Google Git
Sign in
llvm / llvm-project / c9998ec145972714d581e19fd1c9cb276b836387 / . / llvm / test / CodeGen / X86 / vector-interleaved-load-i64-stride-8.ll
blob: 439265452cc47d52c33c3f29f3417841c328541d [file] [log] [blame]
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,FALLBACK0
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-ONLY,FALLBACK1
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-SLOW,FALLBACK2
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST,FALLBACK3
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST-PERLANE,FALLBACK4
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512F-ONLY-SLOW,FALLBACK5
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512F-ONLY-FAST,FALLBACK6
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512DQ-SLOW,FALLBACK7
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512DQ-FAST,FALLBACK8
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512BW-ONLY-SLOW,FALLBACK9
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512BW-ONLY-FAST,FALLBACK10
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512DQBW-SLOW,FALLBACK11
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512DQBW-FAST,FALLBACK12
; These patterns are produced by LoopVectorizer for interleaved loads.
define void @load_i64_stride8_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf2:
; SSE: # %bb.0:
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT: movaps 112(%rdi), %xmm0
; SSE-NEXT: movaps 96(%rdi), %xmm1
; SSE-NEXT: movaps 80(%rdi), %xmm2
; SSE-NEXT: movaps 64(%rdi), %xmm3
; SSE-NEXT: movaps (%rdi), %xmm4
; SSE-NEXT: movaps 16(%rdi), %xmm5
; SSE-NEXT: movaps 32(%rdi), %xmm6
; SSE-NEXT: movaps 48(%rdi), %xmm7
; SSE-NEXT: movaps %xmm4, %xmm8
; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm3[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]
; SSE-NEXT: movaps %xmm5, %xmm3
; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
; SSE-NEXT: movaps %xmm6, %xmm2
; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm1[1]
; SSE-NEXT: movaps %xmm7, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT: movaps %xmm8, (%rsi)
; SSE-NEXT: movaps %xmm4, (%rdx)
; SSE-NEXT: movaps %xmm3, (%rcx)
; SSE-NEXT: movaps %xmm5, (%r8)
; SSE-NEXT: movaps %xmm2, (%r9)
; SSE-NEXT: movaps %xmm6, (%r11)
; SSE-NEXT: movaps %xmm1, (%r10)
; SSE-NEXT: movaps %xmm7, (%rax)
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: load_i64_stride8_vf2:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps 80(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps 16(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm1[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm1[1]
; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm3[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm3[1]
; AVX1-ONLY-NEXT: vmovaps 112(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovaps 48(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm5[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm7[1],xmm5[1]
; AVX1-ONLY-NEXT: vmovaps %xmm2, (%rsi)
; AVX1-ONLY-NEXT: vmovaps %xmm0, (%rdx)
; AVX1-ONLY-NEXT: vmovaps %xmm4, (%rcx)
; AVX1-ONLY-NEXT: vmovaps %xmm1, (%r8)
; AVX1-ONLY-NEXT: vmovaps %xmm6, (%r9)
; AVX1-ONLY-NEXT: vmovaps %xmm3, (%r11)
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%r10)
; AVX1-ONLY-NEXT: vmovaps %xmm5, (%rax)
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: load_i64_stride8_vf2:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps (%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovaps (%rdi), %xmm4
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %xmm5
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %xmm6
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm7 = xmm4[0],xmm6[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %xmm3
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm3[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm3[1]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-ONLY-NEXT: vmovaps %xmm7, (%rsi)
; AVX2-ONLY-NEXT: vmovaps %xmm4, (%rdx)
; AVX2-ONLY-NEXT: vextractf128 $1, %ymm6, (%rcx)
; AVX2-ONLY-NEXT: vextractf128 $1, %ymm2, (%r8)
; AVX2-ONLY-NEXT: vmovaps %xmm8, (%r9)
; AVX2-ONLY-NEXT: vmovaps %xmm3, (%r11)
; AVX2-ONLY-NEXT: vextractf128 $1, %ymm5, (%r10)
; AVX2-ONLY-NEXT: vextractf128 $1, %ymm0, (%rax)
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512-LABEL: load_i64_stride8_vf2:
; AVX512: # %bb.0:
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT: vmovaps 64(%rdi), %xmm0
; AVX512-NEXT: vmovaps (%rdi), %xmm1
; AVX512-NEXT: vmovaps 16(%rdi), %xmm2
; AVX512-NEXT: vmovaps 32(%rdi), %xmm3
; AVX512-NEXT: vmovaps 48(%rdi), %xmm4
; AVX512-NEXT: vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512-NEXT: vmovaps 80(%rdi), %xmm1
; AVX512-NEXT: vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512-NEXT: vmovaps 96(%rdi), %xmm2
; AVX512-NEXT: vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512-NEXT: vmovaps 112(%rdi), %xmm3
; AVX512-NEXT: vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512-NEXT: vmovaps %xmm5, (%rsi)
; AVX512-NEXT: vmovaps %xmm0, (%rdx)
; AVX512-NEXT: vmovaps %xmm6, (%rcx)
; AVX512-NEXT: vmovaps %xmm1, (%r8)
; AVX512-NEXT: vmovaps %xmm7, (%r9)
; AVX512-NEXT: vmovaps %xmm2, (%r11)
; AVX512-NEXT: vmovaps %xmm8, (%r10)
; AVX512-NEXT: vmovaps %xmm3, (%rax)
; AVX512-NEXT: retq
%wide.vec = load <16 x i64>, ptr %in.vec, align 64
%strided.vec0 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 0, i32 8>
%strided.vec1 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 1, i32 9>
%strided.vec2 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 2, i32 10>
%strided.vec3 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 3, i32 11>
%strided.vec4 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 4, i32 12>
%strided.vec5 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 5, i32 13>
%strided.vec6 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 6, i32 14>
%strided.vec7 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 7, i32 15>
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
store <2 x i64> %strided.vec6, ptr %out.vec6, align 64
store <2 x i64> %strided.vec7, ptr %out.vec7, align 64
ret void
}
define void @load_i64_stride8_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf4:
; SSE: # %bb.0:
; SSE-NEXT: movaps 112(%rdi), %xmm5
; SSE-NEXT: movaps 240(%rdi), %xmm0
; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movaps 96(%rdi), %xmm8
; SSE-NEXT: movaps 224(%rdi), %xmm10
; SSE-NEXT: movaps 160(%rdi), %xmm0
; SSE-NEXT: movaps 80(%rdi), %xmm12
; SSE-NEXT: movaps 208(%rdi), %xmm13
; SSE-NEXT: movaps 144(%rdi), %xmm2
; SSE-NEXT: movaps 64(%rdi), %xmm14
; SSE-NEXT: movaps (%rdi), %xmm7
; SSE-NEXT: movaps 16(%rdi), %xmm6
; SSE-NEXT: movaps 32(%rdi), %xmm4
; SSE-NEXT: movaps 48(%rdi), %xmm3
; SSE-NEXT: movaps 192(%rdi), %xmm15
; SSE-NEXT: movaps 128(%rdi), %xmm9
; SSE-NEXT: movaps %xmm9, %xmm11
; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm15[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm15[1]
; SSE-NEXT: movaps %xmm7, %xmm15
; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm14[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm14[1]
; SSE-NEXT: movaps %xmm2, %xmm14
; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm13[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm13[1]
; SSE-NEXT: movaps %xmm6, %xmm13
; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm12[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm12[1]
; SSE-NEXT: movaps %xmm0, %xmm12
; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm10[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm10[1]
; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movaps %xmm4, %xmm10
; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm8[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm8[1]
; SSE-NEXT: movaps %xmm3, %xmm8
; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm5[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm5[1]
; SSE-NEXT: movaps 176(%rdi), %xmm5
; SSE-NEXT: movaps %xmm5, %xmm1
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT: movaps %xmm15, (%rsi)
; SSE-NEXT: movaps %xmm11, 16(%rsi)
; SSE-NEXT: movaps %xmm7, (%rdx)
; SSE-NEXT: movaps %xmm9, 16(%rdx)
; SSE-NEXT: movaps %xmm13, (%rcx)
; SSE-NEXT: movaps %xmm14, 16(%rcx)
; SSE-NEXT: movaps %xmm6, (%r8)
; SSE-NEXT: movaps %xmm2, 16(%r8)
; SSE-NEXT: movaps %xmm10, (%r9)
; SSE-NEXT: movaps %xmm12, 16(%r9)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm4, (%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 16(%rax)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm8, (%rax)
; SSE-NEXT: movaps %xmm1, 16(%rax)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm5, 16(%rax)
; SSE-NEXT: movaps %xmm3, (%rax)
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: load_i64_stride8_vf4:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %ymm4
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %ymm5
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm3[0]
; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm7[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm6[1],xmm3[1]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm8[1],xmm7[1]
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX1-ONLY-NEXT: vmovaps 80(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovaps 16(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm11 = xmm10[0],xmm8[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm10[1],xmm8[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm5[0]
; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm8[1],xmm5[1]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm8 = xmm12[1],xmm11[1]
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 112(%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovaps 48(%rdi), %xmm14
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm12[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm14[1],xmm12[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps %xmm9, (%rsi)
; AVX1-ONLY-NEXT: vmovaps %xmm2, 16(%rsi)
; AVX1-ONLY-NEXT: vmovaps %xmm6, (%rdx)
; AVX1-ONLY-NEXT: vmovaps %xmm3, 16(%rdx)
; AVX1-ONLY-NEXT: vmovaps %ymm7, (%rcx)
; AVX1-ONLY-NEXT: vmovaps %ymm4, (%r8)
; AVX1-ONLY-NEXT: vmovaps %xmm13, (%r9)
; AVX1-ONLY-NEXT: vmovaps %xmm10, 16(%r9)
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%r11)
; AVX1-ONLY-NEXT: vmovaps %xmm5, 16(%r11)
; AVX1-ONLY-NEXT: vmovaps %ymm11, (%r10)
; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: load_i64_stride8_vf4:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %ymm4
; AVX2-ONLY-NEXT: vmovaps (%rdi), %ymm5
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %ymm7
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %xmm8
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %xmm9
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]
; AVX2-ONLY-NEXT: vmovaps (%rdi), %xmm11
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %xmm12
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %xmm13
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm14 = xmm11[0],xmm13[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm8 = xmm9[1],xmm8[1]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm11[1],xmm13[1]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm13[2,3],ymm11[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %xmm5
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %xmm6
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm7 = xmm6[0],xmm5[0]
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %xmm13
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm15 = xmm12[0],xmm13[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm6[1],xmm5[1]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm12[1],xmm13[1]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm13[2,3],ymm12[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
; AVX2-ONLY-NEXT: vmovaps %xmm14, (%rsi)
; AVX2-ONLY-NEXT: vmovaps %xmm10, 16(%rsi)
; AVX2-ONLY-NEXT: vmovaps %xmm9, (%rdx)
; AVX2-ONLY-NEXT: vmovaps %xmm8, 16(%rdx)
; AVX2-ONLY-NEXT: vmovaps %ymm11, (%rcx)
; AVX2-ONLY-NEXT: vmovaps %ymm4, (%r8)
; AVX2-ONLY-NEXT: vmovaps %xmm15, (%r9)
; AVX2-ONLY-NEXT: vmovaps %xmm7, 16(%r9)
; AVX2-ONLY-NEXT: vmovaps %xmm6, (%r11)
; AVX2-ONLY-NEXT: vmovaps %xmm5, 16(%r11)
; AVX2-ONLY-NEXT: vmovaps %ymm12, (%r10)
; AVX2-ONLY-NEXT: vmovaps %ymm0, (%rax)
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512-LABEL: load_i64_stride8_vf4:
; AVX512: # %bb.0:
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm3
; AVX512-NEXT: vmovdqa64 (%rdi), %zmm4
; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm5
; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm6
; AVX512-NEXT: vmovaps 192(%rdi), %xmm0
; AVX512-NEXT: vmovaps 128(%rdi), %xmm1
; AVX512-NEXT: vmovlhps {{.*#+}} xmm7 = xmm1[0],xmm0[0]
; AVX512-NEXT: vmovaps (%rdi), %xmm2
; AVX512-NEXT: vmovaps 64(%rdi), %xmm8
; AVX512-NEXT: vmovlhps {{.*#+}} xmm9 = xmm2[0],xmm8[0]
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm10 = xmm1[1],xmm0[1]
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm8 = xmm2[1],xmm8[1]
; AVX512-NEXT: vmovaps 192(%rdi), %ymm1
; AVX512-NEXT: vmovaps 128(%rdi), %ymm2
; AVX512-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512-NEXT: vmovaps 64(%rdi), %ymm11
; AVX512-NEXT: vmovaps (%rdi), %ymm12
; AVX512-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX512-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,12,4,12]
; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512-NEXT: vpermi2q %zmm5, %zmm6, %zmm2
; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [4,12]
; AVX512-NEXT: vpermi2q %zmm3, %zmm4, %zmm11
; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [5,13,5,13]
; AVX512-NEXT: # ymm11 = mem[0,1,0,1]
; AVX512-NEXT: vpermi2q %zmm5, %zmm6, %zmm11
; AVX512-NEXT: vmovdqa {{.*#+}} xmm12 = [5,13]
; AVX512-NEXT: vpermi2q %zmm3, %zmm4, %zmm12
; AVX512-NEXT: vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3],ymm11[4,5,6,7]
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [6,14,6,14]
; AVX512-NEXT: # ymm12 = mem[0,1,0,1]
; AVX512-NEXT: vpermi2q %zmm5, %zmm6, %zmm12
; AVX512-NEXT: vmovdqa {{.*#+}} xmm13 = [6,14]
; AVX512-NEXT: vpermi2q %zmm3, %zmm4, %zmm13
; AVX512-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [7,15,7,15]
; AVX512-NEXT: # ymm13 = mem[0,1,0,1]
; AVX512-NEXT: vpermi2q %zmm5, %zmm6, %zmm13
; AVX512-NEXT: vmovdqa {{.*#+}} xmm5 = [7,15]
; AVX512-NEXT: vpermi2q %zmm3, %zmm4, %zmm5
; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm13[4,5,6,7]
; AVX512-NEXT: vmovaps %xmm9, (%rsi)
; AVX512-NEXT: vmovaps %xmm7, 16(%rsi)
; AVX512-NEXT: vmovaps %xmm8, (%rdx)
; AVX512-NEXT: vmovaps %xmm10, 16(%rdx)
; AVX512-NEXT: vmovaps %ymm0, (%rcx)
; AVX512-NEXT: vmovaps %ymm1, (%r8)
; AVX512-NEXT: vmovdqa %ymm2, (%r9)
; AVX512-NEXT: vmovdqa %ymm11, (%r11)
; AVX512-NEXT: vmovdqa %ymm12, (%r10)
; AVX512-NEXT: vmovdqa %ymm3, (%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
%wide.vec = load <32 x i64>, ptr %in.vec, align 64
%strided.vec0 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
%strided.vec1 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 1, i32 9, i32 17, i32 25>
%strided.vec2 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 2, i32 10, i32 18, i32 26>
%strided.vec3 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
%strided.vec4 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 4, i32 12, i32 20, i32 28>
%strided.vec5 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 5, i32 13, i32 21, i32 29>
%strided.vec6 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 6, i32 14, i32 22, i32 30>
%strided.vec7 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 7, i32 15, i32 23, i32 31>
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
store <4 x i64> %strided.vec6, ptr %out.vec6, align 64
store <4 x i64> %strided.vec7, ptr %out.vec7, align 64
ret void
}
define void @load_i64_stride8_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf8:
; SSE: # %bb.0:
; SSE-NEXT: subq $152, %rsp
; SSE-NEXT: movaps 336(%rdi), %xmm11
; SSE-NEXT: movaps 464(%rdi), %xmm6
; SSE-NEXT: movaps 400(%rdi), %xmm7
; SSE-NEXT: movaps 80(%rdi), %xmm0
; SSE-NEXT: movaps 208(%rdi), %xmm1
; SSE-NEXT: movaps 144(%rdi), %xmm8
; SSE-NEXT: movaps 320(%rdi), %xmm2
; SSE-NEXT: movaps 256(%rdi), %xmm10
; SSE-NEXT: movaps 448(%rdi), %xmm3
; SSE-NEXT: movaps 384(%rdi), %xmm12
; SSE-NEXT: movaps 64(%rdi), %xmm4
; SSE-NEXT: movaps (%rdi), %xmm13
; SSE-NEXT: movaps 16(%rdi), %xmm9
; SSE-NEXT: movaps 192(%rdi), %xmm5
; SSE-NEXT: movaps 128(%rdi), %xmm14
; SSE-NEXT: movaps %xmm14, %xmm15
; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm5[0]
; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm5[1]
; SSE-NEXT: movaps %xmm14, {{[-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 %xmm12, %xmm4
; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT: movaps %xmm4, (%rsp) # 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 %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 %xmm8, %xmm3
; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]
; SSE-NEXT: movaps %xmm3, {{[-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 %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 %xmm7, %xmm0
; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm6[0]
; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm6[1]
; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movaps 272(%rdi), %xmm15
; SSE-NEXT: movaps %xmm15, %xmm0
; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm11[0]
; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm11[1]
; SSE-NEXT: movaps 96(%rdi), %xmm0
; SSE-NEXT: movaps 32(%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 224(%rdi), %xmm0
; SSE-NEXT: movaps 160(%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 352(%rdi), %xmm0
; SSE-NEXT: movaps 288(%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 480(%rdi), %xmm1
; SSE-NEXT: movaps 416(%rdi), %xmm4
; SSE-NEXT: movaps %xmm4, %xmm8
; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm1[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]
; SSE-NEXT: movaps 112(%rdi), %xmm1
; SSE-NEXT: movaps 48(%rdi), %xmm10
; SSE-NEXT: movaps %xmm10, %xmm14
; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm1[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm1[1]
; SSE-NEXT: movaps 240(%rdi), %xmm0
; SSE-NEXT: movaps 176(%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 368(%rdi), %xmm0
; SSE-NEXT: movaps 304(%rdi), %xmm3
; SSE-NEXT: movaps %xmm3, %xmm5
; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
; SSE-NEXT: movaps 496(%rdi), %xmm0
; SSE-NEXT: movaps 432(%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, 32(%rsi)
; SSE-NEXT: movaps (%rsp), %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, (%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, 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, 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 %xmm15, 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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 16(%r8)
; SSE-NEXT: movaps %xmm11, 32(%r9)
; SSE-NEXT: movaps %xmm8, 48(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, (%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 16(%r9)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm4, 48(%rax)
; SSE-NEXT: movaps %xmm9, 32(%rax)
; SSE-NEXT: movaps %xmm13, 16(%rax)
; SSE-NEXT: movaps %xmm12, (%rax)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm2, 48(%rax)
; SSE-NEXT: movaps %xmm5, 32(%rax)
; SSE-NEXT: movaps %xmm7, 16(%rax)
; SSE-NEXT: movaps %xmm14, (%rax)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm1, 48(%rax)
; SSE-NEXT: movaps %xmm3, 32(%rax)
; SSE-NEXT: movaps %xmm6, 16(%rax)
; SSE-NEXT: movaps %xmm10, (%rax)
; SSE-NEXT: addq $152, %rsp
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: load_i64_stride8_vf8:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: subq $184, %rsp
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovaps 448(%rdi), %ymm2
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %ymm3
; AVX1-ONLY-NEXT: vmovaps 320(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovaps 256(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 448(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm8[0],xmm7[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %xmm9
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm10[0],xmm9[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm12[0],xmm11[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
; AVX1-ONLY-NEXT: vmovaps %xmm4, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm8[1],xmm7[1]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm10[1],xmm9[1]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm12[1],xmm11[1]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX1-ONLY-NEXT: vmovaps 336(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovaps 272(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm11[0],xmm10[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 80(%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovaps 16(%rdi), %xmm13
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm14 = xmm13[0],xmm12[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm14[0,1,2,3],ymm9[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm11[1],xmm10[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm13[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: vmovaps 480(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 352(%rdi), %xmm2
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 288(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm13 = xmm0[0],xmm2[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm15 = xmm2[1],xmm1[1]
; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm11 = xmm2[1],xmm1[1]
; AVX1-ONLY-NEXT: vmovaps 480(%rdi), %ymm5
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %ymm2
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm5[0],ymm2[2],ymm5[2]
; AVX1-ONLY-NEXT: vmovaps 368(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovaps 304(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm4[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm12 = ymm9[0,1,2,3],ymm3[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %ymm9
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %ymm6
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX1-ONLY-NEXT: vmovaps 112(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovaps 48(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm3[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm7 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm4[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, (%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 16(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 48(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 32(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, (%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 16(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 48(%rdx)
; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 32(%rdx)
; 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, 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, 32(%r8)
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%r9)
; AVX1-ONLY-NEXT: vmovaps %xmm10, 16(%r9)
; AVX1-ONLY-NEXT: vmovaps %xmm13, 32(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 48(%r9)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %xmm11, (%rax)
; AVX1-ONLY-NEXT: vmovaps %xmm15, 16(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm2, 48(%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %ymm7, (%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm12, 32(%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm1, 32(%rax)
; AVX1-ONLY-NEXT: addq $184, %rsp
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: load_i64_stride8_vf8:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: subq $184, %rsp
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %ymm11
; AVX2-ONLY-NEXT: vmovaps (%rdi), %ymm12
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %ymm13
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %ymm14
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %ymm15
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %xmm4
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm4[0],xmm2[0]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %xmm5
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %xmm10
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm10[0],xmm5[0]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %xmm7
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %xmm8
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm8[0],xmm7[0]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps (%rdi), %xmm0
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %xmm6
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm10[1],xmm5[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm6[0]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm7[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm6[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm9[0],ymm13[2],ymm9[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm6[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %ymm0
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm8[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm9 = ymm13[1],ymm9[1],ymm13[3],ymm9[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm9[2,3],ymm2[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %xmm2
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm12 = xmm0[0],xmm2[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm15 = xmm2[0],xmm0[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm13 = xmm2[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm2[0],xmm0[0]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm10 = xmm2[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %ymm5
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %ymm4
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %ymm3
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm9[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %ymm0
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm9[0],ymm2[2],ymm9[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm14[2,3],ymm7[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm9[1],ymm2[3],ymm9[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rdx)
; AVX2-ONLY-NEXT: vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 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, 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, 32(%r8)
; AVX2-ONLY-NEXT: vmovaps %xmm8, (%r9)
; AVX2-ONLY-NEXT: vmovaps %xmm15, 16(%r9)
; AVX2-ONLY-NEXT: vmovaps %xmm12, 32(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%r9)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps %xmm10, (%rax)
; AVX2-ONLY-NEXT: vmovaps %xmm13, 16(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps %ymm7, (%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm11, 32(%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps %ymm0, (%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm3, 32(%rax)
; AVX2-ONLY-NEXT: addq $184, %rsp
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-LABEL: load_i64_stride8_vf8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vmovdqa64 192(%rdi), %zmm0
; AVX512F-NEXT: vmovdqa64 128(%rdi), %zmm1
; AVX512F-NEXT: vmovdqa64 64(%rdi), %zmm4
; AVX512F-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512F-NEXT: vmovdqa64 320(%rdi), %zmm8
; AVX512F-NEXT: vmovdqa64 256(%rdi), %zmm3
; AVX512F-NEXT: vmovdqa64 448(%rdi), %zmm6
; AVX512F-NEXT: vmovdqa64 384(%rdi), %zmm9
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [0,8,0,8,0,8,0,8]
; AVX512F-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm18
; AVX512F-NEXT: vpermt2q %zmm6, %zmm5, %zmm18
; AVX512F-NEXT: vpermi2q %zmm8, %zmm3, %zmm5
; AVX512F-NEXT: movb $-64, %al
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [1,9,1,9,1,9,1,9]
; AVX512F-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm11
; AVX512F-NEXT: vpermt2q %zmm6, %zmm10, %zmm11
; AVX512F-NEXT: vpermi2q %zmm8, %zmm3, %zmm10
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
; AVX512F-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm13
; AVX512F-NEXT: vpermt2q %zmm6, %zmm12, %zmm13
; AVX512F-NEXT: vpermi2q %zmm8, %zmm3, %zmm12
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm12 {%k1}
; AVX512F-NEXT: vmovdqa 192(%rdi), %ymm13
; AVX512F-NEXT: vmovdqa 128(%rdi), %ymm14
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm14[0],ymm13[0],ymm14[2],ymm13[2]
; AVX512F-NEXT: vmovdqa64 64(%rdi), %ymm16
; AVX512F-NEXT: vmovdqa64 (%rdi), %ymm17
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm7[2,3],ymm15[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm7, %zmm12, %zmm12
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [3,11,3,11,3,11,3,11]
; AVX512F-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm15
; AVX512F-NEXT: vpermt2q %zmm6, %zmm7, %zmm15
; AVX512F-NEXT: vpermi2q %zmm8, %zmm3, %zmm7
; AVX512F-NEXT: vmovdqa64 %zmm15, %zmm7 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm14[1],ymm13[1],ymm14[3],ymm13[3]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [4,12,4,12,4,12,4,12]
; AVX512F-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm3[0],zmm8[0],zmm3[2],zmm8[2],zmm3[4],zmm8[4],zmm3[6],zmm8[6]
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm16
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm17 = zmm3[1],zmm8[1],zmm3[3],zmm8[3],zmm3[5],zmm8[5],zmm3[7],zmm8[7]
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [6,14,6,14,6,14,6,14]
; AVX512F-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm19
; AVX512F-NEXT: vpermt2q %zmm8, %zmm15, %zmm19
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm19 {%k1} = zmm9[0],zmm6[0],zmm9[2],zmm6[2],zmm9[4],zmm6[4],zmm9[6],zmm6[6]
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [7,15,7,15,7,15,7,15]
; AVX512F-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm8, %zmm20, %zmm3
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm9[1],zmm6[1],zmm9[3],zmm6[3],zmm9[5],zmm6[5],zmm9[7],zmm6[7]
; AVX512F-NEXT: vpermt2q %zmm6, %zmm14, %zmm9
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm7 {%k1}
; AVX512F-NEXT: vpermi2q %zmm4, %zmm2, %zmm14
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [4,12,4,12]
; AVX512F-NEXT: # ymm8 = mem[0,1,0,1]
; AVX512F-NEXT: vpermi2q %zmm0, %zmm1, %zmm8
; AVX512F-NEXT: vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [5,13,5,13,5,13,5,13]
; AVX512F-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm6, %zmm8, %zmm16
; AVX512F-NEXT: vpermi2q %zmm4, %zmm2, %zmm8
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [5,13,5,13]
; AVX512F-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512F-NEXT: vpermi2q %zmm0, %zmm1, %zmm6
; AVX512F-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
; AVX512F-NEXT: vmovdqa64 %zmm18, %zmm5 {%k1}
; AVX512F-NEXT: vmovdqa 192(%rdi), %xmm8
; AVX512F-NEXT: vmovdqa 128(%rdi), %xmm9
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm14 = xmm9[0],xmm8[0]
; AVX512F-NEXT: vmovdqa64 (%rdi), %xmm18
; AVX512F-NEXT: vmovdqa64 64(%rdi), %xmm21
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm22 = xmm18[0],xmm21[0]
; AVX512F-NEXT: vinserti32x4 $1, %xmm14, %ymm22, %ymm14
; AVX512F-NEXT: vinserti64x4 $0, %ymm14, %zmm5, %zmm5
; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm9[1],xmm8[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm18[1],xmm21[1]
; AVX512F-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rdi
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512F-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm8
; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm17 {%k1}
; AVX512F-NEXT: vinserti64x4 $0, %ymm6, %zmm17, %zmm6
; AVX512F-NEXT: vpermi2q %zmm4, %zmm2, %zmm15
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [6,14,6,14]
; AVX512F-NEXT: # ymm9 = mem[0,1,0,1]
; AVX512F-NEXT: vpermi2q %zmm0, %zmm1, %zmm9
; AVX512F-NEXT: vpblendd {{.*#+}} ymm9 = ymm15[0,1,2,3],ymm9[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm9, %zmm19, %zmm9
; AVX512F-NEXT: vpermt2q %zmm4, %zmm20, %zmm2
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [7,15,7,15]
; AVX512F-NEXT: # ymm4 = mem[0,1,0,1]
; AVX512F-NEXT: vpermi2q %zmm0, %zmm1, %zmm4
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm5, (%rsi)
; AVX512F-NEXT: vmovdqa64 %zmm8, (%rdx)
; AVX512F-NEXT: vmovdqa64 %zmm12, (%rcx)
; AVX512F-NEXT: vmovdqa64 %zmm13, (%r8)
; AVX512F-NEXT: vmovdqa64 %zmm7, (%r9)
; AVX512F-NEXT: vmovdqa64 %zmm6, (%r10)
; AVX512F-NEXT: vmovdqa64 %zmm9, (%rdi)
; AVX512F-NEXT: vmovdqa64 %zmm0, (%rax)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_i64_stride8_vf8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm0
; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm1
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm4
; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm8
; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm3
; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm6
; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm9
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm18
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm5, %zmm18
; AVX512BW-NEXT: vpermi2q %zmm8, %zmm3, %zmm5
; AVX512BW-NEXT: movb $-64, %al
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm11
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm10, %zmm11
; AVX512BW-NEXT: vpermi2q %zmm8, %zmm3, %zmm10
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm13
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm12, %zmm13
; AVX512BW-NEXT: vpermi2q %zmm8, %zmm3, %zmm12
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm12 {%k1}
; AVX512BW-NEXT: vmovdqa 192(%rdi), %ymm13
; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm14
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm14[0],ymm13[0],ymm14[2],ymm13[2]
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %ymm16
; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm17
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm7[2,3],ymm15[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm12, %zmm12
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm15
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm7, %zmm15
; AVX512BW-NEXT: vpermi2q %zmm8, %zmm3, %zmm7
; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm7 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm14[1],ymm13[1],ymm14[3],ymm13[3]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm3[0],zmm8[0],zmm3[2],zmm8[2],zmm3[4],zmm8[4],zmm3[6],zmm8[6]
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm16
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm17 = zmm3[1],zmm8[1],zmm3[3],zmm8[3],zmm3[5],zmm8[5],zmm3[7],zmm8[7]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm19
; AVX512BW-NEXT: vpermt2q %zmm8, %zmm15, %zmm19
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm19 {%k1} = zmm9[0],zmm6[0],zmm9[2],zmm6[2],zmm9[4],zmm6[4],zmm9[6],zmm6[6]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm8, %zmm20, %zmm3
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm9[1],zmm6[1],zmm9[3],zmm6[3],zmm9[5],zmm6[5],zmm9[7],zmm6[7]
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm14, %zmm9
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm7 {%k1}
; AVX512BW-NEXT: vpermi2q %zmm4, %zmm2, %zmm14
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [4,12,4,12]
; AVX512BW-NEXT: # ymm8 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm8
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm8, %zmm16
; AVX512BW-NEXT: vpermi2q %zmm4, %zmm2, %zmm8
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [5,13,5,13]
; AVX512BW-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm6
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm5 {%k1}
; AVX512BW-NEXT: vmovdqa 192(%rdi), %xmm8
; AVX512BW-NEXT: vmovdqa 128(%rdi), %xmm9
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm14 = xmm9[0],xmm8[0]
; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm18
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %xmm21
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm22 = xmm18[0],xmm21[0]
; AVX512BW-NEXT: vinserti32x4 $1, %xmm14, %ymm22, %ymm14
; AVX512BW-NEXT: vinserti64x4 $0, %ymm14, %zmm5, %zmm5
; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm9[1],xmm8[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm18[1],xmm21[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rdi
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm8
; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm17 {%k1}
; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm17, %zmm6
; AVX512BW-NEXT: vpermi2q %zmm4, %zmm2, %zmm15
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [6,14,6,14]
; AVX512BW-NEXT: # ymm9 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm9
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm15[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm9, %zmm19, %zmm9
; AVX512BW-NEXT: vpermt2q %zmm4, %zmm20, %zmm2
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [7,15,7,15]
; AVX512BW-NEXT: # ymm4 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm4
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rsi)
; AVX512BW-NEXT: vmovdqa64 %zmm8, (%rdx)
; AVX512BW-NEXT: vmovdqa64 %zmm12, (%rcx)
; AVX512BW-NEXT: vmovdqa64 %zmm13, (%r8)
; AVX512BW-NEXT: vmovdqa64 %zmm7, (%r9)
; AVX512BW-NEXT: vmovdqa64 %zmm6, (%r10)
; AVX512BW-NEXT: vmovdqa64 %zmm9, (%rdi)
; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
%wide.vec = load <64 x i64>, ptr %in.vec, align 64
%strided.vec0 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56>
%strided.vec1 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57>
%strided.vec2 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58>
%strided.vec3 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59>
%strided.vec4 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60>
%strided.vec5 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61>
%strided.vec6 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62>
%strided.vec7 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>
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
store <8 x i64> %strided.vec6, ptr %out.vec6, align 64
store <8 x i64> %strided.vec7, ptr %out.vec7, align 64
ret void
}
define void @load_i64_stride8_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf16:
; SSE: # %bb.0:
; SSE-NEXT: subq $664, %rsp # imm = 0x298
; SSE-NEXT: movaps 832(%rdi), %xmm0
; SSE-NEXT: movaps 320(%rdi), %xmm1
; SSE-NEXT: movaps 256(%rdi), %xmm8
; SSE-NEXT: movaps 960(%rdi), %xmm2
; SSE-NEXT: movaps 896(%rdi), %xmm9
; SSE-NEXT: movaps 448(%rdi), %xmm3
; SSE-NEXT: movaps 384(%rdi), %xmm10
; SSE-NEXT: movaps 576(%rdi), %xmm4
; SSE-NEXT: movaps 512(%rdi), %xmm11
; SSE-NEXT: movaps 64(%rdi), %xmm5
; SSE-NEXT: movaps (%rdi), %xmm12
; SSE-NEXT: movaps 704(%rdi), %xmm6
; SSE-NEXT: movaps 640(%rdi), %xmm13
; SSE-NEXT: movaps 192(%rdi), %xmm7
; SSE-NEXT: movaps 128(%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 80(%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 208(%rdi), %xmm0
; SSE-NEXT: movaps 144(%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 336(%rdi), %xmm0
; SSE-NEXT: movaps 272(%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 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 592(%rdi), %xmm0
; SSE-NEXT: movaps 528(%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 720(%rdi), %xmm0
; SSE-NEXT: movaps 656(%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 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 976(%rdi), %xmm0
; SSE-NEXT: movaps 912(%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 96(%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 224(%rdi), %xmm0
; SSE-NEXT: movaps 160(%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 288(%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 480(%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, (%rsp) # 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 608(%rdi), %xmm0
; SSE-NEXT: movaps 544(%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 672(%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 864(%rdi), %xmm0
; SSE-NEXT: movaps 800(%rdi), %xmm9
; 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 992(%rdi), %xmm0
; SSE-NEXT: movaps 928(%rdi), %xmm7
; 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 112(%rdi), %xmm0
; SSE-NEXT: movaps 48(%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 240(%rdi), %xmm0
; SSE-NEXT: movaps 176(%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 368(%rdi), %xmm0
; SSE-NEXT: movaps 304(%rdi), %xmm10
; SSE-NEXT: movaps %xmm10, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT: movaps 496(%rdi), %xmm0
; SSE-NEXT: movaps 432(%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 624(%rdi), %xmm0
; SSE-NEXT: movaps 560(%rdi), %xmm5
; SSE-NEXT: movaps %xmm5, %xmm11
; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT: movaps 752(%rdi), %xmm0
; SSE-NEXT: movaps 688(%rdi), %xmm2
; SSE-NEXT: movaps %xmm2, %xmm8
; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; SSE-NEXT: movaps 880(%rdi), %xmm0
; SSE-NEXT: movaps 816(%rdi), %xmm1
; SSE-NEXT: movaps %xmm1, %xmm6
; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT: movaps 1008(%rdi), %xmm0
; SSE-NEXT: movaps 944(%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 {{[-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, 32(%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, 48(%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, (%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, 16(%rcx)
; 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 {{[-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 (%rsp), %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 %xmm7, 112(%rax)
; SSE-NEXT: movaps %xmm9, 96(%rax)
; SSE-NEXT: movaps %xmm12, 80(%rax)
; SSE-NEXT: movaps %xmm15, 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: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm4, 112(%rax)
; SSE-NEXT: movaps %xmm6, 96(%rax)
; SSE-NEXT: movaps %xmm8, 80(%rax)
; SSE-NEXT: movaps %xmm11, 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: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm3, 112(%rax)
; SSE-NEXT: movaps %xmm1, 96(%rax)
; SSE-NEXT: movaps %xmm2, 80(%rax)
; SSE-NEXT: movaps %xmm5, 64(%rax)
; SSE-NEXT: movaps %xmm14, 48(%rax)
; SSE-NEXT: movaps %xmm10, 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 $664, %rsp # imm = 0x298
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: load_i64_stride8_vf16:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: subq $792, %rsp # imm = 0x318
; AVX1-ONLY-NEXT: vmovaps 448(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 320(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps 256(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 832(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovaps 768(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 960(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovaps 896(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm6[0]
; AVX1-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 704(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovaps 640(%rdi), %xmm9
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]
; AVX1-ONLY-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 576(%rdi), %xmm10
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 512(%rdi), %xmm0
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm10[0]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm6[1]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm1
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm4[1]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %xmm2
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm9[1],xmm8[1]
; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %xmm3
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm10[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm3[0],xmm2[0]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 448(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 336(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovaps 272(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm3[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm5[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: vmovaps 960(%rdi), %ymm2
; AVX1-ONLY-NEXT: vmovaps 896(%rdi), %ymm5
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
; AVX1-ONLY-NEXT: vmovaps 848(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovaps 784(%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 192(%rdi), %ymm6
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %ymm9
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm6[0],ymm9[2],ymm6[2]
; AVX1-ONLY-NEXT: vmovaps 80(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovaps 16(%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: vmovaps 640(%rdi), %ymm13
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm13[0],ymm10[0],ymm13[2],ymm10[2]
; AVX1-ONLY-NEXT: vmovaps 592(%rdi), %xmm15
; AVX1-ONLY-NEXT: vmovaps 528(%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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX1-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm3[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 {{.*#+}} ymm1 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm8[1],xmm7[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 {{.*#+}} ymm1 = ymm9[1],ymm6[1],ymm9[3],ymm6[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm12[1],xmm11[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 {{.*#+}} ymm1 = 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],ymm1[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 480(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 352(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 288(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 992(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 928(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 864(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 800(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 736(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 672(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 608(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 544(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 480(%rdi), %ymm11
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %ymm8
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm8[0],ymm11[0],ymm8[2],ymm11[2]
; AVX1-ONLY-NEXT: vmovaps 368(%rdi), %xmm9
; AVX1-ONLY-NEXT: vmovaps 304(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm5 = xmm7[0],xmm9[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm5[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: vmovaps 992(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 928(%rdi), %ymm14
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 880(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 816(%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm12[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm10[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 224(%rdi), %ymm10
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %ymm2
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm10[0],ymm2[2],ymm10[2]
; AVX1-ONLY-NEXT: vmovaps 112(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps 48(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm15 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm15[0,1,2,3],ymm6[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 736(%rdi), %ymm6
; AVX1-ONLY-NEXT: vmovaps 672(%rdi), %ymm5
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm15 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX1-ONLY-NEXT: vmovaps 624(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovaps 560(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm13 = xmm3[0],xmm4[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm15 = ymm13[0,1,2,3],ymm15[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm10[1],ymm2[3],ymm10[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm11[1],ymm8[3],ymm11[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm9[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm4[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT: # ymm0 = ymm14[1],mem[1],ymm14[3],mem[3]
; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm3 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm3 = xmm12[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 16(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, (%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 64(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 80(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 112(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 96(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 32(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 48(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, (%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 16(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 64(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 80(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 96(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 112(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 32(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 48(%rdx)
; 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, 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(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm3, (%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, 32(%r8)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 80(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 64(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, (%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 16(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 112(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 96(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 32(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 48(%r9)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 80(%rax)
; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, (%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 16(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 112(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm3, 48(%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %ymm15, 64(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm3, (%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm3, 96(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm3, 32(%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %ymm0, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm1, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm2, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm10, (%rax)
; AVX1-ONLY-NEXT: addq $792, %rsp # imm = 0x318
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: load_i64_stride8_vf16:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: subq $808, %rsp # imm = 0x328
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %xmm3
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 832(%rdi), %xmm4
; AVX2-ONLY-NEXT: vmovaps 768(%rdi), %xmm5
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX2-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 960(%rdi), %xmm6
; AVX2-ONLY-NEXT: vmovaps 896(%rdi), %xmm7
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm6[0]
; AVX2-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 704(%rdi), %xmm8
; AVX2-ONLY-NEXT: vmovaps 640(%rdi), %xmm9
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]
; AVX2-ONLY-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 576(%rdi), %xmm10
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 512(%rdi), %xmm0
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm10[0]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm6[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %xmm1
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm4[1]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %xmm2
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm9[1],xmm8[1]
; AVX2-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm10[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps (%rdi), %xmm0
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %ymm3
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 832(%rdi), %ymm4
; AVX2-ONLY-NEXT: vmovaps 768(%rdi), %ymm5
; AVX2-ONLY-NEXT: vmovaps 960(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 896(%rdi), %ymm7
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %ymm8
; AVX2-ONLY-NEXT: vmovaps (%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %ymm10
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %ymm11
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm12[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 576(%rdi), %ymm12
; AVX2-ONLY-NEXT: vmovaps 512(%rdi), %ymm13
; AVX2-ONLY-NEXT: vmovaps 704(%rdi), %ymm14
; AVX2-ONLY-NEXT: vmovaps 640(%rdi), %ymm15
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm12[0],ymm13[2],ymm12[2]
; AVX2-ONLY-NEXT: vperm2f128 $49, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm0 = ymm0[2,3],mem[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm12[1],ymm13[3],ymm12[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 992(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 928(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 864(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 800(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 736(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 672(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 608(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 544(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm2[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %ymm15
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %ymm10
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %ymm8
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm2[2,3]
; 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: vmovaps 800(%rdi), %ymm11
; AVX2-ONLY-NEXT: vmovaps 992(%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 928(%rdi), %ymm7
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm9[0],ymm7[2],ymm9[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm1[0],ymm11[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %ymm13
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %ymm2
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm6[0],ymm13[0],ymm6[2],ymm13[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 608(%rdi), %ymm5
; AVX2-ONLY-NEXT: vmovaps 544(%rdi), %ymm4
; AVX2-ONLY-NEXT: vmovaps 736(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 672(%rdi), %ymm0
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm14[2,3],ymm12[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm6[1],ymm13[1],ymm6[3],ymm13[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm3[2,3],ymm2[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm3 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm3 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm3[2,3],ymm2[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm9[1],ymm7[3],ymm9[3]
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm3 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm3 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 16(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, (%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 64(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 80(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 112(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 96(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 32(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 48(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, (%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 16(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 64(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 80(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 96(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 112(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 32(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 48(%rdx)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 64(%rcx)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, (%rcx)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 96(%rcx)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 32(%rcx)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 64(%r8)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, (%r8)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 96(%r8)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 32(%r8)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 80(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 64(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, (%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 16(%r9)
; AVX2-ONLY-NEXT: vmovaps (%rsp), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 112(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 96(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 32(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 48(%r9)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 64(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 80(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, (%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 16(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 96(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 112(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 32(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm3, 48(%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps %ymm12, 64(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, (%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 96(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm3, 32(%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps %ymm0, 96(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm1, 64(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm2, 32(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm13, (%rax)
; AVX2-ONLY-NEXT: addq $808, %rsp # imm = 0x328
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-LABEL: load_i64_stride8_vf16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: subq $280, %rsp # imm = 0x118
; AVX512F-NEXT: vmovdqa64 704(%rdi), %zmm19
; AVX512F-NEXT: vmovdqa64 640(%rdi), %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 576(%rdi), %zmm30
; AVX512F-NEXT: vmovaps 512(%rdi), %zmm0
; AVX512F-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 832(%rdi), %zmm16
; AVX512F-NEXT: vmovdqa64 768(%rdi), %zmm8
; AVX512F-NEXT: vmovdqa64 960(%rdi), %zmm17
; AVX512F-NEXT: vmovdqa64 896(%rdi), %zmm9
; AVX512F-NEXT: vmovdqa64 320(%rdi), %zmm11
; AVX512F-NEXT: vmovdqa64 256(%rdi), %zmm7
; AVX512F-NEXT: vmovdqa64 448(%rdi), %zmm10
; AVX512F-NEXT: vmovdqa64 384(%rdi), %zmm12
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [0,8,0,8,0,8,0,8]
; AVX512F-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm20
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm5
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512F-NEXT: vpermt2q %zmm10, %zmm18, %zmm3
; AVX512F-NEXT: vmovdqa64 %zmm7, %zmm13
; AVX512F-NEXT: vpermt2q %zmm11, %zmm18, %zmm13
; AVX512F-NEXT: movb $-64, %al
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm13 {%k1}
; AVX512F-NEXT: vmovdqa 192(%rdi), %xmm1
; AVX512F-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-NEXT: vmovdqa64 128(%rdi), %xmm31
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm31[0],xmm1[0]
; AVX512F-NEXT: vmovdqa64 (%rdi), %xmm23
; AVX512F-NEXT: vmovdqa64 64(%rdi), %xmm21
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm14 = xmm23[0],xmm21[0]
; AVX512F-NEXT: vinserti128 $1, %xmm3, %ymm14, %ymm3
; AVX512F-NEXT: vinserti64x4 $0, %ymm3, %zmm13, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512F-NEXT: vpermt2q %zmm17, %zmm18, %zmm3
; AVX512F-NEXT: vpermi2q %zmm16, %zmm8, %zmm18
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm18 {%k1}
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [2,10,2,10,2,10,2,10]
; AVX512F-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm10, %zmm3, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm7, %zmm13
; AVX512F-NEXT: vpermt2q %zmm11, %zmm3, %zmm13
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512F-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512F-NEXT: vmovdqa 128(%rdi), %ymm14
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX512F-NEXT: vmovdqa64 64(%rdi), %ymm22
; AVX512F-NEXT: vmovdqa64 (%rdi), %ymm24
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm24[0],ymm22[0],ymm24[2],ymm22[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm13, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm1
; AVX512F-NEXT: vpermt2q %zmm17, %zmm3, %zmm1
; AVX512F-NEXT: vpermi2q %zmm16, %zmm8, %zmm3
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm3 {%k1}
; AVX512F-NEXT: vmovdqa 704(%rdi), %ymm1
; AVX512F-NEXT: vmovdqa 640(%rdi), %ymm15
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
; AVX512F-NEXT: vmovdqa64 576(%rdi), %ymm26
; AVX512F-NEXT: vmovdqa64 512(%rdi), %ymm28
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm6[2,3],ymm13[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm6, %zmm3, %zmm3
; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm29 = [3,11,3,11,3,11,3,11]
; AVX512F-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm10, %zmm29, %zmm20
; AVX512F-NEXT: vmovdqa64 %zmm7, %zmm4
; AVX512F-NEXT: vpermt2q %zmm11, %zmm29, %zmm4
; AVX512F-NEXT: vmovdqa64 %zmm20, %zmm4 {%k1}
; AVX512F-NEXT: vmovdqa64 192(%rdi), %zmm25
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm14[1],ymm0[1],ymm14[3],ymm0[3]
; AVX512F-NEXT: vmovdqa64 128(%rdi), %zmm13
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm24[1],ymm22[1],ymm24[3],ymm22[3]
; AVX512F-NEXT: vmovdqa64 64(%rdi), %zmm27
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-NEXT: vmovdqa64 (%rdi), %zmm3
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512F-NEXT: vpermt2q %zmm17, %zmm29, %zmm0
; AVX512F-NEXT: vpermi2q %zmm16, %zmm8, %zmm29
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm22
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512F-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm10, %zmm0, %zmm5
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm7[0],zmm11[0],zmm7[2],zmm11[2],zmm7[4],zmm11[4],zmm7[6],zmm11[6]
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm1 {%k1}
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm4
; AVX512F-NEXT: vpermt2q %zmm27, %zmm0, %zmm4
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,12,4,12]
; AVX512F-NEXT: # ymm5 = mem[0,1,0,1]
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm14
; AVX512F-NEXT: vpermt2q %zmm25, %zmm5, %zmm14
; AVX512F-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm14[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm4, %zmm1, %zmm24
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm1
; AVX512F-NEXT: vpermt2q %zmm17, %zmm0, %zmm1
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512F-NEXT: vpermi2q %zmm30, %zmm6, %zmm0
; AVX512F-NEXT: vpermi2q %zmm19, %zmm2, %zmm5
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm26
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
; AVX512F-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm1
; AVX512F-NEXT: vpermt2q %zmm27, %zmm14, %zmm1
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [5,13,5,13]
; AVX512F-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm4
; AVX512F-NEXT: vpermt2q %zmm25, %zmm2, %zmm4
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm28 = [1,9,1,9,1,9,1,9]
; AVX512F-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm7, %zmm29
; AVX512F-NEXT: vpermt2q %zmm11, %zmm28, %zmm29
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
; AVX512F-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm7, %zmm4
; AVX512F-NEXT: vpermt2q %zmm11, %zmm5, %zmm4
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm15 = zmm7[1],zmm11[1],zmm7[3],zmm11[3],zmm7[5],zmm11[5],zmm7[7],zmm11[7]
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm1 = [7,15,7,15,7,15,7,15]
; AVX512F-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm11, %zmm1, %zmm7
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm12[0],zmm10[0],zmm12[2],zmm10[2],zmm12[4],zmm10[4],zmm12[6],zmm10[6]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm7 {%k1} = zmm12[1],zmm10[1],zmm12[3],zmm10[3],zmm12[5],zmm10[5],zmm12[7],zmm10[7]
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm11
; AVX512F-NEXT: vpermt2q %zmm10, %zmm28, %zmm11
; AVX512F-NEXT: vpermt2q %zmm10, %zmm14, %zmm12
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm15 {%k1}
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm15, %zmm20
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm10
; AVX512F-NEXT: vpermt2q %zmm17, %zmm28, %zmm10
; AVX512F-NEXT: vpermi2q %zmm16, %zmm8, %zmm28
; AVX512F-NEXT: vmovdqa64 %zmm8, %zmm12
; AVX512F-NEXT: vpermt2q %zmm16, %zmm5, %zmm12
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm15 = zmm8[1],zmm16[1],zmm8[3],zmm16[3],zmm8[5],zmm16[5],zmm8[7],zmm16[7]
; AVX512F-NEXT: vpermt2q %zmm16, %zmm1, %zmm8
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm9[0],zmm17[0],zmm9[2],zmm17[2],zmm9[4],zmm17[4],zmm9[6],zmm17[6]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm9[1],zmm17[1],zmm9[3],zmm17[3],zmm9[5],zmm17[5],zmm9[7],zmm17[7]
; AVX512F-NEXT: vpermt2q %zmm17, %zmm14, %zmm9
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm15 {%k1}
; AVX512F-NEXT: vpermi2q %zmm30, %zmm6, %zmm14
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm16
; AVX512F-NEXT: vpermi2q %zmm19, %zmm0, %zmm2
; AVX512F-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm2[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm2, %zmm15, %zmm19
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm9
; AVX512F-NEXT: vpermt2q %zmm27, %zmm5, %zmm9
; AVX512F-NEXT: vpermt2q %zmm27, %zmm1, %zmm3
; AVX512F-NEXT: vpermi2q %zmm30, %zmm6, %zmm5
; AVX512F-NEXT: vpermt2q %zmm30, %zmm1, %zmm6
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512F-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm14
; AVX512F-NEXT: vpermt2q %zmm25, %zmm1, %zmm14
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [7,15,7,15]
; AVX512F-NEXT: # ymm15 = mem[0,1,0,1]
; AVX512F-NEXT: vpermt2q %zmm25, %zmm15, %zmm13
; AVX512F-NEXT: vpermi2q %zmm16, %zmm0, %zmm1
; AVX512F-NEXT: vpermt2q %zmm16, %zmm15, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa 704(%rdi), %xmm15
; AVX512F-NEXT: vmovdqa64 640(%rdi), %xmm16
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm17 = xmm16[0],xmm15[0]
; AVX512F-NEXT: vmovdqa64 576(%rdi), %xmm25
; AVX512F-NEXT: vmovdqa64 512(%rdi), %xmm27
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm27[0],xmm25[0]
; AVX512F-NEXT: vinserti32x4 $1, %xmm17, %ymm30, %ymm17
; AVX512F-NEXT: vinserti64x4 $0, %ymm17, %zmm18, %zmm17
; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm29 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm31, %xmm11 # 16-byte Folded Reload
; AVX512F-NEXT: # xmm11 = xmm31[1],mem[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm18 = xmm23[1],xmm21[1]
; AVX512F-NEXT: vinserti32x4 $1, %xmm11, %ymm18, %ymm11
; AVX512F-NEXT: vinserti64x4 $0, %ymm11, %zmm29, %zmm11
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm28 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm16[1],xmm15[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm15 = xmm27[1],xmm25[1]
; AVX512F-NEXT: vinserti128 $1, %xmm10, %ymm15, %ymm10
; AVX512F-NEXT: vinserti64x4 $0, %ymm10, %zmm28, %zmm10
; AVX512F-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm14[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm9, %zmm4, %zmm4
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512F-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm3, %zmm7, %zmm3
; AVX512F-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm2[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm5, %zmm8, %zmm5
; AVX512F-NEXT: vmovdqa64 %zmm17, 64(%rsi)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%rsi)
; AVX512F-NEXT: vmovdqa64 %zmm10, 64(%rdx)
; AVX512F-NEXT: vmovdqa64 %zmm11, (%rdx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 64(%rcx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%rcx)
; AVX512F-NEXT: vmovdqa64 %zmm22, 64(%r8)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%r8)
; AVX512F-NEXT: vmovdqa64 %zmm26, 64(%r9)
; AVX512F-NEXT: vmovdqa64 %zmm24, (%r9)
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovdqa64 %zmm19, 64(%rax)
; AVX512F-NEXT: vmovdqa64 %zmm20, (%rax)
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovdqa64 %zmm0, 64(%rax)
; AVX512F-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovdqa64 %zmm5, 64(%rax)
; AVX512F-NEXT: vmovdqa64 %zmm3, (%rax)
; AVX512F-NEXT: addq $280, %rsp # imm = 0x118
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_i64_stride8_vf16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: subq $280, %rsp # imm = 0x118
; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm19
; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm30
; AVX512BW-NEXT: vmovaps 512(%rdi), %zmm0
; AVX512BW-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 832(%rdi), %zmm16
; AVX512BW-NEXT: vmovdqa64 768(%rdi), %zmm8
; AVX512BW-NEXT: vmovdqa64 960(%rdi), %zmm17
; AVX512BW-NEXT: vmovdqa64 896(%rdi), %zmm9
; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm11
; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm7
; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm10
; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm12
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm20
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm5
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512BW-NEXT: vpermt2q %zmm10, %zmm18, %zmm3
; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm13
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm18, %zmm13
; AVX512BW-NEXT: movb $-64, %al
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm13 {%k1}
; AVX512BW-NEXT: vmovdqa 192(%rdi), %xmm1
; AVX512BW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT: vmovdqa64 128(%rdi), %xmm31
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm31[0],xmm1[0]
; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm23
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %xmm21
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm14 = xmm23[0],xmm21[0]
; AVX512BW-NEXT: vinserti128 $1, %xmm3, %ymm14, %ymm3
; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm13, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512BW-NEXT: vpermt2q %zmm17, %zmm18, %zmm3
; AVX512BW-NEXT: vpermi2q %zmm16, %zmm8, %zmm18
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm18 {%k1}
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm10, %zmm3, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm13
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm3, %zmm13
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512BW-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm14
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %ymm22
; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm24
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm24[0],ymm22[0],ymm24[2],ymm22[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm13, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm17, %zmm3, %zmm1
; AVX512BW-NEXT: vpermi2q %zmm16, %zmm8, %zmm3
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm3 {%k1}
; AVX512BW-NEXT: vmovdqa 704(%rdi), %ymm1
; AVX512BW-NEXT: vmovdqa 640(%rdi), %ymm15
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
; AVX512BW-NEXT: vmovdqa64 576(%rdi), %ymm26
; AVX512BW-NEXT: vmovdqa64 512(%rdi), %ymm28
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm6[2,3],ymm13[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm3, %zmm3
; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm29 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm10, %zmm29, %zmm20
; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm29, %zmm4
; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm4 {%k1}
; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm25
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm14[1],ymm0[1],ymm14[3],ymm0[3]
; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm13
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm24[1],ymm22[1],ymm24[3],ymm22[3]
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm27
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm3
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512BW-NEXT: vpermt2q %zmm17, %zmm29, %zmm0
; AVX512BW-NEXT: vpermi2q %zmm16, %zmm8, %zmm29
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm22
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm10, %zmm0, %zmm5
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm7[0],zmm11[0],zmm7[2],zmm11[2],zmm7[4],zmm11[4],zmm7[6],zmm11[6]
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm1 {%k1}
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm27, %zmm0, %zmm4
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [4,12,4,12]
; AVX512BW-NEXT: # ymm5 = mem[0,1,0,1]
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm14
; AVX512BW-NEXT: vpermt2q %zmm25, %zmm5, %zmm14
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm14[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm1, %zmm24
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm17, %zmm0, %zmm1
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT: vpermi2q %zmm30, %zmm6, %zmm0
; AVX512BW-NEXT: vpermi2q %zmm19, %zmm2, %zmm5
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm26
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm27, %zmm14, %zmm1
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [5,13,5,13]
; AVX512BW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm25, %zmm2, %zmm4
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm28 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm29
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm28, %zmm29
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm5, %zmm4
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm15 = zmm7[1],zmm11[1],zmm7[3],zmm11[3],zmm7[5],zmm11[5],zmm7[7],zmm11[7]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm1 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm1, %zmm7
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm12[0],zmm10[0],zmm12[2],zmm10[2],zmm12[4],zmm10[4],zmm12[6],zmm10[6]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 {%k1} = zmm12[1],zmm10[1],zmm12[3],zmm10[3],zmm12[5],zmm10[5],zmm12[7],zmm10[7]
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm11
; AVX512BW-NEXT: vpermt2q %zmm10, %zmm28, %zmm11
; AVX512BW-NEXT: vpermt2q %zmm10, %zmm14, %zmm12
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm15 {%k1}
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm15, %zmm20
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm10
; AVX512BW-NEXT: vpermt2q %zmm17, %zmm28, %zmm10
; AVX512BW-NEXT: vpermi2q %zmm16, %zmm8, %zmm28
; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm12
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm5, %zmm12
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm15 = zmm8[1],zmm16[1],zmm8[3],zmm16[3],zmm8[5],zmm16[5],zmm8[7],zmm16[7]
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm1, %zmm8
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm9[0],zmm17[0],zmm9[2],zmm17[2],zmm9[4],zmm17[4],zmm9[6],zmm17[6]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm9[1],zmm17[1],zmm9[3],zmm17[3],zmm9[5],zmm17[5],zmm9[7],zmm17[7]
; AVX512BW-NEXT: vpermt2q %zmm17, %zmm14, %zmm9
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm15 {%k1}
; AVX512BW-NEXT: vpermi2q %zmm30, %zmm6, %zmm14
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm16
; AVX512BW-NEXT: vpermi2q %zmm19, %zmm0, %zmm2
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm15, %zmm19
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm9
; AVX512BW-NEXT: vpermt2q %zmm27, %zmm5, %zmm9
; AVX512BW-NEXT: vpermt2q %zmm27, %zmm1, %zmm3
; AVX512BW-NEXT: vpermi2q %zmm30, %zmm6, %zmm5
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm1, %zmm6
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512BW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm14
; AVX512BW-NEXT: vpermt2q %zmm25, %zmm1, %zmm14
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [7,15,7,15]
; AVX512BW-NEXT: # ymm15 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermt2q %zmm25, %zmm15, %zmm13
; AVX512BW-NEXT: vpermi2q %zmm16, %zmm0, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm15, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa 704(%rdi), %xmm15
; AVX512BW-NEXT: vmovdqa64 640(%rdi), %xmm16
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm17 = xmm16[0],xmm15[0]
; AVX512BW-NEXT: vmovdqa64 576(%rdi), %xmm25
; AVX512BW-NEXT: vmovdqa64 512(%rdi), %xmm27
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm27[0],xmm25[0]
; AVX512BW-NEXT: vinserti32x4 $1, %xmm17, %ymm30, %ymm17
; AVX512BW-NEXT: vinserti64x4 $0, %ymm17, %zmm18, %zmm17
; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm29 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm31, %xmm11 # 16-byte Folded Reload
; AVX512BW-NEXT: # xmm11 = xmm31[1],mem[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm18 = xmm23[1],xmm21[1]
; AVX512BW-NEXT: vinserti32x4 $1, %xmm11, %ymm18, %ymm11
; AVX512BW-NEXT: vinserti64x4 $0, %ymm11, %zmm29, %zmm11
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm28 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm16[1],xmm15[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm15 = xmm27[1],xmm25[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm10, %ymm15, %ymm10
; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm28, %zmm10
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm14[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm9, %zmm4, %zmm4
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm7, %zmm3
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm8, %zmm5
; AVX512BW-NEXT: vmovdqa64 %zmm17, 64(%rsi)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%rsi)
; AVX512BW-NEXT: vmovdqa64 %zmm10, 64(%rdx)
; AVX512BW-NEXT: vmovdqa64 %zmm11, (%rdx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 64(%rcx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%rcx)
; AVX512BW-NEXT: vmovdqa64 %zmm22, 64(%r8)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%r8)
; AVX512BW-NEXT: vmovdqa64 %zmm26, 64(%r9)
; AVX512BW-NEXT: vmovdqa64 %zmm24, (%r9)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa64 %zmm19, 64(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm20, (%rax)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa64 %zmm0, 64(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa64 %zmm5, 64(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm3, (%rax)
; AVX512BW-NEXT: addq $280, %rsp # imm = 0x118
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
%wide.vec = load <128 x i64>, ptr %in.vec, align 64
%strided.vec0 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120>
%strided.vec1 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121>
%strided.vec2 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122>
%strided.vec3 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123>
%strided.vec4 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124>
%strided.vec5 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125>
%strided.vec6 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126>
%strided.vec7 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127>
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
store <16 x i64> %strided.vec6, ptr %out.vec6, align 64
store <16 x i64> %strided.vec7, ptr %out.vec7, align 64
ret void
}
define void @load_i64_stride8_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf32:
; SSE: # %bb.0:
; SSE-NEXT: subq $1688, %rsp # imm = 0x698
; SSE-NEXT: movaps 832(%rdi), %xmm0
; SSE-NEXT: movaps 320(%rdi), %xmm1
; SSE-NEXT: movaps 256(%rdi), %xmm8
; SSE-NEXT: movaps 960(%rdi), %xmm2
; SSE-NEXT: movaps 896(%rdi), %xmm10
; SSE-NEXT: movaps 448(%rdi), %xmm3
; SSE-NEXT: movaps 384(%rdi), %xmm9
; SSE-NEXT: movaps 576(%rdi), %xmm4
; SSE-NEXT: movaps 512(%rdi), %xmm12
; SSE-NEXT: movaps 64(%rdi), %xmm5
; SSE-NEXT: movaps (%rdi), %xmm11
; SSE-NEXT: movaps 704(%rdi), %xmm6
; SSE-NEXT: movaps 640(%rdi), %xmm14
; SSE-NEXT: movaps 192(%rdi), %xmm7
; SSE-NEXT: movaps 128(%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 %xmm11, %xmm7
; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm5[0]
; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm5[1]
; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movaps %xmm9, %xmm5
; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]
; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm3[1]
; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movaps %xmm8, %xmm3
; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]
; SSE-NEXT: movaps %xmm3, {{[-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 %xmm14, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm6[0]
; SSE-NEXT: movaps %xmm1, {{[-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, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
; SSE-NEXT: movaps %xmm1, {{[-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 %xmm10, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSE-NEXT: movaps %xmm1, {{[-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 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 1216(%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 1088(%rdi), %xmm0
; SSE-NEXT: movaps 1024(%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 1472(%rdi), %xmm0
; SSE-NEXT: movaps 1408(%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 1344(%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 1728(%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 1600(%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 1984(%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 1856(%rdi), %xmm0
; SSE-NEXT: movaps 1792(%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 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 208(%rdi), %xmm0
; SSE-NEXT: movaps 144(%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 336(%rdi), %xmm0
; SSE-NEXT: movaps 272(%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 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 592(%rdi), %xmm0
; SSE-NEXT: movaps 528(%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 720(%rdi), %xmm0
; SSE-NEXT: movaps 656(%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 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 976(%rdi), %xmm0
; SSE-NEXT: movaps 912(%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 1040(%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 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 1360(%rdi), %xmm0
; SSE-NEXT: movaps 1296(%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 1424(%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 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 1744(%rdi), %xmm0
; SSE-NEXT: movaps 1680(%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 1808(%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 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 96(%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 224(%rdi), %xmm0
; SSE-NEXT: movaps 160(%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 288(%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 480(%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 608(%rdi), %xmm0
; SSE-NEXT: movaps 544(%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 672(%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 864(%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 992(%rdi), %xmm0
; SSE-NEXT: movaps 928(%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 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 1248(%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 1376(%rdi), %xmm0
; SSE-NEXT: movaps 1312(%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 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 1632(%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 1760(%rdi), %xmm0
; SSE-NEXT: movaps 1696(%rdi), %xmm1
; SSE-NEXT: movaps %xmm1, %xmm2
; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT: movaps %xmm2, (%rsp) # 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 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 2016(%rdi), %xmm0
; SSE-NEXT: movaps 1952(%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 112(%rdi), %xmm0
; SSE-NEXT: movaps 48(%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 240(%rdi), %xmm0
; SSE-NEXT: movaps 176(%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 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 496(%rdi), %xmm0
; SSE-NEXT: movaps 432(%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 624(%rdi), %xmm0
; SSE-NEXT: movaps 560(%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 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 880(%rdi), %xmm0
; SSE-NEXT: movaps 816(%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 944(%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 1136(%rdi), %xmm0
; SSE-NEXT: movaps 1072(%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 1264(%rdi), %xmm0
; SSE-NEXT: movaps 1200(%rdi), %xmm10
; SSE-NEXT: movaps %xmm10, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT: movaps 1392(%rdi), %xmm0
; SSE-NEXT: movaps 1328(%rdi), %xmm6
; SSE-NEXT: movaps %xmm6, %xmm1
; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]
; SSE-NEXT: movaps 1520(%rdi), %xmm0
; SSE-NEXT: movaps 1456(%rdi), %xmm8
; SSE-NEXT: movaps %xmm8, %xmm13
; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
; SSE-NEXT: movaps 1648(%rdi), %xmm0
; SSE-NEXT: movaps 1584(%rdi), %xmm7
; SSE-NEXT: movaps %xmm7, %xmm11
; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT: movaps 1776(%rdi), %xmm0
; SSE-NEXT: movaps 1712(%rdi), %xmm2
; SSE-NEXT: movaps %xmm2, %xmm9
; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; SSE-NEXT: movaps 1904(%rdi), %xmm0
; SSE-NEXT: movaps 1840(%rdi), %xmm1
; SSE-NEXT: movaps %xmm1, %xmm5
; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT: movaps 2032(%rdi), %xmm0
; SSE-NEXT: movaps 1968(%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 {{[-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 {{[-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 (%rsp), %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 %xmm14, 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: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm4, 240(%rax)
; SSE-NEXT: movaps %xmm5, 224(%rax)
; SSE-NEXT: movaps %xmm9, 208(%rax)
; SSE-NEXT: movaps %xmm11, 192(%rax)
; SSE-NEXT: movaps %xmm13, 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: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm3, 240(%rax)
; SSE-NEXT: movaps %xmm1, 224(%rax)
; SSE-NEXT: movaps %xmm2, 208(%rax)
; SSE-NEXT: movaps %xmm7, 192(%rax)
; SSE-NEXT: movaps %xmm8, 176(%rax)
; SSE-NEXT: movaps %xmm6, 160(%rax)
; SSE-NEXT: movaps %xmm10, 144(%rax)
; SSE-NEXT: movaps %xmm12, 128(%rax)
; SSE-NEXT: movaps %xmm15, 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 $1688, %rsp # imm = 0x698
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: load_i64_stride8_vf32:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: subq $2296, %rsp # imm = 0x8F8
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 704(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps 640(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1216(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovaps 1152(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm7[0],xmm6[0]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm4[0]
; AVX1-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 576(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovaps 512(%rdi), %xmm9
; AVX1-ONLY-NEXT: vmovaps 1088(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovaps 1024(%rdi), %xmm11
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm7[1],xmm6[1]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm11[0],xmm10[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm11[1],xmm10[1]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm9[0],xmm8[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1600(%rdi), %xmm2
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm9[1],xmm8[1]
; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1728(%rdi), %xmm3
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1664(%rdi), %xmm0
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm3[0]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1536(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm2[0]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 448(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 320(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 256(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 960(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 896(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 832(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 768(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1472(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1408(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1344(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1280(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1984(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1920(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1856(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1792(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1728(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1616(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1552(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 192(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %ymm13
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm0[0],ymm13[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 80(%rdi), %xmm14
; AVX1-ONLY-NEXT: vmovaps 16(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm11[0],xmm14[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[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: vmovaps 704(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 640(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 592(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 528(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm10[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm4[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: vmovaps 1216(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1152(%rdi), %ymm15
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1104(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1040(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm7[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1984(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1920(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1872(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1808(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1472(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1408(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-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 1360(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1296(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm2[0],xmm1[0]
; 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: vmovaps 960(%rdi), %ymm6
; AVX1-ONLY-NEXT: vmovaps 896(%rdi), %ymm9
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm6[0],ymm9[2],ymm6[2]
; AVX1-ONLY-NEXT: vmovaps 848(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovaps 784(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm5[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm8[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 448(%rdi), %ymm3
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %ymm8
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX1-ONLY-NEXT: vmovaps 336(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps 272(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm12 = xmm1[0],xmm2[0]
; 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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT: # ymm0 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm11 = xmm11[1],xmm14[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm11[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm10 = xmm10[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm10[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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX1-ONLY-NEXT: # ymm0 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm7 = xmm7[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm7[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), %xmm7 # 16-byte Reload
; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm7 = xmm7[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm7[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 = ymm8[1],ymm3[1],ymm8[3],ymm3[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 = ymm9[1],ymm6[1],ymm9[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: 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 480(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 352(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 288(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 992(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 928(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 864(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 800(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1120(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1056(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1248(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1184(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1504(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1440(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1376(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1312(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 736(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 672(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 608(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 544(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1632(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1568(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1760(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1696(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1888(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1824(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2016(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1952(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %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 112(%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: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm2[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[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), %ymm2
; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %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 368(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 304(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm3[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[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), %ymm3
; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 672(%rdi), %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), %xmm10
; AVX1-ONLY-NEXT: vmovaps 560(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm10[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm4[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 992(%rdi), %ymm4
; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 928(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
; AVX1-ONLY-NEXT: vmovaps 880(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovaps 816(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm12 = xmm4[0],xmm11[0]
; 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: vmovaps 1248(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1184(%rdi), %ymm5
; AVX1-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1136(%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovaps 1072(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm13 = xmm5[0],xmm12[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 1504(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1440(%rdi), %ymm6
; AVX1-ONLY-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm0[0],ymm6[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1392(%rdi), %xmm13
; AVX1-ONLY-NEXT: vmovaps 1328(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm14 = xmm6[0],xmm13[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 1760(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1696(%rdi), %ymm7
; AVX1-ONLY-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1648(%rdi), %xmm14
; AVX1-ONLY-NEXT: vmovaps 1584(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm15 = xmm7[0],xmm14[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 2016(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1952(%rdi), %ymm8
; AVX1-ONLY-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm0[0],ymm8[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1904(%rdi), %xmm15
; AVX1-ONLY-NEXT: vmovaps 1840(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm15[0]
; 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: 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 {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm2[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm1[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 {{.*#+}} xmm1 = xmm3[1],xmm10[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm10 = ymm1[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 {{.*#+}} xmm1 = xmm4[1],xmm11[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm1[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 {{.*#+}} xmm1 = xmm5[1],xmm12[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm1[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 {{.*#+}} xmm1 = xmm6[1],xmm13[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm1[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 {{.*#+}} xmm1 = xmm7[1],xmm14[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[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 {{.*#+}} xmm5 = xmm8[1],xmm15[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 240(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 224(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 32(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 160(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 96(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 48(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 176(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 112(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 208(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 192(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 128(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 64(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, (%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 144(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 80(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 16(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 224(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 240(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 160(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 176(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 96(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 112(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 32(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 48(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 192(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 208(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, (%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 16(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 64(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 80(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 128(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 144(%rdx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 32(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 96(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 160(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 224(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 128(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 64(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, (%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 192(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 224(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 160(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 96(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 32(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 192(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 128(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 64(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, (%r8)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 240(%r9)
; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 224(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 48(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 160(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 176(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 96(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 112(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 32(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 16(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, (%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 192(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 208(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 128(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 144(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 64(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 80(%r9)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 240(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 224(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 208(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 192(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, (%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 16(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 80(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 160(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 176(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 144(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 128(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 112(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm5, 48(%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 224(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 192(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 160(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 128(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 96(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 64(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, 32(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm5, (%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %ymm0, 224(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm1, 192(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm2, 160(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm3, 128(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm4, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm10, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm9, 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 $2296, %rsp # imm = 0x8F8
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: load_i64_stride8_vf32:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: subq $2408, %rsp # imm = 0x968
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 704(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovaps 640(%rdi), %xmm3
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1216(%rdi), %xmm4
; AVX2-ONLY-NEXT: vmovaps 1152(%rdi), %xmm6
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm5 = xmm6[0],xmm4[0]
; AVX2-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 576(%rdi), %xmm5
; AVX2-ONLY-NEXT: vmovaps 512(%rdi), %xmm7
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm5[0]
; AVX2-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1088(%rdi), %xmm9
; AVX2-ONLY-NEXT: vmovaps 1024(%rdi), %xmm10
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm10[0],xmm9[0]
; AVX2-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1600(%rdi), %xmm8
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm4[1]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1536(%rdi), %xmm4
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm10[1],xmm9[1]
; AVX2-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm8[0]
; AVX2-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1728(%rdi), %xmm2
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm7[1],xmm5[1]
; AVX2-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1664(%rdi), %xmm3
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm3[0],xmm2[0]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %xmm0
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm8[1]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 960(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 896(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 832(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 768(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1472(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1408(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1344(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1280(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1984(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1920(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1856(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1792(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps (%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1600(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1536(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1728(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1664(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps (%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 576(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 512(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 704(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 640(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm12[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1088(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1024(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1216(%rdi), %ymm14
; AVX2-ONLY-NEXT: vmovaps 1152(%rdi), %ymm15
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1856(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1792(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1984(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1920(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm4[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1344(%rdi), %ymm12
; AVX2-ONLY-NEXT: vmovaps 1280(%rdi), %ymm11
; AVX2-ONLY-NEXT: vmovaps 1472(%rdi), %ymm10
; AVX2-ONLY-NEXT: vmovaps 1408(%rdi), %ymm8
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm6[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 832(%rdi), %ymm7
; AVX2-ONLY-NEXT: vmovaps 768(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 960(%rdi), %ymm4
; AVX2-ONLY-NEXT: vmovaps 896(%rdi), %ymm3
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm2[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %ymm0
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm2[0],ymm9[2],ymm2[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm13[2,3],ymm5[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm5 = ymm13[2,3],ymm5[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm5 = ymm13[2,3],ymm5[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm5 = ymm13[2,3],ymm5[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm5 = ymm13[2,3],ymm5[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm2[1],ymm9[3],ymm2[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 992(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 928(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 864(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 800(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1120(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1056(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1248(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1184(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1504(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1440(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1376(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1312(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 736(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 672(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 608(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 544(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1632(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1568(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1760(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1696(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1888(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1824(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2016(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1952(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm2[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 608(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 544(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 736(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 672(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 864(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 800(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 992(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 928(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1120(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1056(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1248(%rdi), %ymm15
; AVX2-ONLY-NEXT: vmovaps 1184(%rdi), %ymm14
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm14[0],ymm15[0],ymm14[2],ymm15[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1376(%rdi), %ymm13
; AVX2-ONLY-NEXT: vmovaps 1312(%rdi), %ymm12
; AVX2-ONLY-NEXT: vmovaps 1504(%rdi), %ymm11
; AVX2-ONLY-NEXT: vmovaps 1440(%rdi), %ymm10
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm12[0],ymm13[0],ymm12[2],ymm13[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1632(%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 1568(%rdi), %ymm8
; AVX2-ONLY-NEXT: vmovaps 1760(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 1696(%rdi), %ymm5
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm8[0],ymm9[0],ymm8[2],ymm9[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1888(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovaps 1824(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 2016(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovaps 1952(%rdi), %ymm0
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm7 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm7 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm7[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm4 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm4 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm4 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm14[1],ymm15[1],ymm14[3],ymm15[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm14 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm7 = ymm12[1],ymm13[1],ymm12[3],ymm13[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm7[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm5 = ymm8[1],ymm9[1],ymm8[3],ymm9[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm5[2,3],ymm4[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%rdx)
; 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, 96(%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, 224(%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, 192(%rcx)
; 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 {{[-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, 32(%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, 128(%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, (%r8)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%r9)
; AVX2-ONLY-NEXT: vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%r9)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 224(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 192(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 160(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 128(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 96(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 64(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 32(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, (%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps %ymm0, 224(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm5, 192(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm10, 160(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm14, 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 $2408, %rsp # imm = 0x968
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-LABEL: load_i64_stride8_vf32:
; AVX512F: # %bb.0:
; AVX512F-NEXT: subq $2568, %rsp # imm = 0xA08
; AVX512F-NEXT: vmovdqa64 1856(%rdi), %zmm21
; AVX512F-NEXT: vmovdqa64 1984(%rdi), %zmm24
; AVX512F-NEXT: vmovdqa64 1344(%rdi), %zmm19
; AVX512F-NEXT: vmovdqa64 1280(%rdi), %zmm31
; AVX512F-NEXT: vmovdqa64 1472(%rdi), %zmm26
; AVX512F-NEXT: vmovdqa64 1408(%rdi), %zmm27
; AVX512F-NEXT: vmovdqa64 832(%rdi), %zmm22
; AVX512F-NEXT: vmovdqa64 768(%rdi), %zmm9
; AVX512F-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 960(%rdi), %zmm30
; AVX512F-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512F-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512F-NEXT: vmovdqa64 256(%rdi), %zmm6
; AVX512F-NEXT: vmovdqa64 448(%rdi), %zmm20
; AVX512F-NEXT: vmovdqa64 384(%rdi), %zmm29
; AVX512F-NEXT: movb $-64, %al
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512F-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm27, %zmm1
; AVX512F-NEXT: vpermt2q %zmm26, %zmm0, %zmm1
; AVX512F-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-NEXT: vpermt2q %zmm19, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512F-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512F-NEXT: vmovdqa 1152(%rdi), %ymm1
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm1[0],ymm28[0],ymm1[2],ymm28[2]
; AVX512F-NEXT: vmovdqa64 1088(%rdi), %ymm25
; AVX512F-NEXT: vmovdqa 1024(%rdi), %ymm5
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm25[0],ymm5[2],ymm25[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm7[2,3],ymm4[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm2
; AVX512F-NEXT: vpermt2q %zmm30, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm4
; AVX512F-NEXT: vpermt2q %zmm22, %zmm0, %zmm4
; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512F-NEXT: vmovdqa64 704(%rdi), %ymm23
; AVX512F-NEXT: vmovdqa64 640(%rdi), %ymm17
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm17[0],ymm23[0],ymm17[2],ymm23[2]
; AVX512F-NEXT: vmovdqa 576(%rdi), %ymm2
; AVX512F-NEXT: vmovdqa 512(%rdi), %ymm10
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm10[0],ymm2[0],ymm10[2],ymm2[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm11[2,3],ymm8[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm8, %zmm4, %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm29, %zmm4
; AVX512F-NEXT: vpermt2q %zmm20, %zmm0, %zmm4
; AVX512F-NEXT: vmovdqa64 %zmm6, %zmm8
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm7
; AVX512F-NEXT: vpermt2q %zmm12, %zmm0, %zmm8
; AVX512F-NEXT: vmovdqa64 %zmm4, %zmm8 {%k1}
; AVX512F-NEXT: vmovdqa 192(%rdi), %ymm9
; AVX512F-NEXT: vmovdqa 128(%rdi), %ymm11
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
; AVX512F-NEXT: vmovdqa 64(%rdi), %ymm4
; AVX512F-NEXT: vmovdqa64 (%rdi), %ymm16
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm16[0],ymm4[0],ymm16[2],ymm4[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm12 = ymm15[2,3],ymm12[2,3]
; AVX512F-NEXT: vmovdqa64 1920(%rdi), %zmm13
; AVX512F-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vinserti64x4 $0, %ymm12, %zmm8, %zmm8
; AVX512F-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm8
; AVX512F-NEXT: vpermt2q %zmm24, %zmm0, %zmm8
; AVX512F-NEXT: vmovdqa64 1792(%rdi), %zmm12
; AVX512F-NEXT: vmovdqu64 %zmm12, (%rsp) # 64-byte Spill
; AVX512F-NEXT: vpermi2q %zmm21, %zmm12, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512F-NEXT: vmovdqa 1728(%rdi), %ymm14
; AVX512F-NEXT: vmovdqa 1664(%rdi), %ymm12
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm12[0],ymm14[0],ymm12[2],ymm14[2]
; AVX512F-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512F-NEXT: vmovdqa 1536(%rdi), %ymm8
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm8[0],ymm18[0],ymm8[2],ymm18[2]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm15[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm13, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512F-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512F-NEXT: vpermt2q %zmm26, %zmm0, %zmm13
; AVX512F-NEXT: vmovdqa64 %zmm31, %zmm15
; AVX512F-NEXT: vpermt2q %zmm19, %zmm0, %zmm15
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm28[1],ymm1[3],ymm28[3]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm25[1],ymm5[3],ymm25[3]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512F-NEXT: vpermt2q %zmm30, %zmm0, %zmm1
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm13
; AVX512F-NEXT: vpermt2q %zmm22, %zmm0, %zmm3
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm3 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm17[1],ymm23[1],ymm17[3],ymm23[3]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm10[1],ymm2[1],ymm10[3],ymm2[3]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm29, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm20, %zmm0, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm6, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm9[1],ymm11[3],ymm9[3]
; AVX512F-NEXT: vmovdqa64 1216(%rdi), %zmm9
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm16[1],ymm4[1],ymm16[3],ymm4[3]
; AVX512F-NEXT: vmovdqa64 1152(%rdi), %zmm10
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX512F-NEXT: vmovdqa64 1088(%rdi), %zmm11
; AVX512F-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512F-NEXT: vpermt2q %zmm24, %zmm0, %zmm1
; AVX512F-NEXT: vmovdqu64 (%rsp), %zmm23 # 64-byte Reload
; AVX512F-NEXT: vpermi2q %zmm21, %zmm23, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512F-NEXT: vmovdqa64 1024(%rdi), %zmm4
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm14[1],ymm12[3],ymm14[3]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm18[1],ymm8[3],ymm18[3]
; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512F-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm27, %zmm1
; AVX512F-NEXT: vpermt2q %zmm26, %zmm0, %zmm1
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm31[0],zmm19[0],zmm31[2],zmm19[2],zmm31[4],zmm19[4],zmm31[6],zmm19[6]
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512F-NEXT: vmovdqa64 %zmm4, %zmm3
; AVX512F-NEXT: vmovdqa64 %zmm4, %zmm12
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm11, %zmm0, %zmm3
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,12,4,12]
; AVX512F-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm4
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm14
; AVX512F-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm9, %zmm1, %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm2
; AVX512F-NEXT: vpermt2q %zmm30, %zmm0, %zmm2
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm15[0],zmm13[0],zmm15[2],zmm13[2],zmm15[4],zmm13[4],zmm15[6],zmm13[6]
; AVX512F-NEXT: vmovdqa64 %zmm15, %zmm11
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm16
; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-NEXT: vmovdqa64 576(%rdi), %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 512(%rdi), %zmm13
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm2
; AVX512F-NEXT: vpermt2q %zmm4, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa64 704(%rdi), %zmm5
; AVX512F-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm5, %zmm1, %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm20, %zmm0, %zmm29
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
; AVX512F-NEXT: vmovdqa64 %zmm29, %zmm3 {%k1}
; AVX512F-NEXT: vmovdqa64 64(%rdi), %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm4, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa64 192(%rdi), %zmm6
; AVX512F-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm6, %zmm1, %zmm4
; AVX512F-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm2
; AVX512F-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm4
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm23[0],zmm21[0],zmm23[2],zmm21[2],zmm23[4],zmm21[4],zmm23[6],zmm21[6]
; AVX512F-NEXT: vmovdqa64 %zmm21, %zmm6
; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-NEXT: vmovdqa64 1600(%rdi), %zmm7
; AVX512F-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermi2q %zmm7, %zmm2, %zmm0
; AVX512F-NEXT: vmovdqa64 1728(%rdi), %zmm7
; AVX512F-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 1664(%rdi), %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermi2q %zmm7, %zmm2, %zmm1
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [0,8,0,8,0,8,0,8]
; AVX512F-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512F-NEXT: vpermt2q %zmm19, %zmm18, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [1,9,1,9,1,9,1,9]
; AVX512F-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512F-NEXT: vpermt2q %zmm19, %zmm23, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [6,14,6,14,6,14,6,14]
; AVX512F-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-NEXT: vpermt2q %zmm19, %zmm3, %zmm2
; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm1
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm31[1],zmm19[1],zmm31[3],zmm19[3],zmm31[5],zmm19[5],zmm31[7],zmm19[7]
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,7,15,7,15,7,15]
; AVX512F-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm19, %zmm10, %zmm31
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm27[0],zmm26[0],zmm27[2],zmm26[2],zmm27[4],zmm26[4],zmm27[6],zmm26[6]
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm27[1],zmm26[1],zmm27[3],zmm26[3],zmm27[5],zmm26[5],zmm27[7],zmm26[7]
; AVX512F-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm27, %zmm7
; AVX512F-NEXT: vmovdqa64 %zmm27, %zmm1
; AVX512F-NEXT: vpermt2q %zmm26, %zmm18, %zmm27
; AVX512F-NEXT: vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm26, %zmm23, %zmm7
; AVX512F-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512F-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-NEXT: vpermt2q %zmm26, %zmm2, %zmm1
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm8 {%k1}
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm19, %zmm2, %zmm12
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [5,13,5,13]
; AVX512F-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512F-NEXT: vmovdqa64 %zmm14, %zmm15
; AVX512F-NEXT: vpermt2q %zmm9, %zmm7, %zmm15
; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm8, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm27
; AVX512F-NEXT: vpermt2q %zmm16, %zmm18, %zmm27
; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm22
; AVX512F-NEXT: vpermt2q %zmm16, %zmm23, %zmm22
; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm21
; AVX512F-NEXT: vpermt2q %zmm16, %zmm3, %zmm21
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm11[1],zmm16[1],zmm11[3],zmm16[3],zmm11[5],zmm16[5],zmm11[7],zmm16[7]
; AVX512F-NEXT: vpermt2q %zmm16, %zmm10, %zmm11
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm21 {%k1} = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm11 {%k1} = zmm16[1],zmm30[1],zmm16[3],zmm30[3],zmm16[5],zmm30[5],zmm16[7],zmm30[7]
; AVX512F-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm29
; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm11
; AVX512F-NEXT: vpermt2q %zmm30, %zmm18, %zmm16
; AVX512F-NEXT: vpermt2q %zmm30, %zmm23, %zmm29
; AVX512F-NEXT: vpermt2q %zmm30, %zmm2, %zmm11
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm8
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm0, %zmm2, %zmm8
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm5, %zmm7, %zmm15
; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm15[4,5,6,7]
; AVX512F-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm31
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm1, %zmm18, %zmm31
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm20
; AVX512F-NEXT: vpermt2q %zmm1, %zmm23, %zmm20
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm26
; AVX512F-NEXT: vpermt2q %zmm1, %zmm3, %zmm26
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm1[1],zmm5[3],zmm1[3],zmm5[5],zmm1[5],zmm5[7],zmm1[7]
; AVX512F-NEXT: vpermt2q %zmm1, %zmm10, %zmm5
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm12[0],zmm28[0],zmm12[2],zmm28[2],zmm12[4],zmm28[4],zmm12[6],zmm28[6]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm12[1],zmm28[1],zmm12[3],zmm28[3],zmm12[5],zmm28[5],zmm12[7],zmm28[7]
; AVX512F-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm24
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm5
; AVX512F-NEXT: vpermt2q %zmm28, %zmm18, %zmm12
; AVX512F-NEXT: vpermt2q %zmm28, %zmm23, %zmm24
; AVX512F-NEXT: vpermt2q %zmm28, %zmm2, %zmm5
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm30, %zmm25
; AVX512F-NEXT: vpermt2q %zmm4, %zmm18, %zmm25
; AVX512F-NEXT: vmovdqa64 %zmm6, %zmm1
; AVX512F-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512F-NEXT: vpermi2q %zmm1, %zmm6, %zmm18
; AVX512F-NEXT: vmovdqa64 %zmm30, %zmm28
; AVX512F-NEXT: vpermt2q %zmm4, %zmm23, %zmm28
; AVX512F-NEXT: vpermi2q %zmm1, %zmm6, %zmm23
; AVX512F-NEXT: vmovdqa64 %zmm6, %zmm17
; AVX512F-NEXT: vpermt2q %zmm1, %zmm3, %zmm17
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm6[1],zmm1[1],zmm6[3],zmm1[3],zmm6[5],zmm1[5],zmm6[7],zmm1[7]
; AVX512F-NEXT: vpermt2q %zmm1, %zmm10, %zmm6
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} zmm17 {%k1} = zmm30[0],zmm4[0],zmm30[2],zmm4[2],zmm30[4],zmm4[4],zmm30[6],zmm4[6]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm30[1],zmm4[1],zmm30[3],zmm4[3],zmm30[5],zmm4[5],zmm30[7],zmm4[7]
; AVX512F-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm4, %zmm2, %zmm30
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm6
; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm15
; AVX512F-NEXT: vpermt2q %zmm0, %zmm3, %zmm6
; AVX512F-NEXT: vpermt2q %zmm0, %zmm10, %zmm13
; AVX512F-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm4, %zmm3
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm0, %zmm2, %zmm3
; AVX512F-NEXT: vmovdqa64 %zmm4, %zmm1
; AVX512F-NEXT: vpermt2q %zmm0, %zmm15, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm0, %zmm10, %zmm4
; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-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: vpermi2q %zmm1, %zmm0, %zmm2
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm4
; AVX512F-NEXT: vpermt2q %zmm19, %zmm15, %zmm4
; AVX512F-NEXT: vpermi2q %zmm1, %zmm0, %zmm15
; AVX512F-NEXT: vpermt2q %zmm1, %zmm10, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm19, %zmm10, %zmm13
; AVX512F-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
; AVX512F-NEXT: vinserti64x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm0 # 32-byte Folded Reload
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm9 {%k1}
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm5
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm13, %zmm7, %zmm5
; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm5[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm30, %zmm8 {%k1}
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-NEXT: vpermi2q %zmm9, %zmm5, %zmm7
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm7[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [6,14,6,14]
; AVX512F-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm3
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm1, %zmm0, %zmm3
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [7,15,7,15]
; AVX512F-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512F-NEXT: vpermt2q %zmm1, %zmm7, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm1
; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm11
; AVX512F-NEXT: vpermt2q %zmm13, %zmm0, %zmm11
; AVX512F-NEXT: vpermt2q %zmm13, %zmm7, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm8, %zmm10
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vpermt2q %zmm1, %zmm0, %zmm10
; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm2
; AVX512F-NEXT: vpermi2q %zmm9, %zmm5, %zmm0
; AVX512F-NEXT: vpermt2q %zmm9, %zmm7, %zmm2
; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpermt2q %zmm1, %zmm7, %zmm8
; AVX512F-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vinserti64x4 $0, %ymm4, %zmm1, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm3, %zmm21, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm31 {%k1}
; AVX512F-NEXT: vmovdqa 192(%rdi), %xmm14
; AVX512F-NEXT: vmovdqa 128(%rdi), %xmm13
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm8 = xmm13[0],xmm14[0]
; AVX512F-NEXT: vmovdqa (%rdi), %xmm10
; AVX512F-NEXT: vmovdqa 64(%rdi), %xmm7
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm21 = xmm10[0],xmm7[0]
; AVX512F-NEXT: vinserti32x4 $1, %xmm8, %ymm21, %ymm8
; AVX512F-NEXT: vinserti64x4 $0, %ymm8, %zmm31, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm27 {%k1}
; AVX512F-NEXT: vmovdqa 704(%rdi), %xmm12
; AVX512F-NEXT: vmovdqa64 640(%rdi), %xmm21
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm21[0],xmm12[0]
; AVX512F-NEXT: vmovdqa64 576(%rdi), %xmm31
; AVX512F-NEXT: vmovdqa 512(%rdi), %xmm5
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm9 = xmm5[0],xmm31[0]
; AVX512F-NEXT: vinserti32x4 $1, %xmm30, %ymm9, %ymm9
; AVX512F-NEXT: vinserti64x4 $0, %ymm9, %zmm27, %zmm1
; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm3 {%k1}
; AVX512F-NEXT: vmovdqa64 1216(%rdi), %xmm19
; AVX512F-NEXT: vmovdqa64 1152(%rdi), %xmm27
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm27[0],xmm19[0]
; AVX512F-NEXT: vmovdqa 1088(%rdi), %xmm2
; AVX512F-NEXT: vmovdqa 1024(%rdi), %xmm1
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm1[0],xmm2[0]
; AVX512F-NEXT: vinserti32x4 $1, %xmm30, %ymm4, %ymm4
; AVX512F-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm8
; AVX512F-NEXT: vmovdqa64 %zmm25, %zmm18 {%k1}
; AVX512F-NEXT: vmovdqa64 1728(%rdi), %xmm25
; AVX512F-NEXT: vmovdqa64 1664(%rdi), %xmm30
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm30[0],xmm25[0]
; AVX512F-NEXT: vmovdqa 1600(%rdi), %xmm4
; AVX512F-NEXT: vmovdqa 1536(%rdi), %xmm3
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm9 = xmm3[0],xmm4[0]
; AVX512F-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
; AVX512F-NEXT: vinserti64x4 $0, %ymm6, %zmm18, %zmm6
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm16 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm27[1],xmm19[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX512F-NEXT: vinserti128 $1, %xmm9, %ymm1, %ymm1
; AVX512F-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm1
; AVX512F-NEXT: vmovdqa64 %zmm29, %zmm22 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm21[1],xmm12[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm31[1]
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2
; AVX512F-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm2
; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm20 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm13[1],xmm14[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm10[1],xmm7[1]
; AVX512F-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
; AVX512F-NEXT: vinserti64x4 $0, %ymm5, %zmm20, %zmm5
; AVX512F-NEXT: vmovdqa64 %zmm28, %zmm23 {%k1}
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm30[1],xmm25[1]
; AVX512F-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; AVX512F-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3
; AVX512F-NEXT: vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512F-NEXT: vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm4 # 32-byte Folded Reload
; AVX512F-NEXT: # ymm4 = mem[0,1,2,3],ymm11[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm4, %zmm26, %zmm4
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX512F-NEXT: vinserti64x4 $0, %ymm0, %zmm17, %zmm0
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512F-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512F-NEXT: # ymm7 = mem[0,1,2,3],ymm7[4,5,6,7]
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-NEXT: vinsertf64x4 $0, %ymm7, %zmm9, %zmm7
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload
; AVX512F-NEXT: # ymm9 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-NEXT: vinsertf64x4 $0, %ymm9, %zmm10, %zmm9
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
; AVX512F-NEXT: # ymm10 = mem[0,1,2,3],ymm10[4,5,6,7]
; AVX512F-NEXT: vmovups (%rsp), %zmm11 # 64-byte Reload
; AVX512F-NEXT: vinsertf64x4 $0, %ymm10, %zmm11, %zmm10
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512F-NEXT: # ymm11 = mem[0,1,2,3],ymm11[4,5,6,7]
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-NEXT: vinsertf64x4 $0, %ymm11, %zmm12, %zmm11
; AVX512F-NEXT: vmovdqa64 %zmm6, 192(%rsi)
; AVX512F-NEXT: vmovdqa64 %zmm8, 128(%rsi)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm6, 64(%rsi)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm6, (%rsi)
; AVX512F-NEXT: vmovdqa64 %zmm3, 192(%rdx)
; AVX512F-NEXT: vmovdqa64 %zmm5, (%rdx)
; AVX512F-NEXT: vmovdqa64 %zmm2, 64(%rdx)
; AVX512F-NEXT: vmovdqa64 %zmm1, 128(%rdx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 192(%rcx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%rcx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 64(%rcx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 128(%rcx)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 192(%r8)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%r8)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 64(%r8)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 128(%r8)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 192(%r9)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%r9)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 64(%r9)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 128(%r9)
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 192(%rax)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, (%rax)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 64(%rax)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm1, 128(%rax)
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovdqa64 %zmm0, 192(%rax)
; AVX512F-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovaps %zmm11, 128(%rax)
; AVX512F-NEXT: vmovaps %zmm10, 192(%rax)
; AVX512F-NEXT: vmovaps %zmm9, (%rax)
; AVX512F-NEXT: vmovaps %zmm7, 64(%rax)
; AVX512F-NEXT: addq $2568, %rsp # imm = 0xA08
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_i64_stride8_vf32:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: subq $2568, %rsp # imm = 0xA08
; AVX512BW-NEXT: vmovdqa64 1856(%rdi), %zmm21
; AVX512BW-NEXT: vmovdqa64 1984(%rdi), %zmm24
; AVX512BW-NEXT: vmovdqa64 1344(%rdi), %zmm19
; AVX512BW-NEXT: vmovdqa64 1280(%rdi), %zmm31
; AVX512BW-NEXT: vmovdqa64 1472(%rdi), %zmm26
; AVX512BW-NEXT: vmovdqa64 1408(%rdi), %zmm27
; AVX512BW-NEXT: vmovdqa64 832(%rdi), %zmm22
; AVX512BW-NEXT: vmovdqa64 768(%rdi), %zmm9
; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 960(%rdi), %zmm30
; AVX512BW-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm6
; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm20
; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm29
; AVX512BW-NEXT: movb $-64, %al
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm26, %zmm0, %zmm1
; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512BW-NEXT: vmovdqa 1152(%rdi), %ymm1
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm1[0],ymm28[0],ymm1[2],ymm28[2]
; AVX512BW-NEXT: vmovdqa64 1088(%rdi), %ymm25
; AVX512BW-NEXT: vmovdqa 1024(%rdi), %ymm5
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm25[0],ymm5[2],ymm25[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm7[2,3],ymm4[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm2
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm22, %zmm0, %zmm4
; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512BW-NEXT: vmovdqa64 704(%rdi), %ymm23
; AVX512BW-NEXT: vmovdqa64 640(%rdi), %ymm17
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm17[0],ymm23[0],ymm17[2],ymm23[2]
; AVX512BW-NEXT: vmovdqa 576(%rdi), %ymm2
; AVX512BW-NEXT: vmovdqa 512(%rdi), %ymm10
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm10[0],ymm2[0],ymm10[2],ymm2[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm11[2,3],ymm8[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm4, %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm20, %zmm0, %zmm4
; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm8
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm7
; AVX512BW-NEXT: vpermt2q %zmm12, %zmm0, %zmm8
; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm8 {%k1}
; AVX512BW-NEXT: vmovdqa 192(%rdi), %ymm9
; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm11
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
; AVX512BW-NEXT: vmovdqa 64(%rdi), %ymm4
; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm16
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm16[0],ymm4[0],ymm16[2],ymm4[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm12 = ymm15[2,3],ymm12[2,3]
; AVX512BW-NEXT: vmovdqa64 1920(%rdi), %zmm13
; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vinserti64x4 $0, %ymm12, %zmm8, %zmm8
; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm8
; AVX512BW-NEXT: vpermt2q %zmm24, %zmm0, %zmm8
; AVX512BW-NEXT: vmovdqa64 1792(%rdi), %zmm12
; AVX512BW-NEXT: vmovdqu64 %zmm12, (%rsp) # 64-byte Spill
; AVX512BW-NEXT: vpermi2q %zmm21, %zmm12, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512BW-NEXT: vmovdqa 1728(%rdi), %ymm14
; AVX512BW-NEXT: vmovdqa 1664(%rdi), %ymm12
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm12[0],ymm14[0],ymm12[2],ymm14[2]
; AVX512BW-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512BW-NEXT: vmovdqa 1536(%rdi), %ymm8
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm8[0],ymm18[0],ymm8[2],ymm18[2]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm15[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm13, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512BW-NEXT: vpermt2q %zmm26, %zmm0, %zmm13
; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm15
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm0, %zmm15
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm28[1],ymm1[3],ymm28[3]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm25[1],ymm5[3],ymm25[3]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm0, %zmm1
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm13
; AVX512BW-NEXT: vpermt2q %zmm22, %zmm0, %zmm3
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm3 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm17[1],ymm23[1],ymm17[3],ymm23[3]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm10[1],ymm2[1],ymm10[3],ymm2[3]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm20, %zmm0, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm9[1],ymm11[3],ymm9[3]
; AVX512BW-NEXT: vmovdqa64 1216(%rdi), %zmm9
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm16[1],ymm4[1],ymm16[3],ymm4[3]
; AVX512BW-NEXT: vmovdqa64 1152(%rdi), %zmm10
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX512BW-NEXT: vmovdqa64 1088(%rdi), %zmm11
; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm24, %zmm0, %zmm1
; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm23 # 64-byte Reload
; AVX512BW-NEXT: vpermi2q %zmm21, %zmm23, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: vmovdqa64 1024(%rdi), %zmm4
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm14[1],ymm12[3],ymm14[3]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm18[1],ymm8[3],ymm18[3]
; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm26, %zmm0, %zmm1
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm31[0],zmm19[0],zmm31[2],zmm19[2],zmm31[4],zmm19[4],zmm31[6],zmm19[6]
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm3
; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm12
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm11, %zmm0, %zmm3
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,12,4,12]
; AVX512BW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm4
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm14
; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm9, %zmm1, %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm2
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm0, %zmm2
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm15[0],zmm13[0],zmm15[2],zmm13[2],zmm15[4],zmm13[4],zmm15[6],zmm13[6]
; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm11
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm16
; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm13
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm2
; AVX512BW-NEXT: vpermt2q %zmm4, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm5
; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm5, %zmm1, %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm20, %zmm0, %zmm29
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm3 {%k1}
; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm4, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm6
; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm6, %zmm1, %zmm4
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm2
; AVX512BW-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm4
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm23[0],zmm21[0],zmm23[2],zmm21[2],zmm23[4],zmm21[4],zmm23[6],zmm21[6]
; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm6
; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT: vmovdqa64 1600(%rdi), %zmm7
; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermi2q %zmm7, %zmm2, %zmm0
; AVX512BW-NEXT: vmovdqa64 1728(%rdi), %zmm7
; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 1664(%rdi), %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermi2q %zmm7, %zmm2, %zmm1
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm18, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm23, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm3, %zmm2
; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm31[1],zmm19[1],zmm31[3],zmm19[3],zmm31[5],zmm19[5],zmm31[7],zmm19[7]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm10, %zmm31
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm27[0],zmm26[0],zmm27[2],zmm26[2],zmm27[4],zmm26[4],zmm27[6],zmm26[6]
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm27[1],zmm26[1],zmm27[3],zmm26[3],zmm27[5],zmm26[5],zmm27[7],zmm26[7]
; AVX512BW-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm7
; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm26, %zmm18, %zmm27
; AVX512BW-NEXT: vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm26, %zmm23, %zmm7
; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm26, %zmm2, %zmm1
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm8 {%k1}
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm2, %zmm12
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [5,13,5,13]
; AVX512BW-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm15
; AVX512BW-NEXT: vpermt2q %zmm9, %zmm7, %zmm15
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm8, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm27
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm18, %zmm27
; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm22
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm23, %zmm22
; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm21
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm3, %zmm21
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm11[1],zmm16[1],zmm11[3],zmm16[3],zmm11[5],zmm16[5],zmm11[7],zmm16[7]
; AVX512BW-NEXT: vpermt2q %zmm16, %zmm10, %zmm11
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm21 {%k1} = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm11 {%k1} = zmm16[1],zmm30[1],zmm16[3],zmm30[3],zmm16[5],zmm30[5],zmm16[7],zmm30[7]
; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm29
; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm11
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm18, %zmm16
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm23, %zmm29
; AVX512BW-NEXT: vpermt2q %zmm30, %zmm2, %zmm11
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm8
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm2, %zmm8
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm5, %zmm7, %zmm15
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm31
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm18, %zmm31
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm20
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm20
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm26
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm3, %zmm26
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm1[1],zmm5[3],zmm1[3],zmm5[5],zmm1[5],zmm5[7],zmm1[7]
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm5
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm12[0],zmm28[0],zmm12[2],zmm28[2],zmm12[4],zmm28[4],zmm12[6],zmm28[6]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm12[1],zmm28[1],zmm12[3],zmm28[3],zmm12[5],zmm28[5],zmm12[7],zmm28[7]
; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm24
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm5
; AVX512BW-NEXT: vpermt2q %zmm28, %zmm18, %zmm12
; AVX512BW-NEXT: vpermt2q %zmm28, %zmm23, %zmm24
; AVX512BW-NEXT: vpermt2q %zmm28, %zmm2, %zmm5
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm25
; AVX512BW-NEXT: vpermt2q %zmm4, %zmm18, %zmm25
; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm1
; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512BW-NEXT: vpermi2q %zmm1, %zmm6, %zmm18
; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm28
; AVX512BW-NEXT: vpermt2q %zmm4, %zmm23, %zmm28
; AVX512BW-NEXT: vpermi2q %zmm1, %zmm6, %zmm23
; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm17
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm3, %zmm17
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm6[1],zmm1[1],zmm6[3],zmm1[3],zmm6[5],zmm1[5],zmm6[7],zmm1[7]
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm6
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm17 {%k1} = zmm30[0],zmm4[0],zmm30[2],zmm4[2],zmm30[4],zmm4[4],zmm30[6],zmm4[6]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm30[1],zmm4[1],zmm30[3],zmm4[3],zmm30[5],zmm4[5],zmm30[7],zmm4[7]
; AVX512BW-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm4, %zmm2, %zmm30
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm6
; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm15
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm3, %zmm6
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm10, %zmm13
; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm3
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm2, %zmm3
; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm15, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm10, %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-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: vpermi2q %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm4
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm15, %zmm4
; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm15
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm19, %zmm10, %zmm13
; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
; AVX512BW-NEXT: vinserti64x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm0 # 32-byte Folded Reload
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm9 {%k1}
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm5
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm13, %zmm7, %zmm5
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm8 {%k1}
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT: vpermi2q %zmm9, %zmm5, %zmm7
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [6,14,6,14]
; AVX512BW-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm3
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm0, %zmm3
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [7,15,7,15]
; AVX512BW-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm1
; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm11
; AVX512BW-NEXT: vpermt2q %zmm13, %zmm0, %zmm11
; AVX512BW-NEXT: vpermt2q %zmm13, %zmm7, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm0, %zmm10
; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm2
; AVX512BW-NEXT: vpermi2q %zmm9, %zmm5, %zmm0
; AVX512BW-NEXT: vpermt2q %zmm9, %zmm7, %zmm2
; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm8
; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm1, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm21, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm31 {%k1}
; AVX512BW-NEXT: vmovdqa 192(%rdi), %xmm14
; AVX512BW-NEXT: vmovdqa 128(%rdi), %xmm13
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm8 = xmm13[0],xmm14[0]
; AVX512BW-NEXT: vmovdqa (%rdi), %xmm10
; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm7
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm21 = xmm10[0],xmm7[0]
; AVX512BW-NEXT: vinserti32x4 $1, %xmm8, %ymm21, %ymm8
; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm31, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm27 {%k1}
; AVX512BW-NEXT: vmovdqa 704(%rdi), %xmm12
; AVX512BW-NEXT: vmovdqa64 640(%rdi), %xmm21
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm21[0],xmm12[0]
; AVX512BW-NEXT: vmovdqa64 576(%rdi), %xmm31
; AVX512BW-NEXT: vmovdqa 512(%rdi), %xmm5
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm9 = xmm5[0],xmm31[0]
; AVX512BW-NEXT: vinserti32x4 $1, %xmm30, %ymm9, %ymm9
; AVX512BW-NEXT: vinserti64x4 $0, %ymm9, %zmm27, %zmm1
; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm3 {%k1}
; AVX512BW-NEXT: vmovdqa64 1216(%rdi), %xmm19
; AVX512BW-NEXT: vmovdqa64 1152(%rdi), %xmm27
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm27[0],xmm19[0]
; AVX512BW-NEXT: vmovdqa 1088(%rdi), %xmm2
; AVX512BW-NEXT: vmovdqa 1024(%rdi), %xmm1
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm1[0],xmm2[0]
; AVX512BW-NEXT: vinserti32x4 $1, %xmm30, %ymm4, %ymm4
; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm8
; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm18 {%k1}
; AVX512BW-NEXT: vmovdqa64 1728(%rdi), %xmm25
; AVX512BW-NEXT: vmovdqa64 1664(%rdi), %xmm30
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm30[0],xmm25[0]
; AVX512BW-NEXT: vmovdqa 1600(%rdi), %xmm4
; AVX512BW-NEXT: vmovdqa 1536(%rdi), %xmm3
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm9 = xmm3[0],xmm4[0]
; AVX512BW-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm18, %zmm6
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm16 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm27[1],xmm19[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm9, %ymm1, %ymm1
; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm1
; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm22 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm21[1],xmm12[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm31[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2
; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm2
; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm20 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm13[1],xmm14[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm10[1],xmm7[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm20, %zmm5
; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm23 {%k1}
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm30[1],xmm25[1]
; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3
; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512BW-NEXT: vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm4 # 32-byte Folded Reload
; AVX512BW-NEXT: # ymm4 = mem[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm26, %zmm4
; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm17, %zmm0
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-NEXT: # ymm7 = mem[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT: vinsertf64x4 $0, %ymm7, %zmm9, %zmm7
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload
; AVX512BW-NEXT: # ymm9 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT: vinsertf64x4 $0, %ymm9, %zmm10, %zmm9
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
; AVX512BW-NEXT: # ymm10 = mem[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-NEXT: vmovups (%rsp), %zmm11 # 64-byte Reload
; AVX512BW-NEXT: vinsertf64x4 $0, %ymm10, %zmm11, %zmm10
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512BW-NEXT: # ymm11 = mem[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT: vinsertf64x4 $0, %ymm11, %zmm12, %zmm11
; AVX512BW-NEXT: vmovdqa64 %zmm6, 192(%rsi)
; AVX512BW-NEXT: vmovdqa64 %zmm8, 128(%rsi)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm6, 64(%rsi)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm6, (%rsi)
; AVX512BW-NEXT: vmovdqa64 %zmm3, 192(%rdx)
; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rdx)
; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rdx)
; AVX512BW-NEXT: vmovdqa64 %zmm1, 128(%rdx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 192(%rcx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%rcx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 64(%rcx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 128(%rcx)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 192(%r8)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%r8)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 64(%r8)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 128(%r8)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 192(%r9)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%r9)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 64(%r9)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 128(%r9)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 192(%rax)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, (%rax)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 64(%rax)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm1, 128(%rax)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovaps %zmm11, 128(%rax)
; AVX512BW-NEXT: vmovaps %zmm10, 192(%rax)
; AVX512BW-NEXT: vmovaps %zmm9, (%rax)
; AVX512BW-NEXT: vmovaps %zmm7, 64(%rax)
; AVX512BW-NEXT: addq $2568, %rsp # imm = 0xA08
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
%wide.vec = load <256 x i64>, ptr %in.vec, align 64
%strided.vec0 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248>
%strided.vec1 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249>
%strided.vec2 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250>
%strided.vec3 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251>
%strided.vec4 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252>
%strided.vec5 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253>
%strided.vec6 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254>
%strided.vec7 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255>
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
store <32 x i64> %strided.vec6, ptr %out.vec6, align 64
store <32 x i64> %strided.vec7, ptr %out.vec7, align 64
ret void
}
define void @load_i64_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf64:
; SSE: # %bb.0:
; SSE-NEXT: subq $3720, %rsp # imm = 0xE88
; SSE-NEXT: movaps 960(%rdi), %xmm0
; SSE-NEXT: movaps 832(%rdi), %xmm1
; SSE-NEXT: movaps 768(%rdi), %xmm8
; SSE-NEXT: movaps 704(%rdi), %xmm2
; SSE-NEXT: movaps 640(%rdi), %xmm9
; SSE-NEXT: movaps 576(%rdi), %xmm3
; SSE-NEXT: movaps 512(%rdi), %xmm10
; SSE-NEXT: movaps 448(%rdi), %xmm4
; SSE-NEXT: movaps 384(%rdi), %xmm11
; SSE-NEXT: movaps 320(%rdi), %xmm5
; SSE-NEXT: movaps 256(%rdi), %xmm12
; SSE-NEXT: movaps 192(%rdi), %xmm6
; SSE-NEXT: movaps 128(%rdi), %xmm13
; SSE-NEXT: movaps 64(%rdi), %xmm7
; SSE-NEXT: movaps (%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 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 1088(%rdi), %xmm0
; SSE-NEXT: movaps 1024(%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 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 1344(%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 1472(%rdi), %xmm0
; SSE-NEXT: movaps 1408(%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 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 1728(%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 1856(%rdi), %xmm0
; SSE-NEXT: movaps 1792(%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 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 2112(%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 2240(%rdi), %xmm0
; SSE-NEXT: movaps 2176(%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 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 2496(%rdi), %xmm0
; SSE-NEXT: movaps 2432(%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 2624(%rdi), %xmm0
; SSE-NEXT: movaps 2560(%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 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 2880(%rdi), %xmm0
; SSE-NEXT: movaps 2816(%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 3008(%rdi), %xmm0
; SSE-NEXT: movaps 2944(%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 3136(%rdi), %xmm0
; SSE-NEXT: movaps 3072(%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 3264(%rdi), %xmm0
; SSE-NEXT: movaps 3200(%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 3392(%rdi), %xmm0
; SSE-NEXT: movaps 3328(%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 3520(%rdi), %xmm0
; SSE-NEXT: movaps 3456(%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 3648(%rdi), %xmm0
; SSE-NEXT: movaps 3584(%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 3776(%rdi), %xmm0
; SSE-NEXT: movaps 3712(%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 3904(%rdi), %xmm0
; SSE-NEXT: movaps 3840(%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 4032(%rdi), %xmm0
; SSE-NEXT: movaps 3968(%rdi), %xmm1
; SSE-NEXT: movaps %xmm1, %xmm7
; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]
; 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 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 208(%rdi), %xmm0
; SSE-NEXT: movaps 144(%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 336(%rdi), %xmm0
; SSE-NEXT: movaps 272(%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 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 592(%rdi), %xmm0
; SSE-NEXT: movaps 528(%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 720(%rdi), %xmm0
; SSE-NEXT: movaps 656(%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 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 976(%rdi), %xmm0
; SSE-NEXT: movaps 912(%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 1040(%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 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 1360(%rdi), %xmm0
; SSE-NEXT: movaps 1296(%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 1424(%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 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 1744(%rdi), %xmm0
; SSE-NEXT: movaps 1680(%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 1808(%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 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 2128(%rdi), %xmm0
; SSE-NEXT: movaps 2064(%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 2192(%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 2384(%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 2512(%rdi), %xmm0
; SSE-NEXT: movaps 2448(%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 2576(%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 2768(%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 2896(%rdi), %xmm0
; SSE-NEXT: movaps 2832(%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 2960(%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 3152(%rdi), %xmm0
; SSE-NEXT: movaps 3088(%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 3280(%rdi), %xmm0
; SSE-NEXT: movaps 3216(%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 3408(%rdi), %xmm0
; SSE-NEXT: movaps 3344(%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 3536(%rdi), %xmm0
; SSE-NEXT: movaps 3472(%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 3664(%rdi), %xmm0
; SSE-NEXT: movaps 3600(%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 3792(%rdi), %xmm0
; SSE-NEXT: movaps 3728(%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 3920(%rdi), %xmm0
; SSE-NEXT: movaps 3856(%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 4048(%rdi), %xmm0
; SSE-NEXT: movaps 3984(%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 96(%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 224(%rdi), %xmm0
; SSE-NEXT: movaps 160(%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 288(%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 480(%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 608(%rdi), %xmm0
; SSE-NEXT: movaps 544(%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 672(%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 864(%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 992(%rdi), %xmm0
; SSE-NEXT: movaps 928(%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 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 1248(%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 1376(%rdi), %xmm0
; SSE-NEXT: movaps 1312(%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 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 1632(%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 1760(%rdi), %xmm0
; SSE-NEXT: movaps 1696(%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 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 2016(%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 2144(%rdi), %xmm0
; SSE-NEXT: movaps 2080(%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 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 2400(%rdi), %xmm0
; SSE-NEXT: movaps 2336(%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 2528(%rdi), %xmm0
; SSE-NEXT: movaps 2464(%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 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 2784(%rdi), %xmm0
; SSE-NEXT: movaps 2720(%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 2912(%rdi), %xmm0
; SSE-NEXT: movaps 2848(%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 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 3168(%rdi), %xmm0
; SSE-NEXT: movaps 3104(%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 3296(%rdi), %xmm0
; SSE-NEXT: movaps 3232(%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 3424(%rdi), %xmm0
; SSE-NEXT: movaps 3360(%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 3552(%rdi), %xmm0
; SSE-NEXT: movaps 3488(%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 3680(%rdi), %xmm0
; SSE-NEXT: movaps 3616(%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 3808(%rdi), %xmm0
; SSE-NEXT: movaps 3744(%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 3936(%rdi), %xmm0
; SSE-NEXT: movaps 3872(%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 4064(%rdi), %xmm0
; SSE-NEXT: movaps 4000(%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 112(%rdi), %xmm0
; SSE-NEXT: movaps 48(%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 240(%rdi), %xmm0
; SSE-NEXT: movaps 176(%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 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 496(%rdi), %xmm0
; SSE-NEXT: movaps 432(%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 624(%rdi), %xmm0
; SSE-NEXT: movaps 560(%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 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 880(%rdi), %xmm0
; SSE-NEXT: movaps 816(%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 944(%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 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 1264(%rdi), %xmm0
; SSE-NEXT: movaps 1200(%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 1328(%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 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 1648(%rdi), %xmm0
; SSE-NEXT: movaps 1584(%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 1712(%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 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 2032(%rdi), %xmm0
; SSE-NEXT: movaps 1968(%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 2096(%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 2288(%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 2416(%rdi), %xmm0
; SSE-NEXT: movaps 2352(%rdi), %xmm1
; SSE-NEXT: movaps %xmm1, %xmm2
; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT: movaps %xmm2, (%rsp) # 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 2480(%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 2672(%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 2800(%rdi), %xmm0
; SSE-NEXT: movaps 2736(%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 2864(%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 3056(%rdi), %xmm0
; SSE-NEXT: movaps 2992(%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 3184(%rdi), %xmm0
; SSE-NEXT: movaps 3120(%rdi), %xmm9
; 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 3312(%rdi), %xmm0
; SSE-NEXT: movaps 3248(%rdi), %xmm8
; 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 3440(%rdi), %xmm0
; SSE-NEXT: movaps 3376(%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 3568(%rdi), %xmm0
; SSE-NEXT: movaps 3504(%rdi), %xmm10
; SSE-NEXT: movaps %xmm10, %xmm14
; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT: movaps 4016(%rdi), %xmm2
; SSE-NEXT: movaps 3952(%rdi), %xmm4
; SSE-NEXT: movaps 3696(%rdi), %xmm0
; SSE-NEXT: movaps 3632(%rdi), %xmm3
; SSE-NEXT: movaps %xmm3, %xmm11
; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT: movaps 4080(%rdi), %xmm1
; SSE-NEXT: movaps 3888(%rdi), %xmm5
; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
; SSE-NEXT: movaps 3824(%rdi), %xmm6
; SSE-NEXT: movaps 3760(%rdi), %xmm0
; SSE-NEXT: movaps %xmm7, 496(%rsi)
; SSE-NEXT: movaps %xmm0, %xmm7
; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm6[1]
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; SSE-NEXT: movaps %xmm6, 480(%rsi)
; SSE-NEXT: movaps %xmm5, %xmm6
; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm4[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm4[1]
; SSE-NEXT: movaps %xmm2, %xmm4
; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm1[0]
; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm1[1]
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 464(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 448(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 432(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%rsi)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 496(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 480(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 464(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 448(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 432(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%rdx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 496(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 480(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 464(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 448(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 432(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%rcx)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 496(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 480(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 464(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 448(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 432(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%r8)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 496(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 480(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 464(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 448(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 432(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%r9)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%r9)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 496(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 480(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 464(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 448(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 432(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%rax)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm4, 496(%rax)
; SSE-NEXT: movaps %xmm6, 480(%rax)
; SSE-NEXT: movaps %xmm7, 464(%rax)
; SSE-NEXT: movaps %xmm11, 448(%rax)
; SSE-NEXT: movaps %xmm14, 432(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 416(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 400(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 384(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 368(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 352(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 336(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 320(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 304(%rax)
; SSE-NEXT: movaps (%rsp), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 288(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 272(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 256(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 240(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 224(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 208(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 192(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 176(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 160(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 144(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 128(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 112(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 96(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 80(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 64(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 48(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 32(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, 16(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT: movaps %xmm1, (%rax)
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movaps %xmm2, 496(%rax)
; SSE-NEXT: movaps %xmm5, 480(%rax)
; SSE-NEXT: movaps %xmm0, 464(%rax)
; SSE-NEXT: movaps %xmm3, 448(%rax)
; SSE-NEXT: movaps %xmm10, 432(%rax)
; SSE-NEXT: movaps %xmm12, 416(%rax)
; SSE-NEXT: movaps %xmm8, 400(%rax)
; SSE-NEXT: movaps %xmm9, 384(%rax)
; SSE-NEXT: movaps %xmm13, 368(%rax)
; SSE-NEXT: movaps %xmm15, 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 $3720, %rsp # imm = 0xE88
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: load_i64_stride8_vf64:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: subq $4968, %rsp # imm = 0x1368
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 704(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps 640(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1216(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovaps 1152(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2240(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovaps 2176(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm6[0]
; AVX1-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 576(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovaps 1088(%rdi), %xmm9
; AVX1-ONLY-NEXT: vmovaps 2112(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovaps 2048(%rdi), %xmm11
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm7[1],xmm6[1]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm11[0],xmm10[0]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm11[1],xmm10[1]
; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1024(%rdi), %xmm6
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm6[0],xmm9[0]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm9[1]
; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 512(%rdi), %xmm4
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm4[0],xmm8[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm4[1],xmm8[1]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm2
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm2[0]
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1728(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1664(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1600(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1536(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2752(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2688(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2624(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2560(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3264(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3200(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3136(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3072(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3776(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3712(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3648(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3584(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2496(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2432(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2368(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2304(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3008(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2944(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2880(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2816(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3520(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3456(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3392(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3328(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 4032(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3968(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3904(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3840(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 448(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 320(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 256(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 960(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 896(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 832(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 768(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1472(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1408(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1344(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1280(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1984(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1920(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1856(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1792(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 80(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-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: 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 704(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 640(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-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 592(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 528(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, (%rsp) # 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 1216(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1152(%rdi), %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 1104(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1040(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 1728(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1664(%rdi), %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), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1552(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 2240(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2176(%rdi), %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 2128(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2064(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 2752(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2688(%rdi), %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 2640(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2576(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 3264(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3200(%rdi), %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 3152(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3088(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 3776(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3712(%rdi), %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 3664(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3600(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 4032(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3968(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-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 3920(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3856(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 3520(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3456(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-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 3408(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3344(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 3008(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2944(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-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 2896(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2832(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 2496(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2432(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-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 2384(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2320(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-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 1984(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1920(%rdi), %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 1872(%rdi), %xmm14
; AVX1-ONLY-NEXT: vmovaps 1808(%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 1472(%rdi), %ymm12
; AVX1-ONLY-NEXT: vmovaps 1408(%rdi), %ymm11
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX1-ONLY-NEXT: vmovaps 1360(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovaps 1296(%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 960(%rdi), %ymm8
; AVX1-ONLY-NEXT: vmovaps 896(%rdi), %ymm7
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX1-ONLY-NEXT: vmovaps 848(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovaps 784(%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 448(%rdi), %ymm4
; AVX1-ONLY-NEXT: vmovaps 384(%rdi), %ymm3
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX1-ONLY-NEXT: vmovaps 336(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps 272(%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 (%rsp), %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: 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: vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm6[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],ymm12[1],ymm11[3],ymm12[3]
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm9[1],xmm10[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 = xmm13[1],xmm14[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: vmovaps 96(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 352(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 288(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 480(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 608(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 544(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 736(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 672(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 864(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 800(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 992(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 928(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1120(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1056(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1248(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1184(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1376(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1312(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1504(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1440(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1632(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1568(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1760(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1696(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1888(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1824(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2016(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 1952(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2144(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2080(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2272(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2208(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2400(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2336(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2528(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2464(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2656(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2592(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2784(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2720(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2912(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2848(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3040(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 2976(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3168(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3104(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3296(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3232(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3424(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3360(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3552(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3488(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3680(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3616(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3808(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3744(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3936(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 3872(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 4064(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps 4000(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 112(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 48(%rdi), %xmm2
; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[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: vmovaps 480(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 416(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 368(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 304(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm3 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[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: vmovaps 736(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 624(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 560(%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 992(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 880(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 816(%rdi), %xmm3
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm5 = xmm3[0],xmm0[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 1248(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1184(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1136(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1072(%rdi), %xmm4
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm5[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1504(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1392(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1328(%rdi), %xmm5
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm7 = xmm5[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm7[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 1760(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1648(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1584(%rdi), %xmm6
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm8 = xmm6[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm8[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: vmovaps 2016(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1952(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 1904(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 1840(%rdi), %xmm7
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm9 = xmm7[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2272(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 2160(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2096(%rdi), %xmm8
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2528(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm10 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 2416(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2352(%rdi), %xmm9
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm11 = xmm9[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm11[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 2784(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2720(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 2672(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2608(%rdi), %xmm10
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm12 = xmm10[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm11[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3040(%rdi), %ymm0
; 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: vunpcklpd {{.*#+}} ymm12 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 2928(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 2864(%rdi), %xmm11
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm13 = xmm11[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3296(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3232(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 3184(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3120(%rdi), %xmm12
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm14 = xmm12[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3552(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3488(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 3440(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3376(%rdi), %xmm13
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm15 = xmm13[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3808(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3744(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm15 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 3696(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3632(%rdi), %xmm14
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm14[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 4064(%rdi), %ymm0
; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vmovaps 4000(%rdi), %ymm1
; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX1-ONLY-NEXT: vmovaps 3952(%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps 3888(%rdi), %xmm15
; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm15[0],xmm0[0]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[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: 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: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm3[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm4[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm5[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm6[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm7[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm8[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm7 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm9[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm6 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm10[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm5 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm11[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm12[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm13[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm1 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm1 = xmm14[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[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 {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm8 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm8 = xmm15[1],mem[1]
; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 496(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 480(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 432(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 416(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 368(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 352(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 32(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 288(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 224(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 160(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 96(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 48(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 304(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 240(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 176(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 112(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 464(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 448(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 256(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 384(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 320(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 192(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 128(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 64(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 272(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 400(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 336(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 208(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 144(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 80(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 16(%rsi)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 224(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 240(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 160(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 176(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 96(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 112(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 32(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 48(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 480(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 496(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 416(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 432(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 352(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 368(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 288(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 304(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 448(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 464(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 384(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 400(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 320(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 336(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 192(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 208(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 16(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 64(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 80(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 128(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 144(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 256(%rdx)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 272(%rdx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 32(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 96(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 160(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 224(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 288(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 352(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 416(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 480(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 448(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 384(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 320(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 256(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 192(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 128(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 64(%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, (%rcx)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 480(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 416(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 352(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 288(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 224(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 160(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 96(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 32(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 448(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 384(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 320(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 256(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 192(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 128(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 64(%r8)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, (%r8)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 496(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 480(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 464(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 448(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 432(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 416(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 400(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 384(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 368(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 352(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 336(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 320(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 304(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 288(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 272(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 256(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 240(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 224(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 208(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 192(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 176(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 160(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 144(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 128(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 112(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 96(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 80(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 64(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 48(%r9)
; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 32(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 16(%r9)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%r9)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 496(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 480(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 464(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 448(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 432(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 416(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 400(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 384(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 368(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 352(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 336(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 320(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 304(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 288(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 272(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 256(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 240(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 224(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 208(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 192(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 176(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 160(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 144(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 128(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 112(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 80(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 48(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, 16(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm8, (%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 480(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 448(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 416(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 384(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 352(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 320(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 288(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 256(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 224(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 192(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 160(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 128(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 96(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 64(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, 32(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX1-ONLY-NEXT: vmovaps %ymm8, (%rax)
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovaps %ymm0, 480(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm1, 448(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm2, 416(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm3, 384(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm4, 352(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm5, 320(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm6, 288(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm7, 256(%rax)
; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %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 $4968, %rsp # imm = 0x1368
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: load_i64_stride8_vf64:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: subq $5560, %rsp # imm = 0x15B8
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 704(%rdi), %xmm2
; AVX2-ONLY-NEXT: vmovaps 640(%rdi), %xmm3
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1216(%rdi), %xmm4
; AVX2-ONLY-NEXT: vmovaps 1152(%rdi), %xmm5
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX2-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2240(%rdi), %xmm7
; AVX2-ONLY-NEXT: vmovaps 2176(%rdi), %xmm8
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm6 = xmm8[0],xmm7[0]
; AVX2-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 576(%rdi), %xmm6
; AVX2-ONLY-NEXT: vmovaps 1088(%rdi), %xmm9
; AVX2-ONLY-NEXT: vmovaps 2112(%rdi), %xmm10
; AVX2-ONLY-NEXT: vmovaps 2048(%rdi), %xmm11
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm8[1],xmm7[1]
; AVX2-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm7 = xmm11[0],xmm10[0]
; AVX2-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm11[1],xmm10[1]
; AVX2-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1024(%rdi), %xmm7
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm4 = xmm7[0],xmm9[0]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm7[1],xmm9[1]
; AVX2-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 512(%rdi), %xmm4
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm4[0],xmm6[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm4[1],xmm6[1]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1728(%rdi), %xmm2
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1664(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm2[0]
; AVX2-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1600(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1536(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2752(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2688(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2624(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2560(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3264(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3200(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3136(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3072(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3776(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3712(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3648(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3584(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2496(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2432(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2368(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2304(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3008(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2944(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2880(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2816(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3520(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3456(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3392(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3328(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 4032(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3968(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3904(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3840(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 960(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 896(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 832(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 768(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1472(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1408(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1344(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1280(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1984(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1920(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1856(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1792(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps (%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 64(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps (%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 192(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 128(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 576(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 512(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 704(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 640(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1088(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1024(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1216(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1152(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1600(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1536(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1728(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1664(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2112(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2048(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-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 2176(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2624(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2560(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2752(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2688(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3136(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3072(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3264(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3200(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3648(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3584(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3776(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3712(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3904(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3840(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 4032(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3968(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3392(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3328(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3520(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3456(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2880(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2816(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-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 2944(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2368(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2304(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2496(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2432(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm15 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1856(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1792(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1984(%rdi), %ymm11
; AVX2-ONLY-NEXT: vmovaps 1920(%rdi), %ymm10
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1344(%rdi), %ymm9
; AVX2-ONLY-NEXT: vmovaps 1280(%rdi), %ymm14
; AVX2-ONLY-NEXT: vmovaps 1472(%rdi), %ymm8
; AVX2-ONLY-NEXT: vmovaps 1408(%rdi), %ymm7
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 832(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 768(%rdi), %ymm13
; AVX2-ONLY-NEXT: vmovaps 960(%rdi), %ymm5
; AVX2-ONLY-NEXT: vmovaps 896(%rdi), %ymm4
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm13[0],ymm6[0],ymm13[2],ymm6[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 320(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovaps 256(%rdi), %ymm12
; AVX2-ONLY-NEXT: vmovaps 448(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 384(%rdi), %ymm0
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm15 = ymm12[0],ymm3[0],ymm12[2],ymm3[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-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 {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm1 = ymm15[2,3],ymm1[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm12[1],ymm3[1],ymm12[3],ymm3[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm6[1],ymm13[3],ymm6[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-ONLY-NEXT: # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 608(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 544(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 736(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 672(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 864(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 800(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 992(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 928(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1120(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1056(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1248(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1184(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1376(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1312(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1504(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1440(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1632(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1568(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1760(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1696(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1888(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1824(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2016(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 1952(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2144(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2080(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2272(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2208(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2400(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2336(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2528(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2464(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2656(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2592(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2784(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2720(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2912(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2848(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3040(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 2976(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3168(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3104(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3296(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3232(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3424(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3360(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3552(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3488(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3680(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3616(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3808(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3744(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3936(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 3872(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 4064(%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovaps 4000(%rdi), %xmm1
; AVX2-ONLY-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX2-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-ONLY-NEXT: vmovaps 96(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 32(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 224(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 160(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 352(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 288(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 480(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 416(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 608(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 544(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 736(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 672(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: vmovaps 800(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 992(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 928(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1120(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1056(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1248(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1184(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1376(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1312(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1504(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1440(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; 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: vmovaps 1568(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1760(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1696(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1888(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1824(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2016(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 1952(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2144(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2080(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2272(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2208(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2400(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2336(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2528(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2464(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2656(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2592(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2784(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2720(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2912(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2848(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3040(%rdi), %ymm0
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 2976(%rdi), %ymm3
; AVX2-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm15 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3168(%rdi), %ymm1
; AVX2-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3104(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3296(%rdi), %ymm11
; AVX2-ONLY-NEXT: vmovaps 3232(%rdi), %ymm9
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm11[0],ymm9[2],ymm11[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3424(%rdi), %ymm10
; AVX2-ONLY-NEXT: vmovaps 3360(%rdi), %ymm14
; AVX2-ONLY-NEXT: vmovaps 3552(%rdi), %ymm8
; AVX2-ONLY-NEXT: vmovaps 3488(%rdi), %ymm7
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm10[0],ymm14[2],ymm10[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3680(%rdi), %ymm6
; AVX2-ONLY-NEXT: vmovaps 3616(%rdi), %ymm13
; AVX2-ONLY-NEXT: vmovaps 3808(%rdi), %ymm5
; AVX2-ONLY-NEXT: vmovaps 3744(%rdi), %ymm3
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm13[0],ymm6[0],ymm13[2],ymm6[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-ONLY-NEXT: vmovaps 3936(%rdi), %ymm4
; AVX2-ONLY-NEXT: vmovaps 3872(%rdi), %ymm12
; AVX2-ONLY-NEXT: vmovaps 4064(%rdi), %ymm2
; AVX2-ONLY-NEXT: vmovaps 4000(%rdi), %ymm1
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-ONLY-NEXT: vunpcklpd {{.*#+}} ymm15 = ymm12[0],ymm4[0],ymm12[2],ymm4[2]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; 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: 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: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm15 = ymm15[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm11[1],ymm9[3],ymm11[3]
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; 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: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm7 = ymm14[1],ymm10[1],ymm14[3],ymm10[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm13[1],ymm6[1],ymm13[3],ymm6[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm12[1],ymm4[1],ymm12[3],ymm4[3]
; AVX2-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 496(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 480(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 432(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 416(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 368(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 352(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 288(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 304(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 464(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 448(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 256(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 384(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 320(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 272(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 400(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 336(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rsi)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 480(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 496(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 416(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 432(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 352(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 368(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 288(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 304(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 448(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 464(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 384(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 400(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 320(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 336(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 256(%rdx)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 272(%rdx)
; 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, 96(%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, 224(%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, 352(%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, 480(%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, 384(%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, 256(%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, 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, 480(%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, 352(%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, 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 {{[-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, 32(%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, 384(%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, 256(%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, 128(%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, (%r8)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 496(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 480(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 464(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 448(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 432(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 416(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 400(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 384(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 368(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 352(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 336(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 320(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 304(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 288(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 272(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 256(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%r9)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%r9)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 496(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 480(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 464(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 448(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 432(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 416(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 400(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 384(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 368(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 352(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 336(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 320(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 304(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 288(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 272(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 256(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 240(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 224(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 208(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 192(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 176(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 160(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 144(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 128(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 112(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 96(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 80(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 64(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 48(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 32(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, 16(%rax)
; AVX2-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-ONLY-NEXT: vmovaps %xmm1, (%rax)
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 480(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 448(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 416(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 384(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 352(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 320(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 288(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 256(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 224(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 192(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 160(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 128(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 96(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 64(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, 32(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm1, (%rax)
; 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 %ymm7, 416(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm9, 384(%rax)
; AVX2-ONLY-NEXT: vmovaps %ymm15, 352(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm0, 320(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm0, 288(%rax)
; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-ONLY-NEXT: vmovaps %ymm0, 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 $5560, %rsp # imm = 0x15B8
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-ONLY-SLOW-LABEL: load_i64_stride8_vf64:
; AVX512F-ONLY-SLOW: # %bb.0:
; AVX512F-ONLY-SLOW-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: movb $-64, %al
; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k1
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512F-ONLY-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512F-ONLY-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512F-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512F-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512F-ONLY-SLOW-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512F-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512F-ONLY-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512F-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512F-ONLY-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512F-ONLY-SLOW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512F-ONLY-SLOW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512F-ONLY-SLOW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512F-ONLY-SLOW-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512F-ONLY-SLOW-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512F-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512F-ONLY-SLOW-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512F-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512F-ONLY-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rsi)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-SLOW-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-ONLY-SLOW-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512F-ONLY-SLOW-NEXT: vzeroupper
; AVX512F-ONLY-SLOW-NEXT: retq
;
; AVX512F-ONLY-FAST-LABEL: load_i64_stride8_vf64:
; AVX512F-ONLY-FAST: # %bb.0:
; AVX512F-ONLY-FAST-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: movb $-64, %al
; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512F-ONLY-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512F-ONLY-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512F-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512F-ONLY-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512F-ONLY-FAST-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512F-ONLY-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512F-ONLY-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512F-ONLY-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512F-ONLY-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512F-ONLY-FAST-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512F-ONLY-FAST-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512F-ONLY-FAST-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512F-ONLY-FAST-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512F-ONLY-FAST-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512F-ONLY-FAST-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512F-ONLY-FAST-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512F-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512F-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%rsi)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512F-ONLY-FAST-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
; AVX512DQ-SLOW-LABEL: load_i64_stride8_vf64:
; AVX512DQ-SLOW: # %bb.0:
; AVX512DQ-SLOW-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512DQ-SLOW-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512DQ-SLOW-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512DQ-SLOW-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: movb $-64, %al
; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQ-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512DQ-SLOW-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512DQ-SLOW-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-SLOW-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512DQ-SLOW-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512DQ-SLOW-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512DQ-SLOW-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-SLOW-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512DQ-SLOW-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512DQ-SLOW-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512DQ-SLOW-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512DQ-SLOW-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512DQ-SLOW-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512DQ-SLOW-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512DQ-SLOW-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512DQ-SLOW-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512DQ-SLOW-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512DQ-SLOW-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512DQ-SLOW-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQ-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQ-SLOW-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512DQ-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512DQ-SLOW-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512DQ-SLOW-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512DQ-SLOW-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQ-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512DQ-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQ-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512DQ-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512DQ-SLOW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512DQ-SLOW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512DQ-SLOW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512DQ-SLOW-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512DQ-SLOW-NEXT: # ymm16 = mem[0,1,0,1]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512DQ-SLOW-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512DQ-SLOW-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512DQ-SLOW-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512DQ-SLOW-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512DQ-SLOW-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512DQ-SLOW-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512DQ-SLOW-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512DQ-SLOW-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512DQ-SLOW-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512DQ-SLOW-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512DQ-SLOW-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512DQ-SLOW-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512DQ-SLOW-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512DQ-SLOW-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512DQ-SLOW-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512DQ-SLOW-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512DQ-SLOW-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512DQ-SLOW-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512DQ-SLOW-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512DQ-SLOW-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512DQ-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512DQ-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512DQ-SLOW-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512DQ-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512DQ-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512DQ-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%rsi)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-SLOW-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQ-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQ-SLOW-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512DQ-SLOW-NEXT: vzeroupper
; AVX512DQ-SLOW-NEXT: retq
;
; AVX512DQ-FAST-LABEL: load_i64_stride8_vf64:
; AVX512DQ-FAST: # %bb.0:
; AVX512DQ-FAST-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512DQ-FAST-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512DQ-FAST-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512DQ-FAST-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512DQ-FAST-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: movb $-64, %al
; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQ-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512DQ-FAST-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512DQ-FAST-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-FAST-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512DQ-FAST-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512DQ-FAST-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512DQ-FAST-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-FAST-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512DQ-FAST-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512DQ-FAST-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512DQ-FAST-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512DQ-FAST-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512DQ-FAST-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512DQ-FAST-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512DQ-FAST-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512DQ-FAST-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512DQ-FAST-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512DQ-FAST-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512DQ-FAST-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQ-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQ-FAST-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512DQ-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512DQ-FAST-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512DQ-FAST-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512DQ-FAST-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQ-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512DQ-FAST-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512DQ-FAST-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQ-FAST-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQ-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512DQ-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQ-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQ-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512DQ-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512DQ-FAST-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512DQ-FAST-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512DQ-FAST-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512DQ-FAST-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512DQ-FAST-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512DQ-FAST-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512DQ-FAST-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512DQ-FAST-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512DQ-FAST-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512DQ-FAST-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQ-FAST-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512DQ-FAST-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vbroadcasti64x2 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512DQ-FAST-NEXT: # ymm16 = mem[0,1,0,1]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512DQ-FAST-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm0
; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512DQ-FAST-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512DQ-FAST-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512DQ-FAST-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512DQ-FAST-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512DQ-FAST-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512DQ-FAST-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512DQ-FAST-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512DQ-FAST-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512DQ-FAST-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512DQ-FAST-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512DQ-FAST-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512DQ-FAST-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512DQ-FAST-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512DQ-FAST-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512DQ-FAST-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512DQ-FAST-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512DQ-FAST-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512DQ-FAST-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512DQ-FAST-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512DQ-FAST-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512DQ-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512DQ-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512DQ-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%rsi)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FAST-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQ-FAST-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
; AVX512BW-ONLY-SLOW-LABEL: load_i64_stride8_vf64:
; AVX512BW-ONLY-SLOW: # %bb.0:
; AVX512BW-ONLY-SLOW-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: movb $-64, %al
; AVX512BW-ONLY-SLOW-NEXT: kmovd %eax, %k1
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512BW-ONLY-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512BW-ONLY-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512BW-ONLY-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512BW-ONLY-SLOW-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512BW-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512BW-ONLY-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512BW-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512BW-ONLY-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512BW-ONLY-SLOW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512BW-ONLY-SLOW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512BW-ONLY-SLOW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512BW-ONLY-SLOW-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512BW-ONLY-SLOW-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512BW-ONLY-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rsi)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512BW-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512BW-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512BW-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-SLOW-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-ONLY-SLOW-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512BW-ONLY-SLOW-NEXT: vzeroupper
; AVX512BW-ONLY-SLOW-NEXT: retq
;
; AVX512BW-ONLY-FAST-LABEL: load_i64_stride8_vf64:
; AVX512BW-ONLY-FAST: # %bb.0:
; AVX512BW-ONLY-FAST-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: movb $-64, %al
; AVX512BW-ONLY-FAST-NEXT: kmovd %eax, %k1
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512BW-ONLY-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512BW-ONLY-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512BW-ONLY-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-ONLY-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512BW-ONLY-FAST-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512BW-ONLY-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512BW-ONLY-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512BW-ONLY-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512BW-ONLY-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512BW-ONLY-FAST-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512BW-ONLY-FAST-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512BW-ONLY-FAST-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512BW-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512BW-ONLY-FAST-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512BW-ONLY-FAST-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512BW-ONLY-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512BW-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512BW-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512BW-ONLY-FAST-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512BW-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512BW-ONLY-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512BW-ONLY-FAST-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512BW-ONLY-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%rsi)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512BW-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512BW-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512BW-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-FAST-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512BW-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-ONLY-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512BW-ONLY-FAST-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512BW-ONLY-FAST-NEXT: vzeroupper
; AVX512BW-ONLY-FAST-NEXT: retq
;
; AVX512DQBW-SLOW-LABEL: load_i64_stride8_vf64:
; AVX512DQBW-SLOW: # %bb.0:
; AVX512DQBW-SLOW-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512DQBW-SLOW-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512DQBW-SLOW-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: movb $-64, %al
; AVX512DQBW-SLOW-NEXT: kmovd %eax, %k1
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQBW-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512DQBW-SLOW-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512DQBW-SLOW-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512DQBW-SLOW-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512DQBW-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512DQBW-SLOW-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512DQBW-SLOW-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQBW-SLOW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQBW-SLOW-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512DQBW-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQBW-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512DQBW-SLOW-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512DQBW-SLOW-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQBW-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512DQBW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512DQBW-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQBW-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512DQBW-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512DQBW-SLOW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512DQBW-SLOW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512DQBW-SLOW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512DQBW-SLOW-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512DQBW-SLOW-NEXT: # ymm16 = mem[0,1,0,1]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512DQBW-SLOW-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512DQBW-SLOW-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512DQBW-SLOW-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512DQBW-SLOW-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512DQBW-SLOW-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512DQBW-SLOW-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512DQBW-SLOW-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512DQBW-SLOW-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512DQBW-SLOW-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512DQBW-SLOW-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512DQBW-SLOW-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512DQBW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512DQBW-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512DQBW-SLOW-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512DQBW-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512DQBW-SLOW-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512DQBW-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%rsi)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512DQBW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQBW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQBW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-SLOW-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQBW-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-SLOW-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQBW-SLOW-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512DQBW-SLOW-NEXT: vzeroupper
; AVX512DQBW-SLOW-NEXT: retq
;
; AVX512DQBW-FAST-LABEL: load_i64_stride8_vf64:
; AVX512DQBW-FAST: # %bb.0:
; AVX512DQBW-FAST-NEXT: subq $6728, %rsp # imm = 0x1A48
; AVX512DQBW-FAST-NEXT: vmovdqa64 3392(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3328(%rdi), %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqa64 3520(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3456(%rdi), %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1856(%rdi), %zmm11
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1984(%rdi), %zmm15
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 832(%rdi), %zmm6
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 768(%rdi), %zmm7
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 960(%rdi), %zmm24
; AVX512DQBW-FAST-NEXT: vmovdqa64 896(%rdi), %zmm10
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 320(%rdi), %zmm12
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 256(%rdi), %zmm5
; AVX512DQBW-FAST-NEXT: vmovdqa64 448(%rdi), %zmm8
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 384(%rdi), %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: movb $-64, %al
; AVX512DQBW-FAST-NEXT: kmovd %eax, %k1
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQBW-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 3264(%rdi), %ymm21
; AVX512DQBW-FAST-NEXT: vmovdqa 3200(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm21[0],ymm0[2],ymm21[2]
; AVX512DQBW-FAST-NEXT: vmovdqa 3136(%rdi), %ymm4
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 3072(%rdi), %ymm3
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm6, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 704(%rdi), %ymm3
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 640(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQBW-FAST-NEXT: vmovdqa64 576(%rdi), %ymm25
; AVX512DQBW-FAST-NEXT: vmovdqa64 512(%rdi), %ymm23
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm23[0],ymm25[0],ymm23[2],ymm25[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm14, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm5, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 192(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 128(%rdi), %ymm31
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm31[0],ymm0[0],ymm31[2],ymm0[2]
; AVX512DQBW-FAST-NEXT: vmovdqa64 64(%rdi), %ymm20
; AVX512DQBW-FAST-NEXT: vmovdqa64 (%rdi), %ymm19
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm20[0],ymm19[2],ymm20[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vmovdqa64 1920(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 1792(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 1728(%rdi), %ymm3
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 1664(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQBW-FAST-NEXT: vmovdqa64 1600(%rdi), %ymm18
; AVX512DQBW-FAST-NEXT: vmovdqa64 1536(%rdi), %ymm26
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm26[0],ymm18[0],ymm26[2],ymm18[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1472(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1408(%rdi), %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 1344(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1280(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 1216(%rdi), %ymm28
; AVX512DQBW-FAST-NEXT: vmovdqa64 1152(%rdi), %ymm29
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm29[0],ymm28[0],ymm29[2],ymm28[2]
; AVX512DQBW-FAST-NEXT: vmovdqa64 1088(%rdi), %ymm30
; AVX512DQBW-FAST-NEXT: vmovdqa64 1024(%rdi), %ymm27
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm27[0],ymm30[0],ymm27[2],ymm30[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3008(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2944(%rdi), %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 2880(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2816(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 2752(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 2688(%rdi), %ymm11
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]
; AVX512DQBW-FAST-NEXT: vmovdqa64 2624(%rdi), %ymm16
; AVX512DQBW-FAST-NEXT: vmovdqa 2560(%rdi), %ymm9
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm9[0],ymm16[0],ymm9[2],ymm16[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2496(%rdi), %zmm10
; AVX512DQBW-FAST-NEXT: vmovdqa64 2432(%rdi), %zmm7
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm7, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2368(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2304(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 2240(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 2176(%rdi), %ymm3
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX512DQBW-FAST-NEXT: vmovdqa 2112(%rdi), %ymm3
; AVX512DQBW-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 2048(%rdi), %ymm8
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[2],ymm3[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 4032(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3968(%rdi), %zmm6
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm6, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 3904(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 3840(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm1, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 3776(%rdi), %ymm22
; AVX512DQBW-FAST-NEXT: vmovdqa64 3712(%rdi), %ymm17
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm15 = ymm17[0],ymm22[0],ymm17[2],ymm22[2]
; AVX512DQBW-FAST-NEXT: vmovdqa 3648(%rdi), %ymm12
; AVX512DQBW-FAST-NEXT: vmovdqa 3584(%rdi), %ymm0
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm14 = ymm14[2,3],ymm15[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm14, %zmm2, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQBW-FAST-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm4, %zmm2, %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm15
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQBW-FAST-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm13[2,3],ymm14[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm21, %zmm14
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm23[1],ymm25[1],ymm23[3],ymm25[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm25, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm24, %zmm2, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm23, %zmm2, %zmm5
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm31, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm13 = ymm31[1],mem[1],ymm31[3],mem[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm19[1],ymm20[1],ymm19[3],ymm20[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm5, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm31, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm20, %zmm2, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm19, %zmm2, %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm13 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm26[1],ymm18[1],ymm26[3],ymm18[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm18, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm26, %zmm2, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm29[1],ymm28[1],ymm29[3],ymm28[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm15 = ymm27[1],ymm30[1],ymm27[3],ymm30[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm15[2,3],ymm13[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm13, %zmm14, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm27, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm2, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm30, %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm29, %zmm2, %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm16[1],ymm9[3],ymm16[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm11[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm14, %zmm9
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm2, %zmm7
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm16, %zmm10
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm7 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm7 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm5 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm5 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm5[2,3],ymm7[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm10, %zmm5
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm14, %zmm2, %zmm6
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm3, %zmm8, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm6, %zmm2 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm17[1],ymm22[1],ymm17[3],ymm22[3]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm12[1],ymm0[3],ymm12[3]
; AVX512DQBW-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQBW-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, %zmm13
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm4, %zmm11
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm2 = zmm22[0],zmm9[0],zmm22[2],zmm9[2],zmm22[4],zmm9[4],zmm22[6],zmm9[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 3136(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3072(%rdi), %zmm15
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm15, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqa64 3264(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3200(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [4,12,4,12]
; AVX512DQBW-FAST-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm2, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 576(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 512(%rdi), %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 704(%rdi), %zmm21
; AVX512DQBW-FAST-NEXT: vmovdqa64 640(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm21, %zmm6, %zmm4
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm25, %zmm2
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm12[0],zmm23[0],zmm12[2],zmm23[2],zmm12[4],zmm23[4],zmm12[6],zmm23[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 64(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 (%rdi), %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 192(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 128(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm19[0],zmm1[2],zmm19[2],zmm1[4],zmm19[4],zmm1[6],zmm19[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 1600(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1536(%rdi), %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 1728(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1664(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm18, %zmm2
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm26, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm1[0],zmm10[0],zmm1[2],zmm10[2],zmm1[4],zmm10[4],zmm1[6],zmm10[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 1088(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1024(%rdi), %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 1216(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 1152(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm27, %zmm2
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 = zmm30[0],zmm29[0],zmm30[2],zmm29[2],zmm30[4],zmm29[4],zmm30[6],zmm29[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 2624(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2560(%rdi), %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 2752(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2688(%rdi), %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm4
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2112(%rdi), %zmm7
; AVX512DQBW-FAST-NEXT: vmovdqa64 2048(%rdi), %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2240(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 2176(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm30[0],zmm16[2],zmm30[2],zmm16[4],zmm30[4],zmm16[6],zmm30[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqa64 3648(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3584(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 3776(%rdi), %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 3712(%rdi), %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm6
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm1 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # zmm1 = zmm8[0],mem[0],zmm8[2],mem[2],zmm8[4],mem[4],zmm8[6],mem[6]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQBW-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm15
; AVX512DQBW-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [5,13,5,13]
; AVX512DQBW-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm13
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm22[1],zmm9[1],zmm22[3],zmm9[3],zmm22[5],zmm9[5],zmm22[7],zmm9[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm13, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm23, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm17, %zmm19
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm14, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # zmm4 = zmm12[1],mem[1],zmm12[3],mem[3],zmm12[5],mem[5],zmm12[7],mem[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm10, %zmm27
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm4[1],zmm10[1],zmm4[3],zmm10[3],zmm4[5],zmm10[5],zmm4[7],zmm10[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # zmm4 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm8, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm26, %zmm3
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm5[1],zmm30[1],zmm5[3],zmm30[3],zmm5[5],zmm30[5],zmm5[7],zmm30[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm31, %zmm2
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # zmm1 = zmm4[1],mem[1],zmm4[3],mem[3],zmm4[5],mem[5],zmm4[7],mem[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQBW-FAST-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,14,6,14]
; AVX512DQBW-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm24, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm22, %zmm3
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm3 {%k1} # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # zmm3 {%k1} = zmm22[0],mem[0],zmm22[2],mem[2],zmm22[4],mem[4],zmm22[6],mem[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm20, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm21, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm17, %zmm3
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm19, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm16, %zmm0, %zmm14
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm12, %zmm18
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm20[0],zmm14[0],zmm20[2],zmm14[2],zmm20[4],zmm14[4],zmm20[6],zmm14[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm13, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm12, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm26[0],zmm25[0],zmm26[2],zmm25[2],zmm26[4],zmm25[4],zmm26[6],zmm25[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm29, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm27, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm24[0],zmm29[0],zmm24[2],zmm29[2],zmm24[4],zmm29[4],zmm24[6],zmm29[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm9, %zmm0, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm1, %zmm3
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm9, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm22[0],zmm19[0],zmm22[2],zmm19[2],zmm22[4],zmm19[4],zmm22[6],zmm19[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm7, %zmm1, %zmm8
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm0, %zmm5
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm21[0],zmm17[0],zmm21[2],zmm17[2],zmm21[4],zmm17[4],zmm21[6],zmm17[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm31, %zmm6, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm30, %zmm3, %zmm1
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} zmm4 {%k1} = zmm1[0],zmm23[0],zmm1[2],zmm23[2],zmm1[4],zmm23[4],zmm1[6],zmm23[6]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm30
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm5
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm16 = [7,15,7,15,7,15,7,15]
; AVX512DQBW-FAST-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm4[1],zmm1[1],zmm4[3],zmm1[3],zmm4[5],zmm1[5],zmm4[7],zmm1[7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm4, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm4, %zmm6
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,8,0,8,0,8,0,8]
; AVX512DQBW-FAST-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm3, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [1,9,1,9,1,9,1,9]
; AVX512DQBW-FAST-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm2, %zmm3, %zmm30
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm2, %zmm13, %zmm5
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm18, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm18, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm16, %zmm1
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm20[1],zmm14[1],zmm20[3],zmm14[3],zmm20[5],zmm14[5],zmm20[7],zmm14[7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm20, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm14, %zmm3, %zmm20
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm14, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm3, %zmm18
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm28, %zmm13, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm12, %zmm28
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm12, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm16, %zmm12
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm12 {%k1} = zmm26[1],zmm25[1],zmm26[3],zmm25[3],zmm26[5],zmm25[5],zmm26[7],zmm25[7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm26, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm25, %zmm3, %zmm26
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm25, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm3, %zmm28
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm15, %zmm13, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, %zmm25
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm27, %zmm16, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm24, %zmm7
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm24[1],zmm29[1],zmm24[3],zmm29[3],zmm24[5],zmm29[5],zmm24[7],zmm29[7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm29, %zmm3, %zmm7
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm29, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm27, %zmm3, %zmm25
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm27, %zmm13, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm9, %zmm27
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm9, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm16, %zmm9
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm22[1],zmm19[1],zmm22[3],zmm19[3],zmm22[5],zmm19[5],zmm22[7],zmm19[7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm22, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm19, %zmm3, %zmm22
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm19, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm3, %zmm27
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm10, %zmm13, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm31, %zmm29
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm31, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm16, %zmm31
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm31 {%k1} = zmm21[1],zmm17[1],zmm21[3],zmm17[3],zmm21[5],zmm17[5],zmm21[7],zmm17[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm21, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm21, %zmm19
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm17, %zmm3, %zmm19
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm17, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm29
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm3, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm21, %zmm22
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm9, %zmm3, %zmm22
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm24, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm23, %zmm3, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm10, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm8, %zmm13, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm21, %zmm17
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm21, %zmm0
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm9, %zmm13, %zmm17
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm23, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpermi2q %zmm26, %zmm15, %zmm13
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm26, %zmm16, %zmm15
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm24[1],zmm23[1],zmm24[3],zmm23[3],zmm24[5],zmm23[5],zmm24[7],zmm23[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm15, %zmm21
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm15
; AVX512DQBW-FAST-NEXT: vpermt2q %zmm9, %zmm16, %zmm15
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm10[1],zmm8[1],zmm10[3],zmm8[3],zmm10[5],zmm8[5],zmm10[7],zmm8[7]
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm15, %zmm23
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm12 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm5 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm6 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm7 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm8 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm14 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vbroadcasti64x2 {{.*#+}} ymm16 = [7,15,7,15]
; AVX512DQBW-FAST-NEXT: # ymm16 = mem[0,1,0,1]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm15 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm0 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm2 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm11 # 64-byte Folded Reload
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm12[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm12, %zmm12
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm5[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm5, %zmm5
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm8[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm31, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm21, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vpblendd {{.*#+}} ymm15 = ymm14[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm15, %zmm23, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm20, %zmm18 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 192(%rdi), %xmm0
; AVX512DQBW-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 128(%rdi), %xmm1
; AVX512DQBW-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX512DQBW-FAST-NEXT: vmovdqa (%rdi), %xmm0
; AVX512DQBW-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
; AVX512DQBW-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm16 = xmm0[0],xmm1[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm18, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm30 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 704(%rdi), %xmm8
; AVX512DQBW-FAST-NEXT: vmovdqa64 640(%rdi), %xmm20
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm20[0],xmm8[0]
; AVX512DQBW-FAST-NEXT: vmovdqa64 576(%rdi), %xmm18
; AVX512DQBW-FAST-NEXT: vmovdqa64 512(%rdi), %xmm16
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm16[0],xmm18[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm31, %ymm2
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm30, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm25 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 1216(%rdi), %xmm0
; AVX512DQBW-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa 1152(%rdi), %xmm1
; AVX512DQBW-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm31 = xmm1[0],xmm0[0]
; AVX512DQBW-FAST-NEXT: vmovdqa64 1088(%rdi), %xmm23
; AVX512DQBW-FAST-NEXT: vmovdqa 1024(%rdi), %xmm14
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm30 = xmm14[0],xmm23[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm31, %ymm30, %ymm30
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm30, %zmm25, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 1728(%rdi), %xmm30
; AVX512DQBW-FAST-NEXT: vmovdqa64 1664(%rdi), %xmm31
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm31[0],xmm30[0]
; AVX512DQBW-FAST-NEXT: vmovdqa 1600(%rdi), %xmm12
; AVX512DQBW-FAST-NEXT: vmovdqa 1536(%rdi), %xmm11
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm11[0],xmm12[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm19, %zmm29 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 2240(%rdi), %xmm19
; AVX512DQBW-FAST-NEXT: vmovdqa64 2176(%rdi), %xmm21
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm21[0],xmm19[0]
; AVX512DQBW-FAST-NEXT: vmovdqa 2112(%rdi), %xmm15
; AVX512DQBW-FAST-NEXT: vmovdqa 2048(%rdi), %xmm1
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm28 = xmm1[0],xmm15[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm25, %ymm28, %ymm25
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm25, %zmm29, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 2752(%rdi), %xmm28
; AVX512DQBW-FAST-NEXT: vmovdqa64 2688(%rdi), %xmm29
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm29[0],xmm28[0]
; AVX512DQBW-FAST-NEXT: vmovdqa 2624(%rdi), %xmm10
; AVX512DQBW-FAST-NEXT: vmovdqa 2560(%rdi), %xmm9
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm25 = xmm9[0],xmm10[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm25, %ymm2
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm0
; AVX512DQBW-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa64 3264(%rdi), %xmm25
; AVX512DQBW-FAST-NEXT: vmovdqa64 3200(%rdi), %xmm27
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm27[0],xmm25[0]
; AVX512DQBW-FAST-NEXT: vmovdqa 3136(%rdi), %xmm0
; AVX512DQBW-FAST-NEXT: vmovdqa 3072(%rdi), %xmm2
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm26 = xmm2[0],xmm0[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm26, %ymm4
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm22, %zmm22
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa 3776(%rdi), %xmm7
; AVX512DQBW-FAST-NEXT: vmovdqa64 3712(%rdi), %xmm26
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm26[0],xmm7[0]
; AVX512DQBW-FAST-NEXT: vmovdqa 3648(%rdi), %xmm6
; AVX512DQBW-FAST-NEXT: vmovdqa 3584(%rdi), %xmm5
; AVX512DQBW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm24 = xmm5[0],xmm6[0]
; AVX512DQBW-FAST-NEXT: vinserti32x4 $1, %xmm4, %ymm24, %ymm4
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm3, %zmm24
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm27[1],xmm25[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm2
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm3 = xmm20[1],xmm8[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm4 = xmm16[1],xmm18[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm3, %zmm0, %zmm3
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # xmm4 = xmm4[1],mem[1]
; AVX512DQBW-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # xmm8 = xmm8[1],mem[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm4, %zmm0, %zmm4
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm8, %zmm0 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm8 = xmm31[1],xmm30[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm11 = xmm11[1],xmm12[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm8, %zmm0, %zmm8
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm11, %zmm0 {%k1}
; AVX512DQBW-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
; AVX512DQBW-FAST-NEXT: # xmm11 = xmm11[1],mem[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm14[1],xmm23[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm11
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm12, %zmm0 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm29[1],xmm28[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm12, %ymm9, %ymm9
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm9, %zmm0, %zmm9
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm10, %zmm0 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm10 = xmm21[1],xmm19[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm12 = xmm1[1],xmm15[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm10, %zmm0, %zmm10
; AVX512DQBW-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm0, %zmm13 {%k1}
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm7 = xmm26[1],xmm7[1]
; AVX512DQBW-FAST-NEXT: vpunpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX512DQBW-FAST-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5
; AVX512DQBW-FAST-NEXT: vinserti64x4 $0, %ymm5, %zmm13, %zmm1
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm24, 448(%rsi)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm22, 384(%rsi)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%rsi)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%rsi)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%rsi)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%rsi)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%rsi)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%rsi)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm1, 448(%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm10, 256(%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm9, 320(%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm11, 128(%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm8, 192(%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm4, (%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm3, 64(%rdx)
; AVX512DQBW-FAST-NEXT: vmovdqa64 %zmm2, 384(%rdx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 448(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 384(%rcx)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 448(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 384(%r8)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 448(%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%r9)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 384(%r9)
; AVX512DQBW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQBW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQBW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-FAST-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 384(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 448(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 256(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 320(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 128(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 192(%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, (%rax)
; AVX512DQBW-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQBW-FAST-NEXT: vmovaps %zmm0, 64(%rax)
; AVX512DQBW-FAST-NEXT: addq $6728, %rsp # imm = 0x1A48
; AVX512DQBW-FAST-NEXT: vzeroupper
; AVX512DQBW-FAST-NEXT: retq
%wide.vec = load <512 x i64>, ptr %in.vec, align 64
%strided.vec0 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248, i32 256, i32 264, i32 272, i32 280, i32 288, i32 296, i32 304, i32 312, i32 320, i32 328, i32 336, i32 344, i32 352, i32 360, i32 368, i32 376, i32 384, i32 392, i32 400, i32 408, i32 416, i32 424, i32 432, i32 440, i32 448, i32 456, i32 464, i32 472, i32 480, i32 488, i32 496, i32 504>
%strided.vec1 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249, i32 257, i32 265, i32 273, i32 281, i32 289, i32 297, i32 305, i32 313, i32 321, i32 329, i32 337, i32 345, i32 353, i32 361, i32 369, i32 377, i32 385, i32 393, i32 401, i32 409, i32 417, i32 425, i32 433, i32 441, i32 449, i32 457, i32 465, i32 473, i32 481, i32 489, i32 497, i32 505>
%strided.vec2 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250, i32 258, i32 266, i32 274, i32 282, i32 290, i32 298, i32 306, i32 314, i32 322, i32 330, i32 338, i32 346, i32 354, i32 362, i32 370, i32 378, i32 386, i32 394, i32 402, i32 410, i32 418, i32 426, i32 434, i32 442, i32 450, i32 458, i32 466, i32 474, i32 482, i32 490, i32 498, i32 506>
%strided.vec3 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251, i32 259, i32 267, i32 275, i32 283, i32 291, i32 299, i32 307, i32 315, i32 323, i32 331, i32 339, i32 347, i32 355, i32 363, i32 371, i32 379, i32 387, i32 395, i32 403, i32 411, i32 419, i32 427, i32 435, i32 443, i32 451, i32 459, i32 467, i32 475, i32 483, i32 491, i32 499, i32 507>
%strided.vec4 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252, i32 260, i32 268, i32 276, i32 284, i32 292, i32 300, i32 308, i32 316, i32 324, i32 332, i32 340, i32 348, i32 356, i32 364, i32 372, i32 380, i32 388, i32 396, i32 404, i32 412, i32 420, i32 428, i32 436, i32 444, i32 452, i32 460, i32 468, i32 476, i32 484, i32 492, i32 500, i32 508>
%strided.vec5 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253, i32 261, i32 269, i32 277, i32 285, i32 293, i32 301, i32 309, i32 317, i32 325, i32 333, i32 341, i32 349, i32 357, i32 365, i32 373, i32 381, i32 389, i32 397, i32 405, i32 413, i32 421, i32 429, i32 437, i32 445, i32 453, i32 461, i32 469, i32 477, i32 485, i32 493, i32 501, i32 509>
%strided.vec6 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254, i32 262, i32 270, i32 278, i32 286, i32 294, i32 302, i32 310, i32 318, i32 326, i32 334, i32 342, i32 350, i32 358, i32 366, i32 374, i32 382, i32 390, i32 398, i32 406, i32 414, i32 422, i32 430, i32 438, i32 446, i32 454, i32 462, i32 470, i32 478, i32 486, i32 494, i32 502, i32 510>
%strided.vec7 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255, i32 263, i32 271, i32 279, i32 287, i32 295, i32 303, i32 311, i32 319, i32 327, i32 335, i32 343, i32 351, i32 359, i32 367, i32 375, i32 383, i32 391, i32 399, i32 407, i32 415, i32 423, i32 431, i32 439, i32 447, i32 455, i32 463, i32 471, i32 479, i32 487, i32 495, i32 503, i32 511>
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
store <64 x i64> %strided.vec6, ptr %out.vec6, align 64
store <64 x i64> %strided.vec7, ptr %out.vec7, align 64
ret void
}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; AVX: {{.*}}
; AVX1: {{.*}}
; AVX2: {{.*}}
; AVX2-FAST: {{.*}}
; AVX2-FAST-PERLANE: {{.*}}
; AVX2-SLOW: {{.*}}
; AVX512BW-FAST: {{.*}}
; AVX512BW-SLOW: {{.*}}
; AVX512F-FAST: {{.*}}
; AVX512F-SLOW: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
; FALLBACK11: {{.*}}
; FALLBACK12: {{.*}}
; FALLBACK2: {{.*}}
; FALLBACK3: {{.*}}
; FALLBACK4: {{.*}}
; FALLBACK5: {{.*}}
; FALLBACK6: {{.*}}
; FALLBACK7: {{.*}}
; FALLBACK8: {{.*}}
; FALLBACK9: {{.*}}