdi), %xmm0
; SSE-NEXT:    movaps 1072(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1216(%rdi), %xmm0
; SSE-NEXT:    movaps 1168(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1312(%rdi), %xmm0
; SSE-NEXT:    movaps 1264(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1408(%rdi), %xmm0
; SSE-NEXT:    movaps 1360(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1504(%rdi), %xmm0
; SSE-NEXT:    movaps 1456(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 176(%rdi), %xmm0
; SSE-NEXT:    movaps 128(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 272(%rdi), %xmm0
; SSE-NEXT:    movaps 224(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 320(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 464(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 560(%rdi), %xmm0
; SSE-NEXT:    movaps 512(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
; SSE-NEXT:    movaps 656(%rdi), %xmm0
; SSE-NEXT:    movaps 608(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 752(%rdi), %xmm0
; SSE-NEXT:    movaps 704(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
; SSE-NEXT:    movaps 848(%rdi), %xmm0
; SSE-NEXT:    movaps 800(%rdi), %xmm15
; SSE-NEXT:    movaps %xmm15, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
; SSE-NEXT:    movaps 944(%rdi), %xmm0
; SSE-NEXT:    movaps 896(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
; SSE-NEXT:    movaps 1040(%rdi), %xmm0
; SSE-NEXT:    movaps 992(%rdi), %xmm10
; SSE-NEXT:    movaps %xmm10, %xmm12
; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT:    movaps 1136(%rdi), %xmm0
; SSE-NEXT:    movaps 1088(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm11
; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps 1232(%rdi), %xmm0
; SSE-NEXT:    movaps 1184(%rdi), %xmm7
; SSE-NEXT:    movaps %xmm7, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT:    movaps 1328(%rdi), %xmm0
; SSE-NEXT:    movaps 1280(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 1424(%rdi), %xmm0
; SSE-NEXT:    movaps 1376(%rdi), %xmm3
; SSE-NEXT:    movaps %xmm3, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
; SSE-NEXT:    movaps 1520(%rdi), %xmm0
; SSE-NEXT:    movaps 1472(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps %xmm2, 240(%r9)
; SSE-NEXT:    movaps %xmm4, 224(%r9)
; SSE-NEXT:    movaps %xmm6, 208(%r9)
; SSE-NEXT:    movaps %xmm8, 192(%r9)
; SSE-NEXT:    movaps %xmm11, 176(%r9)
; SSE-NEXT:    movaps %xmm12, 160(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm1, 240(%rax)
; SSE-NEXT:    movaps %xmm3, 224(%rax)
; SSE-NEXT:    movaps %xmm5, 208(%rax)
; SSE-NEXT:    movaps %xmm7, 192(%rax)
; SSE-NEXT:    movaps %xmm9, 176(%rax)
; SSE-NEXT:    movaps %xmm10, 160(%rax)
; SSE-NEXT:    movaps %xmm13, 144(%rax)
; SSE-NEXT:    movaps %xmm15, 128(%rax)
; SSE-NEXT:    movaps %xmm14, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps (%rsp), %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 $1176, %rsp # imm = 0x498
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: load_i64_stride6_vf32:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $1624, %rsp # imm = 0x658
; AVX1-ONLY-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX1-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 704(%rdi), %ymm3
; AVX1-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 320(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 240(%rdi), %xmm4
; AVX1-ONLY-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 192(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 672(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX1-ONLY-NEXT:    vmovaps 624(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 576(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1056(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX1-ONLY-NEXT:    vmovaps 1008(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 960(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1392(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1344(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 128(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps (%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovaps 48(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm14[0],xmm13[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 512(%rdi), %ymm12
; AVX1-ONLY-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm11
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX1-ONLY-NEXT:    vmovaps 432(%rdi), %xmm10
; AVX1-ONLY-NEXT:    vmovaps 384(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 896(%rdi), %ymm8
; AVX1-ONLY-NEXT:    vinsertf128 $1, 864(%rdi), %ymm0, %ymm7
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX1-ONLY-NEXT:    vmovaps 816(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps 768(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1280(%rdi), %ymm4
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX1-ONLY-NEXT:    vmovaps 1200(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps 1152(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm14[1],xmm13[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 288(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 256(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 208(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 672(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 640(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 592(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1056(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1120(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 976(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1440(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1408(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovaps 1360(%rdi), %xmm12
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm14[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 96(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 480(%rdi), %ymm11
; AVX1-ONLY-NEXT:    vinsertf128 $1, 544(%rdi), %ymm0, %ymm10
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
; AVX1-ONLY-NEXT:    vmovaps 448(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovaps 400(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 864(%rdi), %ymm7
; AVX1-ONLY-NEXT:    vinsertf128 $1, 928(%rdi), %ymm0, %ymm6
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX1-ONLY-NEXT:    vmovaps 832(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovaps 784(%rdi), %xmm4
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm4[0],xmm5[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1248(%rdi), %ymm3
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1312(%rdi), %ymm0, %ymm13
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm13[0],ymm3[2],ymm13[2]
; AVX1-ONLY-NEXT:    vmovaps 1216(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps 1168(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm12 = xmm12[1],xmm14[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm13[1],ymm3[3],ymm13[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm5[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm9[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 160(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 32(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 352(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 272(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 224(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 544(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 512(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 464(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 656(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 608(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 928(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 896(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 848(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovaps 800(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1120(%rdi), %ymm12
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1088(%rdi), %ymm0, %ymm11
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX1-ONLY-NEXT:    vmovaps 1040(%rdi), %xmm10
; AVX1-ONLY-NEXT:    vmovaps 992(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1312(%rdi), %ymm8
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1280(%rdi), %ymm0, %ymm7
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX1-ONLY-NEXT:    vmovaps 1232(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps 1184(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1504(%rdi), %ymm4
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX1-ONLY-NEXT:    vmovaps 1424(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps 1376(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 192(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 128(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 64(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, (%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 224(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 160(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 96(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 192(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 128(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 64(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, (%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 224(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 160(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 96(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 192(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 128(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 64(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, (%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 224(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 160(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 96(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, (%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 64(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 128(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 192(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 224(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 160(%r8)
; AVX1-ONLY-NEXT:    vmovups (%rsp), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 96(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 224(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 192(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 160(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 128(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 96(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 64(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, 32(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm2, (%r9)
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm9, 160(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm13, 128(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm15, 96(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %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 $1624, %rsp # imm = 0x658
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-ONLY-LABEL: load_i64_stride6_vf32:
; AVX2-ONLY:       # %bb.0:
; AVX2-ONLY-NEXT:    subq $1496, %rsp # imm = 0x5D8
; AVX2-ONLY-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1056(%rdi), %ymm4
; AVX2-ONLY-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 704(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 672(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovups %ymm5, (%rsp) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 320(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 288(%rdi), %ymm7
; AVX2-ONLY-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 240(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 192(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm6[0],xmm0[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[2],ymm1[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 624(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1008(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 960(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1472(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1440(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1392(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 128(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]
; AVX2-ONLY-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps (%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovaps 48(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 512(%rdi), %ymm7
; AVX2-ONLY-NEXT:    vmovaps 480(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]
; AVX2-ONLY-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 432(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovaps 384(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm12[0],xmm11[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 896(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovaps 864(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]
; AVX2-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 816(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovaps 768(%rdi), %xmm10
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm10[0],xmm13[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vmovaps 1248(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX2-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1200(%rdi), %xmm14
; AVX2-ONLY-NEXT:    vmovaps 1152(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm14[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm6[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 296(%rdi), %ymm6
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 680(%rdi), %ymm6
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1064(%rdi), %ymm6
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm4[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1448(%rdi), %ymm4
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm4 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm2[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 104(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm12[1],xmm11[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 488(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm7[1],ymm2[3],ymm7[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm13[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 872(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm14[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1256(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 352(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 256(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps 208(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 736(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 640(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovaps 592(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm6[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1120(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1024(%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovaps 976(%rdi), %xmm9
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm8[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1504(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1408(%rdi), %xmm10
; AVX2-ONLY-NEXT:    vmovaps 1360(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm10[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 160(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 16(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 64(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 544(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-ONLY-NEXT:    vmovaps 448(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 400(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 928(%rdi), %ymm12
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm4[0],ymm12[0],ymm4[2],ymm12[2]
; AVX2-ONLY-NEXT:    vmovaps 832(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps 784(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm13[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1312(%rdi), %ymm14
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm14[0],ymm1[2],ymm14[2]
; AVX2-ONLY-NEXT:    vmovaps 1216(%rdi), %xmm15
; AVX2-ONLY-NEXT:    vmovaps 1168(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm15[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm3[1]
; AVX2-ONLY-NEXT:    vmovaps 352(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm5, %ymm14
; AVX2-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm7[1],xmm6[1]
; AVX2-ONLY-NEXT:    vmovaps 736(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm5, %ymm7
; AVX2-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm9[1],xmm8[1]
; AVX2-ONLY-NEXT:    vmovaps 1120(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm5, %ymm6
; AVX2-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm11[1],xmm10[1]
; AVX2-ONLY-NEXT:    vmovaps 1504(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]
; AVX2-ONLY-NEXT:    vmovaps 1312(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]
; AVX2-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm12[1]
; AVX2-ONLY-NEXT:    vmovaps 928(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm1, %ymm4
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 544(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 160(%rdi), %ymm9
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 32(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 272(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 224(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 464(%rdi), %xmm14
; AVX2-ONLY-NEXT:    vmovaps 416(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 656(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps 608(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 848(%rdi), %xmm10
; AVX2-ONLY-NEXT:    vmovaps 800(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm10[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1040(%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovaps 992(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1232(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovaps 1184(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1424(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps 1376(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 136(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm9[1],ymm0[3],ymm9[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm9 = xmm9[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 328(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm9 = xmm9[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 520(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm13[1],xmm14[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 712(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm11[1],xmm12[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 904(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm10[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 1096(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 1288(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 1480(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm4[1],ymm0[3],ymm4[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-ONLY-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm3, 192(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm5, 160(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm7, 128(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm9, 96(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm13, 64(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm15, 32(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-ONLY-NEXT:    addq $1496, %rsp # imm = 0x5D8
; AVX2-ONLY-NEXT:    vzeroupper
; AVX2-ONLY-NEXT:    retq
;
; AVX512F-LABEL: load_i64_stride6_vf32:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    subq $2632, %rsp # imm = 0xA48
; AVX512F-NEXT:    vmovdqa64 1280(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 1344(%rdi), %zmm21
; AVX512F-NEXT:    vmovdqa64 896(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 960(%rdi), %zmm19
; AVX512F-NEXT:    vmovdqa64 448(%rdi), %zmm18
; AVX512F-NEXT:    vmovdqa64 384(%rdi), %zmm3
; AVX512F-NEXT:    vmovdqa64 512(%rdi), %zmm4
; AVX512F-NEXT:    vmovdqa64 576(%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 128(%rdi), %zmm5
; AVX512F-NEXT:    vmovdqa64 192(%rdi), %zmm25
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]
; AVX512F-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm7
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm6, %zmm7
; AVX512F-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm7 = <0,6,12,u>
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm4, %zmm6, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm18, %zmm7, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm6, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm21, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]
; AVX512F-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm6, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm8
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm6
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm10 = <1,7,13,u>
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm4, %zmm8, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm8, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm21, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]
; AVX512F-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm21, %zmm2, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]
; AVX512F-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm21, %zmm2, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]
; AVX512F-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm8, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]
; AVX512F-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm1, %zmm4
; AVX512F-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm8, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm1, %zmm5
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm8, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm21, %zmm2, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm21, %zmm1, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm18, %zmm10, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm1, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm11 = <10,0,6,u>
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm3, %zmm11, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 {{.*#+}} ymm31 = <11,1,7,u>
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm3, %zmm31, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]
; AVX512F-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm18, %zmm5, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]
; AVX512F-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 (%rdi), %zmm25
; AVX512F-NEXT:    vmovdqa64 64(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm29
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm7, %zmm29
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm31, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm5, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm25
; AVX512F-NEXT:    vmovdqa64 320(%rdi), %zmm9
; AVX512F-NEXT:    vmovdqa64 256(%rdi), %zmm27
; AVX512F-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 704(%rdi), %zmm30
; AVX512F-NEXT:    vmovdqa64 640(%rdi), %zmm26
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm30, %zmm5, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512F-NEXT:    vmovdqa64 1024(%rdi), %zmm20
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm7, %zmm5, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 832(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 768(%rdi), %zmm8
; AVX512F-NEXT:    vmovdqa64 %zmm8, %zmm15
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm2, %zmm15
; AVX512F-NEXT:    vmovdqa64 1472(%rdi), %zmm6
; AVX512F-NEXT:    vmovdqa64 1408(%rdi), %zmm21
; AVX512F-NEXT:    vmovdqa64 %zmm21, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm6, %zmm5, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1216(%rdi), %zmm19
; AVX512F-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512F-NEXT:    vpermi2q %zmm19, %zmm4, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm8, %zmm12
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm10, %zmm12
; AVX512F-NEXT:    vpermi2q %zmm19, %zmm4, %zmm10
; AVX512F-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512F-NEXT:    vpermt2q %zmm8, %zmm11, %zmm13
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm19, %zmm11
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512F-NEXT:    vpermt2q %zmm8, %zmm31, %zmm14
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm19, %zmm31
; AVX512F-NEXT:    vmovdqa64 %zmm8, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm5, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm19, %zmm4, %zmm5
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm0, %zmm4
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm8
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm19
; AVX512F-NEXT:    vpermt2q %zmm7, %zmm0, %zmm19
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm22
; AVX512F-NEXT:    vpermt2q %zmm30, %zmm0, %zmm22
; AVX512F-NEXT:    vmovdqa64 %zmm27, %zmm17
; AVX512F-NEXT:    vpermt2q %zmm9, %zmm0, %zmm17
; AVX512F-NEXT:    vmovdqa64 %zmm21, %zmm18
; AVX512F-NEXT:    vpermt2q %zmm6, %zmm0, %zmm18
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]
; AVX512F-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm24
; AVX512F-NEXT:    vpermt2q %zmm30, %zmm1, %zmm24
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]
; AVX512F-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm30, %zmm2, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]
; AVX512F-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm30, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm3, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]
; AVX512F-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm0, %zmm30
; AVX512F-NEXT:    vmovdqa64 %zmm27, %zmm26
; AVX512F-NEXT:    vpermt2q %zmm9, %zmm1, %zmm26
; AVX512F-NEXT:    vmovdqa64 %zmm27, %zmm23
; AVX512F-NEXT:    vpermt2q %zmm9, %zmm2, %zmm23
; AVX512F-NEXT:    vmovdqa64 %zmm9, %zmm28
; AVX512F-NEXT:    vpermt2q %zmm27, %zmm3, %zmm28
; AVX512F-NEXT:    vpermt2q %zmm27, %zmm0, %zmm9
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm27
; AVX512F-NEXT:    vpermt2q %zmm7, %zmm1, %zmm27
; AVX512F-NEXT:    vpermi2q %zmm6, %zmm21, %zmm1
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm10
; AVX512F-NEXT:    vpermt2q %zmm7, %zmm2, %zmm10
; AVX512F-NEXT:    vpermi2q %zmm6, %zmm21, %zmm2
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm16
; AVX512F-NEXT:    vpermt2q %zmm20, %zmm3, %zmm16
; AVX512F-NEXT:    vpermi2q %zmm21, %zmm6, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm21, %zmm0, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm20, %zmm0, %zmm7
; AVX512F-NEXT:    movb $56, %al
; AVX512F-NEXT:    kmovw %eax, %k1
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}
; AVX512F-NEXT:    movb $-64, %al
; AVX512F-NEXT:    kmovw %eax, %k2
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}
; AVX512F-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm20
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm17
; AVX512F-NEXT:    movb $24, %al
; AVX512F-NEXT:    kmovw %eax, %k2
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}
; AVX512F-NEXT:    movb $-32, %al
; AVX512F-NEXT:    kmovw %eax, %k1
; AVX512F-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm22
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm24
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm23
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm5
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm10
; AVX512F-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm4
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm6
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm20, 192(%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm15, 128(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm7, 64(%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm29, (%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm17, 192(%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm19, (%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm21, 64(%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm12, 128(%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm11, 192(%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm24, (%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm22, 64(%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm13, 128(%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm31, 192(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm23, (%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm18, 64(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm14, 128(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm5, 192(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm2, (%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm1, 64(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm0, 128(%r9)
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT:    vmovdqa64 %zmm6, 128(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm4, 192(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512F-NEXT:    addq $2632, %rsp # imm = 0xA48
; AVX512F-NEXT:    vzeroupper
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride6_vf32:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    subq $2632, %rsp # imm = 0xA48
; AVX512BW-NEXT:    vmovdqa64 1280(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 1344(%rdi), %zmm21
; AVX512BW-NEXT:    vmovdqa64 896(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 960(%rdi), %zmm19
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm18
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm25
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm7
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm6, %zmm7
; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm7 = <0,6,12,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm6, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm6, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm21, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]
; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm6, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm6
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm10 = <1,7,13,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm8, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm21, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]
; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]
; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]
; AVX512BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm8, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]
; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm8, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm8, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm1, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm10, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm11 = <10,0,6,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm11, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} ymm31 = <11,1,7,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm31, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]
; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm5, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm25
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm29
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm29
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm31, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm5, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm25
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm9
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm27
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm30
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm26
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512BW-NEXT:    vmovdqa64 1024(%rdi), %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 832(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 768(%rdi), %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm15
; AVX512BW-NEXT:    vmovdqa64 1472(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 1408(%rdi), %zmm21
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm5, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1216(%rdi), %zmm19
; AVX512BW-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm12
; AVX512BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm11, %zmm13
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm19, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm31, %zmm14
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm19, %zmm31
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm5, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm4
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm22
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm22
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm17
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm0, %zmm17
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm18
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm18
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]
; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm24
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm1, %zmm24
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]
; AVX512BW-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm3, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm30
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm26
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm1, %zmm26
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm23
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm23
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm28
; AVX512BW-NEXT:    vpermt2q %zmm27, %zmm3, %zmm28
; AVX512BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm9
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm27
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm1, %zmm27
; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm21, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm10
; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm21, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm16
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm3, %zmm16
; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm6, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm7
; AVX512BW-NEXT:    movb $56, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}
; AVX512BW-NEXT:    movb $-64, %al
; AVX512BW-NEXT:    kmovd %eax, %k2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm20
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm17
; AVX512BW-NEXT:    movb $24, %al
; AVX512BW-NEXT:    kmovd %eax, %k2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}
; AVX512BW-NEXT:    movb $-32, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm22
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm24
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm23
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm4
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm20, 192(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm15, 128(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm7, 64(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm29, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm17, 192(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm19, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm21, 64(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm12, 128(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm11, 192(%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm24, (%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm22, 64(%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm13, 128(%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm31, 192(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm23, (%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm18, 64(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm14, 128(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, 192(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, 128(%r9)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm6, 128(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm4, 192(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512BW-NEXT:    addq $2632, %rsp # imm = 0xA48
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %wide.vec = load <192 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90, i32 96, i32 102, i32 108, i32 114, i32 120, i32 126, i32 132, i32 138, i32 144, i32 150, i32 156, i32 162, i32 168, i32 174, i32 180, i32 186>
  %strided.vec1 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91, i32 97, i32 103, i32 109, i32 115, i32 121, i32 127, i32 133, i32 139, i32 145, i32 151, i32 157, i32 163, i32 169, i32 175, i32 181, i32 187>
  %strided.vec2 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92, i32 98, i32 104, i32 110, i32 116, i32 122, i32 128, i32 134, i32 140, i32 146, i32 152, i32 158, i32 164, i32 170, i32 176, i32 182, i32 188>
  %strided.vec3 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93, i32 99, i32 105, i32 111, i32 117, i32 123, i32 129, i32 135, i32 141, i32 147, i32 153, i32 159, i32 165, i32 171, i32 177, i32 183, i32 189>
  %strided.vec4 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94, i32 100, i32 106, i32 112, i32 118, i32 124, i32 130, i32 136, i32 142, i32 148, i32 154, i32 160, i32 166, i32 172, i32 178, i32 184, i32 190>
  %strided.vec5 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95, i32 101, i32 107, i32 113, i32 119, i32 125, i32 131, i32 137, i32 143, i32 149, i32 155, i32 161, i32 167, i32 173, i32 179, i32 185, i32 191>
  store <32 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <32 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <32 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <32 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <32 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <32 x i64> %strided.vec5, ptr %out.vec5, align 64
  ret void
}

define void @load_i64_stride6_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {
; SSE-LABEL: load_i64_stride6_vf64:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $2712, %rsp # imm = 0xA98
; SSE-NEXT:    movaps 816(%rdi), %xmm0
; SSE-NEXT:    movaps 720(%rdi), %xmm1
; SSE-NEXT:    movaps 672(%rdi), %xmm8
; SSE-NEXT:    movaps 624(%rdi), %xmm2
; SSE-NEXT:    movaps 576(%rdi), %xmm9
; SSE-NEXT:    movaps 528(%rdi), %xmm3
; SSE-NEXT:    movaps 480(%rdi), %xmm10
; SSE-NEXT:    movaps 432(%rdi), %xmm4
; SSE-NEXT:    movaps 384(%rdi), %xmm11
; SSE-NEXT:    movaps 336(%rdi), %xmm5
; SSE-NEXT:    movaps 288(%rdi), %xmm12
; SSE-NEXT:    movaps 240(%rdi), %xmm6
; SSE-NEXT:    movaps 192(%rdi), %xmm13
; SSE-NEXT:    movaps 144(%rdi), %xmm7
; SSE-NEXT:    movaps 96(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm7[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm13, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm4[1]
; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm3[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm2[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 768(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 912(%rdi), %xmm0
; SSE-NEXT:    movaps 864(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1008(%rdi), %xmm0
; SSE-NEXT:    movaps 960(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1104(%rdi), %xmm0
; SSE-NEXT:    movaps 1056(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1200(%rdi), %xmm0
; SSE-NEXT:    movaps 1152(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1296(%rdi), %xmm0
; SSE-NEXT:    movaps 1248(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1392(%rdi), %xmm0
; SSE-NEXT:    movaps 1344(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1488(%rdi), %xmm0
; SSE-NEXT:    movaps 1440(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1584(%rdi), %xmm0
; SSE-NEXT:    movaps 1536(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1680(%rdi), %xmm0
; SSE-NEXT:    movaps 1632(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1776(%rdi), %xmm0
; SSE-NEXT:    movaps 1728(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1872(%rdi), %xmm0
; SSE-NEXT:    movaps 1824(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1968(%rdi), %xmm0
; SSE-NEXT:    movaps 1920(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2064(%rdi), %xmm0
; SSE-NEXT:    movaps 2016(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2160(%rdi), %xmm0
; SSE-NEXT:    movaps 2112(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2256(%rdi), %xmm0
; SSE-NEXT:    movaps 2208(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2352(%rdi), %xmm0
; SSE-NEXT:    movaps 2304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2448(%rdi), %xmm0
; SSE-NEXT:    movaps 2400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2544(%rdi), %xmm0
; SSE-NEXT:    movaps 2496(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2640(%rdi), %xmm0
; SSE-NEXT:    movaps 2592(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2736(%rdi), %xmm0
; SSE-NEXT:    movaps 2688(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2832(%rdi), %xmm0
; SSE-NEXT:    movaps 2784(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2928(%rdi), %xmm0
; SSE-NEXT:    movaps 2880(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3024(%rdi), %xmm0
; SSE-NEXT:    movaps 2976(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps (%rdi), %xmm1
; SSE-NEXT:    movaps 48(%rdi), %xmm0
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 64(%rdi), %xmm0
; SSE-NEXT:    movaps 16(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 160(%rdi), %xmm0
; SSE-NEXT:    movaps 112(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 256(%rdi), %xmm0
; SSE-NEXT:    movaps 208(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 448(%rdi), %xmm0
; SSE-NEXT:    movaps 400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 544(%rdi), %xmm0
; SSE-NEXT:    movaps 496(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 640(%rdi), %xmm0
; SSE-NEXT:    movaps 592(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 736(%rdi), %xmm0
; SSE-NEXT:    movaps 688(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 832(%rdi), %xmm0
; SSE-NEXT:    movaps 784(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 928(%rdi), %xmm0
; SSE-NEXT:    movaps 880(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1024(%rdi), %xmm0
; SSE-NEXT:    movaps 976(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1120(%rdi), %xmm0
; SSE-NEXT:    movaps 1072(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1216(%rdi), %xmm0
; SSE-NEXT:    movaps 1168(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1312(%rdi), %xmm0
; SSE-NEXT:    movaps 1264(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1408(%rdi), %xmm0
; SSE-NEXT:    movaps 1360(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1504(%rdi), %xmm0
; SSE-NEXT:    movaps 1456(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1600(%rdi), %xmm0
; SSE-NEXT:    movaps 1552(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1696(%rdi), %xmm0
; SSE-NEXT:    movaps 1648(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1792(%rdi), %xmm0
; SSE-NEXT:    movaps 1744(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1888(%rdi), %xmm0
; SSE-NEXT:    movaps 1840(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1984(%rdi), %xmm0
; SSE-NEXT:    movaps 1936(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2080(%rdi), %xmm0
; SSE-NEXT:    movaps 2032(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2176(%rdi), %xmm0
; SSE-NEXT:    movaps 2128(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2272(%rdi), %xmm0
; SSE-NEXT:    movaps 2224(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2368(%rdi), %xmm0
; SSE-NEXT:    movaps 2320(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2464(%rdi), %xmm0
; SSE-NEXT:    movaps 2416(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2560(%rdi), %xmm0
; SSE-NEXT:    movaps 2512(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2656(%rdi), %xmm0
; SSE-NEXT:    movaps 2608(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2752(%rdi), %xmm0
; SSE-NEXT:    movaps 2704(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2848(%rdi), %xmm0
; SSE-NEXT:    movaps 2800(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2944(%rdi), %xmm0
; SSE-NEXT:    movaps 2896(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3040(%rdi), %xmm0
; SSE-NEXT:    movaps 2992(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 176(%rdi), %xmm0
; SSE-NEXT:    movaps 128(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 272(%rdi), %xmm0
; SSE-NEXT:    movaps 224(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 320(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 464(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 560(%rdi), %xmm0
; SSE-NEXT:    movaps 512(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 656(%rdi), %xmm0
; SSE-NEXT:    movaps 608(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 752(%rdi), %xmm0
; SSE-NEXT:    movaps 704(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 848(%rdi), %xmm0
; SSE-NEXT:    movaps 800(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 944(%rdi), %xmm0
; SSE-NEXT:    movaps 896(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1040(%rdi), %xmm0
; SSE-NEXT:    movaps 992(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1136(%rdi), %xmm0
; SSE-NEXT:    movaps 1088(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1232(%rdi), %xmm0
; SSE-NEXT:    movaps 1184(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1328(%rdi), %xmm0
; SSE-NEXT:    movaps 1280(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1424(%rdi), %xmm0
; SSE-NEXT:    movaps 1376(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1520(%rdi), %xmm0
; SSE-NEXT:    movaps 1472(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1616(%rdi), %xmm0
; SSE-NEXT:    movaps 1568(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1712(%rdi), %xmm0
; SSE-NEXT:    movaps 1664(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1808(%rdi), %xmm0
; SSE-NEXT:    movaps 1760(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1904(%rdi), %xmm0
; SSE-NEXT:    movaps 1856(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2000(%rdi), %xmm0
; SSE-NEXT:    movaps 1952(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2096(%rdi), %xmm0
; SSE-NEXT:    movaps 2048(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2192(%rdi), %xmm0
; SSE-NEXT:    movaps 2144(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
; SSE-NEXT:    movaps 2288(%rdi), %xmm0
; SSE-NEXT:    movaps 2240(%rdi), %xmm15
; SSE-NEXT:    movaps %xmm15, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
; SSE-NEXT:    movaps 2384(%rdi), %xmm0
; SSE-NEXT:    movaps 2336(%rdi), %xmm12
; SSE-NEXT:    movaps %xmm12, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]
; SSE-NEXT:    movaps 2480(%rdi), %xmm0
; SSE-NEXT:    movaps 2432(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
; SSE-NEXT:    movaps 2576(%rdi), %xmm0
; SSE-NEXT:    movaps 2528(%rdi), %xmm10
; SSE-NEXT:    movaps %xmm10, %xmm13
; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT:    movaps 2672(%rdi), %xmm0
; SSE-NEXT:    movaps 2624(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm11
; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps 2768(%rdi), %xmm0
; SSE-NEXT:    movaps 2720(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 2864(%rdi), %xmm0
; SSE-NEXT:    movaps 2816(%rdi), %xmm6
; SSE-NEXT:    movaps %xmm6, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]
; SSE-NEXT:    movaps 2960(%rdi), %xmm0
; SSE-NEXT:    movaps 2912(%rdi), %xmm3
; SSE-NEXT:    movaps %xmm3, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
; SSE-NEXT:    movaps 3056(%rdi), %xmm0
; SSE-NEXT:    movaps 3008(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 496(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 480(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 464(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 448(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 432(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 416(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 400(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 384(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 496(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 480(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 464(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 448(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 432(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 416(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 400(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 384(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 496(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 480(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 464(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 448(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 432(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 416(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 400(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 384(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 496(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 480(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 464(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 448(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 432(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 416(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 400(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 384(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps %xmm2, 496(%r9)
; SSE-NEXT:    movaps %xmm4, 480(%r9)
; SSE-NEXT:    movaps %xmm7, 464(%r9)
; SSE-NEXT:    movaps %xmm8, 448(%r9)
; SSE-NEXT:    movaps %xmm11, 432(%r9)
; SSE-NEXT:    movaps %xmm13, 416(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 400(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 384(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 368(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm1, 496(%rax)
; SSE-NEXT:    movaps %xmm3, 480(%rax)
; SSE-NEXT:    movaps %xmm6, 464(%rax)
; SSE-NEXT:    movaps %xmm5, 448(%rax)
; SSE-NEXT:    movaps %xmm9, 432(%rax)
; SSE-NEXT:    movaps %xmm10, 416(%rax)
; SSE-NEXT:    movaps %xmm14, 400(%rax)
; SSE-NEXT:    movaps %xmm12, 384(%rax)
; SSE-NEXT:    movaps %xmm15, 368(%rax)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 352(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 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 $2712, %rsp # imm = 0xA98
; SSE-NEXT:    retq
;
; AVX1-ONLY-LABEL: load_i64_stride6_vf64:
; AVX1-ONLY:       # %bb.0:
; AVX1-ONLY-NEXT:    subq $3768, %rsp # imm = 0xEB8
; AVX1-ONLY-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX1-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 704(%rdi), %ymm3
; AVX1-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 320(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 240(%rdi), %xmm4
; AVX1-ONLY-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 192(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 672(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX1-ONLY-NEXT:    vmovaps 624(%rdi), %xmm3
; AVX1-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 576(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1056(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX1-ONLY-NEXT:    vmovaps 1008(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 960(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1392(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1344(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1856(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1824(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1776(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1728(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2240(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2208(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 2160(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2112(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2624(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2592(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 2544(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2496(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 3008(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2976(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 2928(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2880(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 128(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps (%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 48(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 512(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 432(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 384(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 896(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 864(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 816(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 768(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1280(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1200(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1152(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1664(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1632(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1584(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovaps 1536(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2048(%rdi), %ymm12
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm11
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX1-ONLY-NEXT:    vmovaps 1968(%rdi), %xmm10
; AVX1-ONLY-NEXT:    vmovaps 1920(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2432(%rdi), %ymm8
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2400(%rdi), %ymm0, %ymm7
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX1-ONLY-NEXT:    vmovaps 2352(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps 2304(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2816(%rdi), %ymm4
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2784(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX1-ONLY-NEXT:    vmovaps 2736(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps 2688(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 288(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 256(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 208(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 672(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 640(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 592(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1056(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1120(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 976(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1440(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1408(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1360(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1824(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1888(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1792(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1744(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2208(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2272(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 2176(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2128(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2592(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2656(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 2560(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2512(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2976(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 3040(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 2944(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2896(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2784(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2848(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 2752(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2704(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2400(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2464(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 2368(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2320(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2016(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2080(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1984(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1936(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1632(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1600(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1552(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1312(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT:    vmovaps 1216(%rdi), %xmm12
; AVX1-ONLY-NEXT:    vmovaps 1168(%rdi), %xmm11
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 864(%rdi), %ymm10
; AVX1-ONLY-NEXT:    vinsertf128 $1, 928(%rdi), %ymm0, %ymm9
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX1-ONLY-NEXT:    vmovaps 832(%rdi), %xmm8
; AVX1-ONLY-NEXT:    vmovaps 784(%rdi), %xmm7
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm7[0],xmm8[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 480(%rdi), %ymm6
; AVX1-ONLY-NEXT:    vinsertf128 $1, 544(%rdi), %ymm0, %ymm15
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm15[0],ymm6[2],ymm15[2]
; AVX1-ONLY-NEXT:    vmovaps 448(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovaps 400(%rdi), %xmm4
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm4[0],xmm5[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX1-ONLY-NEXT:    vmovaps 16(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm1
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm14[0],xmm0[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm13[0,1,2,3],ymm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm14[1],xmm0[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm15[1],ymm6[3],ymm15[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm5[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm8[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm11[1],xmm12[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 160(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 32(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 352(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 272(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 224(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 544(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 512(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 464(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 656(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 608(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 928(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 896(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 848(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1120(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1088(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1040(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 992(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1312(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1280(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1232(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1184(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1504(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1424(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1376(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1696(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1616(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1888(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 1856(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 1808(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1760(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2080(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2048(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 2000(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 1952(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2272(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2240(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 2192(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2144(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2464(%rdi), %ymm1
; AVX1-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2432(%rdi), %ymm0, %ymm0
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX1-ONLY-NEXT:    vmovaps 2384(%rdi), %xmm14
; AVX1-ONLY-NEXT:    vmovaps 2336(%rdi), %xmm13
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2656(%rdi), %ymm12
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2624(%rdi), %ymm0, %ymm11
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX1-ONLY-NEXT:    vmovaps 2576(%rdi), %xmm10
; AVX1-ONLY-NEXT:    vmovaps 2528(%rdi), %xmm9
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 2848(%rdi), %ymm8
; AVX1-ONLY-NEXT:    vinsertf128 $1, 2816(%rdi), %ymm0, %ymm7
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX1-ONLY-NEXT:    vmovaps 2768(%rdi), %xmm6
; AVX1-ONLY-NEXT:    vmovaps 2720(%rdi), %xmm5
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovaps 3040(%rdi), %ymm4
; AVX1-ONLY-NEXT:    vinsertf128 $1, 3008(%rdi), %ymm0, %ymm3
; AVX1-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX1-ONLY-NEXT:    vmovaps 2960(%rdi), %xmm2
; AVX1-ONLY-NEXT:    vmovaps 2912(%rdi), %xmm1
; AVX1-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX1-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX1-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX1-ONLY-NEXT:    vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm3[4,5,6,7]
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 448(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 384(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 320(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 256(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 480(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 416(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 352(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 288(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 448(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 384(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 320(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 256(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 480(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 416(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 352(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 288(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 256(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 320(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 384(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 448(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 480(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 416(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 352(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 288(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 480(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 448(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 416(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 384(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 352(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 320(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 288(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 256(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%r8)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 480(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 448(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 416(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 384(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 352(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 320(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 288(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 256(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm1, (%r9)
; AVX1-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT:    vmovaps %ymm4, 480(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm5, 448(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 416(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm13, 384(%rax)
; AVX1-ONLY-NEXT:    vmovaps %ymm15, 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, 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 (%rsp), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 128(%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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX1-ONLY-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX1-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %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 $3768, %rsp # imm = 0xEB8
; AVX1-ONLY-NEXT:    vzeroupper
; AVX1-ONLY-NEXT:    retq
;
; AVX2-ONLY-LABEL: load_i64_stride6_vf64:
; AVX2-ONLY:       # %bb.0:
; AVX2-ONLY-NEXT:    subq $3432, %rsp # imm = 0xD68
; AVX2-ONLY-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1056(%rdi), %ymm4
; AVX2-ONLY-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 704(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 672(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 320(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 288(%rdi), %ymm6
; AVX2-ONLY-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 240(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 192(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm7[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 624(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1008(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 960(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1472(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1440(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1392(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1344(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1856(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1824(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1776(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1728(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2240(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2208(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2160(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2112(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2624(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2592(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2544(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2496(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 3008(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2976(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2928(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2880(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 128(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 96(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps (%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 48(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 512(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 432(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 384(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 864(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 816(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 768(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1280(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1200(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1152(%rdi), %xmm9
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1664(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1632(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1584(%rdi), %xmm14
; AVX2-ONLY-NEXT:    vmovaps 1536(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm14[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2048(%rdi), %ymm12
; AVX2-ONLY-NEXT:    vmovaps 2016(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX2-ONLY-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1968(%rdi), %xmm10
; AVX2-ONLY-NEXT:    vmovaps 1920(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm10[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2432(%rdi), %ymm8
; AVX2-ONLY-NEXT:    vmovaps 2400(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX2-ONLY-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2352(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovaps 2304(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2816(%rdi), %ymm4
; AVX2-ONLY-NEXT:    vmovaps 2784(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
; AVX2-ONLY-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2736(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps 2688(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 296(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 680(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1064(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1448(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1832(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 2216(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 2600(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 2984(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 104(%rdi), %ymm15
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm13[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 488(%rdi), %ymm13
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm11[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 872(%rdi), %ymm11
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm9[1],mem[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1256(%rdi), %ymm9
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm9 = ymm9[1],mem[1],ymm9[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm14[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 1640(%rdi), %ymm7
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm10[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 2024(%rdi), %ymm5
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm12[1],ymm5[3],ymm12[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm6[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 2408(%rdi), %ymm3
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm8[1],ymm3[3],ymm8[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
; AVX2-ONLY-NEXT:    vbroadcastsd 2792(%rdi), %ymm1
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 352(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 256(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 208(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 736(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 640(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 592(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1120(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 976(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1504(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1408(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1360(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1888(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1792(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1744(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 2272(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 2176(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2128(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 2656(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 2560(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2512(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 3040(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 2944(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2896(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 2848(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 2752(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2704(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 2464(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 2368(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2320(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 2080(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1984(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1936(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1696(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1600(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps 1552(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1312(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 1216(%rdi), %xmm9
; AVX2-ONLY-NEXT:    vmovaps 1168(%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 928(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 832(%rdi), %xmm6
; AVX2-ONLY-NEXT:    vmovaps 784(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm5[0],xmm6[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 544(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 448(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovaps 400(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm3[0],xmm13[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 160(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]
; AVX2-ONLY-NEXT:    vmovaps 16(%rdi), %xmm15
; AVX2-ONLY-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm15[0],xmm0[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm15[1],xmm0[1]
; AVX2-ONLY-NEXT:    vmovaps 160(%rdi), %ymm14
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm14[1],ymm2[3],ymm14[3]
; AVX2-ONLY-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm15
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm13[1]
; AVX2-ONLY-NEXT:    vmovaps 544(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm3
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 736(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm4
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm6[1]
; AVX2-ONLY-NEXT:    vmovaps 928(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm2[1],ymm7[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm5
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm6
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm9[1]
; AVX2-ONLY-NEXT:    vmovaps 1312(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm10[1],ymm2[1],ymm10[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm7
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 1504(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm8
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm11[1],xmm12[1]
; AVX2-ONLY-NEXT:    vmovaps 1696(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm9
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 1888(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm10
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 2080(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovaps %ymm2, %ymm11
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 2272(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 2464(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 2656(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 2848(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm0 = xmm0[1],mem[1]
; AVX2-ONLY-NEXT:    vmovaps 3040(%rdi), %ymm2
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 32(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 272(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps %xmm12, (%rsp) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 224(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 464(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 656(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 608(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 848(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1040(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 992(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1232(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1184(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1424(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1376(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1616(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 1808(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1760(%rdi), %xmm13
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm11[0],ymm0[2],ymm11[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2000(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 1952(%rdi), %xmm11
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2192(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT:    vmovaps 2144(%rdi), %xmm9
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2384(%rdi), %xmm12
; AVX2-ONLY-NEXT:    vmovaps 2336(%rdi), %xmm7
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm12[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2576(%rdi), %xmm8
; AVX2-ONLY-NEXT:    vmovaps 2528(%rdi), %xmm5
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2768(%rdi), %xmm4
; AVX2-ONLY-NEXT:    vmovaps 2720(%rdi), %xmm3
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm4[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX2-ONLY-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
; AVX2-ONLY-NEXT:    vmovaps 2960(%rdi), %xmm2
; AVX2-ONLY-NEXT:    vmovaps 2912(%rdi), %xmm1
; AVX2-ONLY-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 136(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 328(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 520(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 712(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 904(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1096(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1288(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1480(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT:    vbroadcastsd 1672(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm15 = xmm15[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 1864(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm13 = xmm13[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 2056(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm11 = xmm11[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 2248(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
; AVX2-ONLY-NEXT:    # xmm9 = xmm9[1],mem[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 2440(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm14[1],ymm0[3],ymm14[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm12[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 2632(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 2824(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm6[1],ymm0[3],ymm6[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vbroadcastsd 3016(%rdi), %ymm0
; AVX2-ONLY-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-ONLY-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX2-ONLY-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 448(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 384(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 320(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 256(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 480(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 416(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 352(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 288(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 448(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 384(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 320(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 256(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 480(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 416(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 352(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 288(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 256(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 320(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 384(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 448(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 480(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 416(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 352(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 288(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 480(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 448(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 416(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 384(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 352(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 320(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 288(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 256(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 480(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 448(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 416(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 384(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 352(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 320(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 288(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 256(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-ONLY-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm3, 448(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm5, 416(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm7, 384(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm9, 352(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm11, 320(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm13, 288(%rax)
; AVX2-ONLY-NEXT:    vmovaps %ymm15, 256(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-ONLY-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-ONLY-NEXT:    addq $3432, %rsp # imm = 0xD68
; AVX2-ONLY-NEXT:    vzeroupper
; AVX2-ONLY-NEXT:    retq
;
; AVX512F-LABEL: load_i64_stride6_vf64:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    subq $7240, %rsp # imm = 0x1C48
; AVX512F-NEXT:    vmovdqa64 2048(%rdi), %zmm3
; AVX512F-NEXT:    vmovdqa64 1280(%rdi), %zmm4
; AVX512F-NEXT:    vmovdqa64 1344(%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 896(%rdi), %zmm5
; AVX512F-NEXT:    vmovdqa64 960(%rdi), %zmm26
; AVX512F-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 576(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512F-NEXT:    vmovdqa64 192(%rdi), %zmm29
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,0,10,0,6,0,10]
; AVX512F-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm29, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm6, %zmm7, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm7, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm7, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm8
; AVX512F-NEXT:    vpermt2q %zmm4, %zmm7, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,0,11,1,7,0,11]
; AVX512F-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm9
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm8, %zmm9
; AVX512F-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm29, %zmm9
; AVX512F-NEXT:    vpermt2q %zmm6, %zmm8, %zmm9
; AVX512F-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm9
; AVX512F-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512F-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm8, %zmm9
; AVX512F-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm9 = [10,4,10,4,10,4,10,4]
; AVX512F-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm10
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm10
; AVX512F-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm10
; AVX512F-NEXT:    vpermt2q %zmm29, %zmm9, %zmm10
; AVX512F-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm10
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm9, %zmm10
; AVX512F-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm10
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm9, %zmm10
; AVX512F-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [11,5,11,5,11,5,11,5]
; AVX512F-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm11
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm10, %zmm11
; AVX512F-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm11
; AVX512F-NEXT:    vpermt2q %zmm29, %zmm10, %zmm11
; AVX512F-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm10, %zmm11
; AVX512F-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm10, %zmm11
; AVX512F-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [12,0,0,6,12,0,0,6]
; AVX512F-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm13
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm12
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm11, %zmm13
; AVX512F-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [13,0,1,7,13,0,1,7]
; AVX512F-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm2, %zmm12
; AVX512F-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm29, %zmm11, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm29, %zmm2, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm11, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm2, %zmm4
; AVX512F-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm11, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm2, %zmm5
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2112(%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm3, %zmm7, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm3, %zmm8, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm9, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm10, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm11, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm2, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 1728(%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm7, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm9, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2432(%rdi), %zmm3
; AVX512F-NEXT:    vmovdqa64 2496(%rdi), %zmm0
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm3, %zmm7, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2816(%rdi), %zmm4
; AVX512F-NEXT:    vmovdqa64 2880(%rdi), %zmm1
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm1, %zmm7
; AVX512F-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5
; AVX512F-NEXT:    vpermt2q %zmm3, %zmm8, %zmm5
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm4, %zmm1, %zmm8
; AVX512F-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm9, %zmm5
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm4, %zmm9
; AVX512F-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm10, %zmm5
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm4, %zmm10
; AVX512F-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm11, %zmm5
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm1, %zmm4, %zmm11
; AVX512F-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm2, %zmm4
; AVX512F-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm0, %zmm2, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 448(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 384(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm8 = <0,6,12,u>
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm9 = <1,7,13,u>
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm10 = <10,0,6,u>
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm10, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm11 = <11,1,7,u>
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm11, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm4 = [4,10,4,10,4,10,4,10]
; AVX512F-NEXT:    # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]
; AVX512F-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 64(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1216(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 1152(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 832(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 768(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1984(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 1920(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1600(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 1536(%rdi), %zmm30
; AVX512F-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm30, %zmm10, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm30, %zmm11, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm30
; AVX512F-NEXT:    vmovdqa64 320(%rdi), %zmm26
; AVX512F-NEXT:    vmovdqa64 256(%rdi), %zmm18
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 704(%rdi), %zmm21
; AVX512F-NEXT:    vmovdqa64 640(%rdi), %zmm13
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm21, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1088(%rdi), %zmm24
; AVX512F-NEXT:    vmovdqa64 1024(%rdi), %zmm19
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm24, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1472(%rdi), %zmm28
; AVX512F-NEXT:    vmovdqa64 1408(%rdi), %zmm16
; AVX512F-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm28, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 1856(%rdi), %zmm25
; AVX512F-NEXT:    vmovdqa64 1792(%rdi), %zmm7
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2240(%rdi), %zmm23
; AVX512F-NEXT:    vmovdqa64 2176(%rdi), %zmm17
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm23, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2624(%rdi), %zmm20
; AVX512F-NEXT:    vmovdqa64 2560(%rdi), %zmm5
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm20, %zmm4, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2368(%rdi), %zmm1
; AVX512F-NEXT:    vmovdqa64 2304(%rdi), %zmm22
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm29
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm8, %zmm29
; AVX512F-NEXT:    vmovdqa64 3008(%rdi), %zmm14
; AVX512F-NEXT:    vmovdqa64 2944(%rdi), %zmm31
; AVX512F-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512F-NEXT:    vpermt2q %zmm14, %zmm4, %zmm2
; AVX512F-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 2752(%rdi), %zmm2
; AVX512F-NEXT:    vmovdqa64 2688(%rdi), %zmm12
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm12, %zmm8
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm12, %zmm9
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm22, %zmm10, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm12, %zmm2, %zmm10
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm22, %zmm11, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm12, %zmm2, %zmm11
; AVX512F-NEXT:    vmovdqa64 %zmm22, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512F-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermi2q %zmm2, %zmm12, %zmm4
; AVX512F-NEXT:    vpermt2q %zmm2, %zmm0, %zmm12
; AVX512F-NEXT:    vpermt2q %zmm1, %zmm0, %zmm22
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm21, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm28, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm24, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm23, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm31, %zmm1
; AVX512F-NEXT:    vpermt2q %zmm14, %zmm0, %zmm1
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]
; AVX512F-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm21, %zmm1, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]
; AVX512F-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm21, %zmm2, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]
; AVX512F-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vmovdqa64 %zmm21, %zmm0
; AVX512F-NEXT:    vpermt2q %zmm13, %zmm3, %zmm0
; AVX512F-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]
; AVX512F-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]
; AVX512F-NEXT:    vpermt2q %zmm13, %zmm0, %zmm21
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm1, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm26, %zmm2, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm18, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm18, %zmm0, %zmm26
; AVX512F-NEXT:    vmovdqa64 %zmm16, %zmm27
; AVX512F-NEXT:    vpermt2q %zmm28, %zmm1, %zmm27
; AVX512F-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm28, %zmm2, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm28, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm16, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm16, %zmm0, %zmm28
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm18
; AVX512F-NEXT:    vpermt2q %zmm24, %zmm1, %zmm18
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm24, %zmm2, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqa64 %zmm24, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm19, %zmm0, %zmm24
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm15
; AVX512F-NEXT:    vpermt2q %zmm23, %zmm1, %zmm15
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm19
; AVX512F-NEXT:    vpermt2q %zmm23, %zmm2, %zmm19
; AVX512F-NEXT:    vmovdqa64 %zmm23, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm17, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm17, %zmm0, %zmm23
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm13
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm1, %zmm13
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm17
; AVX512F-NEXT:    vpermt2q %zmm25, %zmm2, %zmm17
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm7, %zmm3, %zmm6
; AVX512F-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vpermt2q %zmm7, %zmm0, %zmm25
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm6
; AVX512F-NEXT:    vpermt2q %zmm20, %zmm1, %zmm6
; AVX512F-NEXT:    vpermi2q %zmm14, %zmm31, %zmm1
; AVX512F-NEXT:    vmovdqa64 %zmm5, %zmm7
; AVX512F-NEXT:    vpermt2q %zmm20, %zmm2, %zmm7
; AVX512F-NEXT:    vpermi2q %zmm14, %zmm31, %zmm2
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm16
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm3, %zmm16
; AVX512F-NEXT:    vpermi2q %zmm31, %zmm14, %zmm3
; AVX512F-NEXT:    vpermt2q %zmm31, %zmm0, %zmm14
; AVX512F-NEXT:    vpermt2q %zmm5, %zmm0, %zmm20
; AVX512F-NEXT:    movb $56, %al
; AVX512F-NEXT:    kmovw %eax, %k1
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    movb $-64, %al
; AVX512F-NEXT:    kmovw %eax, %k2
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}
; AVX512F-NEXT:    movb $24, %al
; AVX512F-NEXT:    kmovw %eax, %k2
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    movb $-32, %al
; AVX512F-NEXT:    kmovw %eax, %k1
; AVX512F-NEXT:    vmovdqa64 %zmm6, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm27, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm18, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm15, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm7, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm19, %zmm31 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm17, %zmm27 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm11 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm5 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm6 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm7 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm7 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm13 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm13 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vmovdqa64 %zmm2, %zmm15 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm4, %zmm2, %zmm4
; AVX512F-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm21, %zmm2 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm26, %zmm3 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm16 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm28, %zmm16 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm24, %zmm17 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm18 # 16-byte Folded Reload
; AVX512F-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm30, %zmm19, %zmm19
; AVX512F-NEXT:    vmovdqa64 %zmm25, %zmm19 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm12, %zmm21, %zmm12
; AVX512F-NEXT:    vmovdqa64 %zmm14, %zmm12 {%k1}
; AVX512F-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-NEXT:    vinserti32x4 $0, %xmm22, %zmm14, %zmm14
; AVX512F-NEXT:    vmovdqa64 %zmm20, %zmm14 {%k1}
; AVX512F-NEXT:    vmovdqa64 %zmm8, 448(%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm29, 384(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 320(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 256(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 192(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 128(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 64(%rsi)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, (%rsi)
; AVX512F-NEXT:    vmovdqa64 %zmm9, 448(%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 256(%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 320(%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 128(%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 192(%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, (%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 64(%rdx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 384(%rdx)
; AVX512F-NEXT:    vmovdqa64 %zmm10, 448(%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 256(%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 320(%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 128(%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 192(%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, (%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 64(%rcx)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 384(%rcx)
; AVX512F-NEXT:    vmovdqa64 %zmm11, 448(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm27, 256(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm31, 320(%r8)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 128(%r8)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 192(%r8)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, (%r8)
; AVX512F-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 64(%r8)
; AVX512F-NEXT:    vmovups (%rsp), %zmm8 # 64-byte Reload
; AVX512F-NEXT:    vmovaps %zmm8, 384(%r8)
; AVX512F-NEXT:    vmovdqa64 %zmm4, 448(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm15, 256(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm13, 320(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm7, 128(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm6, 192(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm5, (%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm1, 64(%r9)
; AVX512F-NEXT:    vmovdqa64 %zmm0, 384(%r9)
; AVX512F-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT:    vmovdqa64 %zmm14, 384(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm12, 448(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm19, 256(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm18, 320(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm17, 128(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm16, 192(%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512F-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512F-NEXT:    addq $7240, %rsp # imm = 0x1C48
; AVX512F-NEXT:    vzeroupper
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride6_vf64:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    subq $7240, %rsp # imm = 0x1C48
; AVX512BW-NEXT:    vmovdqa64 2048(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 1280(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqa64 1344(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 896(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 960(%rdi), %zmm26
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm29
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,0,10,0,6,0,10]
; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm7, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,0,11,1,7,0,11]
; AVX512BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm8, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm8, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm8, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm9 = [10,4,10,4,10,4,10,4]
; AVX512BW-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm9, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm9, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm9, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [11,5,11,5,11,5,11,5]
; AVX512BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm10, %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm10, %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm10, %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [12,0,0,6,12,0,0,6]
; AVX512BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm11, %zmm13
; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [13,0,1,7,13,0,1,7]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm12
; AVX512BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm11, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm11, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm11, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm2, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2112(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm7, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm8, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm9, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm10, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm11, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 1728(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm9, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2432(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 2496(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm7, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2816(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqa64 2880(%rdi), %zmm1
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm1, %zmm7
; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm8, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm1, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm9, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm4, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm10, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm4, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm11, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm4, %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm8 = <0,6,12,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm9 = <1,7,13,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm10 = <10,0,6,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm10, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa {{.*#+}} ymm11 = <11,1,7,u>
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm4 = [4,10,4,10,4,10,4,10]
; AVX512BW-NEXT:    # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1216(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 1152(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 832(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 768(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1984(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 1920(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1600(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 1536(%rdi), %zmm30
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm10, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm11, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm30
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm26
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm18
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm21
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1088(%rdi), %zmm24
; AVX512BW-NEXT:    vmovdqa64 1024(%rdi), %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1472(%rdi), %zmm28
; AVX512BW-NEXT:    vmovdqa64 1408(%rdi), %zmm16
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1856(%rdi), %zmm25
; AVX512BW-NEXT:    vmovdqa64 1792(%rdi), %zmm7
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2240(%rdi), %zmm23
; AVX512BW-NEXT:    vmovdqa64 2176(%rdi), %zmm17
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2624(%rdi), %zmm20
; AVX512BW-NEXT:    vmovdqa64 2560(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2368(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 2304(%rdi), %zmm22
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm29
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm8, %zmm29
; AVX512BW-NEXT:    vmovdqa64 3008(%rdi), %zmm14
; AVX512BW-NEXT:    vmovdqa64 2944(%rdi), %zmm31
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm14, %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2752(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 2688(%rdi), %zmm12
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm12, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm9, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm12, %zmm9
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm22, %zmm10, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm12, %zmm2, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm22, %zmm11, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm12, %zmm2, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm4, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm12, %zmm4
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm22
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm14, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]
; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]
; AVX512BW-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm3, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm21
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm1, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm26
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm27
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm1, %zmm27
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm28
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm18
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm1, %zmm18
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm24
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm1, %zmm15
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm2, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm17, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm17, %zmm0, %zmm23
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm17
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm2, %zmm17
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm3, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm25
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm1, %zmm6
; AVX512BW-NEXT:    vpermi2q %zmm14, %zmm31, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm7
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm2, %zmm7
; AVX512BW-NEXT:    vpermi2q %zmm14, %zmm31, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm16
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm3, %zmm16
; AVX512BW-NEXT:    vpermi2q %zmm31, %zmm14, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm14
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm20
; AVX512BW-NEXT:    movb $56, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    movb $-64, %al
; AVX512BW-NEXT:    kmovd %eax, %k2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}
; AVX512BW-NEXT:    movb $24, %al
; AVX512BW-NEXT:    kmovd %eax, %k2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    movb $-32, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm31 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm27 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm11 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm5 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm6 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm7 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm7 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm13 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm13 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm15 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm4, %zmm2, %zmm4
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm2 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm3 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm16 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm16 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm17 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm18 # 16-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm30, %zmm19, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm19 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm12, %zmm21, %zmm12
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm12 {%k1}
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-NEXT:    vinserti32x4 $0, %xmm22, %zmm14, %zmm14
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm14 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %zmm8, 448(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm29, 384(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 320(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 256(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 192(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 128(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 64(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm9, 448(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 256(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 320(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 128(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 192(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, (%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 64(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 384(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm10, 448(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 256(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 320(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 128(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 192(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, (%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 64(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 384(%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm11, 448(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm27, 256(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm31, 320(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 128(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 192(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, (%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 64(%r8)
; AVX512BW-NEXT:    vmovups (%rsp), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm8, 384(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm4, 448(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm15, 256(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm13, 320(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm7, 128(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm6, 192(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, (%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, 384(%r9)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm14, 384(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm12, 448(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm19, 256(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm18, 320(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm17, 128(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm16, 192(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512BW-NEXT:    addq $7240, %rsp # imm = 0x1C48
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
  %wide.vec = load <384 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <384 x i64> %wide.vec, <384 x i64> poison, <64 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90, i32 96, i32 102, i32 108, i32 114, i32 120, i32 126, i32 132, i32 138, i32 144, i32 150, i32 156, i32 162, i32 168, i32 174, i32 180, i32 186, i32 192, i32 198, i32 204, i32 210, i32 216, i32 222, i32 228, i32 234, i32 240, i32 246, i32 252, i32 258, i32 264, i32 270, i32 276, i32 282, i32 288, i32 294, i32 300, i32 306, i32 312, i32 318, i32 324, i32 330, i32 336, i32 342, i32 348, i32 354, i32 360, i32 366, i32 372, i32 378>
  %strided.vec1 = shufflevector <384 x i64> %wide.vec, <384 x i64> poison, <64 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91, i32 97, i32 103, i32 109, i32 115, i32 121, i32 127, i32 133, i32 139, i32 145, i32 151, i32 157, i32 163, i32 169, i32 175, i32 181, i32 187, i32 193, i32 199, i32 205, i32 211, i32 217, i32 223, i32 229, i32 235, i32 241, i32 247, i32 253, i32 259, i32 265, i32 271, i32 277, i32 283, i32 289, i32 295, i32 301, i32 307, i32 313, i32 319, i32 325, i32 331, i32 337, i32 343, i32 349, i32 355, i32 361, i32 367, i32 373, i32 379>
  %strided.vec2 = shufflevector <384 x i64> %wide.vec, <384 x i64> poison, <64 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92, i32 98, i32 104, i32 110, i32 116, i32 122, i32 128, i32 134, i32 140, i32 146, i32 152, i32 158, i32 164, i32 170, i32 176, i32 182, i32 188, i32 194, i32 200, i32 206, i32 212, i32 218, i32 224, i32 230, i32 236, i32 242, i32 248, i32 254, i32 260, i32 266, i32 272, i32 278, i32 284, i32 290, i32 296, i32 302, i32 308, i32 314, i32 320, i32 326, i32 332, i32 338, i32 344, i32 350, i32 356, i32 362, i32 368, i32 374, i32 380>
  %strided.vec3 = shufflevector <384 x i64> %wide.vec, <384 x i64> poison, <64 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93, i32 99, i32 105, i32 111, i32 117, i32 123, i32 129, i32 135, i32 141, i32 147, i32 153, i32 159, i32 165, i32 171, i32 177, i32 183, i32 189, i32 195, i32 201, i32 207, i32 213, i32 219, i32 225, i32 231, i32 237, i32 243, i32 249, i32 255, i32 261, i32 267, i32 273, i32 279, i32 285, i32 291, i32 297, i32 303, i32 309, i32 315, i32 321, i32 327, i32 333, i32 339, i32 345, i32 351, i32 357, i32 363, i32 369, i32 375, i32 381>
  %strided.vec4 = shufflevector <384 x i64> %wide.vec, <384 x i64> poison, <64 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94, i32 100, i32 106, i32 112, i32 118, i32 124, i32 130, i32 136, i32 142, i32 148, i32 154, i32 160, i32 166, i32 172, i32 178, i32 184, i32 190, i32 196, i32 202, i32 208, i32 214, i32 220, i32 226, i32 232, i32 238, i32 244, i32 250, i32 256, i32 262, i32 268, i32 274, i32 280, i32 286, i32 292, i32 298, i32 304, i32 310, i32 316, i32 322, i32 328, i32 334, i32 340, i32 346, i32 352, i32 358, i32 364, i32 370, i32 376, i32 382>
  %strided.vec5 = shufflevector <384 x i64> %wide.vec, <384 x i64> poison, <64 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95, i32 101, i32 107, i32 113, i32 119, i32 125, i32 131, i32 137, i32 143, i32 149, i32 155, i32 161, i32 167, i32 173, i32 179, i32 185, i32 191, i32 197, i32 203, i32 209, i32 215, i32 221, i32 227, i32 233, i32 239, i32 245, i32 251, i32 257, i32 263, i32 269, i32 275, i32 281, i32 287, i32 293, i32 299, i32 305, i32 311, i32 317, i32 323, i32 329, i32 335, i32 341, i32 347, i32 353, i32 359, i32 365, i32 371, i32 377, i32 383>
  store <64 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <64 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <64 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <64 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <64 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <64 x i64> %strided.vec5, ptr %out.vec5, align 64
  ret void
}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; AVX1: {{.*}}
; AVX2: {{.*}}
; AVX2-FAST: {{.*}}
; AVX2-FAST-PERLANE: {{.*}}
; AVX2-SLOW: {{.*}}
; AVX512: {{.*}}
; AVX512BW-FAST: {{.*}}
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512BW-ONLY-SLOW: {{.*}}
; AVX512BW-SLOW: {{.*}}
; AVX512DQ-FAST: {{.*}}
; AVX512DQ-SLOW: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512DQBW-SLOW: {{.*}}
; AVX512F-FAST: {{.*}}
; AVX512F-ONLY-FAST: {{.*}}
; AVX512F-ONLY-SLOW: {{.*}}
; AVX512F-SLOW: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
; FALLBACK11: {{.*}}
; FALLBACK12: {{.*}}
; FALLBACK2: {{.*}}
; FALLBACK3: {{.*}}
; FALLBACK4: {{.*}}
; FALLBACK5: {{.*}}
; FALLBACK6: {{.*}}
; FALLBACK7: {{.*}}
; FALLBACK8: {{.*}}
; FALLBACK9: {{.*}}