brintos

brintos / llvm-project-archived public Read only

0
0
Text · 1.2 MiB · 7d32093 Raw
17066 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE3; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx  | FileCheck %s --check-prefixes=AVX4; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX25; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FP6; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FCP7; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX5128; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512-FCP9; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512DQ10; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-FCP11; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512BW12; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW-FCP13; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512DQ-BW14; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-BW-FCP15 16; These patterns are produced by LoopVectorizer for interleaved loads.17 18define void @load_i64_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {19; SSE-LABEL: load_i64_stride6_vf2:20; SSE:       # %bb.0:21; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax22; SSE-NEXT:    movaps 80(%rdi), %xmm023; SSE-NEXT:    movaps 64(%rdi), %xmm124; SSE-NEXT:    movaps (%rdi), %xmm225; SSE-NEXT:    movaps 16(%rdi), %xmm326; SSE-NEXT:    movaps 32(%rdi), %xmm427; SSE-NEXT:    movaps 48(%rdi), %xmm528; SSE-NEXT:    movaps %xmm2, %xmm629; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]30; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]31; SSE-NEXT:    movaps %xmm3, %xmm532; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]33; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm1[1]34; SSE-NEXT:    movaps %xmm4, %xmm135; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]36; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]37; SSE-NEXT:    movaps %xmm6, (%rsi)38; SSE-NEXT:    movaps %xmm2, (%rdx)39; SSE-NEXT:    movaps %xmm5, (%rcx)40; SSE-NEXT:    movaps %xmm3, (%r8)41; SSE-NEXT:    movaps %xmm1, (%r9)42; SSE-NEXT:    movaps %xmm4, (%rax)43; SSE-NEXT:    retq44;45; AVX-LABEL: load_i64_stride6_vf2:46; AVX:       # %bb.0:47; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax48; AVX-NEXT:    vmovaps (%rdi), %xmm049; AVX-NEXT:    vmovaps 16(%rdi), %xmm150; AVX-NEXT:    vmovaps 32(%rdi), %xmm251; AVX-NEXT:    vmovaps 48(%rdi), %xmm352; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]53; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]54; AVX-NEXT:    vmovaps 64(%rdi), %xmm355; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]56; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]57; AVX-NEXT:    vmovaps 80(%rdi), %xmm358; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]59; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]60; AVX-NEXT:    vmovaps %xmm4, (%rsi)61; AVX-NEXT:    vmovaps %xmm0, (%rdx)62; AVX-NEXT:    vmovaps %xmm5, (%rcx)63; AVX-NEXT:    vmovaps %xmm1, (%r8)64; AVX-NEXT:    vmovaps %xmm6, (%r9)65; AVX-NEXT:    vmovaps %xmm2, (%rax)66; AVX-NEXT:    retq67;68; AVX2-LABEL: load_i64_stride6_vf2:69; AVX2:       # %bb.0:70; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax71; AVX2-NEXT:    vmovaps (%rdi), %xmm072; AVX2-NEXT:    vmovaps 16(%rdi), %xmm173; AVX2-NEXT:    vmovaps 32(%rdi), %xmm274; AVX2-NEXT:    vmovaps 48(%rdi), %xmm375; AVX2-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]76; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]77; AVX2-NEXT:    vmovaps 64(%rdi), %xmm378; AVX2-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]79; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]80; AVX2-NEXT:    vmovaps 80(%rdi), %xmm381; AVX2-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]82; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]83; AVX2-NEXT:    vmovaps %xmm4, (%rsi)84; AVX2-NEXT:    vmovaps %xmm0, (%rdx)85; AVX2-NEXT:    vmovaps %xmm5, (%rcx)86; AVX2-NEXT:    vmovaps %xmm1, (%r8)87; AVX2-NEXT:    vmovaps %xmm6, (%r9)88; AVX2-NEXT:    vmovaps %xmm2, (%rax)89; AVX2-NEXT:    retq90;91; AVX2-FP-LABEL: load_i64_stride6_vf2:92; AVX2-FP:       # %bb.0:93; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax94; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm095; AVX2-FP-NEXT:    vmovaps 16(%rdi), %xmm196; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm297; AVX2-FP-NEXT:    vmovaps 48(%rdi), %xmm398; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]99; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]100; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm3101; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]102; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]103; AVX2-FP-NEXT:    vmovaps 80(%rdi), %xmm3104; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]105; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]106; AVX2-FP-NEXT:    vmovaps %xmm4, (%rsi)107; AVX2-FP-NEXT:    vmovaps %xmm0, (%rdx)108; AVX2-FP-NEXT:    vmovaps %xmm5, (%rcx)109; AVX2-FP-NEXT:    vmovaps %xmm1, (%r8)110; AVX2-FP-NEXT:    vmovaps %xmm6, (%r9)111; AVX2-FP-NEXT:    vmovaps %xmm2, (%rax)112; AVX2-FP-NEXT:    retq113;114; AVX2-FCP-LABEL: load_i64_stride6_vf2:115; AVX2-FCP:       # %bb.0:116; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax117; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm0118; AVX2-FCP-NEXT:    vmovaps 16(%rdi), %xmm1119; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm2120; AVX2-FCP-NEXT:    vmovaps 48(%rdi), %xmm3121; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]122; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]123; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm3124; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]125; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]126; AVX2-FCP-NEXT:    vmovaps 80(%rdi), %xmm3127; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]128; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]129; AVX2-FCP-NEXT:    vmovaps %xmm4, (%rsi)130; AVX2-FCP-NEXT:    vmovaps %xmm0, (%rdx)131; AVX2-FCP-NEXT:    vmovaps %xmm5, (%rcx)132; AVX2-FCP-NEXT:    vmovaps %xmm1, (%r8)133; AVX2-FCP-NEXT:    vmovaps %xmm6, (%r9)134; AVX2-FCP-NEXT:    vmovaps %xmm2, (%rax)135; AVX2-FCP-NEXT:    retq136;137; AVX512-LABEL: load_i64_stride6_vf2:138; AVX512:       # %bb.0:139; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax140; AVX512-NEXT:    vmovaps (%rdi), %xmm0141; AVX512-NEXT:    vmovaps 16(%rdi), %xmm1142; AVX512-NEXT:    vmovaps 32(%rdi), %xmm2143; AVX512-NEXT:    vmovaps 48(%rdi), %xmm3144; AVX512-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]145; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]146; AVX512-NEXT:    vmovaps 64(%rdi), %xmm3147; AVX512-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]148; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]149; AVX512-NEXT:    vmovaps 80(%rdi), %xmm3150; AVX512-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]151; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]152; AVX512-NEXT:    vmovaps %xmm4, (%rsi)153; AVX512-NEXT:    vmovaps %xmm0, (%rdx)154; AVX512-NEXT:    vmovaps %xmm5, (%rcx)155; AVX512-NEXT:    vmovaps %xmm1, (%r8)156; AVX512-NEXT:    vmovaps %xmm6, (%r9)157; AVX512-NEXT:    vmovaps %xmm2, (%rax)158; AVX512-NEXT:    retq159;160; AVX512-FCP-LABEL: load_i64_stride6_vf2:161; AVX512-FCP:       # %bb.0:162; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax163; AVX512-FCP-NEXT:    vmovaps (%rdi), %xmm0164; AVX512-FCP-NEXT:    vmovaps 16(%rdi), %xmm1165; AVX512-FCP-NEXT:    vmovaps 32(%rdi), %xmm2166; AVX512-FCP-NEXT:    vmovaps 48(%rdi), %xmm3167; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]168; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]169; AVX512-FCP-NEXT:    vmovaps 64(%rdi), %xmm3170; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]171; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]172; AVX512-FCP-NEXT:    vmovaps 80(%rdi), %xmm3173; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]174; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]175; AVX512-FCP-NEXT:    vmovaps %xmm4, (%rsi)176; AVX512-FCP-NEXT:    vmovaps %xmm0, (%rdx)177; AVX512-FCP-NEXT:    vmovaps %xmm5, (%rcx)178; AVX512-FCP-NEXT:    vmovaps %xmm1, (%r8)179; AVX512-FCP-NEXT:    vmovaps %xmm6, (%r9)180; AVX512-FCP-NEXT:    vmovaps %xmm2, (%rax)181; AVX512-FCP-NEXT:    retq182;183; AVX512DQ-LABEL: load_i64_stride6_vf2:184; AVX512DQ:       # %bb.0:185; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax186; AVX512DQ-NEXT:    vmovaps (%rdi), %xmm0187; AVX512DQ-NEXT:    vmovaps 16(%rdi), %xmm1188; AVX512DQ-NEXT:    vmovaps 32(%rdi), %xmm2189; AVX512DQ-NEXT:    vmovaps 48(%rdi), %xmm3190; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]191; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]192; AVX512DQ-NEXT:    vmovaps 64(%rdi), %xmm3193; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]194; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]195; AVX512DQ-NEXT:    vmovaps 80(%rdi), %xmm3196; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]197; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]198; AVX512DQ-NEXT:    vmovaps %xmm4, (%rsi)199; AVX512DQ-NEXT:    vmovaps %xmm0, (%rdx)200; AVX512DQ-NEXT:    vmovaps %xmm5, (%rcx)201; AVX512DQ-NEXT:    vmovaps %xmm1, (%r8)202; AVX512DQ-NEXT:    vmovaps %xmm6, (%r9)203; AVX512DQ-NEXT:    vmovaps %xmm2, (%rax)204; AVX512DQ-NEXT:    retq205;206; AVX512DQ-FCP-LABEL: load_i64_stride6_vf2:207; AVX512DQ-FCP:       # %bb.0:208; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax209; AVX512DQ-FCP-NEXT:    vmovaps (%rdi), %xmm0210; AVX512DQ-FCP-NEXT:    vmovaps 16(%rdi), %xmm1211; AVX512DQ-FCP-NEXT:    vmovaps 32(%rdi), %xmm2212; AVX512DQ-FCP-NEXT:    vmovaps 48(%rdi), %xmm3213; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]214; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]215; AVX512DQ-FCP-NEXT:    vmovaps 64(%rdi), %xmm3216; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]217; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]218; AVX512DQ-FCP-NEXT:    vmovaps 80(%rdi), %xmm3219; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]220; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]221; AVX512DQ-FCP-NEXT:    vmovaps %xmm4, (%rsi)222; AVX512DQ-FCP-NEXT:    vmovaps %xmm0, (%rdx)223; AVX512DQ-FCP-NEXT:    vmovaps %xmm5, (%rcx)224; AVX512DQ-FCP-NEXT:    vmovaps %xmm1, (%r8)225; AVX512DQ-FCP-NEXT:    vmovaps %xmm6, (%r9)226; AVX512DQ-FCP-NEXT:    vmovaps %xmm2, (%rax)227; AVX512DQ-FCP-NEXT:    retq228;229; AVX512BW-LABEL: load_i64_stride6_vf2:230; AVX512BW:       # %bb.0:231; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax232; AVX512BW-NEXT:    vmovaps (%rdi), %xmm0233; AVX512BW-NEXT:    vmovaps 16(%rdi), %xmm1234; AVX512BW-NEXT:    vmovaps 32(%rdi), %xmm2235; AVX512BW-NEXT:    vmovaps 48(%rdi), %xmm3236; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]237; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]238; AVX512BW-NEXT:    vmovaps 64(%rdi), %xmm3239; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]240; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]241; AVX512BW-NEXT:    vmovaps 80(%rdi), %xmm3242; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]243; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]244; AVX512BW-NEXT:    vmovaps %xmm4, (%rsi)245; AVX512BW-NEXT:    vmovaps %xmm0, (%rdx)246; AVX512BW-NEXT:    vmovaps %xmm5, (%rcx)247; AVX512BW-NEXT:    vmovaps %xmm1, (%r8)248; AVX512BW-NEXT:    vmovaps %xmm6, (%r9)249; AVX512BW-NEXT:    vmovaps %xmm2, (%rax)250; AVX512BW-NEXT:    retq251;252; AVX512BW-FCP-LABEL: load_i64_stride6_vf2:253; AVX512BW-FCP:       # %bb.0:254; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax255; AVX512BW-FCP-NEXT:    vmovaps (%rdi), %xmm0256; AVX512BW-FCP-NEXT:    vmovaps 16(%rdi), %xmm1257; AVX512BW-FCP-NEXT:    vmovaps 32(%rdi), %xmm2258; AVX512BW-FCP-NEXT:    vmovaps 48(%rdi), %xmm3259; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]260; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]261; AVX512BW-FCP-NEXT:    vmovaps 64(%rdi), %xmm3262; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]263; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]264; AVX512BW-FCP-NEXT:    vmovaps 80(%rdi), %xmm3265; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]266; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]267; AVX512BW-FCP-NEXT:    vmovaps %xmm4, (%rsi)268; AVX512BW-FCP-NEXT:    vmovaps %xmm0, (%rdx)269; AVX512BW-FCP-NEXT:    vmovaps %xmm5, (%rcx)270; AVX512BW-FCP-NEXT:    vmovaps %xmm1, (%r8)271; AVX512BW-FCP-NEXT:    vmovaps %xmm6, (%r9)272; AVX512BW-FCP-NEXT:    vmovaps %xmm2, (%rax)273; AVX512BW-FCP-NEXT:    retq274;275; AVX512DQ-BW-LABEL: load_i64_stride6_vf2:276; AVX512DQ-BW:       # %bb.0:277; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax278; AVX512DQ-BW-NEXT:    vmovaps (%rdi), %xmm0279; AVX512DQ-BW-NEXT:    vmovaps 16(%rdi), %xmm1280; AVX512DQ-BW-NEXT:    vmovaps 32(%rdi), %xmm2281; AVX512DQ-BW-NEXT:    vmovaps 48(%rdi), %xmm3282; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]283; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]284; AVX512DQ-BW-NEXT:    vmovaps 64(%rdi), %xmm3285; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]286; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]287; AVX512DQ-BW-NEXT:    vmovaps 80(%rdi), %xmm3288; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]289; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]290; AVX512DQ-BW-NEXT:    vmovaps %xmm4, (%rsi)291; AVX512DQ-BW-NEXT:    vmovaps %xmm0, (%rdx)292; AVX512DQ-BW-NEXT:    vmovaps %xmm5, (%rcx)293; AVX512DQ-BW-NEXT:    vmovaps %xmm1, (%r8)294; AVX512DQ-BW-NEXT:    vmovaps %xmm6, (%r9)295; AVX512DQ-BW-NEXT:    vmovaps %xmm2, (%rax)296; AVX512DQ-BW-NEXT:    retq297;298; AVX512DQ-BW-FCP-LABEL: load_i64_stride6_vf2:299; AVX512DQ-BW-FCP:       # %bb.0:300; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax301; AVX512DQ-BW-FCP-NEXT:    vmovaps (%rdi), %xmm0302; AVX512DQ-BW-FCP-NEXT:    vmovaps 16(%rdi), %xmm1303; AVX512DQ-BW-FCP-NEXT:    vmovaps 32(%rdi), %xmm2304; AVX512DQ-BW-FCP-NEXT:    vmovaps 48(%rdi), %xmm3305; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm3[0]306; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]307; AVX512DQ-BW-FCP-NEXT:    vmovaps 64(%rdi), %xmm3308; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm3[0]309; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]310; AVX512DQ-BW-FCP-NEXT:    vmovaps 80(%rdi), %xmm3311; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm3[0]312; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]313; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm4, (%rsi)314; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm0, (%rdx)315; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm5, (%rcx)316; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm1, (%r8)317; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm6, (%r9)318; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm2, (%rax)319; AVX512DQ-BW-FCP-NEXT:    retq320  %wide.vec = load <12 x i64>, ptr %in.vec, align 64321  %strided.vec0 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 0, i32 6>322  %strided.vec1 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 1, i32 7>323  %strided.vec2 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 2, i32 8>324  %strided.vec3 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 3, i32 9>325  %strided.vec4 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 4, i32 10>326  %strided.vec5 = shufflevector <12 x i64> %wide.vec, <12 x i64> poison, <2 x i32> <i32 5, i32 11>327  store <2 x i64> %strided.vec0, ptr %out.vec0, align 64328  store <2 x i64> %strided.vec1, ptr %out.vec1, align 64329  store <2 x i64> %strided.vec2, ptr %out.vec2, align 64330  store <2 x i64> %strided.vec3, ptr %out.vec3, align 64331  store <2 x i64> %strided.vec4, ptr %out.vec4, align 64332  store <2 x i64> %strided.vec5, ptr %out.vec5, align 64333  ret void334}335 336define void @load_i64_stride6_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {337; SSE-LABEL: load_i64_stride6_vf4:338; SSE:       # %bb.0:339; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax340; SSE-NEXT:    movaps 80(%rdi), %xmm3341; SSE-NEXT:    movaps 176(%rdi), %xmm5342; SSE-NEXT:    movaps 128(%rdi), %xmm0343; SSE-NEXT:    movaps 64(%rdi), %xmm6344; SSE-NEXT:    movaps 160(%rdi), %xmm7345; SSE-NEXT:    movaps 112(%rdi), %xmm1346; SSE-NEXT:    movaps (%rdi), %xmm8347; SSE-NEXT:    movaps 16(%rdi), %xmm4348; SSE-NEXT:    movaps 32(%rdi), %xmm2349; SSE-NEXT:    movaps 48(%rdi), %xmm9350; SSE-NEXT:    movaps 144(%rdi), %xmm10351; SSE-NEXT:    movaps 96(%rdi), %xmm11352; SSE-NEXT:    movaps %xmm11, %xmm12353; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm10[0]354; SSE-NEXT:    movaps %xmm8, %xmm13355; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm9[0]356; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm10[1]357; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm9[1]358; SSE-NEXT:    movaps %xmm1, %xmm9359; SSE-NEXT:    movlhps {{.*#+}} xmm9 = xmm9[0],xmm7[0]360; SSE-NEXT:    movaps %xmm4, %xmm10361; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm6[0]362; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm7[1]363; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]364; SSE-NEXT:    movaps %xmm0, %xmm6365; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]366; SSE-NEXT:    movaps %xmm2, %xmm7367; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm3[0]368; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm5[1]369; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]370; SSE-NEXT:    movaps %xmm13, (%rsi)371; SSE-NEXT:    movaps %xmm12, 16(%rsi)372; SSE-NEXT:    movaps %xmm8, (%rdx)373; SSE-NEXT:    movaps %xmm11, 16(%rdx)374; SSE-NEXT:    movaps %xmm10, (%rcx)375; SSE-NEXT:    movaps %xmm9, 16(%rcx)376; SSE-NEXT:    movaps %xmm4, (%r8)377; SSE-NEXT:    movaps %xmm1, 16(%r8)378; SSE-NEXT:    movaps %xmm7, (%r9)379; SSE-NEXT:    movaps %xmm6, 16(%r9)380; SSE-NEXT:    movaps %xmm2, (%rax)381; SSE-NEXT:    movaps %xmm0, 16(%rax)382; SSE-NEXT:    retq383;384; AVX-LABEL: load_i64_stride6_vf4:385; AVX:       # %bb.0:386; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax387; AVX-NEXT:    vmovaps 160(%rdi), %ymm0388; AVX-NEXT:    vmovaps 96(%rdi), %ymm1389; AVX-NEXT:    vmovaps 128(%rdi), %ymm2390; AVX-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm3391; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]392; AVX-NEXT:    vmovaps (%rdi), %xmm5393; AVX-NEXT:    vmovaps 16(%rdi), %xmm6394; AVX-NEXT:    vmovaps 32(%rdi), %xmm7395; AVX-NEXT:    vmovaps 48(%rdi), %xmm8396; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm5[0],xmm8[0]397; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]398; AVX-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]399; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm8[1]400; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]401; AVX-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm3402; AVX-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm3[0],ymm1[2],ymm3[2]403; AVX-NEXT:    vmovaps 64(%rdi), %xmm8404; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm6[0],xmm8[0]405; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]406; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]407; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm6[1],xmm8[1]408; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]409; AVX-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm3410; AVX-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]411; AVX-NEXT:    vmovaps 80(%rdi), %xmm8412; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm7[0],xmm8[0]413; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]414; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm0[1],ymm3[3],ymm0[3]415; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm7[1],xmm8[1]416; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]417; AVX-NEXT:    vmovaps %ymm4, (%rsi)418; AVX-NEXT:    vmovaps %ymm2, (%rdx)419; AVX-NEXT:    vmovaps %ymm5, (%rcx)420; AVX-NEXT:    vmovaps %ymm1, (%r8)421; AVX-NEXT:    vmovaps %ymm6, (%r9)422; AVX-NEXT:    vmovaps %ymm0, (%rax)423; AVX-NEXT:    vzeroupper424; AVX-NEXT:    retq425;426; AVX2-LABEL: load_i64_stride6_vf4:427; AVX2:       # %bb.0:428; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax429; AVX2-NEXT:    vmovaps 160(%rdi), %ymm0430; AVX2-NEXT:    vmovaps 128(%rdi), %ymm1431; AVX2-NEXT:    vmovaps 96(%rdi), %ymm2432; AVX2-NEXT:    vmovaps (%rdi), %xmm3433; AVX2-NEXT:    vmovaps 16(%rdi), %xmm4434; AVX2-NEXT:    vmovaps 32(%rdi), %xmm5435; AVX2-NEXT:    vmovaps 48(%rdi), %xmm6436; AVX2-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm6[0]437; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]438; AVX2-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]439; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]440; AVX2-NEXT:    vbroadcastsd 104(%rdi), %ymm8441; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm1[1],ymm8[3],ymm1[3]442; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]443; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5,6,7]444; AVX2-NEXT:    vbroadcastsd 160(%rdi), %ymm6445; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm6[0],ymm2[2],ymm6[2]446; AVX2-NEXT:    vmovaps 64(%rdi), %xmm8447; AVX2-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm4[0],xmm8[0]448; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]449; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm8[1]450; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm0[1],ymm2[3],ymm0[3]451; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]452; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]453; AVX2-NEXT:    vmovaps 80(%rdi), %xmm4454; AVX2-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm4[0]455; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]456; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]457; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]458; AVX2-NEXT:    vbroadcastsd 136(%rdi), %ymm8459; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm0[1],ymm8[3],ymm0[3]460; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]461; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]462; AVX2-NEXT:    vmovaps %ymm7, (%rsi)463; AVX2-NEXT:    vmovaps %ymm3, (%rdx)464; AVX2-NEXT:    vmovaps %ymm6, (%rcx)465; AVX2-NEXT:    vmovaps %ymm2, (%r8)466; AVX2-NEXT:    vmovaps %ymm1, (%r9)467; AVX2-NEXT:    vmovaps %ymm0, (%rax)468; AVX2-NEXT:    vzeroupper469; AVX2-NEXT:    retq470;471; AVX2-FP-LABEL: load_i64_stride6_vf4:472; AVX2-FP:       # %bb.0:473; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax474; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm0475; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm1476; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm2477; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm3478; AVX2-FP-NEXT:    vmovaps 16(%rdi), %xmm4479; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm5480; AVX2-FP-NEXT:    vmovaps 48(%rdi), %xmm6481; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm6[0]482; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]483; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]484; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]485; AVX2-FP-NEXT:    vbroadcastsd 104(%rdi), %ymm8486; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm1[1],ymm8[3],ymm1[3]487; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]488; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5,6,7]489; AVX2-FP-NEXT:    vbroadcastsd 160(%rdi), %ymm6490; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm6[0],ymm2[2],ymm6[2]491; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm8492; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm4[0],xmm8[0]493; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]494; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm8[1]495; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm0[1],ymm2[3],ymm0[3]496; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]497; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]498; AVX2-FP-NEXT:    vmovaps 80(%rdi), %xmm4499; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm4[0]500; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]501; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]502; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]503; AVX2-FP-NEXT:    vbroadcastsd 136(%rdi), %ymm8504; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm0[1],ymm8[3],ymm0[3]505; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]506; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]507; AVX2-FP-NEXT:    vmovaps %ymm7, (%rsi)508; AVX2-FP-NEXT:    vmovaps %ymm3, (%rdx)509; AVX2-FP-NEXT:    vmovaps %ymm6, (%rcx)510; AVX2-FP-NEXT:    vmovaps %ymm2, (%r8)511; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)512; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)513; AVX2-FP-NEXT:    vzeroupper514; AVX2-FP-NEXT:    retq515;516; AVX2-FCP-LABEL: load_i64_stride6_vf4:517; AVX2-FCP:       # %bb.0:518; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax519; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm0520; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm1521; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm2522; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm3523; AVX2-FCP-NEXT:    vmovaps 16(%rdi), %xmm4524; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm5525; AVX2-FCP-NEXT:    vmovaps 48(%rdi), %xmm6526; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm6[0]527; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]528; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]529; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]530; AVX2-FCP-NEXT:    vbroadcastsd 104(%rdi), %ymm8531; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm1[1],ymm8[3],ymm1[3]532; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]533; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5,6,7]534; AVX2-FCP-NEXT:    vbroadcastsd 160(%rdi), %ymm6535; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm6[0],ymm2[2],ymm6[2]536; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm8537; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm4[0],xmm8[0]538; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]539; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm8[1]540; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm0[1],ymm2[3],ymm0[3]541; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]542; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]543; AVX2-FCP-NEXT:    vmovaps 80(%rdi), %xmm4544; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm4[0]545; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]546; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]547; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]548; AVX2-FCP-NEXT:    vbroadcastsd 136(%rdi), %ymm8549; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm0[1],ymm8[3],ymm0[3]550; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]551; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]552; AVX2-FCP-NEXT:    vmovaps %ymm7, (%rsi)553; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rdx)554; AVX2-FCP-NEXT:    vmovaps %ymm6, (%rcx)555; AVX2-FCP-NEXT:    vmovaps %ymm2, (%r8)556; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r9)557; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)558; AVX2-FCP-NEXT:    vzeroupper559; AVX2-FCP-NEXT:    retq560;561; AVX512-LABEL: load_i64_stride6_vf4:562; AVX512:       # %bb.0:563; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax564; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm2565; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm3566; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]567; AVX512-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0568; AVX512-NEXT:    vpbroadcastq 144(%rdi), %ymm1569; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]570; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]571; AVX512-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1572; AVX512-NEXT:    vmovdqa 128(%rdi), %ymm4573; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm4[6,7]574; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [10,0,6,0]575; AVX512-NEXT:    vpermi2q %zmm2, %zmm3, %zmm5576; AVX512-NEXT:    vmovdqa 160(%rdi), %xmm6577; AVX512-NEXT:    vpbroadcastq %xmm6, %ymm7578; AVX512-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]579; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6580; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]581; AVX512-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7582; AVX512-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]583; AVX512-NEXT:    vmovdqa 160(%rdi), %ymm7584; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]585; AVX512-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,3]586; AVX512-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]587; AVX512-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8588; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]589; AVX512-NEXT:    vpbroadcastq 136(%rdi), %ymm8590; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm8[1],ymm7[1],ymm8[3],ymm7[3]591; AVX512-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]592; AVX512-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8593; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm7[4,5,6,7]594; AVX512-NEXT:    vmovdqa %ymm0, (%rsi)595; AVX512-NEXT:    vmovdqa %ymm1, (%rdx)596; AVX512-NEXT:    vmovdqa %ymm5, (%rcx)597; AVX512-NEXT:    vmovdqa %ymm6, (%r8)598; AVX512-NEXT:    vmovdqa %ymm4, (%r9)599; AVX512-NEXT:    vmovdqa %ymm2, (%rax)600; AVX512-NEXT:    vzeroupper601; AVX512-NEXT:    retq602;603; AVX512-FCP-LABEL: load_i64_stride6_vf4:604; AVX512-FCP:       # %bb.0:605; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax606; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2607; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3608; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]609; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0610; AVX512-FCP-NEXT:    vpbroadcastq 144(%rdi), %ymm1611; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]612; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]613; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1614; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]615; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm4 = [10,0,6,0]616; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm4617; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [0,1,2,4]618; AVX512-FCP-NEXT:    vmovdqa 160(%rdi), %ymm6619; AVX512-FCP-NEXT:    vpermi2q %ymm6, %ymm4, %ymm5620; AVX512-FCP-NEXT:    vinserti128 $1, 160(%rdi), %ymm0, %ymm4621; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]622; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7623; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm7[0,1,2,3,4,5],ymm4[6,7]624; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,4,0,6]625; AVX512-FCP-NEXT:    vpermq 128(%rdi), %zmm7, %zmm7626; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]627; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8628; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]629; AVX512-FCP-NEXT:    vpbroadcastq 136(%rdi), %ymm8630; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm8[1],ymm6[1],ymm8[3],ymm6[3]631; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]632; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8633; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm6[4,5,6,7]634; AVX512-FCP-NEXT:    vmovdqa %ymm0, (%rsi)635; AVX512-FCP-NEXT:    vmovdqa %ymm1, (%rdx)636; AVX512-FCP-NEXT:    vmovdqa %ymm5, (%rcx)637; AVX512-FCP-NEXT:    vmovdqa %ymm4, (%r8)638; AVX512-FCP-NEXT:    vmovdqa %ymm7, (%r9)639; AVX512-FCP-NEXT:    vmovdqa %ymm2, (%rax)640; AVX512-FCP-NEXT:    vzeroupper641; AVX512-FCP-NEXT:    retq642;643; AVX512DQ-LABEL: load_i64_stride6_vf4:644; AVX512DQ:       # %bb.0:645; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax646; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm2647; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm3648; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]649; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0650; AVX512DQ-NEXT:    vpbroadcastq 144(%rdi), %ymm1651; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]652; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]653; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1654; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %ymm4655; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm4[6,7]656; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [10,0,6,0]657; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm3, %zmm5658; AVX512DQ-NEXT:    vmovdqa 160(%rdi), %xmm6659; AVX512DQ-NEXT:    vpbroadcastq %xmm6, %ymm7660; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]661; AVX512DQ-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6662; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]663; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7664; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]665; AVX512DQ-NEXT:    vmovdqa 160(%rdi), %ymm7666; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]667; AVX512DQ-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,3]668; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]669; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8670; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]671; AVX512DQ-NEXT:    vpbroadcastq 136(%rdi), %ymm8672; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm8[1],ymm7[1],ymm8[3],ymm7[3]673; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]674; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8675; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm7[4,5,6,7]676; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rsi)677; AVX512DQ-NEXT:    vmovdqa %ymm1, (%rdx)678; AVX512DQ-NEXT:    vmovdqa %ymm5, (%rcx)679; AVX512DQ-NEXT:    vmovdqa %ymm6, (%r8)680; AVX512DQ-NEXT:    vmovdqa %ymm4, (%r9)681; AVX512DQ-NEXT:    vmovdqa %ymm2, (%rax)682; AVX512DQ-NEXT:    vzeroupper683; AVX512DQ-NEXT:    retq684;685; AVX512DQ-FCP-LABEL: load_i64_stride6_vf4:686; AVX512DQ-FCP:       # %bb.0:687; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax688; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2689; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3690; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]691; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0692; AVX512DQ-FCP-NEXT:    vpbroadcastq 144(%rdi), %ymm1693; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]694; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]695; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1696; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]697; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm4 = [10,0,6,0]698; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm4699; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [0,1,2,4]700; AVX512DQ-FCP-NEXT:    vmovdqa 160(%rdi), %ymm6701; AVX512DQ-FCP-NEXT:    vpermi2q %ymm6, %ymm4, %ymm5702; AVX512DQ-FCP-NEXT:    vinserti128 $1, 160(%rdi), %ymm0, %ymm4703; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]704; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7705; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm7[0,1,2,3,4,5],ymm4[6,7]706; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,4,0,6]707; AVX512DQ-FCP-NEXT:    vpermq 128(%rdi), %zmm7, %zmm7708; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]709; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8710; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]711; AVX512DQ-FCP-NEXT:    vpbroadcastq 136(%rdi), %ymm8712; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm8[1],ymm6[1],ymm8[3],ymm6[3]713; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]714; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8715; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm6[4,5,6,7]716; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, (%rsi)717; AVX512DQ-FCP-NEXT:    vmovdqa %ymm1, (%rdx)718; AVX512DQ-FCP-NEXT:    vmovdqa %ymm5, (%rcx)719; AVX512DQ-FCP-NEXT:    vmovdqa %ymm4, (%r8)720; AVX512DQ-FCP-NEXT:    vmovdqa %ymm7, (%r9)721; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, (%rax)722; AVX512DQ-FCP-NEXT:    vzeroupper723; AVX512DQ-FCP-NEXT:    retq724;725; AVX512BW-LABEL: load_i64_stride6_vf4:726; AVX512BW:       # %bb.0:727; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax728; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm2729; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm3730; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]731; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0732; AVX512BW-NEXT:    vpbroadcastq 144(%rdi), %ymm1733; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]734; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]735; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1736; AVX512BW-NEXT:    vmovdqa 128(%rdi), %ymm4737; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm4[6,7]738; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [10,0,6,0]739; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm3, %zmm5740; AVX512BW-NEXT:    vmovdqa 160(%rdi), %xmm6741; AVX512BW-NEXT:    vpbroadcastq %xmm6, %ymm7742; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]743; AVX512BW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6744; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]745; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7746; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]747; AVX512BW-NEXT:    vmovdqa 160(%rdi), %ymm7748; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]749; AVX512BW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,3]750; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]751; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8752; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]753; AVX512BW-NEXT:    vpbroadcastq 136(%rdi), %ymm8754; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm8[1],ymm7[1],ymm8[3],ymm7[3]755; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]756; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8757; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm7[4,5,6,7]758; AVX512BW-NEXT:    vmovdqa %ymm0, (%rsi)759; AVX512BW-NEXT:    vmovdqa %ymm1, (%rdx)760; AVX512BW-NEXT:    vmovdqa %ymm5, (%rcx)761; AVX512BW-NEXT:    vmovdqa %ymm6, (%r8)762; AVX512BW-NEXT:    vmovdqa %ymm4, (%r9)763; AVX512BW-NEXT:    vmovdqa %ymm2, (%rax)764; AVX512BW-NEXT:    vzeroupper765; AVX512BW-NEXT:    retq766;767; AVX512BW-FCP-LABEL: load_i64_stride6_vf4:768; AVX512BW-FCP:       # %bb.0:769; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax770; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2771; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3772; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]773; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0774; AVX512BW-FCP-NEXT:    vpbroadcastq 144(%rdi), %ymm1775; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]776; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]777; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1778; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]779; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm4 = [10,0,6,0]780; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm4781; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [0,1,2,4]782; AVX512BW-FCP-NEXT:    vmovdqa 160(%rdi), %ymm6783; AVX512BW-FCP-NEXT:    vpermi2q %ymm6, %ymm4, %ymm5784; AVX512BW-FCP-NEXT:    vinserti128 $1, 160(%rdi), %ymm0, %ymm4785; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]786; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7787; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm7[0,1,2,3,4,5],ymm4[6,7]788; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,4,0,6]789; AVX512BW-FCP-NEXT:    vpermq 128(%rdi), %zmm7, %zmm7790; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]791; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8792; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]793; AVX512BW-FCP-NEXT:    vpbroadcastq 136(%rdi), %ymm8794; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm8[1],ymm6[1],ymm8[3],ymm6[3]795; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]796; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8797; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm6[4,5,6,7]798; AVX512BW-FCP-NEXT:    vmovdqa %ymm0, (%rsi)799; AVX512BW-FCP-NEXT:    vmovdqa %ymm1, (%rdx)800; AVX512BW-FCP-NEXT:    vmovdqa %ymm5, (%rcx)801; AVX512BW-FCP-NEXT:    vmovdqa %ymm4, (%r8)802; AVX512BW-FCP-NEXT:    vmovdqa %ymm7, (%r9)803; AVX512BW-FCP-NEXT:    vmovdqa %ymm2, (%rax)804; AVX512BW-FCP-NEXT:    vzeroupper805; AVX512BW-FCP-NEXT:    retq806;807; AVX512DQ-BW-LABEL: load_i64_stride6_vf4:808; AVX512DQ-BW:       # %bb.0:809; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax810; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm2811; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm3812; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]813; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0814; AVX512DQ-BW-NEXT:    vpbroadcastq 144(%rdi), %ymm1815; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]816; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]817; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1818; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %ymm4819; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm4[6,7]820; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [10,0,6,0]821; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm3, %zmm5822; AVX512DQ-BW-NEXT:    vmovdqa 160(%rdi), %xmm6823; AVX512DQ-BW-NEXT:    vpbroadcastq %xmm6, %ymm7824; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]825; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6826; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]827; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7828; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]829; AVX512DQ-BW-NEXT:    vmovdqa 160(%rdi), %ymm7830; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]831; AVX512DQ-BW-NEXT:    vpermq {{.*#+}} ymm4 = ymm4[0,1,0,3]832; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]833; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8834; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]835; AVX512DQ-BW-NEXT:    vpbroadcastq 136(%rdi), %ymm8836; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm8[1],ymm7[1],ymm8[3],ymm7[3]837; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]838; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8839; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm7[4,5,6,7]840; AVX512DQ-BW-NEXT:    vmovdqa %ymm0, (%rsi)841; AVX512DQ-BW-NEXT:    vmovdqa %ymm1, (%rdx)842; AVX512DQ-BW-NEXT:    vmovdqa %ymm5, (%rcx)843; AVX512DQ-BW-NEXT:    vmovdqa %ymm6, (%r8)844; AVX512DQ-BW-NEXT:    vmovdqa %ymm4, (%r9)845; AVX512DQ-BW-NEXT:    vmovdqa %ymm2, (%rax)846; AVX512DQ-BW-NEXT:    vzeroupper847; AVX512DQ-BW-NEXT:    retq848;849; AVX512DQ-BW-FCP-LABEL: load_i64_stride6_vf4:850; AVX512DQ-BW-FCP:       # %bb.0:851; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax852; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2853; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3854; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm0 = [0,6,12,0]855; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm0856; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq 144(%rdi), %ymm1857; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]858; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm1 = [1,7,13,0]859; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm1860; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]861; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm4 = [10,0,6,0]862; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm4863; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm5 = [0,1,2,4]864; AVX512DQ-BW-FCP-NEXT:    vmovdqa 160(%rdi), %ymm6865; AVX512DQ-BW-FCP-NEXT:    vpermi2q %ymm6, %ymm4, %ymm5866; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 160(%rdi), %ymm0, %ymm4867; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [11,1,7,0]868; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm3, %zmm7869; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm7[0,1,2,3,4,5],ymm4[6,7]870; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,4,0,6]871; AVX512DQ-BW-FCP-NEXT:    vpermq 128(%rdi), %zmm7, %zmm7872; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [4,10]873; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8874; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]875; AVX512DQ-BW-FCP-NEXT:    vpbroadcastq 136(%rdi), %ymm8876; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm8[1],ymm6[1],ymm8[3],ymm6[3]877; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm8 = [5,11]878; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm2, %zmm8879; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm6[4,5,6,7]880; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm0, (%rsi)881; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm1, (%rdx)882; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm5, (%rcx)883; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm4, (%r8)884; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm7, (%r9)885; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm2, (%rax)886; AVX512DQ-BW-FCP-NEXT:    vzeroupper887; AVX512DQ-BW-FCP-NEXT:    retq888  %wide.vec = load <24 x i64>, ptr %in.vec, align 64889  %strided.vec0 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 0, i32 6, i32 12, i32 18>890  %strided.vec1 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 1, i32 7, i32 13, i32 19>891  %strided.vec2 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 2, i32 8, i32 14, i32 20>892  %strided.vec3 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 3, i32 9, i32 15, i32 21>893  %strided.vec4 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 4, i32 10, i32 16, i32 22>894  %strided.vec5 = shufflevector <24 x i64> %wide.vec, <24 x i64> poison, <4 x i32> <i32 5, i32 11, i32 17, i32 23>895  store <4 x i64> %strided.vec0, ptr %out.vec0, align 64896  store <4 x i64> %strided.vec1, ptr %out.vec1, align 64897  store <4 x i64> %strided.vec2, ptr %out.vec2, align 64898  store <4 x i64> %strided.vec3, ptr %out.vec3, align 64899  store <4 x i64> %strided.vec4, ptr %out.vec4, align 64900  store <4 x i64> %strided.vec5, ptr %out.vec5, align 64901  ret void902}903 904define void @load_i64_stride6_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {905; SSE-LABEL: load_i64_stride6_vf8:906; SSE:       # %bb.0:907; SSE-NEXT:    subq $24, %rsp908; SSE-NEXT:    movaps 160(%rdi), %xmm10909; SSE-NEXT:    movaps 256(%rdi), %xmm13910; SSE-NEXT:    movaps 208(%rdi), %xmm5911; SSE-NEXT:    movaps 352(%rdi), %xmm15912; SSE-NEXT:    movaps 304(%rdi), %xmm6913; SSE-NEXT:    movaps 64(%rdi), %xmm0914; SSE-NEXT:    movaps (%rdi), %xmm8915; SSE-NEXT:    movaps 16(%rdi), %xmm7916; SSE-NEXT:    movaps 48(%rdi), %xmm1917; SSE-NEXT:    movaps 144(%rdi), %xmm2918; SSE-NEXT:    movaps 96(%rdi), %xmm11919; SSE-NEXT:    movaps 240(%rdi), %xmm3920; SSE-NEXT:    movaps 192(%rdi), %xmm12921; SSE-NEXT:    movaps 336(%rdi), %xmm4922; SSE-NEXT:    movaps 288(%rdi), %xmm9923; SSE-NEXT:    movaps %xmm9, %xmm14924; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm4[0]925; SSE-NEXT:    movaps %xmm14, (%rsp) # 16-byte Spill926; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm4[1]927; SSE-NEXT:    movaps %xmm12, %xmm4928; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]929; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill930; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm3[1]931; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill932; SSE-NEXT:    movaps %xmm11, %xmm3933; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm2[0]934; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm2[1]935; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill936; SSE-NEXT:    movaps %xmm8, %xmm12937; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm1[0]938; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]939; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill940; SSE-NEXT:    movaps %xmm7, %xmm2941; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]942; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill943; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]944; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill945; SSE-NEXT:    movaps %xmm6, %xmm14946; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm15[0]947; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm15[1]948; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill949; SSE-NEXT:    movaps %xmm5, %xmm15950; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm13[0]951; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm13[1]952; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill953; SSE-NEXT:    movaps 112(%rdi), %xmm13954; SSE-NEXT:    movaps %xmm13, %xmm7955; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm10[0]956; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm10[1]957; SSE-NEXT:    movaps 80(%rdi), %xmm0958; SSE-NEXT:    movaps 32(%rdi), %xmm10959; SSE-NEXT:    movaps %xmm10, %xmm8960; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]961; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]962; SSE-NEXT:    movaps 368(%rdi), %xmm0963; SSE-NEXT:    movaps 320(%rdi), %xmm5964; SSE-NEXT:    movaps %xmm5, %xmm6965; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]966; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]967; SSE-NEXT:    movaps 272(%rdi), %xmm0968; SSE-NEXT:    movaps 224(%rdi), %xmm3969; SSE-NEXT:    movaps %xmm3, %xmm4970; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]971; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]972; SSE-NEXT:    movaps 176(%rdi), %xmm0973; SSE-NEXT:    movaps 128(%rdi), %xmm1974; SSE-NEXT:    movaps %xmm1, %xmm2975; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]976; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]977; SSE-NEXT:    movaps %xmm11, 16(%rsi)978; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload979; SSE-NEXT:    movaps %xmm0, 32(%rsi)980; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload981; SSE-NEXT:    movaps %xmm0, 48(%rsi)982; SSE-NEXT:    movaps %xmm12, (%rsi)983; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload984; SSE-NEXT:    movaps %xmm0, 16(%rdx)985; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload986; SSE-NEXT:    movaps %xmm0, 32(%rdx)987; SSE-NEXT:    movaps %xmm9, 48(%rdx)988; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload989; SSE-NEXT:    movaps %xmm0, (%rdx)990; SSE-NEXT:    movaps %xmm7, 16(%rcx)991; SSE-NEXT:    movaps %xmm15, 32(%rcx)992; SSE-NEXT:    movaps %xmm14, 48(%rcx)993; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload994; SSE-NEXT:    movaps %xmm0, (%rcx)995; SSE-NEXT:    movaps %xmm13, 16(%r8)996; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload997; SSE-NEXT:    movaps %xmm0, 32(%r8)998; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload999; SSE-NEXT:    movaps %xmm0, 48(%r8)1000; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1001; SSE-NEXT:    movaps %xmm0, (%r8)1002; SSE-NEXT:    movaps %xmm2, 16(%r9)1003; SSE-NEXT:    movaps %xmm4, 32(%r9)1004; SSE-NEXT:    movaps %xmm6, 48(%r9)1005; SSE-NEXT:    movaps %xmm8, (%r9)1006; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax1007; SSE-NEXT:    movaps %xmm1, 16(%rax)1008; SSE-NEXT:    movaps %xmm3, 32(%rax)1009; SSE-NEXT:    movaps %xmm5, 48(%rax)1010; SSE-NEXT:    movaps %xmm10, (%rax)1011; SSE-NEXT:    addq $24, %rsp1012; SSE-NEXT:    retq1013;1014; AVX-LABEL: load_i64_stride6_vf8:1015; AVX:       # %bb.0:1016; AVX-NEXT:    vmovaps 352(%rdi), %ymm01017; AVX-NEXT:    vmovaps 288(%rdi), %ymm41018; AVX-NEXT:    vmovaps 96(%rdi), %ymm71019; AVX-NEXT:    vmovaps 320(%rdi), %ymm51020; AVX-NEXT:    vmovaps 128(%rdi), %ymm31021; AVX-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm61022; AVX-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm6[0],ymm3[0],ymm6[2],ymm3[2]1023; AVX-NEXT:    vmovaps (%rdi), %xmm91024; AVX-NEXT:    vmovaps 16(%rdi), %xmm81025; AVX-NEXT:    vmovaps 48(%rdi), %xmm101026; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm9[0],xmm10[0]1027; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]1028; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1029; AVX-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm111030; AVX-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]1031; AVX-NEXT:    vmovaps 240(%rdi), %xmm121032; AVX-NEXT:    vmovaps 192(%rdi), %xmm131033; AVX-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm13[0],xmm12[0]1034; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3],ymm2[4,5,6,7]1035; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1036; AVX-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm6[1],ymm3[1],ymm6[3],ymm3[3]1037; AVX-NEXT:    vunpckhpd {{.*#+}} xmm6 = xmm9[1],xmm10[1]1038; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm3[4,5,6,7]1039; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1040; AVX-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]1041; AVX-NEXT:    vunpckhpd {{.*#+}} xmm6 = xmm13[1],xmm12[1]1042; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1043; AVX-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm101044; AVX-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm7[0],ymm10[0],ymm7[2],ymm10[2]1045; AVX-NEXT:    vmovaps 64(%rdi), %xmm111046; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm11[0]1047; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]1048; AVX-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm121049; AVX-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm4[0],ymm12[0],ymm4[2],ymm12[2]1050; AVX-NEXT:    vmovaps 256(%rdi), %xmm131051; AVX-NEXT:    vmovaps 208(%rdi), %xmm141052; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm13[0]1053; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3],ymm9[4,5,6,7]1054; AVX-NEXT:    vmovaps 160(%rdi), %ymm151055; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm10[1],ymm7[3],ymm10[3]1056; AVX-NEXT:    vmovaps 32(%rdi), %xmm101057; AVX-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm8[1],xmm11[1]1058; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1059; AVX-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm4[1],ymm12[1],ymm4[3],ymm12[3]1060; AVX-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm14[1],xmm13[1]1061; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]1062; AVX-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm81063; AVX-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm8[0],ymm15[0],ymm8[2],ymm15[2]1064; AVX-NEXT:    vmovaps 80(%rdi), %xmm121065; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm10[0],xmm12[0]1066; AVX-NEXT:    vblendps {{.*#+}} ymm11 = ymm13[0,1,2,3],ymm11[4,5,6,7]1067; AVX-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm131068; AVX-NEXT:    vmovaps %ymm0, %ymm31069; AVX-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm13[0],ymm0[0],ymm13[2],ymm0[2]1070; AVX-NEXT:    vmovaps 272(%rdi), %xmm11071; AVX-NEXT:    vmovaps 224(%rdi), %xmm01072; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]1073; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4,5,6,7]1074; AVX-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm15[1],ymm8[3],ymm15[3]1075; AVX-NEXT:    vunpckhpd {{.*#+}} xmm10 = xmm10[1],xmm12[1]1076; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]1077; AVX-NEXT:    vunpckhpd {{.*#+}} ymm10 = ymm13[1],ymm3[1],ymm13[3],ymm3[3]1078; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]1079; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm10[4,5,6,7]1080; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1081; AVX-NEXT:    vmovaps %ymm1, 32(%rsi)1082; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1083; AVX-NEXT:    vmovaps %ymm1, (%rsi)1084; AVX-NEXT:    vmovaps %ymm5, 32(%rdx)1085; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1086; AVX-NEXT:    vmovaps %ymm1, (%rdx)1087; AVX-NEXT:    vmovaps %ymm9, 32(%rcx)1088; AVX-NEXT:    vmovaps %ymm6, (%rcx)1089; AVX-NEXT:    vmovaps %ymm4, 32(%r8)1090; AVX-NEXT:    vmovaps %ymm7, (%r8)1091; AVX-NEXT:    vmovaps %ymm2, 32(%r9)1092; AVX-NEXT:    vmovaps %ymm11, (%r9)1093; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax1094; AVX-NEXT:    vmovaps %ymm0, 32(%rax)1095; AVX-NEXT:    vmovaps %ymm8, (%rax)1096; AVX-NEXT:    vzeroupper1097; AVX-NEXT:    retq1098;1099; AVX2-LABEL: load_i64_stride6_vf8:1100; AVX2:       # %bb.0:1101; AVX2-NEXT:    vmovaps 352(%rdi), %ymm01102; AVX2-NEXT:    vmovaps 320(%rdi), %ymm21103; AVX2-NEXT:    vmovaps 288(%rdi), %ymm71104; AVX2-NEXT:    vmovaps 128(%rdi), %ymm51105; AVX2-NEXT:    vmovaps 96(%rdi), %ymm91106; AVX2-NEXT:    vmovaps (%rdi), %xmm41107; AVX2-NEXT:    vmovaps 16(%rdi), %xmm111108; AVX2-NEXT:    vmovaps 48(%rdi), %xmm61109; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm6[0]1110; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm9[0],ymm5[0],ymm9[2],ymm5[2]1111; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]1112; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]1113; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1114; AVX2-NEXT:    vmovaps 240(%rdi), %xmm81115; AVX2-NEXT:    vmovaps 192(%rdi), %xmm101116; AVX2-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm10[0],xmm8[0]1117; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm7[0],ymm2[0],ymm7[2],ymm2[2]1118; AVX2-NEXT:    vpermpd {{.*#+}} ymm12 = ymm12[0,1,0,3]1119; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm12[4,5,6,7]1120; AVX2-NEXT:    vbroadcastsd 104(%rdi), %ymm121121; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm12 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]1122; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]1123; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm12[4,5,6,7]1124; AVX2-NEXT:    vbroadcastsd 296(%rdi), %ymm61125; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm2[1],ymm6[3],ymm2[3]1126; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm10[1],xmm8[1]1127; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]1128; AVX2-NEXT:    vbroadcastsd 160(%rdi), %ymm81129; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]1130; AVX2-NEXT:    vmovaps 64(%rdi), %xmm121131; AVX2-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm11[0],xmm12[0]1132; AVX2-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]1133; AVX2-NEXT:    vbroadcastsd 352(%rdi), %ymm101134; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm10[0],ymm7[2],ymm10[2]1135; AVX2-NEXT:    vmovaps 256(%rdi), %xmm131136; AVX2-NEXT:    vmovaps 208(%rdi), %xmm141137; AVX2-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm13[0]1138; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm15[0,1,2,3],ymm10[4,5,6,7]1139; AVX2-NEXT:    vmovaps 160(%rdi), %ymm151140; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm11[1],xmm12[1]1141; AVX2-NEXT:    vmovaps 32(%rdi), %xmm121142; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm15[1],ymm9[3],ymm15[3]1143; AVX2-NEXT:    vpermpd {{.*#+}} ymm9 = ymm9[0,1,2,1]1144; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3],ymm9[4,5,6,7]1145; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]1146; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm0[1],ymm7[3],ymm0[3]1147; AVX2-NEXT:    vpermpd {{.*#+}} ymm7 = ymm7[0,1,2,1]1148; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]1149; AVX2-NEXT:    vmovaps 80(%rdi), %xmm111150; AVX2-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]1151; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm15[0],ymm5[2],ymm15[2]1152; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]1153; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]1154; AVX2-NEXT:    vmovaps 272(%rdi), %xmm131155; AVX2-NEXT:    vmovaps 224(%rdi), %xmm141156; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]1157; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]1158; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm14[0],xmm13[0]1159; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]1160; AVX2-NEXT:    vbroadcastsd 136(%rdi), %ymm21161; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]1162; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm12[1],xmm11[1]1163; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm2[4,5,6,7]1164; AVX2-NEXT:    vbroadcastsd 328(%rdi), %ymm111165; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm0[1],ymm11[3],ymm0[3]1166; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]1167; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]1168; AVX2-NEXT:    vmovaps %ymm3, 32(%rsi)1169; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload1170; AVX2-NEXT:    vmovaps %ymm3, (%rsi)1171; AVX2-NEXT:    vmovaps %ymm6, 32(%rdx)1172; AVX2-NEXT:    vmovaps %ymm4, (%rdx)1173; AVX2-NEXT:    vmovaps %ymm10, 32(%rcx)1174; AVX2-NEXT:    vmovaps %ymm8, (%rcx)1175; AVX2-NEXT:    vmovaps %ymm7, 32(%r8)1176; AVX2-NEXT:    vmovaps %ymm9, (%r8)1177; AVX2-NEXT:    vmovaps %ymm1, 32(%r9)1178; AVX2-NEXT:    vmovaps %ymm5, (%r9)1179; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax1180; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)1181; AVX2-NEXT:    vmovaps %ymm2, (%rax)1182; AVX2-NEXT:    vzeroupper1183; AVX2-NEXT:    retq1184;1185; AVX2-FP-LABEL: load_i64_stride6_vf8:1186; AVX2-FP:       # %bb.0:1187; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm01188; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm21189; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm71190; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm51191; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm91192; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm41193; AVX2-FP-NEXT:    vmovaps 16(%rdi), %xmm111194; AVX2-FP-NEXT:    vmovaps 48(%rdi), %xmm61195; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm6[0]1196; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm9[0],ymm5[0],ymm9[2],ymm5[2]1197; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]1198; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]1199; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1200; AVX2-FP-NEXT:    vmovaps 240(%rdi), %xmm81201; AVX2-FP-NEXT:    vmovaps 192(%rdi), %xmm101202; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm10[0],xmm8[0]1203; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm7[0],ymm2[0],ymm7[2],ymm2[2]1204; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm12 = ymm12[0,1,0,3]1205; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm12[4,5,6,7]1206; AVX2-FP-NEXT:    vbroadcastsd 104(%rdi), %ymm121207; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm12 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]1208; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]1209; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm12[4,5,6,7]1210; AVX2-FP-NEXT:    vbroadcastsd 296(%rdi), %ymm61211; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm2[1],ymm6[3],ymm2[3]1212; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm10[1],xmm8[1]1213; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]1214; AVX2-FP-NEXT:    vbroadcastsd 160(%rdi), %ymm81215; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]1216; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm121217; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm11[0],xmm12[0]1218; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]1219; AVX2-FP-NEXT:    vbroadcastsd 352(%rdi), %ymm101220; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm10[0],ymm7[2],ymm10[2]1221; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm131222; AVX2-FP-NEXT:    vmovaps 208(%rdi), %xmm141223; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm13[0]1224; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm15[0,1,2,3],ymm10[4,5,6,7]1225; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm151226; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm11[1],xmm12[1]1227; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm121228; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm15[1],ymm9[3],ymm15[3]1229; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm9 = ymm9[0,1,2,1]1230; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3],ymm9[4,5,6,7]1231; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]1232; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm0[1],ymm7[3],ymm0[3]1233; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm7[0,1,2,1]1234; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]1235; AVX2-FP-NEXT:    vmovaps 80(%rdi), %xmm111236; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]1237; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm15[0],ymm5[2],ymm15[2]1238; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]1239; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]1240; AVX2-FP-NEXT:    vmovaps 272(%rdi), %xmm131241; AVX2-FP-NEXT:    vmovaps 224(%rdi), %xmm141242; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]1243; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]1244; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm14[0],xmm13[0]1245; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]1246; AVX2-FP-NEXT:    vbroadcastsd 136(%rdi), %ymm21247; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]1248; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm12[1],xmm11[1]1249; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm2[4,5,6,7]1250; AVX2-FP-NEXT:    vbroadcastsd 328(%rdi), %ymm111251; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm0[1],ymm11[3],ymm0[3]1252; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]1253; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]1254; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%rsi)1255; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload1256; AVX2-FP-NEXT:    vmovaps %ymm3, (%rsi)1257; AVX2-FP-NEXT:    vmovaps %ymm6, 32(%rdx)1258; AVX2-FP-NEXT:    vmovaps %ymm4, (%rdx)1259; AVX2-FP-NEXT:    vmovaps %ymm10, 32(%rcx)1260; AVX2-FP-NEXT:    vmovaps %ymm8, (%rcx)1261; AVX2-FP-NEXT:    vmovaps %ymm7, 32(%r8)1262; AVX2-FP-NEXT:    vmovaps %ymm9, (%r8)1263; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r9)1264; AVX2-FP-NEXT:    vmovaps %ymm5, (%r9)1265; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1266; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)1267; AVX2-FP-NEXT:    vmovaps %ymm2, (%rax)1268; AVX2-FP-NEXT:    vzeroupper1269; AVX2-FP-NEXT:    retq1270;1271; AVX2-FCP-LABEL: load_i64_stride6_vf8:1272; AVX2-FCP:       # %bb.0:1273; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm01274; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm21275; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm71276; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm51277; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm91278; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm41279; AVX2-FCP-NEXT:    vmovaps 16(%rdi), %xmm111280; AVX2-FCP-NEXT:    vmovaps 48(%rdi), %xmm61281; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm6[0]1282; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm9[0],ymm5[0],ymm9[2],ymm5[2]1283; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]1284; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]1285; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1286; AVX2-FCP-NEXT:    vmovaps 240(%rdi), %xmm81287; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %xmm101288; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm10[0],xmm8[0]1289; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm7[0],ymm2[0],ymm7[2],ymm2[2]1290; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm12 = ymm12[0,1,0,3]1291; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm12[4,5,6,7]1292; AVX2-FCP-NEXT:    vbroadcastsd 104(%rdi), %ymm121293; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm12 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]1294; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]1295; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm12[4,5,6,7]1296; AVX2-FCP-NEXT:    vbroadcastsd 296(%rdi), %ymm61297; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm2[1],ymm6[3],ymm2[3]1298; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm10[1],xmm8[1]1299; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]1300; AVX2-FCP-NEXT:    vbroadcastsd 160(%rdi), %ymm81301; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]1302; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm121303; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm11[0],xmm12[0]1304; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]1305; AVX2-FCP-NEXT:    vbroadcastsd 352(%rdi), %ymm101306; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm10[0],ymm7[2],ymm10[2]1307; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm131308; AVX2-FCP-NEXT:    vmovaps 208(%rdi), %xmm141309; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm13[0]1310; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm15[0,1,2,3],ymm10[4,5,6,7]1311; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm151312; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm11[1],xmm12[1]1313; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm121314; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm15[1],ymm9[3],ymm15[3]1315; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm9 = ymm9[0,1,2,1]1316; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3],ymm9[4,5,6,7]1317; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]1318; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm0[1],ymm7[3],ymm0[3]1319; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm7[0,1,2,1]1320; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]1321; AVX2-FCP-NEXT:    vmovaps 80(%rdi), %xmm111322; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]1323; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm15[0],ymm5[2],ymm15[2]1324; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]1325; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]1326; AVX2-FCP-NEXT:    vmovaps 272(%rdi), %xmm131327; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %xmm141328; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]1329; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]1330; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm14[0],xmm13[0]1331; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]1332; AVX2-FCP-NEXT:    vbroadcastsd 136(%rdi), %ymm21333; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]1334; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm12[1],xmm11[1]1335; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm2[4,5,6,7]1336; AVX2-FCP-NEXT:    vbroadcastsd 328(%rdi), %ymm111337; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm0[1],ymm11[3],ymm0[3]1338; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm14[1],xmm13[1]1339; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]1340; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%rsi)1341; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload1342; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rsi)1343; AVX2-FCP-NEXT:    vmovaps %ymm6, 32(%rdx)1344; AVX2-FCP-NEXT:    vmovaps %ymm4, (%rdx)1345; AVX2-FCP-NEXT:    vmovaps %ymm10, 32(%rcx)1346; AVX2-FCP-NEXT:    vmovaps %ymm8, (%rcx)1347; AVX2-FCP-NEXT:    vmovaps %ymm7, 32(%r8)1348; AVX2-FCP-NEXT:    vmovaps %ymm9, (%r8)1349; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r9)1350; AVX2-FCP-NEXT:    vmovaps %ymm5, (%r9)1351; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1352; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)1353; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rax)1354; AVX2-FCP-NEXT:    vzeroupper1355; AVX2-FCP-NEXT:    retq1356;1357; AVX512-LABEL: load_i64_stride6_vf8:1358; AVX512:       # %bb.0:1359; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax1360; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm61361; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm71362; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm01363; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm11364; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm31365; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm41366; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1367; AVX512-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1368; AVX512-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51369; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1370; AVX512-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21371; AVX512-NEXT:    movb $56, %dil1372; AVX512-NEXT:    kmovw %edi, %k11373; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1374; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1375; AVX512-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1376; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm81377; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1378; AVX512-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1379; AVX512-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91380; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1381; AVX512-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1382; AVX512-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101383; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1384; AVX512-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1385; AVX512-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111386; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1387; AVX512-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1388; AVX512-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121389; AVX512-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71390; AVX512-NEXT:    movb $-64, %dil1391; AVX512-NEXT:    kmovw %edi, %k21392; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1393; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1394; AVX512-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1395; AVX512-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71396; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1397; AVX512-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131398; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1399; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1400; AVX512-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1401; AVX512-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81402; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1403; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1404; AVX512-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1405; AVX512-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61406; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1407; AVX512-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81408; AVX512-NEXT:    movb $24, %dil1409; AVX512-NEXT:    kmovw %edi, %k21410; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1411; AVX512-NEXT:    movb $-32, %dil1412; AVX512-NEXT:    kmovw %edi, %k11413; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1414; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1415; AVX512-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1416; AVX512-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61417; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1418; AVX512-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91419; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1420; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1421; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1422; AVX512-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1423; AVX512-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61424; AVX512-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51425; AVX512-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51426; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1427; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1428; AVX512-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1429; AVX512-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61430; AVX512-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01431; AVX512-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01432; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1433; AVX512-NEXT:    vmovdqa64 %zmm2, (%rsi)1434; AVX512-NEXT:    vmovdqa64 %zmm13, (%rdx)1435; AVX512-NEXT:    vmovdqa64 %zmm8, (%rcx)1436; AVX512-NEXT:    vmovdqa64 %zmm9, (%r8)1437; AVX512-NEXT:    vmovdqa64 %zmm5, (%r9)1438; AVX512-NEXT:    vmovdqa64 %zmm0, (%rax)1439; AVX512-NEXT:    vzeroupper1440; AVX512-NEXT:    retq1441;1442; AVX512-FCP-LABEL: load_i64_stride6_vf8:1443; AVX512-FCP:       # %bb.0:1444; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1445; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm61446; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm71447; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm01448; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm11449; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm31450; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm41451; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1452; AVX512-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1453; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51454; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1455; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21456; AVX512-FCP-NEXT:    movb $56, %dil1457; AVX512-FCP-NEXT:    kmovw %edi, %k11458; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1459; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1460; AVX512-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1461; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm81462; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1463; AVX512-FCP-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1464; AVX512-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91465; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1466; AVX512-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1467; AVX512-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101468; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1469; AVX512-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1470; AVX512-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111471; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1472; AVX512-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1473; AVX512-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121474; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71475; AVX512-FCP-NEXT:    movb $-64, %dil1476; AVX512-FCP-NEXT:    kmovw %edi, %k21477; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1478; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1479; AVX512-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1480; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71481; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1482; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131483; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1484; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1485; AVX512-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1486; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81487; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1488; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1489; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1490; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61491; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1492; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81493; AVX512-FCP-NEXT:    movb $24, %dil1494; AVX512-FCP-NEXT:    kmovw %edi, %k21495; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1496; AVX512-FCP-NEXT:    movb $-32, %dil1497; AVX512-FCP-NEXT:    kmovw %edi, %k11498; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1499; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1500; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1501; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61502; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1503; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91504; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1505; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1506; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1507; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1508; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61509; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51510; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51511; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1512; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1513; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1514; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61515; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01516; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01517; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1518; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)1519; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, (%rdx)1520; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, (%rcx)1521; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, (%r8)1522; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)1523; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)1524; AVX512-FCP-NEXT:    vzeroupper1525; AVX512-FCP-NEXT:    retq1526;1527; AVX512DQ-LABEL: load_i64_stride6_vf8:1528; AVX512DQ:       # %bb.0:1529; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax1530; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm61531; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm71532; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm01533; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm11534; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm31535; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm41536; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1537; AVX512DQ-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1538; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51539; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1540; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21541; AVX512DQ-NEXT:    movb $56, %dil1542; AVX512DQ-NEXT:    kmovw %edi, %k11543; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1544; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1545; AVX512DQ-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1546; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm81547; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1548; AVX512DQ-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1549; AVX512DQ-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91550; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1551; AVX512DQ-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1552; AVX512DQ-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101553; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1554; AVX512DQ-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1555; AVX512DQ-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111556; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1557; AVX512DQ-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1558; AVX512DQ-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121559; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71560; AVX512DQ-NEXT:    movb $-64, %dil1561; AVX512DQ-NEXT:    kmovw %edi, %k21562; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1563; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1564; AVX512DQ-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1565; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71566; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1567; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131568; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1569; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1570; AVX512DQ-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1571; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81572; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1573; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1574; AVX512DQ-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1575; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61576; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1577; AVX512DQ-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81578; AVX512DQ-NEXT:    movb $24, %dil1579; AVX512DQ-NEXT:    kmovw %edi, %k21580; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1581; AVX512DQ-NEXT:    movb $-32, %dil1582; AVX512DQ-NEXT:    kmovw %edi, %k11583; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1584; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1585; AVX512DQ-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1586; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61587; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1588; AVX512DQ-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91589; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1590; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1591; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1592; AVX512DQ-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1593; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61594; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51595; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51596; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1597; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1598; AVX512DQ-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1599; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61600; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01601; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01602; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1603; AVX512DQ-NEXT:    vmovdqa64 %zmm2, (%rsi)1604; AVX512DQ-NEXT:    vmovdqa64 %zmm13, (%rdx)1605; AVX512DQ-NEXT:    vmovdqa64 %zmm8, (%rcx)1606; AVX512DQ-NEXT:    vmovdqa64 %zmm9, (%r8)1607; AVX512DQ-NEXT:    vmovdqa64 %zmm5, (%r9)1608; AVX512DQ-NEXT:    vmovdqa64 %zmm0, (%rax)1609; AVX512DQ-NEXT:    vzeroupper1610; AVX512DQ-NEXT:    retq1611;1612; AVX512DQ-FCP-LABEL: load_i64_stride6_vf8:1613; AVX512DQ-FCP:       # %bb.0:1614; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1615; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm61616; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm71617; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm01618; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm11619; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm31620; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm41621; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1622; AVX512DQ-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1623; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51624; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1625; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21626; AVX512DQ-FCP-NEXT:    movb $56, %dil1627; AVX512DQ-FCP-NEXT:    kmovw %edi, %k11628; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1629; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1630; AVX512DQ-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1631; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm81632; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1633; AVX512DQ-FCP-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1634; AVX512DQ-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91635; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1636; AVX512DQ-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1637; AVX512DQ-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101638; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1639; AVX512DQ-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1640; AVX512DQ-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111641; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1642; AVX512DQ-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1643; AVX512DQ-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121644; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71645; AVX512DQ-FCP-NEXT:    movb $-64, %dil1646; AVX512DQ-FCP-NEXT:    kmovw %edi, %k21647; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1648; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1649; AVX512DQ-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1650; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71651; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1652; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131653; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1654; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1655; AVX512DQ-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1656; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81657; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1658; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1659; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1660; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61661; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1662; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81663; AVX512DQ-FCP-NEXT:    movb $24, %dil1664; AVX512DQ-FCP-NEXT:    kmovw %edi, %k21665; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1666; AVX512DQ-FCP-NEXT:    movb $-32, %dil1667; AVX512DQ-FCP-NEXT:    kmovw %edi, %k11668; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1669; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1670; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1671; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61672; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1673; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91674; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1675; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1676; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1677; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1678; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61679; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51680; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51681; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1682; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1683; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1684; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61685; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01686; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01687; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1688; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)1689; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, (%rdx)1690; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, (%rcx)1691; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, (%r8)1692; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)1693; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)1694; AVX512DQ-FCP-NEXT:    vzeroupper1695; AVX512DQ-FCP-NEXT:    retq1696;1697; AVX512BW-LABEL: load_i64_stride6_vf8:1698; AVX512BW:       # %bb.0:1699; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax1700; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm61701; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm71702; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm01703; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm11704; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm31705; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm41706; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1707; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1708; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51709; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1710; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21711; AVX512BW-NEXT:    movb $56, %dil1712; AVX512BW-NEXT:    kmovd %edi, %k11713; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1714; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1715; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1716; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm81717; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1718; AVX512BW-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1719; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91720; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1721; AVX512BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1722; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101723; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1724; AVX512BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1725; AVX512BW-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111726; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1727; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1728; AVX512BW-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121729; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71730; AVX512BW-NEXT:    movb $-64, %dil1731; AVX512BW-NEXT:    kmovd %edi, %k21732; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1733; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1734; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1735; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71736; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1737; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131738; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1739; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1740; AVX512BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1741; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81742; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1743; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1744; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1745; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61746; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1747; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81748; AVX512BW-NEXT:    movb $24, %dil1749; AVX512BW-NEXT:    kmovd %edi, %k21750; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1751; AVX512BW-NEXT:    movb $-32, %dil1752; AVX512BW-NEXT:    kmovd %edi, %k11753; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1754; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1755; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1756; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61757; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1758; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91759; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1760; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1761; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1762; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1763; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61764; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51765; AVX512BW-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51766; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1767; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1768; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1769; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61770; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01771; AVX512BW-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01772; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1773; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rsi)1774; AVX512BW-NEXT:    vmovdqa64 %zmm13, (%rdx)1775; AVX512BW-NEXT:    vmovdqa64 %zmm8, (%rcx)1776; AVX512BW-NEXT:    vmovdqa64 %zmm9, (%r8)1777; AVX512BW-NEXT:    vmovdqa64 %zmm5, (%r9)1778; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rax)1779; AVX512BW-NEXT:    vzeroupper1780; AVX512BW-NEXT:    retq1781;1782; AVX512BW-FCP-LABEL: load_i64_stride6_vf8:1783; AVX512BW-FCP:       # %bb.0:1784; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1785; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm61786; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm71787; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm01788; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm11789; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm31790; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm41791; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1792; AVX512BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1793; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51794; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1795; AVX512BW-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21796; AVX512BW-FCP-NEXT:    movb $56, %dil1797; AVX512BW-FCP-NEXT:    kmovd %edi, %k11798; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1799; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1800; AVX512BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1801; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm81802; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1803; AVX512BW-FCP-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1804; AVX512BW-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91805; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1806; AVX512BW-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1807; AVX512BW-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101808; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1809; AVX512BW-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1810; AVX512BW-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111811; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1812; AVX512BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1813; AVX512BW-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121814; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71815; AVX512BW-FCP-NEXT:    movb $-64, %dil1816; AVX512BW-FCP-NEXT:    kmovd %edi, %k21817; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1818; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1819; AVX512BW-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1820; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71821; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1822; AVX512BW-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131823; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1824; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1825; AVX512BW-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1826; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81827; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1828; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1829; AVX512BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1830; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61831; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1832; AVX512BW-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81833; AVX512BW-FCP-NEXT:    movb $24, %dil1834; AVX512BW-FCP-NEXT:    kmovd %edi, %k21835; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1836; AVX512BW-FCP-NEXT:    movb $-32, %dil1837; AVX512BW-FCP-NEXT:    kmovd %edi, %k11838; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1839; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1840; AVX512BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1841; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61842; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1843; AVX512BW-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91844; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1845; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1846; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1847; AVX512BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1848; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61849; AVX512BW-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51850; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51851; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1852; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1853; AVX512BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1854; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61855; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01856; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01857; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1858; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)1859; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, (%rdx)1860; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rcx)1861; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, (%r8)1862; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)1863; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)1864; AVX512BW-FCP-NEXT:    vzeroupper1865; AVX512BW-FCP-NEXT:    retq1866;1867; AVX512DQ-BW-LABEL: load_i64_stride6_vf8:1868; AVX512DQ-BW:       # %bb.0:1869; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax1870; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm61871; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm71872; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm01873; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm11874; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm31875; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm41876; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1877; AVX512DQ-BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1878; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51879; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1880; AVX512DQ-BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21881; AVX512DQ-BW-NEXT:    movb $56, %dil1882; AVX512DQ-BW-NEXT:    kmovd %edi, %k11883; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1884; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1885; AVX512DQ-BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1886; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm81887; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1888; AVX512DQ-BW-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1889; AVX512DQ-BW-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91890; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1891; AVX512DQ-BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1892; AVX512DQ-BW-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101893; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1894; AVX512DQ-BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1895; AVX512DQ-BW-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111896; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1897; AVX512DQ-BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1898; AVX512DQ-BW-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121899; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71900; AVX512DQ-BW-NEXT:    movb $-64, %dil1901; AVX512DQ-BW-NEXT:    kmovd %edi, %k21902; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1903; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1904; AVX512DQ-BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1905; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71906; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1907; AVX512DQ-BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131908; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1909; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1910; AVX512DQ-BW-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1911; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81912; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1913; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1914; AVX512DQ-BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1915; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61916; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]1917; AVX512DQ-BW-NEXT:    vpermi2q %zmm0, %zmm1, %zmm81918; AVX512DQ-BW-NEXT:    movb $24, %dil1919; AVX512DQ-BW-NEXT:    kmovd %edi, %k21920; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}1921; AVX512DQ-BW-NEXT:    movb $-32, %dil1922; AVX512DQ-BW-NEXT:    kmovd %edi, %k11923; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}1924; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]1925; AVX512DQ-BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1926; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61927; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]1928; AVX512DQ-BW-NEXT:    vpermi2q %zmm0, %zmm1, %zmm91929; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}1930; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}1931; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]1932; AVX512DQ-BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1933; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61934; AVX512DQ-BW-NEXT:    vpermi2q %zmm1, %zmm0, %zmm51935; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm51936; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}1937; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]1938; AVX512DQ-BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]1939; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm3, %zmm61940; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm01941; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm01942; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}1943; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rsi)1944; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, (%rdx)1945; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, (%rcx)1946; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, (%r8)1947; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, (%r9)1948; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rax)1949; AVX512DQ-BW-NEXT:    vzeroupper1950; AVX512DQ-BW-NEXT:    retq1951;1952; AVX512DQ-BW-FCP-LABEL: load_i64_stride6_vf8:1953; AVX512DQ-BW-FCP:       # %bb.0:1954; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1955; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm61956; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm71957; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm01958; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm11959; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm31960; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm41961; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,0,10,0,6,0,10]1962; AVX512DQ-BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3]1963; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm51964; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,6,12,0]1965; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm21966; AVX512DQ-BW-FCP-NEXT:    movb $56, %dil1967; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k11968; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k1}1969; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]1970; AVX512DQ-BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1971; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm81972; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm9 = [0,0,6,12,0,0,6,12]1973; AVX512DQ-BW-FCP-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3]1974; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm91975; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm10 = [0,1,7,13,0,1,7,13]1976; AVX512DQ-BW-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3]1977; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm6, %zmm7, %zmm101978; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [0,10,0,6,0,10,0,6]1979; AVX512DQ-BW-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]1980; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm111981; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [0,11,1,7,0,11,1,7]1982; AVX512DQ-BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]1983; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm7, %zmm6, %zmm121984; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm71985; AVX512DQ-BW-FCP-NEXT:    movb $-64, %dil1986; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k21987; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k2}1988; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [1,7,0,11,1,7,0,11]1989; AVX512DQ-BW-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]1990; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm71991; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm13 = [1,7,13,0]1992; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm131993; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm13 {%k1}1994; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm7 = [5,11,5,11,5,11,5,11]1995; AVX512DQ-BW-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]1996; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm81997; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm13 {%k2}1998; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [10,4,10,4,10,4,10,4]1999; AVX512DQ-BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]2000; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm62001; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [10,0,6,0]2002; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm82003; AVX512DQ-BW-FCP-NEXT:    movb $24, %dil2004; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k22005; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm8 {%k2}2006; AVX512DQ-BW-FCP-NEXT:    movb $-32, %dil2007; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k12008; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}2009; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm6 = [11,5,11,5,11,5,11,5]2010; AVX512DQ-BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]2011; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm62012; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [11,1,7,0]2013; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm0, %zmm1, %zmm92014; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k2}2015; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k1}2016; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [12,0,0,6,12,0,0,6]2017; AVX512DQ-BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]2018; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm62019; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm1, %zmm0, %zmm52020; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm6, %zmm52021; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm5 {%k1}2022; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [13,0,1,7,13,0,1,7]2023; AVX512DQ-BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]2024; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm3, %zmm62025; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm02026; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm0, %zmm6, %zmm02027; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm0 {%k1}2028; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)2029; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, (%rdx)2030; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rcx)2031; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, (%r8)2032; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)2033; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)2034; AVX512DQ-BW-FCP-NEXT:    vzeroupper2035; AVX512DQ-BW-FCP-NEXT:    retq2036  %wide.vec = load <48 x i64>, ptr %in.vec, align 642037  %strided.vec0 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42>2038  %strided.vec1 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43>2039  %strided.vec2 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44>2040  %strided.vec3 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45>2041  %strided.vec4 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46>2042  %strided.vec5 = shufflevector <48 x i64> %wide.vec, <48 x i64> poison, <8 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47>2043  store <8 x i64> %strided.vec0, ptr %out.vec0, align 642044  store <8 x i64> %strided.vec1, ptr %out.vec1, align 642045  store <8 x i64> %strided.vec2, ptr %out.vec2, align 642046  store <8 x i64> %strided.vec3, ptr %out.vec3, align 642047  store <8 x i64> %strided.vec4, ptr %out.vec4, align 642048  store <8 x i64> %strided.vec5, ptr %out.vec5, align 642049  ret void2050}2051 2052define void @load_i64_stride6_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {2053; SSE-LABEL: load_i64_stride6_vf16:2054; SSE:       # %bb.0:2055; SSE-NEXT:    subq $408, %rsp # imm = 0x1982056; SSE-NEXT:    movaps (%rdi), %xmm72057; SSE-NEXT:    movaps 624(%rdi), %xmm02058; SSE-NEXT:    movaps 576(%rdi), %xmm82059; SSE-NEXT:    movaps 240(%rdi), %xmm12060; SSE-NEXT:    movaps 192(%rdi), %xmm92061; SSE-NEXT:    movaps 720(%rdi), %xmm22062; SSE-NEXT:    movaps 672(%rdi), %xmm102063; SSE-NEXT:    movaps 336(%rdi), %xmm32064; SSE-NEXT:    movaps 288(%rdi), %xmm112065; SSE-NEXT:    movaps 432(%rdi), %xmm42066; SSE-NEXT:    movaps 384(%rdi), %xmm132067; SSE-NEXT:    movaps 528(%rdi), %xmm52068; SSE-NEXT:    movaps 480(%rdi), %xmm122069; SSE-NEXT:    movaps 144(%rdi), %xmm62070; SSE-NEXT:    movaps 96(%rdi), %xmm142071; SSE-NEXT:    movaps %xmm14, %xmm152072; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm6[0]2073; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2074; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm6[1]2075; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2076; SSE-NEXT:    movaps %xmm12, %xmm62077; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]2078; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2079; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]2080; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2081; SSE-NEXT:    movaps %xmm13, %xmm52082; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]2083; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2084; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm4[1]2085; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2086; SSE-NEXT:    movaps %xmm11, %xmm42087; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]2088; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2089; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm3[1]2090; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2091; SSE-NEXT:    movaps %xmm10, %xmm32092; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]2093; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2094; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm2[1]2095; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2096; SSE-NEXT:    movaps %xmm9, %xmm22097; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]2098; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2099; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm1[1]2100; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2101; SSE-NEXT:    movaps %xmm8, %xmm12102; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2103; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2104; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]2105; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2106; SSE-NEXT:    movaps 48(%rdi), %xmm02107; SSE-NEXT:    movaps %xmm7, %xmm12108; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2109; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2110; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]2111; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2112; SSE-NEXT:    movaps 64(%rdi), %xmm02113; SSE-NEXT:    movaps 16(%rdi), %xmm12114; SSE-NEXT:    movaps %xmm1, %xmm22115; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2116; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2117; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2118; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2119; SSE-NEXT:    movaps 160(%rdi), %xmm02120; SSE-NEXT:    movaps 112(%rdi), %xmm12121; SSE-NEXT:    movaps %xmm1, %xmm22122; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2123; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2124; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2125; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2126; SSE-NEXT:    movaps 256(%rdi), %xmm02127; SSE-NEXT:    movaps 208(%rdi), %xmm12128; SSE-NEXT:    movaps %xmm1, %xmm22129; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2130; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2131; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2132; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2133; SSE-NEXT:    movaps 352(%rdi), %xmm02134; SSE-NEXT:    movaps 304(%rdi), %xmm12135; SSE-NEXT:    movaps %xmm1, %xmm22136; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2137; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2138; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2139; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2140; SSE-NEXT:    movaps 448(%rdi), %xmm02141; SSE-NEXT:    movaps 400(%rdi), %xmm12142; SSE-NEXT:    movaps %xmm1, %xmm22143; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2144; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2145; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2146; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill2147; SSE-NEXT:    movaps 544(%rdi), %xmm02148; SSE-NEXT:    movaps 496(%rdi), %xmm12149; SSE-NEXT:    movaps %xmm1, %xmm22150; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2151; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2152; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2153; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2154; SSE-NEXT:    movaps 640(%rdi), %xmm02155; SSE-NEXT:    movaps 592(%rdi), %xmm152156; SSE-NEXT:    movaps %xmm15, %xmm12157; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2158; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2159; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]2160; SSE-NEXT:    movaps 736(%rdi), %xmm02161; SSE-NEXT:    movaps 688(%rdi), %xmm112162; SSE-NEXT:    movaps %xmm11, %xmm12163; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2164; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2165; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]2166; SSE-NEXT:    movaps 80(%rdi), %xmm02167; SSE-NEXT:    movaps 32(%rdi), %xmm12168; SSE-NEXT:    movaps %xmm1, %xmm22169; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2170; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2171; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2172; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2173; SSE-NEXT:    movaps 176(%rdi), %xmm02174; SSE-NEXT:    movaps 128(%rdi), %xmm132175; SSE-NEXT:    movaps %xmm13, %xmm12176; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2177; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2178; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]2179; SSE-NEXT:    movaps 272(%rdi), %xmm02180; SSE-NEXT:    movaps 224(%rdi), %xmm122181; SSE-NEXT:    movaps %xmm12, %xmm142182; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]2183; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]2184; SSE-NEXT:    movaps 368(%rdi), %xmm02185; SSE-NEXT:    movaps 320(%rdi), %xmm92186; SSE-NEXT:    movaps %xmm9, %xmm102187; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]2188; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]2189; SSE-NEXT:    movaps 464(%rdi), %xmm02190; SSE-NEXT:    movaps 416(%rdi), %xmm52191; SSE-NEXT:    movaps %xmm5, %xmm82192; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]2193; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]2194; SSE-NEXT:    movaps 560(%rdi), %xmm02195; SSE-NEXT:    movaps 512(%rdi), %xmm62196; SSE-NEXT:    movaps %xmm6, %xmm72197; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]2198; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]2199; SSE-NEXT:    movaps 656(%rdi), %xmm02200; SSE-NEXT:    movaps 608(%rdi), %xmm32201; SSE-NEXT:    movaps %xmm3, %xmm42202; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]2203; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]2204; SSE-NEXT:    movaps 752(%rdi), %xmm02205; SSE-NEXT:    movaps 704(%rdi), %xmm12206; SSE-NEXT:    movaps %xmm1, %xmm22207; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2208; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2209; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2210; SSE-NEXT:    movaps %xmm0, 96(%rsi)2211; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2212; SSE-NEXT:    movaps %xmm0, 32(%rsi)2213; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2214; SSE-NEXT:    movaps %xmm0, 112(%rsi)2215; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2216; SSE-NEXT:    movaps %xmm0, 48(%rsi)2217; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2218; SSE-NEXT:    movaps %xmm0, 64(%rsi)2219; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2220; SSE-NEXT:    movaps %xmm0, (%rsi)2221; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2222; SSE-NEXT:    movaps %xmm0, 80(%rsi)2223; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2224; SSE-NEXT:    movaps %xmm0, 16(%rsi)2225; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2226; SSE-NEXT:    movaps %xmm0, 96(%rdx)2227; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2228; SSE-NEXT:    movaps %xmm0, 32(%rdx)2229; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2230; SSE-NEXT:    movaps %xmm0, 112(%rdx)2231; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2232; SSE-NEXT:    movaps %xmm0, 48(%rdx)2233; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2234; SSE-NEXT:    movaps %xmm0, 64(%rdx)2235; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2236; SSE-NEXT:    movaps %xmm0, (%rdx)2237; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2238; SSE-NEXT:    movaps %xmm0, 80(%rdx)2239; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2240; SSE-NEXT:    movaps %xmm0, 16(%rdx)2241; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2242; SSE-NEXT:    movaps %xmm0, 96(%rcx)2243; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2244; SSE-NEXT:    movaps %xmm0, 112(%rcx)2245; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2246; SSE-NEXT:    movaps %xmm0, 64(%rcx)2247; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2248; SSE-NEXT:    movaps %xmm0, 80(%rcx)2249; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2250; SSE-NEXT:    movaps %xmm0, 32(%rcx)2251; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2252; SSE-NEXT:    movaps %xmm0, 48(%rcx)2253; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2254; SSE-NEXT:    movaps %xmm0, (%rcx)2255; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2256; SSE-NEXT:    movaps %xmm0, 16(%rcx)2257; SSE-NEXT:    movaps %xmm11, 112(%r8)2258; SSE-NEXT:    movaps %xmm15, 96(%r8)2259; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2260; SSE-NEXT:    movaps %xmm0, 80(%r8)2261; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload2262; SSE-NEXT:    movaps %xmm0, 64(%r8)2263; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2264; SSE-NEXT:    movaps %xmm0, 48(%r8)2265; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2266; SSE-NEXT:    movaps %xmm0, 32(%r8)2267; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2268; SSE-NEXT:    movaps %xmm0, 16(%r8)2269; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2270; SSE-NEXT:    movaps %xmm0, (%r8)2271; SSE-NEXT:    movaps %xmm2, 112(%r9)2272; SSE-NEXT:    movaps %xmm4, 96(%r9)2273; SSE-NEXT:    movaps %xmm7, 80(%r9)2274; SSE-NEXT:    movaps %xmm8, 64(%r9)2275; SSE-NEXT:    movaps %xmm10, 48(%r9)2276; SSE-NEXT:    movaps %xmm14, 32(%r9)2277; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2278; SSE-NEXT:    movaps %xmm0, 16(%r9)2279; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2280; SSE-NEXT:    movaps %xmm0, (%r9)2281; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax2282; SSE-NEXT:    movaps %xmm1, 112(%rax)2283; SSE-NEXT:    movaps %xmm3, 96(%rax)2284; SSE-NEXT:    movaps %xmm6, 80(%rax)2285; SSE-NEXT:    movaps %xmm5, 64(%rax)2286; SSE-NEXT:    movaps %xmm9, 48(%rax)2287; SSE-NEXT:    movaps %xmm12, 32(%rax)2288; SSE-NEXT:    movaps %xmm13, 16(%rax)2289; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2290; SSE-NEXT:    movaps %xmm0, (%rax)2291; SSE-NEXT:    addq $408, %rsp # imm = 0x1982292; SSE-NEXT:    retq2293;2294; AVX-LABEL: load_i64_stride6_vf16:2295; AVX:       # %bb.0:2296; AVX-NEXT:    subq $552, %rsp # imm = 0x2282297; AVX-NEXT:    vmovaps 320(%rdi), %ymm12298; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2299; AVX-NEXT:    vmovaps 512(%rdi), %ymm82300; AVX-NEXT:    vmovups %ymm8, (%rsp) # 32-byte Spill2301; AVX-NEXT:    vmovaps 128(%rdi), %ymm22302; AVX-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm32303; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]2304; AVX-NEXT:    vmovaps (%rdi), %xmm52305; AVX-NEXT:    vmovaps 48(%rdi), %xmm62306; AVX-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm5[0],xmm6[0]2307; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm4[4,5,6,7]2308; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2309; AVX-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm42310; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm4[0],ymm8[0],ymm4[2],ymm8[2]2311; AVX-NEXT:    vmovaps 432(%rdi), %xmm82312; AVX-NEXT:    vmovaps 384(%rdi), %xmm92313; AVX-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]2314; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm7[4,5,6,7]2315; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2316; AVX-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm72317; AVX-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm1[0],ymm7[2],ymm1[2]2318; AVX-NEXT:    vmovaps 240(%rdi), %xmm112319; AVX-NEXT:    vmovaps 192(%rdi), %xmm122320; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]2321; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm10[4,5,6,7]2322; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2323; AVX-NEXT:    vmovaps 704(%rdi), %ymm102324; AVX-NEXT:    vinsertf128 $1, 672(%rdi), %ymm0, %ymm132325; AVX-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm13[0],ymm10[0],ymm13[2],ymm10[2]2326; AVX-NEXT:    vmovaps 624(%rdi), %xmm152327; AVX-NEXT:    vmovaps 576(%rdi), %xmm02328; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm15[0]2329; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]2330; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2331; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]2332; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm6[1]2333; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]2334; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2335; AVX-NEXT:    vunpckhpd (%rsp), %ymm4, %ymm1 # 32-byte Folded Reload2336; AVX-NEXT:    # ymm1 = ymm4[1],mem[1],ymm4[3],mem[3]2337; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm9[1],xmm8[1]2338; AVX-NEXT:    vmovaps 480(%rdi), %ymm32339; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2340; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]2341; AVX-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill2342; AVX-NEXT:    vmovaps 96(%rdi), %ymm22343; AVX-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2344; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm4 # 32-byte Folded Reload2345; AVX-NEXT:    # ymm4 = ymm7[1],mem[1],ymm7[3],mem[3]2346; AVX-NEXT:    vmovaps 16(%rdi), %xmm62347; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2348; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm12[1],xmm11[1]2349; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm4[4,5,6,7]2350; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2351; AVX-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm13[1],ymm10[1],ymm13[3],ymm10[3]2352; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]2353; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]2354; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2355; AVX-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm142356; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm14[0],ymm2[2],ymm14[2]2357; AVX-NEXT:    vmovaps 64(%rdi), %xmm132358; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm6[0],xmm13[0]2359; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm4[4,5,6,7]2360; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2361; AVX-NEXT:    vinsertf128 $1, 544(%rdi), %ymm0, %ymm112362; AVX-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm11[0],ymm3[2],ymm11[2]2363; AVX-NEXT:    vmovaps 448(%rdi), %xmm72364; AVX-NEXT:    vmovaps 400(%rdi), %xmm82365; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm7[0]2366; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm6[4,5,6,7]2367; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2368; AVX-NEXT:    vmovaps 288(%rdi), %ymm62369; AVX-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm152370; AVX-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm6[0],ymm15[0],ymm6[2],ymm15[2]2371; AVX-NEXT:    vmovaps 256(%rdi), %xmm42372; AVX-NEXT:    vmovaps 208(%rdi), %xmm32373; AVX-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm3[0],xmm4[0]2374; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm10[4,5,6,7]2375; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2376; AVX-NEXT:    vmovaps 672(%rdi), %ymm52377; AVX-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm22378; AVX-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]2379; AVX-NEXT:    vmovaps 640(%rdi), %xmm12380; AVX-NEXT:    vmovaps 592(%rdi), %xmm02381; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm0[0],xmm1[0]2382; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm12[4,5,6,7]2383; AVX-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2384; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2385; AVX-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm14[1],ymm9[3],ymm14[3]2386; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload2387; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm10[1],xmm13[1]2388; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]2389; AVX-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2390; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2391; AVX-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm9[1],ymm11[1],ymm9[3],ymm11[3]2392; AVX-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm8[1],xmm7[1]2393; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm11[4,5,6,7]2394; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2395; AVX-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]2396; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]2397; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]2398; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2399; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm15[1],ymm6[3],ymm15[3]2400; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm4[1]2401; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2402; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2403; AVX-NEXT:    vmovaps 544(%rdi), %ymm02404; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2405; AVX-NEXT:    vinsertf128 $1, 512(%rdi), %ymm0, %ymm12406; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2407; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]2408; AVX-NEXT:    vmovaps 464(%rdi), %xmm142409; AVX-NEXT:    vmovaps 416(%rdi), %xmm132410; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm13[0],xmm14[0]2411; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]2412; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2413; AVX-NEXT:    vmovaps 160(%rdi), %ymm112414; AVX-NEXT:    vmovaps 32(%rdi), %xmm122415; AVX-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm102416; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]2417; AVX-NEXT:    vmovaps 80(%rdi), %xmm92418; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm12[0],xmm9[0]2419; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]2420; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2421; AVX-NEXT:    vmovaps 352(%rdi), %ymm82422; AVX-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm72423; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]2424; AVX-NEXT:    vmovaps 272(%rdi), %xmm62425; AVX-NEXT:    vmovaps 224(%rdi), %xmm32426; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm3[0],xmm6[0]2427; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]2428; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2429; AVX-NEXT:    vmovaps 736(%rdi), %ymm52430; AVX-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm22431; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm5[0],ymm2[2],ymm5[2]2432; AVX-NEXT:    vmovaps 656(%rdi), %xmm12433; AVX-NEXT:    vmovaps 608(%rdi), %xmm02434; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm0[0],xmm1[0]2435; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm15[0,1,2,3],ymm4[4,5,6,7]2436; AVX-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]2437; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]2438; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3],ymm2[4,5,6,7]2439; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2440; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload2441; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]2442; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm13[1],xmm14[1]2443; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2444; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]2445; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm12[1],xmm9[1]2446; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]2447; AVX-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]2448; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]2449; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]2450; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2451; AVX-NEXT:    vmovaps %ymm3, 96(%rsi)2452; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2453; AVX-NEXT:    vmovaps %ymm3, 32(%rsi)2454; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2455; AVX-NEXT:    vmovaps %ymm3, 64(%rsi)2456; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2457; AVX-NEXT:    vmovaps %ymm3, (%rsi)2458; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2459; AVX-NEXT:    vmovaps %ymm3, 96(%rdx)2460; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2461; AVX-NEXT:    vmovaps %ymm3, 32(%rdx)2462; AVX-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload2463; AVX-NEXT:    vmovaps %ymm3, 64(%rdx)2464; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2465; AVX-NEXT:    vmovaps %ymm3, (%rdx)2466; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2467; AVX-NEXT:    vmovaps %ymm3, 96(%rcx)2468; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2469; AVX-NEXT:    vmovaps %ymm3, 32(%rcx)2470; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2471; AVX-NEXT:    vmovaps %ymm3, 64(%rcx)2472; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2473; AVX-NEXT:    vmovaps %ymm3, (%rcx)2474; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2475; AVX-NEXT:    vmovaps %ymm3, 32(%r8)2476; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2477; AVX-NEXT:    vmovaps %ymm3, 96(%r8)2478; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2479; AVX-NEXT:    vmovaps %ymm3, 64(%r8)2480; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2481; AVX-NEXT:    vmovaps %ymm3, (%r8)2482; AVX-NEXT:    vmovaps %ymm4, 96(%r9)2483; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2484; AVX-NEXT:    vmovaps %ymm3, 32(%r9)2485; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2486; AVX-NEXT:    vmovaps %ymm3, (%r9)2487; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2488; AVX-NEXT:    vmovaps %ymm3, 64(%r9)2489; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax2490; AVX-NEXT:    vmovaps %ymm2, 32(%rax)2491; AVX-NEXT:    vmovaps %ymm1, (%rax)2492; AVX-NEXT:    vmovaps %ymm0, 64(%rax)2493; AVX-NEXT:    vmovaps %ymm5, 96(%rax)2494; AVX-NEXT:    addq $552, %rsp # imm = 0x2282495; AVX-NEXT:    vzeroupper2496; AVX-NEXT:    retq2497;2498; AVX2-LABEL: load_i64_stride6_vf16:2499; AVX2:       # %bb.0:2500; AVX2-NEXT:    subq $488, %rsp # imm = 0x1E82501; AVX2-NEXT:    vmovaps 320(%rdi), %ymm102502; AVX2-NEXT:    vmovaps 288(%rdi), %ymm122503; AVX2-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2504; AVX2-NEXT:    vmovaps 512(%rdi), %ymm72505; AVX2-NEXT:    vmovaps 480(%rdi), %ymm42506; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2507; AVX2-NEXT:    vmovaps 128(%rdi), %ymm152508; AVX2-NEXT:    vmovaps 96(%rdi), %ymm32509; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2510; AVX2-NEXT:    vmovaps (%rdi), %xmm12511; AVX2-NEXT:    vmovaps 48(%rdi), %xmm52512; AVX2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm5[0]2513; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm15[0],ymm3[2],ymm15[2]2514; AVX2-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2515; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]2516; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm3[4,5,6,7]2517; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2518; AVX2-NEXT:    vmovaps 432(%rdi), %xmm32519; AVX2-NEXT:    vmovaps 384(%rdi), %xmm62520; AVX2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm3[0]2521; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]2522; AVX2-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2523; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,3]2524; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm4[4,5,6,7]2525; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2526; AVX2-NEXT:    vmovaps 240(%rdi), %xmm112527; AVX2-NEXT:    vmovaps 192(%rdi), %xmm92528; AVX2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm9[0],xmm11[0]2529; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]2530; AVX2-NEXT:    vmovaps %ymm10, %ymm42531; AVX2-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2532; AVX2-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2533; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm8[4,5,6,7]2534; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2535; AVX2-NEXT:    vmovaps 704(%rdi), %ymm102536; AVX2-NEXT:    vmovaps 672(%rdi), %ymm82537; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]2538; AVX2-NEXT:    vpermpd {{.*#+}} ymm13 = ymm13[0,1,0,3]2539; AVX2-NEXT:    vmovaps 624(%rdi), %xmm142540; AVX2-NEXT:    vmovaps 576(%rdi), %xmm02541; AVX2-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm0[0],xmm14[0]2542; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm12[0,1,2,3],ymm13[4,5,6,7]2543; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2544; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm5[1]2545; AVX2-NEXT:    vbroadcastsd 104(%rdi), %ymm52546; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm15[1],ymm5[3],ymm15[3]2547; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]2548; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2549; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm3[1]2550; AVX2-NEXT:    vbroadcastsd 488(%rdi), %ymm32551; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]2552; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]2553; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2554; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm9[1],xmm11[1]2555; AVX2-NEXT:    vbroadcastsd 296(%rdi), %ymm32556; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]2557; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]2558; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2559; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm14[1]2560; AVX2-NEXT:    vbroadcastsd 680(%rdi), %ymm12561; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]2562; AVX2-NEXT:    vmovaps %ymm10, %ymm142563; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2564; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2565; AVX2-NEXT:    vbroadcastsd 160(%rdi), %ymm02566; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload2567; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]2568; AVX2-NEXT:    vmovaps 16(%rdi), %xmm72569; AVX2-NEXT:    vmovaps 64(%rdi), %xmm32570; AVX2-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm7[0],xmm3[0]2571; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]2572; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2573; AVX2-NEXT:    vbroadcastsd 544(%rdi), %ymm02574; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload2575; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]2576; AVX2-NEXT:    vmovaps 448(%rdi), %xmm42577; AVX2-NEXT:    vmovaps 400(%rdi), %xmm52578; AVX2-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]2579; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]2580; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2581; AVX2-NEXT:    vbroadcastsd 352(%rdi), %ymm02582; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2583; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]2584; AVX2-NEXT:    vmovaps 256(%rdi), %xmm122585; AVX2-NEXT:    vmovaps 208(%rdi), %xmm22586; AVX2-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm12[0]2587; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]2588; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2589; AVX2-NEXT:    vbroadcastsd 736(%rdi), %ymm02590; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm0[0],ymm8[2],ymm0[2]2591; AVX2-NEXT:    vmovaps 640(%rdi), %xmm132592; AVX2-NEXT:    vmovaps 592(%rdi), %xmm12593; AVX2-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm1[0],xmm13[0]2594; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]2595; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2596; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm3[1]2597; AVX2-NEXT:    vmovaps 160(%rdi), %ymm152598; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm15[1],ymm10[3],ymm15[3]2599; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]2600; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]2601; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2602; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]2603; AVX2-NEXT:    vmovaps 544(%rdi), %ymm102604; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]2605; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]2606; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]2607; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2608; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm13[1]2609; AVX2-NEXT:    vmovaps 736(%rdi), %ymm42610; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm4[1],ymm8[3],ymm4[3]2611; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]2612; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm3[4,5,6,7]2613; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill2614; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm12[1]2615; AVX2-NEXT:    vmovaps 352(%rdi), %ymm52616; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm9[1],ymm5[1],ymm9[3],ymm5[3]2617; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]2618; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]2619; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2620; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2621; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]2622; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]2623; AVX2-NEXT:    vmovaps 464(%rdi), %xmm72624; AVX2-NEXT:    vmovaps 416(%rdi), %xmm62625; AVX2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm7[0]2626; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm1[4,5,6,7]2627; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2628; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2629; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]2630; AVX2-NEXT:    vpermpd {{.*#+}} ymm8 = ymm1[0,1,0,3]2631; AVX2-NEXT:    vmovaps 32(%rdi), %xmm32632; AVX2-NEXT:    vmovaps 80(%rdi), %xmm22633; AVX2-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm3[0],xmm2[0]2634; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm8[4,5,6,7]2635; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2636; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]2637; AVX2-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2638; AVX2-NEXT:    vmovaps 272(%rdi), %xmm122639; AVX2-NEXT:    vmovaps 224(%rdi), %xmm12640; AVX2-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm12[0]2641; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm8[4,5,6,7]2642; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]2643; AVX2-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2644; AVX2-NEXT:    vmovaps 656(%rdi), %xmm92645; AVX2-NEXT:    vmovaps 608(%rdi), %xmm02646; AVX2-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm9[0]2647; AVX2-NEXT:    vblendps {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]2648; AVX2-NEXT:    vbroadcastsd 712(%rdi), %ymm142649; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]2650; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]2651; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm4[4,5,6,7]2652; AVX2-NEXT:    vbroadcastsd 520(%rdi), %ymm02653; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]2654; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm7[1]2655; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]2656; AVX2-NEXT:    vbroadcastsd 136(%rdi), %ymm02657; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm15[1],ymm0[3],ymm15[3]2658; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]2659; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]2660; AVX2-NEXT:    vbroadcastsd 328(%rdi), %ymm22661; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]2662; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm12[1]2663; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2664; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2665; AVX2-NEXT:    vmovaps %ymm2, 96(%rsi)2666; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2667; AVX2-NEXT:    vmovaps %ymm2, 32(%rsi)2668; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2669; AVX2-NEXT:    vmovaps %ymm2, 64(%rsi)2670; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2671; AVX2-NEXT:    vmovaps %ymm2, (%rsi)2672; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2673; AVX2-NEXT:    vmovaps %ymm2, 96(%rdx)2674; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2675; AVX2-NEXT:    vmovaps %ymm2, 32(%rdx)2676; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2677; AVX2-NEXT:    vmovaps %ymm2, 64(%rdx)2678; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2679; AVX2-NEXT:    vmovaps %ymm2, (%rdx)2680; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2681; AVX2-NEXT:    vmovaps %ymm2, 96(%rcx)2682; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2683; AVX2-NEXT:    vmovaps %ymm2, 32(%rcx)2684; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2685; AVX2-NEXT:    vmovaps %ymm2, 64(%rcx)2686; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2687; AVX2-NEXT:    vmovaps %ymm2, (%rcx)2688; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2689; AVX2-NEXT:    vmovaps %ymm2, 32(%r8)2690; AVX2-NEXT:    vmovups (%rsp), %ymm2 # 32-byte Reload2691; AVX2-NEXT:    vmovaps %ymm2, 96(%r8)2692; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2693; AVX2-NEXT:    vmovaps %ymm2, 64(%r8)2694; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2695; AVX2-NEXT:    vmovaps %ymm2, (%r8)2696; AVX2-NEXT:    vmovaps %ymm8, 96(%r9)2697; AVX2-NEXT:    vmovaps %ymm11, 32(%r9)2698; AVX2-NEXT:    vmovaps %ymm13, (%r9)2699; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2700; AVX2-NEXT:    vmovaps %ymm2, 64(%r9)2701; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax2702; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)2703; AVX2-NEXT:    vmovaps %ymm0, (%rax)2704; AVX2-NEXT:    vmovaps %ymm4, 64(%rax)2705; AVX2-NEXT:    vmovaps %ymm9, 96(%rax)2706; AVX2-NEXT:    addq $488, %rsp # imm = 0x1E82707; AVX2-NEXT:    vzeroupper2708; AVX2-NEXT:    retq2709;2710; AVX2-FP-LABEL: load_i64_stride6_vf16:2711; AVX2-FP:       # %bb.0:2712; AVX2-FP-NEXT:    subq $488, %rsp # imm = 0x1E82713; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm102714; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm122715; AVX2-FP-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2716; AVX2-FP-NEXT:    vmovaps 512(%rdi), %ymm72717; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm42718; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2719; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm152720; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm32721; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2722; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm12723; AVX2-FP-NEXT:    vmovaps 48(%rdi), %xmm52724; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm5[0]2725; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm15[0],ymm3[2],ymm15[2]2726; AVX2-FP-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2727; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]2728; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm3[4,5,6,7]2729; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2730; AVX2-FP-NEXT:    vmovaps 432(%rdi), %xmm32731; AVX2-FP-NEXT:    vmovaps 384(%rdi), %xmm62732; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm3[0]2733; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]2734; AVX2-FP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2735; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,3]2736; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm4[4,5,6,7]2737; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2738; AVX2-FP-NEXT:    vmovaps 240(%rdi), %xmm112739; AVX2-FP-NEXT:    vmovaps 192(%rdi), %xmm92740; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm9[0],xmm11[0]2741; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]2742; AVX2-FP-NEXT:    vmovaps %ymm10, %ymm42743; AVX2-FP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2744; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2745; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm8[4,5,6,7]2746; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2747; AVX2-FP-NEXT:    vmovaps 704(%rdi), %ymm102748; AVX2-FP-NEXT:    vmovaps 672(%rdi), %ymm82749; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]2750; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm13 = ymm13[0,1,0,3]2751; AVX2-FP-NEXT:    vmovaps 624(%rdi), %xmm142752; AVX2-FP-NEXT:    vmovaps 576(%rdi), %xmm02753; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm0[0],xmm14[0]2754; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm12[0,1,2,3],ymm13[4,5,6,7]2755; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2756; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm5[1]2757; AVX2-FP-NEXT:    vbroadcastsd 104(%rdi), %ymm52758; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm15[1],ymm5[3],ymm15[3]2759; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]2760; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2761; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm3[1]2762; AVX2-FP-NEXT:    vbroadcastsd 488(%rdi), %ymm32763; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]2764; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]2765; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2766; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm9[1],xmm11[1]2767; AVX2-FP-NEXT:    vbroadcastsd 296(%rdi), %ymm32768; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]2769; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]2770; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2771; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm14[1]2772; AVX2-FP-NEXT:    vbroadcastsd 680(%rdi), %ymm12773; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]2774; AVX2-FP-NEXT:    vmovaps %ymm10, %ymm142775; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2776; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2777; AVX2-FP-NEXT:    vbroadcastsd 160(%rdi), %ymm02778; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload2779; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]2780; AVX2-FP-NEXT:    vmovaps 16(%rdi), %xmm72781; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm32782; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm7[0],xmm3[0]2783; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]2784; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2785; AVX2-FP-NEXT:    vbroadcastsd 544(%rdi), %ymm02786; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload2787; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]2788; AVX2-FP-NEXT:    vmovaps 448(%rdi), %xmm42789; AVX2-FP-NEXT:    vmovaps 400(%rdi), %xmm52790; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]2791; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]2792; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2793; AVX2-FP-NEXT:    vbroadcastsd 352(%rdi), %ymm02794; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2795; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]2796; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm122797; AVX2-FP-NEXT:    vmovaps 208(%rdi), %xmm22798; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm12[0]2799; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]2800; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2801; AVX2-FP-NEXT:    vbroadcastsd 736(%rdi), %ymm02802; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm0[0],ymm8[2],ymm0[2]2803; AVX2-FP-NEXT:    vmovaps 640(%rdi), %xmm132804; AVX2-FP-NEXT:    vmovaps 592(%rdi), %xmm12805; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm1[0],xmm13[0]2806; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]2807; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2808; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm3[1]2809; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm152810; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm15[1],ymm10[3],ymm15[3]2811; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]2812; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]2813; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2814; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]2815; AVX2-FP-NEXT:    vmovaps 544(%rdi), %ymm102816; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]2817; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]2818; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]2819; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2820; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm13[1]2821; AVX2-FP-NEXT:    vmovaps 736(%rdi), %ymm42822; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm4[1],ymm8[3],ymm4[3]2823; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]2824; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm3[4,5,6,7]2825; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill2826; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm12[1]2827; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm52828; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm9[1],ymm5[1],ymm9[3],ymm5[3]2829; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]2830; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]2831; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2832; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2833; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]2834; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]2835; AVX2-FP-NEXT:    vmovaps 464(%rdi), %xmm72836; AVX2-FP-NEXT:    vmovaps 416(%rdi), %xmm62837; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm7[0]2838; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm1[4,5,6,7]2839; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2840; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2841; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]2842; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm1[0,1,0,3]2843; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm32844; AVX2-FP-NEXT:    vmovaps 80(%rdi), %xmm22845; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm3[0],xmm2[0]2846; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm8[4,5,6,7]2847; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2848; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]2849; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2850; AVX2-FP-NEXT:    vmovaps 272(%rdi), %xmm122851; AVX2-FP-NEXT:    vmovaps 224(%rdi), %xmm12852; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm12[0]2853; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm8[4,5,6,7]2854; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]2855; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2856; AVX2-FP-NEXT:    vmovaps 656(%rdi), %xmm92857; AVX2-FP-NEXT:    vmovaps 608(%rdi), %xmm02858; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm9[0]2859; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]2860; AVX2-FP-NEXT:    vbroadcastsd 712(%rdi), %ymm142861; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]2862; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]2863; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm4[4,5,6,7]2864; AVX2-FP-NEXT:    vbroadcastsd 520(%rdi), %ymm02865; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]2866; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm7[1]2867; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]2868; AVX2-FP-NEXT:    vbroadcastsd 136(%rdi), %ymm02869; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm15[1],ymm0[3],ymm15[3]2870; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]2871; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]2872; AVX2-FP-NEXT:    vbroadcastsd 328(%rdi), %ymm22873; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]2874; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm12[1]2875; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2876; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2877; AVX2-FP-NEXT:    vmovaps %ymm2, 96(%rsi)2878; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2879; AVX2-FP-NEXT:    vmovaps %ymm2, 32(%rsi)2880; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2881; AVX2-FP-NEXT:    vmovaps %ymm2, 64(%rsi)2882; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2883; AVX2-FP-NEXT:    vmovaps %ymm2, (%rsi)2884; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2885; AVX2-FP-NEXT:    vmovaps %ymm2, 96(%rdx)2886; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2887; AVX2-FP-NEXT:    vmovaps %ymm2, 32(%rdx)2888; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2889; AVX2-FP-NEXT:    vmovaps %ymm2, 64(%rdx)2890; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2891; AVX2-FP-NEXT:    vmovaps %ymm2, (%rdx)2892; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2893; AVX2-FP-NEXT:    vmovaps %ymm2, 96(%rcx)2894; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2895; AVX2-FP-NEXT:    vmovaps %ymm2, 32(%rcx)2896; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2897; AVX2-FP-NEXT:    vmovaps %ymm2, 64(%rcx)2898; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2899; AVX2-FP-NEXT:    vmovaps %ymm2, (%rcx)2900; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2901; AVX2-FP-NEXT:    vmovaps %ymm2, 32(%r8)2902; AVX2-FP-NEXT:    vmovups (%rsp), %ymm2 # 32-byte Reload2903; AVX2-FP-NEXT:    vmovaps %ymm2, 96(%r8)2904; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2905; AVX2-FP-NEXT:    vmovaps %ymm2, 64(%r8)2906; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2907; AVX2-FP-NEXT:    vmovaps %ymm2, (%r8)2908; AVX2-FP-NEXT:    vmovaps %ymm8, 96(%r9)2909; AVX2-FP-NEXT:    vmovaps %ymm11, 32(%r9)2910; AVX2-FP-NEXT:    vmovaps %ymm13, (%r9)2911; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2912; AVX2-FP-NEXT:    vmovaps %ymm2, 64(%r9)2913; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax2914; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)2915; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)2916; AVX2-FP-NEXT:    vmovaps %ymm4, 64(%rax)2917; AVX2-FP-NEXT:    vmovaps %ymm9, 96(%rax)2918; AVX2-FP-NEXT:    addq $488, %rsp # imm = 0x1E82919; AVX2-FP-NEXT:    vzeroupper2920; AVX2-FP-NEXT:    retq2921;2922; AVX2-FCP-LABEL: load_i64_stride6_vf16:2923; AVX2-FCP:       # %bb.0:2924; AVX2-FCP-NEXT:    subq $488, %rsp # imm = 0x1E82925; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm102926; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm122927; AVX2-FCP-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2928; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %ymm72929; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm42930; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2931; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm152932; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm32933; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2934; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm12935; AVX2-FCP-NEXT:    vmovaps 48(%rdi), %xmm52936; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm5[0]2937; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm15[0],ymm3[2],ymm15[2]2938; AVX2-FCP-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2939; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]2940; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm3[4,5,6,7]2941; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2942; AVX2-FCP-NEXT:    vmovaps 432(%rdi), %xmm32943; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %xmm62944; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm3[0]2945; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[2]2946; AVX2-FCP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2947; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,3]2948; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm4[4,5,6,7]2949; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2950; AVX2-FCP-NEXT:    vmovaps 240(%rdi), %xmm112951; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %xmm92952; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm9[0],xmm11[0]2953; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]2954; AVX2-FCP-NEXT:    vmovaps %ymm10, %ymm42955; AVX2-FCP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2956; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2957; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm8[4,5,6,7]2958; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2959; AVX2-FCP-NEXT:    vmovaps 704(%rdi), %ymm102960; AVX2-FCP-NEXT:    vmovaps 672(%rdi), %ymm82961; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]2962; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm13 = ymm13[0,1,0,3]2963; AVX2-FCP-NEXT:    vmovaps 624(%rdi), %xmm142964; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %xmm02965; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm0[0],xmm14[0]2966; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm12[0,1,2,3],ymm13[4,5,6,7]2967; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2968; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm5[1]2969; AVX2-FCP-NEXT:    vbroadcastsd 104(%rdi), %ymm52970; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm15[1],ymm5[3],ymm15[3]2971; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]2972; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2973; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm3[1]2974; AVX2-FCP-NEXT:    vbroadcastsd 488(%rdi), %ymm32975; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]2976; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]2977; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2978; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm9[1],xmm11[1]2979; AVX2-FCP-NEXT:    vbroadcastsd 296(%rdi), %ymm32980; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]2981; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]2982; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2983; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm14[1]2984; AVX2-FCP-NEXT:    vbroadcastsd 680(%rdi), %ymm12985; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]2986; AVX2-FCP-NEXT:    vmovaps %ymm10, %ymm142987; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2988; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2989; AVX2-FCP-NEXT:    vbroadcastsd 160(%rdi), %ymm02990; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload2991; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]2992; AVX2-FCP-NEXT:    vmovaps 16(%rdi), %xmm72993; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm32994; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm7[0],xmm3[0]2995; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]2996; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2997; AVX2-FCP-NEXT:    vbroadcastsd 544(%rdi), %ymm02998; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload2999; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[2],ymm0[2]3000; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %xmm43001; AVX2-FCP-NEXT:    vmovaps 400(%rdi), %xmm53002; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]3003; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]3004; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3005; AVX2-FCP-NEXT:    vbroadcastsd 352(%rdi), %ymm03006; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload3007; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]3008; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm123009; AVX2-FCP-NEXT:    vmovaps 208(%rdi), %xmm23010; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm12[0]3011; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]3012; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3013; AVX2-FCP-NEXT:    vbroadcastsd 736(%rdi), %ymm03014; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm0[0],ymm8[2],ymm0[2]3015; AVX2-FCP-NEXT:    vmovaps 640(%rdi), %xmm133016; AVX2-FCP-NEXT:    vmovaps 592(%rdi), %xmm13017; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm1[0],xmm13[0]3018; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]3019; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3020; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm3[1]3021; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm153022; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm15[1],ymm10[3],ymm15[3]3023; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]3024; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]3025; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3026; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]3027; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %ymm103028; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]3029; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]3030; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]3031; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3032; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm13[1]3033; AVX2-FCP-NEXT:    vmovaps 736(%rdi), %ymm43034; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm4[1],ymm8[3],ymm4[3]3035; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]3036; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm3[4,5,6,7]3037; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill3038; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm12[1]3039; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm53040; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm9[1],ymm5[1],ymm9[3],ymm5[3]3041; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]3042; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]3043; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3044; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3045; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]3046; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]3047; AVX2-FCP-NEXT:    vmovaps 464(%rdi), %xmm73048; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %xmm63049; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm7[0]3050; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm1[4,5,6,7]3051; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3052; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3053; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]3054; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm1[0,1,0,3]3055; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm33056; AVX2-FCP-NEXT:    vmovaps 80(%rdi), %xmm23057; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm3[0],xmm2[0]3058; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm8[4,5,6,7]3059; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3060; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]3061; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]3062; AVX2-FCP-NEXT:    vmovaps 272(%rdi), %xmm123063; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %xmm13064; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm12[0]3065; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm8[4,5,6,7]3066; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]3067; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]3068; AVX2-FCP-NEXT:    vmovaps 656(%rdi), %xmm93069; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %xmm03070; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm9[0]3071; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]3072; AVX2-FCP-NEXT:    vbroadcastsd 712(%rdi), %ymm143073; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]3074; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]3075; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm4[4,5,6,7]3076; AVX2-FCP-NEXT:    vbroadcastsd 520(%rdi), %ymm03077; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]3078; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm7[1]3079; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]3080; AVX2-FCP-NEXT:    vbroadcastsd 136(%rdi), %ymm03081; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm15[1],ymm0[3],ymm15[3]3082; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]3083; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]3084; AVX2-FCP-NEXT:    vbroadcastsd 328(%rdi), %ymm23085; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]3086; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm12[1]3087; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]3088; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3089; AVX2-FCP-NEXT:    vmovaps %ymm2, 96(%rsi)3090; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3091; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rsi)3092; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3093; AVX2-FCP-NEXT:    vmovaps %ymm2, 64(%rsi)3094; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3095; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rsi)3096; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3097; AVX2-FCP-NEXT:    vmovaps %ymm2, 96(%rdx)3098; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3099; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rdx)3100; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3101; AVX2-FCP-NEXT:    vmovaps %ymm2, 64(%rdx)3102; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3103; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rdx)3104; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3105; AVX2-FCP-NEXT:    vmovaps %ymm2, 96(%rcx)3106; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3107; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rcx)3108; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3109; AVX2-FCP-NEXT:    vmovaps %ymm2, 64(%rcx)3110; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3111; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rcx)3112; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3113; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%r8)3114; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm2 # 32-byte Reload3115; AVX2-FCP-NEXT:    vmovaps %ymm2, 96(%r8)3116; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3117; AVX2-FCP-NEXT:    vmovaps %ymm2, 64(%r8)3118; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3119; AVX2-FCP-NEXT:    vmovaps %ymm2, (%r8)3120; AVX2-FCP-NEXT:    vmovaps %ymm8, 96(%r9)3121; AVX2-FCP-NEXT:    vmovaps %ymm11, 32(%r9)3122; AVX2-FCP-NEXT:    vmovaps %ymm13, (%r9)3123; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload3124; AVX2-FCP-NEXT:    vmovaps %ymm2, 64(%r9)3125; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3126; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rax)3127; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)3128; AVX2-FCP-NEXT:    vmovaps %ymm4, 64(%rax)3129; AVX2-FCP-NEXT:    vmovaps %ymm9, 96(%rax)3130; AVX2-FCP-NEXT:    addq $488, %rsp # imm = 0x1E83131; AVX2-FCP-NEXT:    vzeroupper3132; AVX2-FCP-NEXT:    retq3133;3134; AVX512-LABEL: load_i64_stride6_vf16:3135; AVX512:       # %bb.0:3136; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax3137; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm23138; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm13139; AVX512-NEXT:    vmovdqa64 512(%rdi), %zmm33140; AVX512-NEXT:    vmovdqa64 576(%rdi), %zmm53141; AVX512-NEXT:    vmovdqa64 704(%rdi), %zmm03142; AVX512-NEXT:    vmovdqa64 640(%rdi), %zmm43143; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm83144; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm123145; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm93146; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm113147; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm103148; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm133149; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]3150; AVX512-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]3151; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm153152; AVX512-NEXT:    vpermt2q %zmm10, %zmm14, %zmm153153; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]3154; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm73155; AVX512-NEXT:    vpermt2q %zmm11, %zmm6, %zmm73156; AVX512-NEXT:    movb $56, %dil3157; AVX512-NEXT:    kmovw %edi, %k13158; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}3159; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]3160; AVX512-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3161; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm153162; AVX512-NEXT:    vpermt2q %zmm8, %zmm16, %zmm153163; AVX512-NEXT:    movb $-64, %dil3164; AVX512-NEXT:    kmovw %edi, %k23165; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}3166; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm153167; AVX512-NEXT:    vpermt2q %zmm0, %zmm16, %zmm153168; AVX512-NEXT:    vpermi2q %zmm3, %zmm5, %zmm143169; AVX512-NEXT:    vpermi2q %zmm2, %zmm1, %zmm63170; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}3171; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}3172; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]3173; AVX512-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]3174; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm173175; AVX512-NEXT:    vpermt2q %zmm10, %zmm18, %zmm173176; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]3177; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm153178; AVX512-NEXT:    vpermt2q %zmm11, %zmm14, %zmm153179; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}3180; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]3181; AVX512-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3182; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm193183; AVX512-NEXT:    vpermt2q %zmm8, %zmm17, %zmm193184; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}3185; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm193186; AVX512-NEXT:    vpermt2q %zmm0, %zmm17, %zmm193187; AVX512-NEXT:    vpermi2q %zmm3, %zmm5, %zmm183188; AVX512-NEXT:    vpermi2q %zmm2, %zmm1, %zmm143189; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}3190; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}3191; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]3192; AVX512-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3193; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm213194; AVX512-NEXT:    vpermt2q %zmm13, %zmm20, %zmm213195; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]3196; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm193197; AVX512-NEXT:    vpermt2q %zmm9, %zmm18, %zmm193198; AVX512-NEXT:    movb $24, %dil3199; AVX512-NEXT:    kmovw %edi, %k23200; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}3201; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]3202; AVX512-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]3203; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm223204; AVX512-NEXT:    vpermt2q %zmm8, %zmm21, %zmm223205; AVX512-NEXT:    movb $-32, %dil3206; AVX512-NEXT:    kmovw %edi, %k13207; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}3208; AVX512-NEXT:    vpermi2q %zmm0, %zmm4, %zmm213209; AVX512-NEXT:    vpermi2q %zmm5, %zmm3, %zmm203210; AVX512-NEXT:    vpermi2q %zmm1, %zmm2, %zmm183211; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}3212; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}3213; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]3214; AVX512-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3215; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm233216; AVX512-NEXT:    vpermt2q %zmm13, %zmm22, %zmm233217; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]3218; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm213219; AVX512-NEXT:    vpermt2q %zmm9, %zmm20, %zmm213220; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}3221; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]3222; AVX512-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]3223; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm243224; AVX512-NEXT:    vpermt2q %zmm8, %zmm23, %zmm243225; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}3226; AVX512-NEXT:    vpermi2q %zmm0, %zmm4, %zmm233227; AVX512-NEXT:    vpermi2q %zmm5, %zmm3, %zmm223228; AVX512-NEXT:    vpermi2q %zmm1, %zmm2, %zmm203229; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}3230; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}3231; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]3232; AVX512-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3233; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm233234; AVX512-NEXT:    vpermt2q %zmm12, %zmm22, %zmm233235; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]3236; AVX512-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]3237; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm253238; AVX512-NEXT:    vpermt2q %zmm13, %zmm24, %zmm253239; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm263240; AVX512-NEXT:    vpermt2q %zmm11, %zmm16, %zmm263241; AVX512-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm253242; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}3243; AVX512-NEXT:    vpermi2q %zmm5, %zmm3, %zmm243244; AVX512-NEXT:    vpermi2q %zmm2, %zmm1, %zmm163245; AVX512-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm163246; AVX512-NEXT:    vpermi2q %zmm4, %zmm0, %zmm223247; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}3248; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]3249; AVX512-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3250; AVX512-NEXT:    vpermt2q %zmm12, %zmm22, %zmm83251; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]3252; AVX512-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]3253; AVX512-NEXT:    vpermt2q %zmm13, %zmm12, %zmm103254; AVX512-NEXT:    vpermt2q %zmm11, %zmm17, %zmm93255; AVX512-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm93256; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}3257; AVX512-NEXT:    vpermt2q %zmm5, %zmm12, %zmm33258; AVX512-NEXT:    vpermt2q %zmm2, %zmm17, %zmm13259; AVX512-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm13260; AVX512-NEXT:    vpermt2q %zmm4, %zmm22, %zmm03261; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}3262; AVX512-NEXT:    vmovdqa64 %zmm6, 64(%rsi)3263; AVX512-NEXT:    vmovdqa64 %zmm7, (%rsi)3264; AVX512-NEXT:    vmovdqa64 %zmm14, 64(%rdx)3265; AVX512-NEXT:    vmovdqa64 %zmm15, (%rdx)3266; AVX512-NEXT:    vmovdqa64 %zmm18, 64(%rcx)3267; AVX512-NEXT:    vmovdqa64 %zmm19, (%rcx)3268; AVX512-NEXT:    vmovdqa64 %zmm20, 64(%r8)3269; AVX512-NEXT:    vmovdqa64 %zmm21, (%r8)3270; AVX512-NEXT:    vmovdqa64 %zmm16, 64(%r9)3271; AVX512-NEXT:    vmovdqa64 %zmm25, (%r9)3272; AVX512-NEXT:    vmovdqa64 %zmm1, 64(%rax)3273; AVX512-NEXT:    vmovdqa64 %zmm9, (%rax)3274; AVX512-NEXT:    vzeroupper3275; AVX512-NEXT:    retq3276;3277; AVX512-FCP-LABEL: load_i64_stride6_vf16:3278; AVX512-FCP:       # %bb.0:3279; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3280; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm23281; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm13282; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm33283; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm53284; AVX512-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm03285; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm43286; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm83287; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm123288; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm93289; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm113290; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm103291; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm133292; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]3293; AVX512-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]3294; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm153295; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm14, %zmm153296; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]3297; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm73298; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm6, %zmm73299; AVX512-FCP-NEXT:    movb $56, %dil3300; AVX512-FCP-NEXT:    kmovw %edi, %k13301; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}3302; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]3303; AVX512-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3304; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm153305; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm16, %zmm153306; AVX512-FCP-NEXT:    movb $-64, %dil3307; AVX512-FCP-NEXT:    kmovw %edi, %k23308; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}3309; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm153310; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm153311; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm143312; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm63313; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}3314; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}3315; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]3316; AVX512-FCP-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]3317; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm173318; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm18, %zmm173319; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]3320; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm153321; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm14, %zmm153322; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}3323; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]3324; AVX512-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3325; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm193326; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm17, %zmm193327; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}3328; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm193329; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm17, %zmm193330; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm183331; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm143332; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}3333; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}3334; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]3335; AVX512-FCP-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3336; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm213337; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm20, %zmm213338; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]3339; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm193340; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm18, %zmm193341; AVX512-FCP-NEXT:    movb $24, %dil3342; AVX512-FCP-NEXT:    kmovw %edi, %k23343; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}3344; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]3345; AVX512-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]3346; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm223347; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm21, %zmm223348; AVX512-FCP-NEXT:    movb $-32, %dil3349; AVX512-FCP-NEXT:    kmovw %edi, %k13350; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}3351; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm213352; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm203353; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm183354; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}3355; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}3356; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]3357; AVX512-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3358; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm233359; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm22, %zmm233360; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]3361; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm213362; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm20, %zmm213363; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}3364; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]3365; AVX512-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]3366; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm243367; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm23, %zmm243368; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}3369; AVX512-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm233370; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm223371; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm203372; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}3373; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}3374; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]3375; AVX512-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3376; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm233377; AVX512-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm233378; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]3379; AVX512-FCP-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]3380; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm253381; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm24, %zmm253382; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm263383; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm16, %zmm263384; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm253385; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}3386; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm243387; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm163388; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm163389; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm0, %zmm223390; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}3391; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]3392; AVX512-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3393; AVX512-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm83394; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]3395; AVX512-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]3396; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm12, %zmm103397; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm17, %zmm93398; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm93399; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}3400; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm12, %zmm33401; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm13402; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm13403; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm22, %zmm03404; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}3405; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rsi)3406; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, (%rsi)3407; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 64(%rdx)3408; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, (%rdx)3409; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rcx)3410; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, (%rcx)3411; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 64(%r8)3412; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, (%r8)3413; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, 64(%r9)3414; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, (%r9)3415; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)3416; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, (%rax)3417; AVX512-FCP-NEXT:    vzeroupper3418; AVX512-FCP-NEXT:    retq3419;3420; AVX512DQ-LABEL: load_i64_stride6_vf16:3421; AVX512DQ:       # %bb.0:3422; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax3423; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm23424; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm13425; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %zmm33426; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %zmm53427; AVX512DQ-NEXT:    vmovdqa64 704(%rdi), %zmm03428; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %zmm43429; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm83430; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm123431; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm93432; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm113433; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm103434; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm133435; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]3436; AVX512DQ-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]3437; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm153438; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm14, %zmm153439; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]3440; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm73441; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm6, %zmm73442; AVX512DQ-NEXT:    movb $56, %dil3443; AVX512DQ-NEXT:    kmovw %edi, %k13444; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}3445; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]3446; AVX512DQ-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3447; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm153448; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm16, %zmm153449; AVX512DQ-NEXT:    movb $-64, %dil3450; AVX512DQ-NEXT:    kmovw %edi, %k23451; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}3452; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm153453; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm16, %zmm153454; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm5, %zmm143455; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm1, %zmm63456; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}3457; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}3458; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]3459; AVX512DQ-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]3460; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm173461; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm18, %zmm173462; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]3463; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm153464; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm14, %zmm153465; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}3466; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]3467; AVX512DQ-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3468; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm193469; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm17, %zmm193470; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}3471; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm193472; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm17, %zmm193473; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm5, %zmm183474; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm1, %zmm143475; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}3476; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}3477; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]3478; AVX512DQ-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3479; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm213480; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm20, %zmm213481; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]3482; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm193483; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm18, %zmm193484; AVX512DQ-NEXT:    movb $24, %dil3485; AVX512DQ-NEXT:    kmovw %edi, %k23486; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}3487; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]3488; AVX512DQ-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]3489; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm223490; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm21, %zmm223491; AVX512DQ-NEXT:    movb $-32, %dil3492; AVX512DQ-NEXT:    kmovw %edi, %k13493; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}3494; AVX512DQ-NEXT:    vpermi2q %zmm0, %zmm4, %zmm213495; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm3, %zmm203496; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm2, %zmm183497; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}3498; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}3499; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]3500; AVX512DQ-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3501; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm233502; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm22, %zmm233503; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]3504; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm213505; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm20, %zmm213506; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}3507; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]3508; AVX512DQ-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]3509; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm243510; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm23, %zmm243511; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}3512; AVX512DQ-NEXT:    vpermi2q %zmm0, %zmm4, %zmm233513; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm3, %zmm223514; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm2, %zmm203515; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}3516; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}3517; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]3518; AVX512DQ-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3519; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm233520; AVX512DQ-NEXT:    vpermt2q %zmm12, %zmm22, %zmm233521; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]3522; AVX512DQ-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]3523; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm253524; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm24, %zmm253525; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm263526; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm16, %zmm263527; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm253528; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}3529; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm3, %zmm243530; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm1, %zmm163531; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm163532; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm0, %zmm223533; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}3534; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]3535; AVX512DQ-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3536; AVX512DQ-NEXT:    vpermt2q %zmm12, %zmm22, %zmm83537; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]3538; AVX512DQ-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]3539; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm12, %zmm103540; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm17, %zmm93541; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm93542; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}3543; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm12, %zmm33544; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm17, %zmm13545; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm13546; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm22, %zmm03547; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}3548; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 64(%rsi)3549; AVX512DQ-NEXT:    vmovdqa64 %zmm7, (%rsi)3550; AVX512DQ-NEXT:    vmovdqa64 %zmm14, 64(%rdx)3551; AVX512DQ-NEXT:    vmovdqa64 %zmm15, (%rdx)3552; AVX512DQ-NEXT:    vmovdqa64 %zmm18, 64(%rcx)3553; AVX512DQ-NEXT:    vmovdqa64 %zmm19, (%rcx)3554; AVX512DQ-NEXT:    vmovdqa64 %zmm20, 64(%r8)3555; AVX512DQ-NEXT:    vmovdqa64 %zmm21, (%r8)3556; AVX512DQ-NEXT:    vmovdqa64 %zmm16, 64(%r9)3557; AVX512DQ-NEXT:    vmovdqa64 %zmm25, (%r9)3558; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 64(%rax)3559; AVX512DQ-NEXT:    vmovdqa64 %zmm9, (%rax)3560; AVX512DQ-NEXT:    vzeroupper3561; AVX512DQ-NEXT:    retq3562;3563; AVX512DQ-FCP-LABEL: load_i64_stride6_vf16:3564; AVX512DQ-FCP:       # %bb.0:3565; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3566; AVX512DQ-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm23567; AVX512DQ-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm13568; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm33569; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm53570; AVX512DQ-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm03571; AVX512DQ-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm43572; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm83573; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm123574; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm93575; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm113576; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm103577; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm133578; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]3579; AVX512DQ-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]3580; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm153581; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm14, %zmm153582; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]3583; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm73584; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm6, %zmm73585; AVX512DQ-FCP-NEXT:    movb $56, %dil3586; AVX512DQ-FCP-NEXT:    kmovw %edi, %k13587; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}3588; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]3589; AVX512DQ-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3590; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm153591; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm16, %zmm153592; AVX512DQ-FCP-NEXT:    movb $-64, %dil3593; AVX512DQ-FCP-NEXT:    kmovw %edi, %k23594; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}3595; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm153596; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm153597; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm143598; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm63599; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}3600; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}3601; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]3602; AVX512DQ-FCP-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]3603; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm173604; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm18, %zmm173605; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]3606; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm153607; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm14, %zmm153608; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}3609; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]3610; AVX512DQ-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3611; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm193612; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm17, %zmm193613; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}3614; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm193615; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm17, %zmm193616; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm183617; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm143618; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}3619; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}3620; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]3621; AVX512DQ-FCP-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3622; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm213623; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm20, %zmm213624; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]3625; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm193626; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm18, %zmm193627; AVX512DQ-FCP-NEXT:    movb $24, %dil3628; AVX512DQ-FCP-NEXT:    kmovw %edi, %k23629; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}3630; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]3631; AVX512DQ-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]3632; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm223633; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm21, %zmm223634; AVX512DQ-FCP-NEXT:    movb $-32, %dil3635; AVX512DQ-FCP-NEXT:    kmovw %edi, %k13636; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}3637; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm213638; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm203639; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm183640; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}3641; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}3642; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]3643; AVX512DQ-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3644; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm233645; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm22, %zmm233646; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]3647; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm213648; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm20, %zmm213649; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}3650; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]3651; AVX512DQ-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]3652; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm243653; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm23, %zmm243654; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}3655; AVX512DQ-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm233656; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm223657; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm203658; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}3659; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}3660; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]3661; AVX512DQ-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3662; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm233663; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm233664; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]3665; AVX512DQ-FCP-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]3666; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm253667; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm24, %zmm253668; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm263669; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm16, %zmm263670; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm253671; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}3672; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm243673; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm163674; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm163675; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm0, %zmm223676; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}3677; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]3678; AVX512DQ-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3679; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm83680; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]3681; AVX512DQ-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]3682; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm12, %zmm103683; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm17, %zmm93684; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm93685; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}3686; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm12, %zmm33687; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm13688; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm13689; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm22, %zmm03690; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}3691; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rsi)3692; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, (%rsi)3693; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 64(%rdx)3694; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, (%rdx)3695; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rcx)3696; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, (%rcx)3697; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 64(%r8)3698; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, (%r8)3699; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, 64(%r9)3700; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, (%r9)3701; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)3702; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, (%rax)3703; AVX512DQ-FCP-NEXT:    vzeroupper3704; AVX512DQ-FCP-NEXT:    retq3705;3706; AVX512BW-LABEL: load_i64_stride6_vf16:3707; AVX512BW:       # %bb.0:3708; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax3709; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm23710; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm13711; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm33712; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm53713; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm03714; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm43715; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm83716; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm123717; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm93718; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm113719; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm103720; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm133721; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]3722; AVX512BW-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]3723; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm153724; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm14, %zmm153725; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]3726; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm73727; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm6, %zmm73728; AVX512BW-NEXT:    movb $56, %dil3729; AVX512BW-NEXT:    kmovd %edi, %k13730; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}3731; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]3732; AVX512BW-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3733; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm153734; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm16, %zmm153735; AVX512BW-NEXT:    movb $-64, %dil3736; AVX512BW-NEXT:    kmovd %edi, %k23737; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}3738; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm153739; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm16, %zmm153740; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm5, %zmm143741; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm63742; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}3743; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}3744; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]3745; AVX512BW-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]3746; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm173747; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm18, %zmm173748; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]3749; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm153750; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm14, %zmm153751; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}3752; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]3753; AVX512BW-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3754; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm193755; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm17, %zmm193756; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}3757; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm193758; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm17, %zmm193759; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm5, %zmm183760; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm143761; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}3762; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}3763; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]3764; AVX512BW-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3765; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm213766; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm20, %zmm213767; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]3768; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm193769; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm18, %zmm193770; AVX512BW-NEXT:    movb $24, %dil3771; AVX512BW-NEXT:    kmovd %edi, %k23772; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}3773; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]3774; AVX512BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]3775; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm223776; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm21, %zmm223777; AVX512BW-NEXT:    movb $-32, %dil3778; AVX512BW-NEXT:    kmovd %edi, %k13779; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}3780; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm4, %zmm213781; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm203782; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm2, %zmm183783; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}3784; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}3785; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]3786; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3787; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm233788; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm22, %zmm233789; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]3790; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm213791; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm20, %zmm213792; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}3793; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]3794; AVX512BW-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]3795; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm243796; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm23, %zmm243797; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}3798; AVX512BW-NEXT:    vpermi2q %zmm0, %zmm4, %zmm233799; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm223800; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm2, %zmm203801; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}3802; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}3803; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]3804; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3805; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm233806; AVX512BW-NEXT:    vpermt2q %zmm12, %zmm22, %zmm233807; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]3808; AVX512BW-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]3809; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm253810; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm24, %zmm253811; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm263812; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm16, %zmm263813; AVX512BW-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm253814; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}3815; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm243816; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm163817; AVX512BW-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm163818; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm0, %zmm223819; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}3820; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]3821; AVX512BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3822; AVX512BW-NEXT:    vpermt2q %zmm12, %zmm22, %zmm83823; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]3824; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]3825; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm12, %zmm103826; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm17, %zmm93827; AVX512BW-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm93828; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}3829; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm12, %zmm33830; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm17, %zmm13831; AVX512BW-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm13832; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm22, %zmm03833; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}3834; AVX512BW-NEXT:    vmovdqa64 %zmm6, 64(%rsi)3835; AVX512BW-NEXT:    vmovdqa64 %zmm7, (%rsi)3836; AVX512BW-NEXT:    vmovdqa64 %zmm14, 64(%rdx)3837; AVX512BW-NEXT:    vmovdqa64 %zmm15, (%rdx)3838; AVX512BW-NEXT:    vmovdqa64 %zmm18, 64(%rcx)3839; AVX512BW-NEXT:    vmovdqa64 %zmm19, (%rcx)3840; AVX512BW-NEXT:    vmovdqa64 %zmm20, 64(%r8)3841; AVX512BW-NEXT:    vmovdqa64 %zmm21, (%r8)3842; AVX512BW-NEXT:    vmovdqa64 %zmm16, 64(%r9)3843; AVX512BW-NEXT:    vmovdqa64 %zmm25, (%r9)3844; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%rax)3845; AVX512BW-NEXT:    vmovdqa64 %zmm9, (%rax)3846; AVX512BW-NEXT:    vzeroupper3847; AVX512BW-NEXT:    retq3848;3849; AVX512BW-FCP-LABEL: load_i64_stride6_vf16:3850; AVX512BW-FCP:       # %bb.0:3851; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3852; AVX512BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm23853; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm13854; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm33855; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm53856; AVX512BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm03857; AVX512BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm43858; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm83859; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm123860; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm93861; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm113862; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm103863; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm133864; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]3865; AVX512BW-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]3866; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm153867; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm14, %zmm153868; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]3869; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm73870; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm6, %zmm73871; AVX512BW-FCP-NEXT:    movb $56, %dil3872; AVX512BW-FCP-NEXT:    kmovd %edi, %k13873; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}3874; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]3875; AVX512BW-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3876; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm153877; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm16, %zmm153878; AVX512BW-FCP-NEXT:    movb $-64, %dil3879; AVX512BW-FCP-NEXT:    kmovd %edi, %k23880; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}3881; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm153882; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm153883; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm143884; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm63885; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}3886; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}3887; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]3888; AVX512BW-FCP-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]3889; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm173890; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm18, %zmm173891; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]3892; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm153893; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm14, %zmm153894; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}3895; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]3896; AVX512BW-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3897; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm193898; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm17, %zmm193899; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}3900; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm193901; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm17, %zmm193902; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm183903; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm143904; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}3905; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}3906; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]3907; AVX512BW-FCP-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3908; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm213909; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm20, %zmm213910; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]3911; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm193912; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm18, %zmm193913; AVX512BW-FCP-NEXT:    movb $24, %dil3914; AVX512BW-FCP-NEXT:    kmovd %edi, %k23915; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}3916; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]3917; AVX512BW-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]3918; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm223919; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm21, %zmm223920; AVX512BW-FCP-NEXT:    movb $-32, %dil3921; AVX512BW-FCP-NEXT:    kmovd %edi, %k13922; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}3923; AVX512BW-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm213924; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm203925; AVX512BW-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm183926; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}3927; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}3928; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]3929; AVX512BW-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3930; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm233931; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm22, %zmm233932; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]3933; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm213934; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm20, %zmm213935; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}3936; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]3937; AVX512BW-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]3938; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm243939; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm23, %zmm243940; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}3941; AVX512BW-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm233942; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm223943; AVX512BW-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm203944; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}3945; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}3946; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]3947; AVX512BW-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3948; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm233949; AVX512BW-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm233950; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]3951; AVX512BW-FCP-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]3952; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm253953; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm24, %zmm253954; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm263955; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm16, %zmm263956; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm253957; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}3958; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm243959; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm163960; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm163961; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm0, %zmm223962; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}3963; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]3964; AVX512BW-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]3965; AVX512BW-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm83966; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]3967; AVX512BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]3968; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm12, %zmm103969; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm17, %zmm93970; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm93971; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}3972; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm12, %zmm33973; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm13974; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm13975; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm22, %zmm03976; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}3977; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rsi)3978; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, (%rsi)3979; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, 64(%rdx)3980; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, (%rdx)3981; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rcx)3982; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, (%rcx)3983; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, 64(%r8)3984; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, (%r8)3985; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, 64(%r9)3986; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, (%r9)3987; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)3988; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, (%rax)3989; AVX512BW-FCP-NEXT:    vzeroupper3990; AVX512BW-FCP-NEXT:    retq3991;3992; AVX512DQ-BW-LABEL: load_i64_stride6_vf16:3993; AVX512DQ-BW:       # %bb.0:3994; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax3995; AVX512DQ-BW-NEXT:    vmovdqa64 448(%rdi), %zmm23996; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm13997; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %zmm33998; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %zmm53999; AVX512DQ-BW-NEXT:    vmovdqa64 704(%rdi), %zmm04000; AVX512DQ-BW-NEXT:    vmovdqa64 640(%rdi), %zmm44001; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm84002; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm124003; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm94004; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm114005; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm104006; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm134007; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]4008; AVX512DQ-BW-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]4009; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm154010; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm14, %zmm154011; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]4012; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm74013; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm6, %zmm74014; AVX512DQ-BW-NEXT:    movb $56, %dil4015; AVX512DQ-BW-NEXT:    kmovd %edi, %k14016; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}4017; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]4018; AVX512DQ-BW-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4019; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm154020; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm16, %zmm154021; AVX512DQ-BW-NEXT:    movb $-64, %dil4022; AVX512DQ-BW-NEXT:    kmovd %edi, %k24023; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}4024; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm154025; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm16, %zmm154026; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm5, %zmm144027; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm64028; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}4029; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}4030; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]4031; AVX512DQ-BW-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]4032; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm174033; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm18, %zmm174034; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]4035; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm154036; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm14, %zmm154037; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}4038; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]4039; AVX512DQ-BW-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4040; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm194041; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm17, %zmm194042; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}4043; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm194044; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm17, %zmm194045; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm5, %zmm184046; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm144047; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}4048; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}4049; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]4050; AVX512DQ-BW-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4051; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm214052; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm20, %zmm214053; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]4054; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm194055; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm18, %zmm194056; AVX512DQ-BW-NEXT:    movb $24, %dil4057; AVX512DQ-BW-NEXT:    kmovd %edi, %k24058; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}4059; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]4060; AVX512DQ-BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]4061; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm224062; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm21, %zmm224063; AVX512DQ-BW-NEXT:    movb $-32, %dil4064; AVX512DQ-BW-NEXT:    kmovd %edi, %k14065; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}4066; AVX512DQ-BW-NEXT:    vpermi2q %zmm0, %zmm4, %zmm214067; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm204068; AVX512DQ-BW-NEXT:    vpermi2q %zmm1, %zmm2, %zmm184069; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}4070; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}4071; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]4072; AVX512DQ-BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4073; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm234074; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm22, %zmm234075; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]4076; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm214077; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm20, %zmm214078; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}4079; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]4080; AVX512DQ-BW-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]4081; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm244082; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm23, %zmm244083; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}4084; AVX512DQ-BW-NEXT:    vpermi2q %zmm0, %zmm4, %zmm234085; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm224086; AVX512DQ-BW-NEXT:    vpermi2q %zmm1, %zmm2, %zmm204087; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}4088; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}4089; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]4090; AVX512DQ-BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]4091; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm234092; AVX512DQ-BW-NEXT:    vpermt2q %zmm12, %zmm22, %zmm234093; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]4094; AVX512DQ-BW-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]4095; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm254096; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm24, %zmm254097; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm264098; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm16, %zmm264099; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm254100; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}4101; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm3, %zmm244102; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm1, %zmm164103; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm164104; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm0, %zmm224105; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}4106; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]4107; AVX512DQ-BW-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]4108; AVX512DQ-BW-NEXT:    vpermt2q %zmm12, %zmm22, %zmm84109; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]4110; AVX512DQ-BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]4111; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm12, %zmm104112; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm17, %zmm94113; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm94114; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}4115; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm12, %zmm34116; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm17, %zmm14117; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm14118; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm22, %zmm04119; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}4120; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, 64(%rsi)4121; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, (%rsi)4122; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, 64(%rdx)4123; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, (%rdx)4124; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, 64(%rcx)4125; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, (%rcx)4126; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, 64(%r8)4127; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, (%r8)4128; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, 64(%r9)4129; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, (%r9)4130; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 64(%rax)4131; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, (%rax)4132; AVX512DQ-BW-NEXT:    vzeroupper4133; AVX512DQ-BW-NEXT:    retq4134;4135; AVX512DQ-BW-FCP-LABEL: load_i64_stride6_vf16:4136; AVX512DQ-BW-FCP:       # %bb.0:4137; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax4138; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm24139; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm14140; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm34141; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm54142; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm04143; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm44144; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm84145; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm124146; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm94147; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm114148; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm104149; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm134150; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,0,10,0,6,0,10]4151; AVX512DQ-BW-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3]4152; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm154153; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm14, %zmm154154; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm6 = [0,6,12,0]4155; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm74156; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm6, %zmm74157; AVX512DQ-BW-FCP-NEXT:    movb $56, %dil4158; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k14159; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k1}4160; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [4,10,4,10,4,10,4,10]4161; AVX512DQ-BW-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4162; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm154163; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm16, %zmm154164; AVX512DQ-BW-FCP-NEXT:    movb $-64, %dil4165; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k24166; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm7 {%k2}4167; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm154168; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm16, %zmm154169; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm144170; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm64171; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm6 {%k1}4172; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm6 {%k2}4173; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm18 = [1,7,0,11,1,7,0,11]4174; AVX512DQ-BW-FCP-NEXT:    # zmm18 = mem[0,1,2,3,0,1,2,3]4175; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm174176; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm18, %zmm174177; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm14 = [1,7,13,0]4178; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm154179; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm14, %zmm154180; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm15 {%k1}4181; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [5,11,5,11,5,11,5,11]4182; AVX512DQ-BW-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4183; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm194184; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm17, %zmm194185; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k2}4186; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm194187; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm17, %zmm194188; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm5, %zmm184189; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm144190; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm14 {%k1}4191; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k2}4192; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm20 = [10,4,10,4,10,4,10,4]4193; AVX512DQ-BW-FCP-NEXT:    # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4194; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm214195; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm20, %zmm214196; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm18 = [10,0,6,0]4197; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm194198; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm18, %zmm194199; AVX512DQ-BW-FCP-NEXT:    movb $24, %dil4200; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k24201; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm19 {%k2}4202; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,6,12,0,0,6,12]4203; AVX512DQ-BW-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3]4204; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm224205; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm21, %zmm224206; AVX512DQ-BW-FCP-NEXT:    movb $-32, %dil4207; AVX512DQ-BW-FCP-NEXT:    kmovd %edi, %k14208; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm19 {%k1}4209; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm214210; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm204211; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm184212; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}4213; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm18 {%k1}4214; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm22 = [11,5,11,5,11,5,11,5]4215; AVX512DQ-BW-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4216; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm234217; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm22, %zmm234218; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm20 = [11,1,7,0]4219; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm214220; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm20, %zmm214221; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm21 {%k2}4222; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm23 = [0,1,7,13,0,1,7,13]4223; AVX512DQ-BW-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3]4224; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm244225; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm23, %zmm244226; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm21 {%k1}4227; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm0, %zmm4, %zmm234228; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm224229; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm1, %zmm2, %zmm204230; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm20 {%k2}4231; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm20 {%k1}4232; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,10,0,6,0,10,0,6]4233; AVX512DQ-BW-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]4234; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm234235; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm234236; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm24 = [12,0,0,6,12,0,0,6]4237; AVX512DQ-BW-FCP-NEXT:    # zmm24 = mem[0,1,2,3,0,1,2,3]4238; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm254239; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm24, %zmm254240; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm264241; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm16, %zmm264242; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm26, %zmm25, %zmm254243; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm25 {%k1}4244; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm3, %zmm244245; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm1, %zmm164246; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm16, %zmm24, %zmm164247; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm0, %zmm224248; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm16 {%k1}4249; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm22 = [0,11,1,7,0,11,1,7]4250; AVX512DQ-BW-FCP-NEXT:    # zmm22 = mem[0,1,2,3,0,1,2,3]4251; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm12, %zmm22, %zmm84252; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm12 = [13,0,1,7,13,0,1,7]4253; AVX512DQ-BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3]4254; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm12, %zmm104255; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm17, %zmm94256; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm9, %zmm10, %zmm94257; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm9 {%k1}4258; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm12, %zmm34259; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm17, %zmm14260; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm1, %zmm3, %zmm14261; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm22, %zmm04262; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}4263; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rsi)4264; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, (%rsi)4265; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, 64(%rdx)4266; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, (%rdx)4267; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rcx)4268; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, (%rcx)4269; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, 64(%r8)4270; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, (%r8)4271; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, 64(%r9)4272; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, (%r9)4273; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rax)4274; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, (%rax)4275; AVX512DQ-BW-FCP-NEXT:    vzeroupper4276; AVX512DQ-BW-FCP-NEXT:    retq4277  %wide.vec = load <96 x i64>, ptr %in.vec, align 644278  %strided.vec0 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90>4279  %strided.vec1 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91>4280  %strided.vec2 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92>4281  %strided.vec3 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93>4282  %strided.vec4 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94>4283  %strided.vec5 = shufflevector <96 x i64> %wide.vec, <96 x i64> poison, <16 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95>4284  store <16 x i64> %strided.vec0, ptr %out.vec0, align 644285  store <16 x i64> %strided.vec1, ptr %out.vec1, align 644286  store <16 x i64> %strided.vec2, ptr %out.vec2, align 644287  store <16 x i64> %strided.vec3, ptr %out.vec3, align 644288  store <16 x i64> %strided.vec4, ptr %out.vec4, align 644289  store <16 x i64> %strided.vec5, ptr %out.vec5, align 644290  ret void4291}4292 4293define void @load_i64_stride6_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {4294; SSE-LABEL: load_i64_stride6_vf32:4295; SSE:       # %bb.0:4296; SSE-NEXT:    subq $1176, %rsp # imm = 0x4984297; SSE-NEXT:    movaps 624(%rdi), %xmm04298; SSE-NEXT:    movaps 576(%rdi), %xmm94299; SSE-NEXT:    movaps 240(%rdi), %xmm34300; SSE-NEXT:    movaps 192(%rdi), %xmm84301; SSE-NEXT:    movaps 720(%rdi), %xmm14302; SSE-NEXT:    movaps 672(%rdi), %xmm114303; SSE-NEXT:    movaps 336(%rdi), %xmm54304; SSE-NEXT:    movaps 288(%rdi), %xmm104305; SSE-NEXT:    movaps 432(%rdi), %xmm44306; SSE-NEXT:    movaps 384(%rdi), %xmm124307; SSE-NEXT:    movaps 912(%rdi), %xmm24308; SSE-NEXT:    movaps 528(%rdi), %xmm64309; SSE-NEXT:    movaps 480(%rdi), %xmm144310; SSE-NEXT:    movaps 144(%rdi), %xmm74311; SSE-NEXT:    movaps 96(%rdi), %xmm134312; SSE-NEXT:    movaps %xmm13, %xmm154313; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]4314; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4315; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm7[1]4316; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4317; SSE-NEXT:    movaps %xmm10, %xmm74318; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm5[0]4319; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4320; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm5[1]4321; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4322; SSE-NEXT:    movaps %xmm8, %xmm54323; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]4324; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4325; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm3[1]4326; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4327; SSE-NEXT:    movaps %xmm14, %xmm34328; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm6[0]4329; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4330; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm6[1]4331; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4332; SSE-NEXT:    movaps %xmm12, %xmm34333; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm4[0]4334; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4335; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm4[1]4336; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4337; SSE-NEXT:    movaps %xmm11, %xmm34338; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]4339; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4340; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm1[1]4341; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4342; SSE-NEXT:    movaps %xmm9, %xmm14343; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4344; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4345; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]4346; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4347; SSE-NEXT:    movaps 864(%rdi), %xmm04348; SSE-NEXT:    movaps %xmm0, %xmm14349; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4350; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4351; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]4352; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4353; SSE-NEXT:    movaps 816(%rdi), %xmm04354; SSE-NEXT:    movaps 768(%rdi), %xmm14355; SSE-NEXT:    movaps %xmm1, %xmm24356; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4357; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4358; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4359; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4360; SSE-NEXT:    movaps 1104(%rdi), %xmm04361; SSE-NEXT:    movaps 1056(%rdi), %xmm14362; SSE-NEXT:    movaps %xmm1, %xmm24363; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4364; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4365; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4366; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4367; SSE-NEXT:    movaps 1008(%rdi), %xmm04368; SSE-NEXT:    movaps 960(%rdi), %xmm14369; SSE-NEXT:    movaps %xmm1, %xmm24370; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4371; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4372; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4373; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4374; SSE-NEXT:    movaps 1296(%rdi), %xmm04375; SSE-NEXT:    movaps 1248(%rdi), %xmm14376; SSE-NEXT:    movaps %xmm1, %xmm24377; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4378; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4379; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4380; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4381; SSE-NEXT:    movaps 1200(%rdi), %xmm04382; SSE-NEXT:    movaps 1152(%rdi), %xmm14383; SSE-NEXT:    movaps %xmm1, %xmm24384; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4385; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4386; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4387; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4388; SSE-NEXT:    movaps 1488(%rdi), %xmm04389; SSE-NEXT:    movaps 1440(%rdi), %xmm14390; SSE-NEXT:    movaps %xmm1, %xmm24391; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4392; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4393; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4394; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4395; SSE-NEXT:    movaps 1392(%rdi), %xmm04396; SSE-NEXT:    movaps 1344(%rdi), %xmm14397; SSE-NEXT:    movaps %xmm1, %xmm24398; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4399; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4400; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4401; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4402; SSE-NEXT:    movaps (%rdi), %xmm14403; SSE-NEXT:    movaps 48(%rdi), %xmm04404; SSE-NEXT:    movaps %xmm1, %xmm24405; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4406; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4407; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4408; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4409; SSE-NEXT:    movaps 64(%rdi), %xmm04410; SSE-NEXT:    movaps 16(%rdi), %xmm14411; SSE-NEXT:    movaps %xmm1, %xmm24412; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4413; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4414; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4415; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4416; SSE-NEXT:    movaps 160(%rdi), %xmm04417; SSE-NEXT:    movaps 112(%rdi), %xmm14418; SSE-NEXT:    movaps %xmm1, %xmm24419; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4420; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4421; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4422; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4423; SSE-NEXT:    movaps 256(%rdi), %xmm04424; SSE-NEXT:    movaps 208(%rdi), %xmm14425; SSE-NEXT:    movaps %xmm1, %xmm24426; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4427; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4428; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4429; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4430; SSE-NEXT:    movaps 352(%rdi), %xmm04431; SSE-NEXT:    movaps 304(%rdi), %xmm14432; SSE-NEXT:    movaps %xmm1, %xmm24433; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4434; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4435; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4436; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4437; SSE-NEXT:    movaps 448(%rdi), %xmm04438; SSE-NEXT:    movaps 400(%rdi), %xmm14439; SSE-NEXT:    movaps %xmm1, %xmm24440; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4441; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4442; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4443; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4444; SSE-NEXT:    movaps 544(%rdi), %xmm04445; SSE-NEXT:    movaps 496(%rdi), %xmm14446; SSE-NEXT:    movaps %xmm1, %xmm24447; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4448; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4449; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4450; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4451; SSE-NEXT:    movaps 640(%rdi), %xmm04452; SSE-NEXT:    movaps 592(%rdi), %xmm14453; SSE-NEXT:    movaps %xmm1, %xmm24454; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4455; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4456; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4457; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4458; SSE-NEXT:    movaps 736(%rdi), %xmm04459; SSE-NEXT:    movaps 688(%rdi), %xmm14460; SSE-NEXT:    movaps %xmm1, %xmm24461; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4462; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4463; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4464; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4465; SSE-NEXT:    movaps 832(%rdi), %xmm04466; SSE-NEXT:    movaps 784(%rdi), %xmm14467; SSE-NEXT:    movaps %xmm1, %xmm24468; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4469; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4470; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4471; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4472; SSE-NEXT:    movaps 928(%rdi), %xmm04473; SSE-NEXT:    movaps 880(%rdi), %xmm14474; SSE-NEXT:    movaps %xmm1, %xmm24475; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4476; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4477; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4478; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4479; SSE-NEXT:    movaps 1024(%rdi), %xmm04480; SSE-NEXT:    movaps 976(%rdi), %xmm14481; SSE-NEXT:    movaps %xmm1, %xmm24482; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4483; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4484; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4485; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4486; SSE-NEXT:    movaps 1120(%rdi), %xmm04487; SSE-NEXT:    movaps 1072(%rdi), %xmm14488; SSE-NEXT:    movaps %xmm1, %xmm24489; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4490; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4491; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4492; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4493; SSE-NEXT:    movaps 1216(%rdi), %xmm04494; SSE-NEXT:    movaps 1168(%rdi), %xmm14495; SSE-NEXT:    movaps %xmm1, %xmm24496; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4497; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4498; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4499; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4500; SSE-NEXT:    movaps 1312(%rdi), %xmm04501; SSE-NEXT:    movaps 1264(%rdi), %xmm14502; SSE-NEXT:    movaps %xmm1, %xmm24503; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4504; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4505; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4506; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4507; SSE-NEXT:    movaps 1408(%rdi), %xmm04508; SSE-NEXT:    movaps 1360(%rdi), %xmm14509; SSE-NEXT:    movaps %xmm1, %xmm24510; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4511; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4512; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4513; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4514; SSE-NEXT:    movaps 1504(%rdi), %xmm04515; SSE-NEXT:    movaps 1456(%rdi), %xmm14516; SSE-NEXT:    movaps %xmm1, %xmm24517; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4518; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4519; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4520; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4521; SSE-NEXT:    movaps 80(%rdi), %xmm04522; SSE-NEXT:    movaps 32(%rdi), %xmm14523; SSE-NEXT:    movaps %xmm1, %xmm24524; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4525; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4526; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4527; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4528; SSE-NEXT:    movaps 176(%rdi), %xmm04529; SSE-NEXT:    movaps 128(%rdi), %xmm14530; SSE-NEXT:    movaps %xmm1, %xmm24531; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4532; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4533; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4534; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4535; SSE-NEXT:    movaps 272(%rdi), %xmm04536; SSE-NEXT:    movaps 224(%rdi), %xmm14537; SSE-NEXT:    movaps %xmm1, %xmm24538; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4539; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4540; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4541; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4542; SSE-NEXT:    movaps 368(%rdi), %xmm04543; SSE-NEXT:    movaps 320(%rdi), %xmm14544; SSE-NEXT:    movaps %xmm1, %xmm24545; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4546; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4547; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4548; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4549; SSE-NEXT:    movaps 464(%rdi), %xmm04550; SSE-NEXT:    movaps 416(%rdi), %xmm14551; SSE-NEXT:    movaps %xmm1, %xmm24552; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4553; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4554; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4555; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4556; SSE-NEXT:    movaps 560(%rdi), %xmm04557; SSE-NEXT:    movaps 512(%rdi), %xmm14558; SSE-NEXT:    movaps %xmm1, %xmm24559; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4560; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4561; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4562; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill4563; SSE-NEXT:    movaps 656(%rdi), %xmm04564; SSE-NEXT:    movaps 608(%rdi), %xmm14565; SSE-NEXT:    movaps %xmm1, %xmm24566; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4567; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4568; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4569; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4570; SSE-NEXT:    movaps 752(%rdi), %xmm04571; SSE-NEXT:    movaps 704(%rdi), %xmm144572; SSE-NEXT:    movaps %xmm14, %xmm14573; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4574; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4575; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]4576; SSE-NEXT:    movaps 848(%rdi), %xmm04577; SSE-NEXT:    movaps 800(%rdi), %xmm154578; SSE-NEXT:    movaps %xmm15, %xmm14579; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4580; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4581; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]4582; SSE-NEXT:    movaps 944(%rdi), %xmm04583; SSE-NEXT:    movaps 896(%rdi), %xmm134584; SSE-NEXT:    movaps %xmm13, %xmm14585; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4586; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4587; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]4588; SSE-NEXT:    movaps 1040(%rdi), %xmm04589; SSE-NEXT:    movaps 992(%rdi), %xmm104590; SSE-NEXT:    movaps %xmm10, %xmm124591; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm0[0]4592; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]4593; SSE-NEXT:    movaps 1136(%rdi), %xmm04594; SSE-NEXT:    movaps 1088(%rdi), %xmm94595; SSE-NEXT:    movaps %xmm9, %xmm114596; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]4597; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]4598; SSE-NEXT:    movaps 1232(%rdi), %xmm04599; SSE-NEXT:    movaps 1184(%rdi), %xmm74600; SSE-NEXT:    movaps %xmm7, %xmm84601; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]4602; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]4603; SSE-NEXT:    movaps 1328(%rdi), %xmm04604; SSE-NEXT:    movaps 1280(%rdi), %xmm54605; SSE-NEXT:    movaps %xmm5, %xmm64606; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]4607; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]4608; SSE-NEXT:    movaps 1424(%rdi), %xmm04609; SSE-NEXT:    movaps 1376(%rdi), %xmm34610; SSE-NEXT:    movaps %xmm3, %xmm44611; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]4612; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]4613; SSE-NEXT:    movaps 1520(%rdi), %xmm04614; SSE-NEXT:    movaps 1472(%rdi), %xmm14615; SSE-NEXT:    movaps %xmm1, %xmm24616; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4617; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4618; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4619; SSE-NEXT:    movaps %xmm0, 224(%rsi)4620; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4621; SSE-NEXT:    movaps %xmm0, 160(%rsi)4622; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4623; SSE-NEXT:    movaps %xmm0, 96(%rsi)4624; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4625; SSE-NEXT:    movaps %xmm0, 32(%rsi)4626; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4627; SSE-NEXT:    movaps %xmm0, 240(%rsi)4628; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4629; SSE-NEXT:    movaps %xmm0, 176(%rsi)4630; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4631; SSE-NEXT:    movaps %xmm0, 112(%rsi)4632; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4633; SSE-NEXT:    movaps %xmm0, 48(%rsi)4634; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4635; SSE-NEXT:    movaps %xmm0, 192(%rsi)4636; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4637; SSE-NEXT:    movaps %xmm0, 128(%rsi)4638; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4639; SSE-NEXT:    movaps %xmm0, 64(%rsi)4640; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4641; SSE-NEXT:    movaps %xmm0, (%rsi)4642; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4643; SSE-NEXT:    movaps %xmm0, 208(%rsi)4644; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4645; SSE-NEXT:    movaps %xmm0, 144(%rsi)4646; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4647; SSE-NEXT:    movaps %xmm0, 80(%rsi)4648; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4649; SSE-NEXT:    movaps %xmm0, 16(%rsi)4650; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4651; SSE-NEXT:    movaps %xmm0, 224(%rdx)4652; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4653; SSE-NEXT:    movaps %xmm0, 240(%rdx)4654; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4655; SSE-NEXT:    movaps %xmm0, 192(%rdx)4656; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4657; SSE-NEXT:    movaps %xmm0, 208(%rdx)4658; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4659; SSE-NEXT:    movaps %xmm0, 160(%rdx)4660; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4661; SSE-NEXT:    movaps %xmm0, 176(%rdx)4662; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4663; SSE-NEXT:    movaps %xmm0, 128(%rdx)4664; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4665; SSE-NEXT:    movaps %xmm0, 144(%rdx)4666; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4667; SSE-NEXT:    movaps %xmm0, 96(%rdx)4668; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4669; SSE-NEXT:    movaps %xmm0, 112(%rdx)4670; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4671; SSE-NEXT:    movaps %xmm0, 64(%rdx)4672; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4673; SSE-NEXT:    movaps %xmm0, 80(%rdx)4674; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4675; SSE-NEXT:    movaps %xmm0, 32(%rdx)4676; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4677; SSE-NEXT:    movaps %xmm0, 48(%rdx)4678; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4679; SSE-NEXT:    movaps %xmm0, (%rdx)4680; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4681; SSE-NEXT:    movaps %xmm0, 16(%rdx)4682; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4683; SSE-NEXT:    movaps %xmm0, 240(%rcx)4684; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4685; SSE-NEXT:    movaps %xmm0, 224(%rcx)4686; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4687; SSE-NEXT:    movaps %xmm0, 208(%rcx)4688; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4689; SSE-NEXT:    movaps %xmm0, 192(%rcx)4690; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4691; SSE-NEXT:    movaps %xmm0, 176(%rcx)4692; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4693; SSE-NEXT:    movaps %xmm0, 160(%rcx)4694; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4695; SSE-NEXT:    movaps %xmm0, 144(%rcx)4696; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4697; SSE-NEXT:    movaps %xmm0, 128(%rcx)4698; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4699; SSE-NEXT:    movaps %xmm0, 112(%rcx)4700; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4701; SSE-NEXT:    movaps %xmm0, 96(%rcx)4702; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4703; SSE-NEXT:    movaps %xmm0, 80(%rcx)4704; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4705; SSE-NEXT:    movaps %xmm0, 64(%rcx)4706; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4707; SSE-NEXT:    movaps %xmm0, 48(%rcx)4708; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4709; SSE-NEXT:    movaps %xmm0, 32(%rcx)4710; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4711; SSE-NEXT:    movaps %xmm0, 16(%rcx)4712; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4713; SSE-NEXT:    movaps %xmm0, (%rcx)4714; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4715; SSE-NEXT:    movaps %xmm0, 240(%r8)4716; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4717; SSE-NEXT:    movaps %xmm0, 224(%r8)4718; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4719; SSE-NEXT:    movaps %xmm0, 208(%r8)4720; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4721; SSE-NEXT:    movaps %xmm0, 192(%r8)4722; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4723; SSE-NEXT:    movaps %xmm0, 176(%r8)4724; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4725; SSE-NEXT:    movaps %xmm0, 160(%r8)4726; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4727; SSE-NEXT:    movaps %xmm0, 144(%r8)4728; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4729; SSE-NEXT:    movaps %xmm0, 128(%r8)4730; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4731; SSE-NEXT:    movaps %xmm0, 112(%r8)4732; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4733; SSE-NEXT:    movaps %xmm0, 96(%r8)4734; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4735; SSE-NEXT:    movaps %xmm0, 80(%r8)4736; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4737; SSE-NEXT:    movaps %xmm0, 64(%r8)4738; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4739; SSE-NEXT:    movaps %xmm0, 48(%r8)4740; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4741; SSE-NEXT:    movaps %xmm0, 32(%r8)4742; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4743; SSE-NEXT:    movaps %xmm0, 16(%r8)4744; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4745; SSE-NEXT:    movaps %xmm0, (%r8)4746; SSE-NEXT:    movaps %xmm2, 240(%r9)4747; SSE-NEXT:    movaps %xmm4, 224(%r9)4748; SSE-NEXT:    movaps %xmm6, 208(%r9)4749; SSE-NEXT:    movaps %xmm8, 192(%r9)4750; SSE-NEXT:    movaps %xmm11, 176(%r9)4751; SSE-NEXT:    movaps %xmm12, 160(%r9)4752; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4753; SSE-NEXT:    movaps %xmm0, 144(%r9)4754; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4755; SSE-NEXT:    movaps %xmm0, 128(%r9)4756; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4757; SSE-NEXT:    movaps %xmm0, 112(%r9)4758; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4759; SSE-NEXT:    movaps %xmm0, 96(%r9)4760; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4761; SSE-NEXT:    movaps %xmm0, 80(%r9)4762; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4763; SSE-NEXT:    movaps %xmm0, 64(%r9)4764; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4765; SSE-NEXT:    movaps %xmm0, 48(%r9)4766; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4767; SSE-NEXT:    movaps %xmm0, 32(%r9)4768; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4769; SSE-NEXT:    movaps %xmm0, 16(%r9)4770; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4771; SSE-NEXT:    movaps %xmm0, (%r9)4772; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax4773; SSE-NEXT:    movaps %xmm1, 240(%rax)4774; SSE-NEXT:    movaps %xmm3, 224(%rax)4775; SSE-NEXT:    movaps %xmm5, 208(%rax)4776; SSE-NEXT:    movaps %xmm7, 192(%rax)4777; SSE-NEXT:    movaps %xmm9, 176(%rax)4778; SSE-NEXT:    movaps %xmm10, 160(%rax)4779; SSE-NEXT:    movaps %xmm13, 144(%rax)4780; SSE-NEXT:    movaps %xmm15, 128(%rax)4781; SSE-NEXT:    movaps %xmm14, 112(%rax)4782; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4783; SSE-NEXT:    movaps %xmm0, 96(%rax)4784; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload4785; SSE-NEXT:    movaps %xmm0, 80(%rax)4786; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4787; SSE-NEXT:    movaps %xmm0, 64(%rax)4788; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4789; SSE-NEXT:    movaps %xmm0, 48(%rax)4790; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4791; SSE-NEXT:    movaps %xmm0, 32(%rax)4792; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4793; SSE-NEXT:    movaps %xmm0, 16(%rax)4794; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4795; SSE-NEXT:    movaps %xmm0, (%rax)4796; SSE-NEXT:    addq $1176, %rsp # imm = 0x4984797; SSE-NEXT:    retq4798;4799; AVX-LABEL: load_i64_stride6_vf32:4800; AVX:       # %bb.0:4801; AVX-NEXT:    subq $1624, %rsp # imm = 0x6584802; AVX-NEXT:    vmovaps 1088(%rdi), %ymm24803; AVX-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4804; AVX-NEXT:    vmovaps 704(%rdi), %ymm34805; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4806; AVX-NEXT:    vmovaps 320(%rdi), %ymm14807; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4808; AVX-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm04809; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4810; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]4811; AVX-NEXT:    vmovaps 240(%rdi), %xmm44812; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4813; AVX-NEXT:    vmovaps 192(%rdi), %xmm14814; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4815; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]4816; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4817; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4818; AVX-NEXT:    vinsertf128 $1, 672(%rdi), %ymm0, %ymm04819; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4820; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]4821; AVX-NEXT:    vmovaps 624(%rdi), %xmm34822; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4823; AVX-NEXT:    vmovaps 576(%rdi), %xmm14824; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4825; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]4826; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4827; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4828; AVX-NEXT:    vinsertf128 $1, 1056(%rdi), %ymm0, %ymm04829; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4830; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]4831; AVX-NEXT:    vmovaps 1008(%rdi), %xmm24832; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4833; AVX-NEXT:    vmovaps 960(%rdi), %xmm14834; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4835; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4836; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4837; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4838; AVX-NEXT:    vmovaps 1472(%rdi), %ymm14839; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4840; AVX-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm0, %ymm04841; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4842; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]4843; AVX-NEXT:    vmovaps 1392(%rdi), %xmm24844; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4845; AVX-NEXT:    vmovaps 1344(%rdi), %xmm14846; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4847; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4848; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4849; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4850; AVX-NEXT:    vmovaps 128(%rdi), %ymm14851; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4852; AVX-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm04853; AVX-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill4854; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]4855; AVX-NEXT:    vmovaps (%rdi), %xmm144856; AVX-NEXT:    vmovaps 48(%rdi), %xmm134857; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm14[0],xmm13[0]4858; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4859; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4860; AVX-NEXT:    vmovaps 512(%rdi), %ymm124861; AVX-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm114862; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]4863; AVX-NEXT:    vmovaps 432(%rdi), %xmm104864; AVX-NEXT:    vmovaps 384(%rdi), %xmm94865; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]4866; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4867; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4868; AVX-NEXT:    vmovaps 896(%rdi), %ymm84869; AVX-NEXT:    vinsertf128 $1, 864(%rdi), %ymm0, %ymm74870; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]4871; AVX-NEXT:    vmovaps 816(%rdi), %xmm64872; AVX-NEXT:    vmovaps 768(%rdi), %xmm54873; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]4874; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4875; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4876; AVX-NEXT:    vmovaps 1280(%rdi), %ymm44877; AVX-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm34878; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]4879; AVX-NEXT:    vmovaps 1200(%rdi), %xmm24880; AVX-NEXT:    vmovaps 1152(%rdi), %xmm14881; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]4882; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]4883; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4884; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4885; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4886; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]4887; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload4888; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload4889; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]4890; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]4891; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4892; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4893; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4894; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]4895; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload4896; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload4897; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]4898; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]4899; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4900; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4901; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4902; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]4903; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload4904; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload4905; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]4906; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]4907; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4908; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4909; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4910; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]4911; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload4912; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload4913; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]4914; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]4915; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4916; AVX-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload4917; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4918; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]4919; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm14[1],xmm13[1]4920; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]4921; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4922; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]4923; AVX-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]4924; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]4925; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4926; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]4927; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]4928; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]4929; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4930; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]4931; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]4932; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4933; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4934; AVX-NEXT:    vmovaps 288(%rdi), %ymm14935; AVX-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill4936; AVX-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm04937; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4938; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]4939; AVX-NEXT:    vmovaps 256(%rdi), %xmm24940; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4941; AVX-NEXT:    vmovaps 208(%rdi), %xmm14942; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4943; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4944; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4945; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4946; AVX-NEXT:    vmovaps 672(%rdi), %ymm14947; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4948; AVX-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm04949; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4950; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]4951; AVX-NEXT:    vmovaps 640(%rdi), %xmm24952; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4953; AVX-NEXT:    vmovaps 592(%rdi), %xmm14954; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4955; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4956; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4957; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4958; AVX-NEXT:    vmovaps 1056(%rdi), %ymm14959; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4960; AVX-NEXT:    vinsertf128 $1, 1120(%rdi), %ymm0, %ymm04961; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4962; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]4963; AVX-NEXT:    vmovaps 1024(%rdi), %xmm24964; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4965; AVX-NEXT:    vmovaps 976(%rdi), %xmm14966; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4967; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4968; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4969; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4970; AVX-NEXT:    vmovaps 1440(%rdi), %ymm14971; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4972; AVX-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm04973; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4974; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]4975; AVX-NEXT:    vmovaps 1408(%rdi), %xmm144976; AVX-NEXT:    vmovaps 1360(%rdi), %xmm124977; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm14[0]4978; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4979; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4980; AVX-NEXT:    vmovaps 96(%rdi), %ymm14981; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4982; AVX-NEXT:    vmovaps 16(%rdi), %xmm24983; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4984; AVX-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm04985; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4986; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]4987; AVX-NEXT:    vmovaps 64(%rdi), %xmm14988; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4989; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]4990; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4991; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4992; AVX-NEXT:    vmovaps 480(%rdi), %ymm114993; AVX-NEXT:    vinsertf128 $1, 544(%rdi), %ymm0, %ymm104994; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]4995; AVX-NEXT:    vmovaps 448(%rdi), %xmm94996; AVX-NEXT:    vmovaps 400(%rdi), %xmm84997; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]4998; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4999; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5000; AVX-NEXT:    vmovaps 864(%rdi), %ymm75001; AVX-NEXT:    vinsertf128 $1, 928(%rdi), %ymm0, %ymm65002; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]5003; AVX-NEXT:    vmovaps 832(%rdi), %xmm55004; AVX-NEXT:    vmovaps 784(%rdi), %xmm45005; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm4[0],xmm5[0]5006; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]5007; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5008; AVX-NEXT:    vmovaps 1248(%rdi), %ymm35009; AVX-NEXT:    vinsertf128 $1, 1312(%rdi), %ymm0, %ymm135010; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm13[0],ymm3[2],ymm13[2]5011; AVX-NEXT:    vmovaps 1216(%rdi), %xmm25012; AVX-NEXT:    vmovaps 1168(%rdi), %xmm15013; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]5014; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5015; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5016; AVX-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload5017; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5018; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5019; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5020; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5021; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5022; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5023; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5024; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5025; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5026; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5027; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5028; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5029; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5030; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5031; AVX-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill5032; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5033; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5034; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5035; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5036; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5037; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5038; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5039; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5040; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5041; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5042; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5043; AVX-NEXT:    vunpckhpd {{.*#+}} xmm12 = xmm12[1],xmm14[1]5044; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm0[4,5,6,7]5045; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5046; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm13[1],ymm3[3],ymm13[3]5047; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]5048; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5049; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5050; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]5051; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm5[1]5052; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5053; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5054; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]5055; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm9[1]5056; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5057; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5058; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5059; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5060; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5061; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5062; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload5063; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]5064; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5065; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5066; AVX-NEXT:    vmovaps 160(%rdi), %ymm15067; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5068; AVX-NEXT:    vmovaps 32(%rdi), %xmm25069; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5070; AVX-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm05071; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5072; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5073; AVX-NEXT:    vmovaps 80(%rdi), %xmm15074; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5075; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]5076; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5077; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5078; AVX-NEXT:    vmovaps 352(%rdi), %ymm15079; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5080; AVX-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm05081; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5082; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5083; AVX-NEXT:    vmovaps 272(%rdi), %xmm25084; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5085; AVX-NEXT:    vmovaps 224(%rdi), %xmm15086; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5087; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]5088; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5089; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5090; AVX-NEXT:    vmovaps 544(%rdi), %ymm15091; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5092; AVX-NEXT:    vinsertf128 $1, 512(%rdi), %ymm0, %ymm05093; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5094; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5095; AVX-NEXT:    vmovaps 464(%rdi), %xmm25096; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5097; AVX-NEXT:    vmovaps 416(%rdi), %xmm15098; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5099; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]5100; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5101; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5102; AVX-NEXT:    vmovaps 736(%rdi), %ymm15103; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5104; AVX-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm05105; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5106; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5107; AVX-NEXT:    vmovaps 656(%rdi), %xmm25108; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5109; AVX-NEXT:    vmovaps 608(%rdi), %xmm15110; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5111; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]5112; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5113; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5114; AVX-NEXT:    vmovaps 928(%rdi), %ymm15115; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5116; AVX-NEXT:    vinsertf128 $1, 896(%rdi), %ymm0, %ymm05117; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5118; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5119; AVX-NEXT:    vmovaps 848(%rdi), %xmm145120; AVX-NEXT:    vmovaps 800(%rdi), %xmm135121; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]5122; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5123; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5124; AVX-NEXT:    vmovaps 1120(%rdi), %ymm125125; AVX-NEXT:    vinsertf128 $1, 1088(%rdi), %ymm0, %ymm115126; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]5127; AVX-NEXT:    vmovaps 1040(%rdi), %xmm105128; AVX-NEXT:    vmovaps 992(%rdi), %xmm95129; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]5130; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5131; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5132; AVX-NEXT:    vmovaps 1312(%rdi), %ymm85133; AVX-NEXT:    vinsertf128 $1, 1280(%rdi), %ymm0, %ymm75134; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]5135; AVX-NEXT:    vmovaps 1232(%rdi), %xmm65136; AVX-NEXT:    vmovaps 1184(%rdi), %xmm55137; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]5138; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5139; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5140; AVX-NEXT:    vmovaps 1504(%rdi), %ymm45141; AVX-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm0, %ymm35142; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]5143; AVX-NEXT:    vmovaps 1424(%rdi), %xmm25144; AVX-NEXT:    vmovaps 1376(%rdi), %xmm15145; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]5146; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5147; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5148; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5149; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5150; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5151; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5152; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5153; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5154; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5155; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5156; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5157; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5158; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5159; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5160; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5161; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5162; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5163; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5164; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5165; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5166; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5167; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5168; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5169; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5170; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5171; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5172; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5173; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5174; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5175; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5176; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5177; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]5178; AVX-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]5179; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5180; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5181; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5182; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]5183; AVX-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]5184; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]5185; AVX-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]5186; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]5187; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]5188; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]5189; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]5190; AVX-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]5191; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]5192; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]5193; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5194; AVX-NEXT:    vmovaps %ymm2, 192(%rsi)5195; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5196; AVX-NEXT:    vmovaps %ymm2, 128(%rsi)5197; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5198; AVX-NEXT:    vmovaps %ymm2, 64(%rsi)5199; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5200; AVX-NEXT:    vmovaps %ymm2, (%rsi)5201; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5202; AVX-NEXT:    vmovaps %ymm2, 224(%rsi)5203; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5204; AVX-NEXT:    vmovaps %ymm2, 160(%rsi)5205; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5206; AVX-NEXT:    vmovaps %ymm2, 96(%rsi)5207; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5208; AVX-NEXT:    vmovaps %ymm2, 32(%rsi)5209; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5210; AVX-NEXT:    vmovaps %ymm2, 192(%rdx)5211; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5212; AVX-NEXT:    vmovaps %ymm2, 128(%rdx)5213; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5214; AVX-NEXT:    vmovaps %ymm2, 64(%rdx)5215; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5216; AVX-NEXT:    vmovaps %ymm2, (%rdx)5217; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5218; AVX-NEXT:    vmovaps %ymm2, 224(%rdx)5219; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5220; AVX-NEXT:    vmovaps %ymm2, 160(%rdx)5221; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5222; AVX-NEXT:    vmovaps %ymm2, 96(%rdx)5223; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5224; AVX-NEXT:    vmovaps %ymm2, 32(%rdx)5225; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5226; AVX-NEXT:    vmovaps %ymm2, 192(%rcx)5227; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5228; AVX-NEXT:    vmovaps %ymm2, 128(%rcx)5229; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5230; AVX-NEXT:    vmovaps %ymm2, 64(%rcx)5231; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5232; AVX-NEXT:    vmovaps %ymm2, (%rcx)5233; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5234; AVX-NEXT:    vmovaps %ymm2, 224(%rcx)5235; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5236; AVX-NEXT:    vmovaps %ymm2, 160(%rcx)5237; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5238; AVX-NEXT:    vmovaps %ymm2, 96(%rcx)5239; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5240; AVX-NEXT:    vmovaps %ymm2, 32(%rcx)5241; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5242; AVX-NEXT:    vmovaps %ymm2, (%r8)5243; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5244; AVX-NEXT:    vmovaps %ymm2, 64(%r8)5245; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5246; AVX-NEXT:    vmovaps %ymm2, 128(%r8)5247; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5248; AVX-NEXT:    vmovaps %ymm2, 192(%r8)5249; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5250; AVX-NEXT:    vmovaps %ymm2, 224(%r8)5251; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5252; AVX-NEXT:    vmovaps %ymm2, 160(%r8)5253; AVX-NEXT:    vmovups (%rsp), %ymm2 # 32-byte Reload5254; AVX-NEXT:    vmovaps %ymm2, 96(%r8)5255; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5256; AVX-NEXT:    vmovaps %ymm2, 32(%r8)5257; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5258; AVX-NEXT:    vmovaps %ymm2, 224(%r9)5259; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5260; AVX-NEXT:    vmovaps %ymm2, 192(%r9)5261; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5262; AVX-NEXT:    vmovaps %ymm2, 160(%r9)5263; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5264; AVX-NEXT:    vmovaps %ymm2, 128(%r9)5265; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5266; AVX-NEXT:    vmovaps %ymm2, 96(%r9)5267; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5268; AVX-NEXT:    vmovaps %ymm2, 64(%r9)5269; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5270; AVX-NEXT:    vmovaps %ymm2, 32(%r9)5271; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5272; AVX-NEXT:    vmovaps %ymm2, (%r9)5273; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax5274; AVX-NEXT:    vmovaps %ymm1, 224(%rax)5275; AVX-NEXT:    vmovaps %ymm0, 192(%rax)5276; AVX-NEXT:    vmovaps %ymm9, 160(%rax)5277; AVX-NEXT:    vmovaps %ymm13, 128(%rax)5278; AVX-NEXT:    vmovaps %ymm15, 96(%rax)5279; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5280; AVX-NEXT:    vmovaps %ymm0, 64(%rax)5281; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5282; AVX-NEXT:    vmovaps %ymm0, 32(%rax)5283; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5284; AVX-NEXT:    vmovaps %ymm0, (%rax)5285; AVX-NEXT:    addq $1624, %rsp # imm = 0x6585286; AVX-NEXT:    vzeroupper5287; AVX-NEXT:    retq5288;5289; AVX2-LABEL: load_i64_stride6_vf32:5290; AVX2:       # %bb.0:5291; AVX2-NEXT:    subq $1496, %rsp # imm = 0x5D85292; AVX2-NEXT:    vmovaps 1088(%rdi), %ymm25293; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5294; AVX2-NEXT:    vmovaps 1056(%rdi), %ymm45295; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5296; AVX2-NEXT:    vmovaps 704(%rdi), %ymm35297; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5298; AVX2-NEXT:    vmovaps 672(%rdi), %ymm55299; AVX2-NEXT:    vmovups %ymm5, (%rsp) # 32-byte Spill5300; AVX2-NEXT:    vmovaps 320(%rdi), %ymm15301; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5302; AVX2-NEXT:    vmovaps 288(%rdi), %ymm75303; AVX2-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5304; AVX2-NEXT:    vmovaps 240(%rdi), %xmm05305; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5306; AVX2-NEXT:    vmovaps 192(%rdi), %xmm65307; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm6[0],xmm0[0]5308; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[2],ymm1[2]5309; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5310; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5311; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5312; AVX2-NEXT:    vmovaps 624(%rdi), %xmm15313; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5314; AVX2-NEXT:    vmovaps 576(%rdi), %xmm05315; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5316; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]5317; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]5318; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5319; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5320; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5321; AVX2-NEXT:    vmovaps 1008(%rdi), %xmm15322; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5323; AVX2-NEXT:    vmovaps 960(%rdi), %xmm05324; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5325; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]5326; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]5327; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5328; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5329; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5330; AVX2-NEXT:    vmovaps 1472(%rdi), %ymm05331; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5332; AVX2-NEXT:    vmovaps 1440(%rdi), %ymm15333; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5334; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5335; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5336; AVX2-NEXT:    vmovaps 1392(%rdi), %xmm15337; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5338; AVX2-NEXT:    vmovaps 1344(%rdi), %xmm45339; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm1[0]5340; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5341; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5342; AVX2-NEXT:    vmovaps 128(%rdi), %ymm155343; AVX2-NEXT:    vmovaps 96(%rdi), %ymm05344; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5345; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]5346; AVX2-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5347; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5348; AVX2-NEXT:    vmovaps (%rdi), %xmm85349; AVX2-NEXT:    vmovaps 48(%rdi), %xmm25350; AVX2-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm2[0]5351; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]5352; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5353; AVX2-NEXT:    vmovaps 512(%rdi), %ymm75354; AVX2-NEXT:    vmovaps 480(%rdi), %ymm05355; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5356; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]5357; AVX2-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5358; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5359; AVX2-NEXT:    vmovaps 432(%rdi), %xmm115360; AVX2-NEXT:    vmovaps 384(%rdi), %xmm125361; AVX2-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm12[0],xmm11[0]5362; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]5363; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5364; AVX2-NEXT:    vmovaps 896(%rdi), %ymm55365; AVX2-NEXT:    vmovaps 864(%rdi), %ymm05366; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5367; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]5368; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5369; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5370; AVX2-NEXT:    vmovaps 816(%rdi), %xmm135371; AVX2-NEXT:    vmovaps 768(%rdi), %xmm105372; AVX2-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm10[0],xmm13[0]5373; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm0[4,5,6,7]5374; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5375; AVX2-NEXT:    vmovaps 1280(%rdi), %ymm35376; AVX2-NEXT:    vmovaps 1248(%rdi), %ymm05377; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5378; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]5379; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5380; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5381; AVX2-NEXT:    vmovaps 1200(%rdi), %xmm145382; AVX2-NEXT:    vmovaps 1152(%rdi), %xmm15383; AVX2-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm14[0]5384; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]5385; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5386; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload5387; AVX2-NEXT:    # xmm0 = xmm6[1],mem[1]5388; AVX2-NEXT:    vbroadcastsd 296(%rdi), %ymm65389; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload5390; AVX2-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]5391; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]5392; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5393; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5394; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5395; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]5396; AVX2-NEXT:    vbroadcastsd 680(%rdi), %ymm65397; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload5398; AVX2-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]5399; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]5400; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5401; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5402; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5403; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]5404; AVX2-NEXT:    vbroadcastsd 1064(%rdi), %ymm65405; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload5406; AVX2-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]5407; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]5408; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5409; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm0 # 16-byte Folded Reload5410; AVX2-NEXT:    # xmm0 = xmm4[1],mem[1]5411; AVX2-NEXT:    vbroadcastsd 1448(%rdi), %ymm45412; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload5413; AVX2-NEXT:    # ymm4 = ymm4[1],mem[1],ymm4[3],mem[3]5414; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]5415; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5416; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm2[1]5417; AVX2-NEXT:    vbroadcastsd 104(%rdi), %ymm25418; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]5419; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5420; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5421; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm12[1],xmm11[1]5422; AVX2-NEXT:    vbroadcastsd 488(%rdi), %ymm25423; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm7[1],ymm2[3],ymm7[3]5424; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5425; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5426; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm13[1]5427; AVX2-NEXT:    vbroadcastsd 872(%rdi), %ymm25428; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]5429; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5430; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5431; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm14[1]5432; AVX2-NEXT:    vbroadcastsd 1256(%rdi), %ymm15433; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]5434; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5435; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5436; AVX2-NEXT:    vbroadcastsd 352(%rdi), %ymm05437; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5438; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5439; AVX2-NEXT:    vmovaps 256(%rdi), %xmm35440; AVX2-NEXT:    vmovaps 208(%rdi), %xmm55441; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm3[0]5442; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5443; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5444; AVX2-NEXT:    vbroadcastsd 736(%rdi), %ymm05445; AVX2-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload5446; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5447; AVX2-NEXT:    vmovaps 640(%rdi), %xmm65448; AVX2-NEXT:    vmovaps 592(%rdi), %xmm75449; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm6[0]5450; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5451; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5452; AVX2-NEXT:    vbroadcastsd 1120(%rdi), %ymm05453; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5454; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5455; AVX2-NEXT:    vmovaps 1024(%rdi), %xmm85456; AVX2-NEXT:    vmovaps 976(%rdi), %xmm95457; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm8[0]5458; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5459; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5460; AVX2-NEXT:    vbroadcastsd 1504(%rdi), %ymm05461; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5462; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5463; AVX2-NEXT:    vmovaps 1408(%rdi), %xmm105464; AVX2-NEXT:    vmovaps 1360(%rdi), %xmm115465; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm10[0]5466; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5467; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5468; AVX2-NEXT:    vbroadcastsd 160(%rdi), %ymm05469; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5470; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5471; AVX2-NEXT:    vmovaps 16(%rdi), %xmm05472; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5473; AVX2-NEXT:    vmovaps 64(%rdi), %xmm25474; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5475; AVX2-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm2[0]5476; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm1[4,5,6,7]5477; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5478; AVX2-NEXT:    vbroadcastsd 544(%rdi), %ymm15479; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5480; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5481; AVX2-NEXT:    vmovaps 448(%rdi), %xmm05482; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5483; AVX2-NEXT:    vmovaps 400(%rdi), %xmm15484; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5485; AVX2-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm1[0],xmm0[0]5486; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm12[4,5,6,7]5487; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5488; AVX2-NEXT:    vbroadcastsd 928(%rdi), %ymm125489; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5490; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm4[0],ymm12[0],ymm4[2],ymm12[2]5491; AVX2-NEXT:    vmovaps 832(%rdi), %xmm125492; AVX2-NEXT:    vmovaps 784(%rdi), %xmm135493; AVX2-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm13[0],xmm12[0]5494; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm14[4,5,6,7]5495; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5496; AVX2-NEXT:    vbroadcastsd 1312(%rdi), %ymm145497; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5498; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm14[0],ymm1[2],ymm14[2]5499; AVX2-NEXT:    vmovaps 1216(%rdi), %xmm155500; AVX2-NEXT:    vmovaps 1168(%rdi), %xmm05501; AVX2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm15[0]5502; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4,5,6,7]5503; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5504; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm3[1]5505; AVX2-NEXT:    vmovaps 352(%rdi), %ymm145506; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5507; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm14[1],ymm3[3],ymm14[3]5508; AVX2-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5509; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]5510; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5511; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5512; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm7[1],xmm6[1]5513; AVX2-NEXT:    vmovaps 736(%rdi), %ymm75514; AVX2-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload5515; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]5516; AVX2-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5517; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]5518; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5519; AVX2-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill5520; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm9[1],xmm8[1]5521; AVX2-NEXT:    vmovaps 1120(%rdi), %ymm65522; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5523; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm6[1],ymm3[3],ymm6[3]5524; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5525; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]5526; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5527; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5528; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm11[1],xmm10[1]5529; AVX2-NEXT:    vmovaps 1504(%rdi), %ymm55530; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5531; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5532; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]5533; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]5534; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5535; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5536; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]5537; AVX2-NEXT:    vmovaps 1312(%rdi), %ymm35538; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]5539; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5540; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]5541; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5542; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5543; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm12[1]5544; AVX2-NEXT:    vmovaps 928(%rdi), %ymm15545; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]5546; AVX2-NEXT:    vmovaps %ymm1, %ymm45547; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5548; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]5549; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5550; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5551; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5552; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5553; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]5554; AVX2-NEXT:    vmovaps 544(%rdi), %ymm25555; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5556; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]5557; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5558; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]5559; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5560; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5561; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5562; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5563; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]5564; AVX2-NEXT:    vmovaps 160(%rdi), %ymm95565; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5566; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]5567; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]5568; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5569; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5570; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5571; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]5572; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5573; AVX2-NEXT:    vmovaps 32(%rdi), %xmm55574; AVX2-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5575; AVX2-NEXT:    vmovaps 80(%rdi), %xmm15576; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5577; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm1[0]5578; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5579; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5580; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5581; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]5582; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5583; AVX2-NEXT:    vmovaps 272(%rdi), %xmm55584; AVX2-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5585; AVX2-NEXT:    vmovaps 224(%rdi), %xmm15586; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5587; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]5588; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5589; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5590; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5591; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]5592; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5593; AVX2-NEXT:    vmovaps 464(%rdi), %xmm145594; AVX2-NEXT:    vmovaps 416(%rdi), %xmm135595; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]5596; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5597; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5598; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5599; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]5600; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5601; AVX2-NEXT:    vmovaps 656(%rdi), %xmm125602; AVX2-NEXT:    vmovaps 608(%rdi), %xmm115603; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]5604; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5605; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5606; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5607; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]5608; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5609; AVX2-NEXT:    vmovaps 848(%rdi), %xmm105610; AVX2-NEXT:    vmovaps 800(%rdi), %xmm75611; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm10[0]5612; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5613; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5614; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5615; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]5616; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5617; AVX2-NEXT:    vmovaps 1040(%rdi), %xmm85618; AVX2-NEXT:    vmovaps 992(%rdi), %xmm55619; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]5620; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5621; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5622; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5623; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]5624; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5625; AVX2-NEXT:    vmovaps 1232(%rdi), %xmm65626; AVX2-NEXT:    vmovaps 1184(%rdi), %xmm35627; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]5628; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5629; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5630; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5631; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5632; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]5633; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5634; AVX2-NEXT:    vmovaps 1424(%rdi), %xmm25635; AVX2-NEXT:    vmovaps 1376(%rdi), %xmm15636; AVX2-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]5637; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]5638; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5639; AVX2-NEXT:    vbroadcastsd 136(%rdi), %ymm05640; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm9[1],ymm0[3],ymm9[3]5641; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload5642; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload5643; AVX2-NEXT:    # xmm9 = xmm9[1],mem[1]5644; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]5645; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5646; AVX2-NEXT:    vbroadcastsd 328(%rdi), %ymm05647; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5648; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5649; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload5650; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload5651; AVX2-NEXT:    # xmm9 = xmm9[1],mem[1]5652; AVX2-NEXT:    vblendps {{.*#+}} ymm15 = ymm9[0,1,2,3],ymm0[4,5,6,7]5653; AVX2-NEXT:    vbroadcastsd 520(%rdi), %ymm05654; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5655; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5656; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm13[1],xmm14[1]5657; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm0[4,5,6,7]5658; AVX2-NEXT:    vbroadcastsd 712(%rdi), %ymm05659; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5660; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5661; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm11[1],xmm12[1]5662; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]5663; AVX2-NEXT:    vbroadcastsd 904(%rdi), %ymm05664; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5665; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5666; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm10[1]5667; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]5668; AVX2-NEXT:    vbroadcastsd 1096(%rdi), %ymm05669; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5670; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5671; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]5672; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]5673; AVX2-NEXT:    vbroadcastsd 1288(%rdi), %ymm05674; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5675; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]5676; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]5677; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]5678; AVX2-NEXT:    vbroadcastsd 1480(%rdi), %ymm05679; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm4[1],ymm0[3],ymm4[3]5680; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]5681; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5682; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5683; AVX2-NEXT:    vmovaps %ymm1, 192(%rsi)5684; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5685; AVX2-NEXT:    vmovaps %ymm1, 128(%rsi)5686; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5687; AVX2-NEXT:    vmovaps %ymm1, 64(%rsi)5688; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5689; AVX2-NEXT:    vmovaps %ymm1, (%rsi)5690; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5691; AVX2-NEXT:    vmovaps %ymm1, 224(%rsi)5692; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5693; AVX2-NEXT:    vmovaps %ymm1, 160(%rsi)5694; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5695; AVX2-NEXT:    vmovaps %ymm1, 96(%rsi)5696; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5697; AVX2-NEXT:    vmovaps %ymm1, 32(%rsi)5698; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5699; AVX2-NEXT:    vmovaps %ymm1, 192(%rdx)5700; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5701; AVX2-NEXT:    vmovaps %ymm1, 128(%rdx)5702; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5703; AVX2-NEXT:    vmovaps %ymm1, 64(%rdx)5704; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5705; AVX2-NEXT:    vmovaps %ymm1, (%rdx)5706; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5707; AVX2-NEXT:    vmovaps %ymm1, 224(%rdx)5708; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5709; AVX2-NEXT:    vmovaps %ymm1, 160(%rdx)5710; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5711; AVX2-NEXT:    vmovaps %ymm1, 96(%rdx)5712; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5713; AVX2-NEXT:    vmovaps %ymm1, 32(%rdx)5714; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5715; AVX2-NEXT:    vmovaps %ymm1, 192(%rcx)5716; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5717; AVX2-NEXT:    vmovaps %ymm1, 128(%rcx)5718; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5719; AVX2-NEXT:    vmovaps %ymm1, 64(%rcx)5720; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5721; AVX2-NEXT:    vmovaps %ymm1, (%rcx)5722; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5723; AVX2-NEXT:    vmovaps %ymm1, 224(%rcx)5724; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5725; AVX2-NEXT:    vmovaps %ymm1, 160(%rcx)5726; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5727; AVX2-NEXT:    vmovaps %ymm1, 96(%rcx)5728; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5729; AVX2-NEXT:    vmovaps %ymm1, 32(%rcx)5730; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5731; AVX2-NEXT:    vmovaps %ymm1, (%r8)5732; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5733; AVX2-NEXT:    vmovaps %ymm1, 64(%r8)5734; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5735; AVX2-NEXT:    vmovaps %ymm1, 128(%r8)5736; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5737; AVX2-NEXT:    vmovaps %ymm1, 192(%r8)5738; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5739; AVX2-NEXT:    vmovaps %ymm1, 224(%r8)5740; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5741; AVX2-NEXT:    vmovaps %ymm1, 160(%r8)5742; AVX2-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload5743; AVX2-NEXT:    vmovaps %ymm1, 96(%r8)5744; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5745; AVX2-NEXT:    vmovaps %ymm1, 32(%r8)5746; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5747; AVX2-NEXT:    vmovaps %ymm1, 224(%r9)5748; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5749; AVX2-NEXT:    vmovaps %ymm1, 192(%r9)5750; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5751; AVX2-NEXT:    vmovaps %ymm1, 160(%r9)5752; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5753; AVX2-NEXT:    vmovaps %ymm1, 128(%r9)5754; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5755; AVX2-NEXT:    vmovaps %ymm1, 96(%r9)5756; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5757; AVX2-NEXT:    vmovaps %ymm1, 64(%r9)5758; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5759; AVX2-NEXT:    vmovaps %ymm1, 32(%r9)5760; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5761; AVX2-NEXT:    vmovaps %ymm1, (%r9)5762; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax5763; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)5764; AVX2-NEXT:    vmovaps %ymm3, 192(%rax)5765; AVX2-NEXT:    vmovaps %ymm5, 160(%rax)5766; AVX2-NEXT:    vmovaps %ymm7, 128(%rax)5767; AVX2-NEXT:    vmovaps %ymm9, 96(%rax)5768; AVX2-NEXT:    vmovaps %ymm13, 64(%rax)5769; AVX2-NEXT:    vmovaps %ymm15, 32(%rax)5770; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5771; AVX2-NEXT:    vmovaps %ymm0, (%rax)5772; AVX2-NEXT:    addq $1496, %rsp # imm = 0x5D85773; AVX2-NEXT:    vzeroupper5774; AVX2-NEXT:    retq5775;5776; AVX2-FP-LABEL: load_i64_stride6_vf32:5777; AVX2-FP:       # %bb.0:5778; AVX2-FP-NEXT:    subq $1496, %rsp # imm = 0x5D85779; AVX2-FP-NEXT:    vmovaps 1088(%rdi), %ymm25780; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5781; AVX2-FP-NEXT:    vmovaps 1056(%rdi), %ymm45782; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5783; AVX2-FP-NEXT:    vmovaps 704(%rdi), %ymm35784; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5785; AVX2-FP-NEXT:    vmovaps 672(%rdi), %ymm55786; AVX2-FP-NEXT:    vmovups %ymm5, (%rsp) # 32-byte Spill5787; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm15788; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5789; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm75790; AVX2-FP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5791; AVX2-FP-NEXT:    vmovaps 240(%rdi), %xmm05792; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5793; AVX2-FP-NEXT:    vmovaps 192(%rdi), %xmm65794; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm6[0],xmm0[0]5795; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[2],ymm1[2]5796; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5797; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5798; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5799; AVX2-FP-NEXT:    vmovaps 624(%rdi), %xmm15800; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5801; AVX2-FP-NEXT:    vmovaps 576(%rdi), %xmm05802; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5803; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]5804; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]5805; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5806; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5807; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5808; AVX2-FP-NEXT:    vmovaps 1008(%rdi), %xmm15809; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5810; AVX2-FP-NEXT:    vmovaps 960(%rdi), %xmm05811; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5812; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]5813; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]5814; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5815; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5816; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5817; AVX2-FP-NEXT:    vmovaps 1472(%rdi), %ymm05818; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5819; AVX2-FP-NEXT:    vmovaps 1440(%rdi), %ymm15820; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5821; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5822; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5823; AVX2-FP-NEXT:    vmovaps 1392(%rdi), %xmm15824; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5825; AVX2-FP-NEXT:    vmovaps 1344(%rdi), %xmm45826; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm1[0]5827; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5828; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5829; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm155830; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm05831; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5832; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]5833; AVX2-FP-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5834; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5835; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm85836; AVX2-FP-NEXT:    vmovaps 48(%rdi), %xmm25837; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm2[0]5838; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]5839; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5840; AVX2-FP-NEXT:    vmovaps 512(%rdi), %ymm75841; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm05842; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5843; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]5844; AVX2-FP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5845; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5846; AVX2-FP-NEXT:    vmovaps 432(%rdi), %xmm115847; AVX2-FP-NEXT:    vmovaps 384(%rdi), %xmm125848; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm12[0],xmm11[0]5849; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]5850; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5851; AVX2-FP-NEXT:    vmovaps 896(%rdi), %ymm55852; AVX2-FP-NEXT:    vmovaps 864(%rdi), %ymm05853; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5854; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]5855; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5856; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5857; AVX2-FP-NEXT:    vmovaps 816(%rdi), %xmm135858; AVX2-FP-NEXT:    vmovaps 768(%rdi), %xmm105859; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm10[0],xmm13[0]5860; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm0[4,5,6,7]5861; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5862; AVX2-FP-NEXT:    vmovaps 1280(%rdi), %ymm35863; AVX2-FP-NEXT:    vmovaps 1248(%rdi), %ymm05864; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5865; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]5866; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5867; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]5868; AVX2-FP-NEXT:    vmovaps 1200(%rdi), %xmm145869; AVX2-FP-NEXT:    vmovaps 1152(%rdi), %xmm15870; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm14[0]5871; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]5872; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5873; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload5874; AVX2-FP-NEXT:    # xmm0 = xmm6[1],mem[1]5875; AVX2-FP-NEXT:    vbroadcastsd 296(%rdi), %ymm65876; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload5877; AVX2-FP-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]5878; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]5879; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5880; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5881; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5882; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]5883; AVX2-FP-NEXT:    vbroadcastsd 680(%rdi), %ymm65884; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload5885; AVX2-FP-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]5886; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]5887; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5888; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5889; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5890; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]5891; AVX2-FP-NEXT:    vbroadcastsd 1064(%rdi), %ymm65892; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload5893; AVX2-FP-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]5894; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]5895; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5896; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm0 # 16-byte Folded Reload5897; AVX2-FP-NEXT:    # xmm0 = xmm4[1],mem[1]5898; AVX2-FP-NEXT:    vbroadcastsd 1448(%rdi), %ymm45899; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload5900; AVX2-FP-NEXT:    # ymm4 = ymm4[1],mem[1],ymm4[3],mem[3]5901; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]5902; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5903; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm2[1]5904; AVX2-FP-NEXT:    vbroadcastsd 104(%rdi), %ymm25905; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]5906; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5907; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5908; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm12[1],xmm11[1]5909; AVX2-FP-NEXT:    vbroadcastsd 488(%rdi), %ymm25910; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm7[1],ymm2[3],ymm7[3]5911; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5912; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5913; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm13[1]5914; AVX2-FP-NEXT:    vbroadcastsd 872(%rdi), %ymm25915; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]5916; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5917; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5918; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm14[1]5919; AVX2-FP-NEXT:    vbroadcastsd 1256(%rdi), %ymm15920; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]5921; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5922; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5923; AVX2-FP-NEXT:    vbroadcastsd 352(%rdi), %ymm05924; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5925; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5926; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm35927; AVX2-FP-NEXT:    vmovaps 208(%rdi), %xmm55928; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm3[0]5929; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5930; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5931; AVX2-FP-NEXT:    vbroadcastsd 736(%rdi), %ymm05932; AVX2-FP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload5933; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5934; AVX2-FP-NEXT:    vmovaps 640(%rdi), %xmm65935; AVX2-FP-NEXT:    vmovaps 592(%rdi), %xmm75936; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm6[0]5937; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5938; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5939; AVX2-FP-NEXT:    vbroadcastsd 1120(%rdi), %ymm05940; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5941; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5942; AVX2-FP-NEXT:    vmovaps 1024(%rdi), %xmm85943; AVX2-FP-NEXT:    vmovaps 976(%rdi), %xmm95944; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm8[0]5945; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5946; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5947; AVX2-FP-NEXT:    vbroadcastsd 1504(%rdi), %ymm05948; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5949; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5950; AVX2-FP-NEXT:    vmovaps 1408(%rdi), %xmm105951; AVX2-FP-NEXT:    vmovaps 1360(%rdi), %xmm115952; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm10[0]5953; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5954; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5955; AVX2-FP-NEXT:    vbroadcastsd 160(%rdi), %ymm05956; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5957; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]5958; AVX2-FP-NEXT:    vmovaps 16(%rdi), %xmm05959; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5960; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm25961; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5962; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm2[0]5963; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm1[4,5,6,7]5964; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5965; AVX2-FP-NEXT:    vbroadcastsd 544(%rdi), %ymm15966; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5967; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5968; AVX2-FP-NEXT:    vmovaps 448(%rdi), %xmm05969; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5970; AVX2-FP-NEXT:    vmovaps 400(%rdi), %xmm15971; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5972; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm1[0],xmm0[0]5973; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm12[4,5,6,7]5974; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5975; AVX2-FP-NEXT:    vbroadcastsd 928(%rdi), %ymm125976; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5977; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm4[0],ymm12[0],ymm4[2],ymm12[2]5978; AVX2-FP-NEXT:    vmovaps 832(%rdi), %xmm125979; AVX2-FP-NEXT:    vmovaps 784(%rdi), %xmm135980; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm13[0],xmm12[0]5981; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm14[4,5,6,7]5982; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5983; AVX2-FP-NEXT:    vbroadcastsd 1312(%rdi), %ymm145984; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5985; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm14[0],ymm1[2],ymm14[2]5986; AVX2-FP-NEXT:    vmovaps 1216(%rdi), %xmm155987; AVX2-FP-NEXT:    vmovaps 1168(%rdi), %xmm05988; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm15[0]5989; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4,5,6,7]5990; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5991; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm3[1]5992; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm145993; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5994; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm14[1],ymm3[3],ymm14[3]5995; AVX2-FP-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5996; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]5997; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5998; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5999; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm7[1],xmm6[1]6000; AVX2-FP-NEXT:    vmovaps 736(%rdi), %ymm76001; AVX2-FP-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload6002; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]6003; AVX2-FP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6004; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6005; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6006; AVX2-FP-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill6007; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm9[1],xmm8[1]6008; AVX2-FP-NEXT:    vmovaps 1120(%rdi), %ymm66009; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6010; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm6[1],ymm3[3],ymm6[3]6011; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6012; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6013; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6014; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6015; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm11[1],xmm10[1]6016; AVX2-FP-NEXT:    vmovaps 1504(%rdi), %ymm56017; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6018; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6019; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]6020; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6021; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6022; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6023; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]6024; AVX2-FP-NEXT:    vmovaps 1312(%rdi), %ymm36025; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]6026; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6027; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]6028; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6029; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6030; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm12[1]6031; AVX2-FP-NEXT:    vmovaps 928(%rdi), %ymm16032; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]6033; AVX2-FP-NEXT:    vmovaps %ymm1, %ymm46034; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6035; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]6036; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6037; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6038; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6039; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6040; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]6041; AVX2-FP-NEXT:    vmovaps 544(%rdi), %ymm26042; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6043; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]6044; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6045; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6046; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6047; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6048; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6049; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6050; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]6051; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm96052; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6053; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]6054; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6055; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6056; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6057; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6058; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]6059; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6060; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm56061; AVX2-FP-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6062; AVX2-FP-NEXT:    vmovaps 80(%rdi), %xmm16063; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6064; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm1[0]6065; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6066; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6067; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6068; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]6069; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6070; AVX2-FP-NEXT:    vmovaps 272(%rdi), %xmm56071; AVX2-FP-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6072; AVX2-FP-NEXT:    vmovaps 224(%rdi), %xmm16073; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6074; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]6075; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6076; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6077; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6078; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]6079; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6080; AVX2-FP-NEXT:    vmovaps 464(%rdi), %xmm146081; AVX2-FP-NEXT:    vmovaps 416(%rdi), %xmm136082; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]6083; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6084; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6085; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6086; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]6087; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6088; AVX2-FP-NEXT:    vmovaps 656(%rdi), %xmm126089; AVX2-FP-NEXT:    vmovaps 608(%rdi), %xmm116090; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]6091; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6092; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6093; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6094; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]6095; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6096; AVX2-FP-NEXT:    vmovaps 848(%rdi), %xmm106097; AVX2-FP-NEXT:    vmovaps 800(%rdi), %xmm76098; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm10[0]6099; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6100; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6101; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6102; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]6103; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6104; AVX2-FP-NEXT:    vmovaps 1040(%rdi), %xmm86105; AVX2-FP-NEXT:    vmovaps 992(%rdi), %xmm56106; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]6107; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6108; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6109; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6110; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]6111; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6112; AVX2-FP-NEXT:    vmovaps 1232(%rdi), %xmm66113; AVX2-FP-NEXT:    vmovaps 1184(%rdi), %xmm36114; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]6115; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6116; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6117; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6118; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload6119; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]6120; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6121; AVX2-FP-NEXT:    vmovaps 1424(%rdi), %xmm26122; AVX2-FP-NEXT:    vmovaps 1376(%rdi), %xmm16123; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]6124; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]6125; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6126; AVX2-FP-NEXT:    vbroadcastsd 136(%rdi), %ymm06127; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm9[1],ymm0[3],ymm9[3]6128; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload6129; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload6130; AVX2-FP-NEXT:    # xmm9 = xmm9[1],mem[1]6131; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]6132; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6133; AVX2-FP-NEXT:    vbroadcastsd 328(%rdi), %ymm06134; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6135; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6136; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload6137; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload6138; AVX2-FP-NEXT:    # xmm9 = xmm9[1],mem[1]6139; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm15 = ymm9[0,1,2,3],ymm0[4,5,6,7]6140; AVX2-FP-NEXT:    vbroadcastsd 520(%rdi), %ymm06141; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6142; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6143; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm13[1],xmm14[1]6144; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm0[4,5,6,7]6145; AVX2-FP-NEXT:    vbroadcastsd 712(%rdi), %ymm06146; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6147; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6148; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm11[1],xmm12[1]6149; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]6150; AVX2-FP-NEXT:    vbroadcastsd 904(%rdi), %ymm06151; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6152; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6153; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm10[1]6154; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]6155; AVX2-FP-NEXT:    vbroadcastsd 1096(%rdi), %ymm06156; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6157; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6158; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]6159; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]6160; AVX2-FP-NEXT:    vbroadcastsd 1288(%rdi), %ymm06161; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6162; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6163; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]6164; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]6165; AVX2-FP-NEXT:    vbroadcastsd 1480(%rdi), %ymm06166; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm4[1],ymm0[3],ymm4[3]6167; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]6168; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6169; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6170; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rsi)6171; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6172; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rsi)6173; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6174; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rsi)6175; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6176; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)6177; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6178; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rsi)6179; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6180; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rsi)6181; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6182; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rsi)6183; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6184; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rsi)6185; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6186; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rdx)6187; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6188; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rdx)6189; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6190; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rdx)6191; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6192; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)6193; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6194; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rdx)6195; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6196; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rdx)6197; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6198; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rdx)6199; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6200; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rdx)6201; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6202; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rcx)6203; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6204; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rcx)6205; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6206; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rcx)6207; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6208; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)6209; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6210; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rcx)6211; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6212; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rcx)6213; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6214; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rcx)6215; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6216; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rcx)6217; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6218; AVX2-FP-NEXT:    vmovaps %ymm1, (%r8)6219; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6220; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r8)6221; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6222; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r8)6223; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6224; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r8)6225; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6226; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r8)6227; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6228; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r8)6229; AVX2-FP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload6230; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r8)6231; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6232; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r8)6233; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6234; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r9)6235; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6236; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r9)6237; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6238; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r9)6239; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6240; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r9)6241; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6242; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r9)6243; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6244; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r9)6245; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6246; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r9)6247; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6248; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)6249; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax6250; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)6251; AVX2-FP-NEXT:    vmovaps %ymm3, 192(%rax)6252; AVX2-FP-NEXT:    vmovaps %ymm5, 160(%rax)6253; AVX2-FP-NEXT:    vmovaps %ymm7, 128(%rax)6254; AVX2-FP-NEXT:    vmovaps %ymm9, 96(%rax)6255; AVX2-FP-NEXT:    vmovaps %ymm13, 64(%rax)6256; AVX2-FP-NEXT:    vmovaps %ymm15, 32(%rax)6257; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6258; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)6259; AVX2-FP-NEXT:    addq $1496, %rsp # imm = 0x5D86260; AVX2-FP-NEXT:    vzeroupper6261; AVX2-FP-NEXT:    retq6262;6263; AVX2-FCP-LABEL: load_i64_stride6_vf32:6264; AVX2-FCP:       # %bb.0:6265; AVX2-FCP-NEXT:    subq $1496, %rsp # imm = 0x5D86266; AVX2-FCP-NEXT:    vmovaps 1088(%rdi), %ymm26267; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6268; AVX2-FCP-NEXT:    vmovaps 1056(%rdi), %ymm46269; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6270; AVX2-FCP-NEXT:    vmovaps 704(%rdi), %ymm36271; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6272; AVX2-FCP-NEXT:    vmovaps 672(%rdi), %ymm56273; AVX2-FCP-NEXT:    vmovups %ymm5, (%rsp) # 32-byte Spill6274; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm16275; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6276; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm76277; AVX2-FCP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6278; AVX2-FCP-NEXT:    vmovaps 240(%rdi), %xmm06279; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6280; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %xmm66281; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm6[0],xmm0[0]6282; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[2],ymm1[2]6283; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]6284; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6285; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6286; AVX2-FCP-NEXT:    vmovaps 624(%rdi), %xmm16287; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6288; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %xmm06289; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6290; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]6291; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]6292; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]6293; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6294; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6295; AVX2-FCP-NEXT:    vmovaps 1008(%rdi), %xmm16296; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6297; AVX2-FCP-NEXT:    vmovaps 960(%rdi), %xmm06298; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6299; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]6300; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]6301; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]6302; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6303; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6304; AVX2-FCP-NEXT:    vmovaps 1472(%rdi), %ymm06305; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6306; AVX2-FCP-NEXT:    vmovaps 1440(%rdi), %ymm16307; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6308; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]6309; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6310; AVX2-FCP-NEXT:    vmovaps 1392(%rdi), %xmm16311; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6312; AVX2-FCP-NEXT:    vmovaps 1344(%rdi), %xmm46313; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm4[0],xmm1[0]6314; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6315; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6316; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm156317; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm06318; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6319; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]6320; AVX2-FCP-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6321; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6322; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm86323; AVX2-FCP-NEXT:    vmovaps 48(%rdi), %xmm26324; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm2[0]6325; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]6326; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6327; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %ymm76328; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm06329; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6330; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]6331; AVX2-FCP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6332; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6333; AVX2-FCP-NEXT:    vmovaps 432(%rdi), %xmm116334; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %xmm126335; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm12[0],xmm11[0]6336; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3],ymm0[4,5,6,7]6337; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6338; AVX2-FCP-NEXT:    vmovaps 896(%rdi), %ymm56339; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %ymm06340; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6341; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]6342; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6343; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6344; AVX2-FCP-NEXT:    vmovaps 816(%rdi), %xmm136345; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %xmm106346; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm10[0],xmm13[0]6347; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm0[4,5,6,7]6348; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6349; AVX2-FCP-NEXT:    vmovaps 1280(%rdi), %ymm36350; AVX2-FCP-NEXT:    vmovaps 1248(%rdi), %ymm06351; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6352; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]6353; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6354; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6355; AVX2-FCP-NEXT:    vmovaps 1200(%rdi), %xmm146356; AVX2-FCP-NEXT:    vmovaps 1152(%rdi), %xmm16357; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm1[0],xmm14[0]6358; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]6359; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6360; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm0 # 16-byte Folded Reload6361; AVX2-FCP-NEXT:    # xmm0 = xmm6[1],mem[1]6362; AVX2-FCP-NEXT:    vbroadcastsd 296(%rdi), %ymm66363; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload6364; AVX2-FCP-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]6365; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]6366; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6367; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6368; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6369; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]6370; AVX2-FCP-NEXT:    vbroadcastsd 680(%rdi), %ymm66371; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload6372; AVX2-FCP-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]6373; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]6374; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6375; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6376; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6377; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]6378; AVX2-FCP-NEXT:    vbroadcastsd 1064(%rdi), %ymm66379; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload6380; AVX2-FCP-NEXT:    # ymm6 = ymm6[1],mem[1],ymm6[3],mem[3]6381; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]6382; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6383; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm0 # 16-byte Folded Reload6384; AVX2-FCP-NEXT:    # xmm0 = xmm4[1],mem[1]6385; AVX2-FCP-NEXT:    vbroadcastsd 1448(%rdi), %ymm46386; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload6387; AVX2-FCP-NEXT:    # ymm4 = ymm4[1],mem[1],ymm4[3],mem[3]6388; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]6389; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6390; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm2[1]6391; AVX2-FCP-NEXT:    vbroadcastsd 104(%rdi), %ymm26392; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm15[1],ymm2[3],ymm15[3]6393; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6394; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6395; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm12[1],xmm11[1]6396; AVX2-FCP-NEXT:    vbroadcastsd 488(%rdi), %ymm26397; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm7[1],ymm2[3],ymm7[3]6398; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6399; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6400; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm13[1]6401; AVX2-FCP-NEXT:    vbroadcastsd 872(%rdi), %ymm26402; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm5[1],ymm2[3],ymm5[3]6403; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6404; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6405; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm14[1]6406; AVX2-FCP-NEXT:    vbroadcastsd 1256(%rdi), %ymm16407; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]6408; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6409; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6410; AVX2-FCP-NEXT:    vbroadcastsd 352(%rdi), %ymm06411; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6412; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]6413; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm36414; AVX2-FCP-NEXT:    vmovaps 208(%rdi), %xmm56415; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm3[0]6416; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6417; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6418; AVX2-FCP-NEXT:    vbroadcastsd 736(%rdi), %ymm06419; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload6420; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]6421; AVX2-FCP-NEXT:    vmovaps 640(%rdi), %xmm66422; AVX2-FCP-NEXT:    vmovaps 592(%rdi), %xmm76423; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm6[0]6424; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6425; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6426; AVX2-FCP-NEXT:    vbroadcastsd 1120(%rdi), %ymm06427; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6428; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]6429; AVX2-FCP-NEXT:    vmovaps 1024(%rdi), %xmm86430; AVX2-FCP-NEXT:    vmovaps 976(%rdi), %xmm96431; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm8[0]6432; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6433; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6434; AVX2-FCP-NEXT:    vbroadcastsd 1504(%rdi), %ymm06435; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6436; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]6437; AVX2-FCP-NEXT:    vmovaps 1408(%rdi), %xmm106438; AVX2-FCP-NEXT:    vmovaps 1360(%rdi), %xmm116439; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm10[0]6440; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6441; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6442; AVX2-FCP-NEXT:    vbroadcastsd 160(%rdi), %ymm06443; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6444; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]6445; AVX2-FCP-NEXT:    vmovaps 16(%rdi), %xmm06446; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6447; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm26448; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6449; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm2[0]6450; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm1[4,5,6,7]6451; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6452; AVX2-FCP-NEXT:    vbroadcastsd 544(%rdi), %ymm16453; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6454; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]6455; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %xmm06456; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6457; AVX2-FCP-NEXT:    vmovaps 400(%rdi), %xmm16458; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6459; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm1[0],xmm0[0]6460; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm12[4,5,6,7]6461; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6462; AVX2-FCP-NEXT:    vbroadcastsd 928(%rdi), %ymm126463; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload6464; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm4[0],ymm12[0],ymm4[2],ymm12[2]6465; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %xmm126466; AVX2-FCP-NEXT:    vmovaps 784(%rdi), %xmm136467; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm13[0],xmm12[0]6468; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm14[4,5,6,7]6469; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6470; AVX2-FCP-NEXT:    vbroadcastsd 1312(%rdi), %ymm146471; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6472; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm14[0],ymm1[2],ymm14[2]6473; AVX2-FCP-NEXT:    vmovaps 1216(%rdi), %xmm156474; AVX2-FCP-NEXT:    vmovaps 1168(%rdi), %xmm06475; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm15[0]6476; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4,5,6,7]6477; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6478; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm3[1]6479; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm146480; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6481; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm14[1],ymm3[3],ymm14[3]6482; AVX2-FCP-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6483; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6484; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6485; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6486; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm7[1],xmm6[1]6487; AVX2-FCP-NEXT:    vmovaps 736(%rdi), %ymm76488; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload6489; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm7[1],ymm3[3],ymm7[3]6490; AVX2-FCP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6491; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6492; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6493; AVX2-FCP-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill6494; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm9[1],xmm8[1]6495; AVX2-FCP-NEXT:    vmovaps 1120(%rdi), %ymm66496; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6497; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm6[1],ymm3[3],ymm6[3]6498; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6499; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6500; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6501; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6502; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm11[1],xmm10[1]6503; AVX2-FCP-NEXT:    vmovaps 1504(%rdi), %ymm56504; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6505; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6506; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]6507; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm3 = ymm3[0,1,2,1]6508; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6509; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6510; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]6511; AVX2-FCP-NEXT:    vmovaps 1312(%rdi), %ymm36512; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]6513; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6514; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]6515; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6516; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6517; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm12[1]6518; AVX2-FCP-NEXT:    vmovaps 928(%rdi), %ymm16519; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]6520; AVX2-FCP-NEXT:    vmovaps %ymm1, %ymm46521; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6522; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1]6523; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6524; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6525; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6526; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6527; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]6528; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %ymm26529; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6530; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]6531; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6532; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6533; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6534; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6535; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6536; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6537; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]6538; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm96539; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6540; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]6541; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6542; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6543; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6544; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6545; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]6546; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6547; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm56548; AVX2-FCP-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6549; AVX2-FCP-NEXT:    vmovaps 80(%rdi), %xmm16550; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6551; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm1[0]6552; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6553; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6554; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6555; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]6556; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6557; AVX2-FCP-NEXT:    vmovaps 272(%rdi), %xmm56558; AVX2-FCP-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6559; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %xmm16560; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6561; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]6562; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6563; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6564; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6565; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]6566; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6567; AVX2-FCP-NEXT:    vmovaps 464(%rdi), %xmm146568; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %xmm136569; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]6570; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6571; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6572; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6573; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]6574; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6575; AVX2-FCP-NEXT:    vmovaps 656(%rdi), %xmm126576; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %xmm116577; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]6578; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6579; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6580; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6581; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]6582; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6583; AVX2-FCP-NEXT:    vmovaps 848(%rdi), %xmm106584; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %xmm76585; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm10[0]6586; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6587; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6588; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6589; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]6590; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6591; AVX2-FCP-NEXT:    vmovaps 1040(%rdi), %xmm86592; AVX2-FCP-NEXT:    vmovaps 992(%rdi), %xmm56593; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]6594; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6595; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6596; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6597; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]6598; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6599; AVX2-FCP-NEXT:    vmovaps 1232(%rdi), %xmm66600; AVX2-FCP-NEXT:    vmovaps 1184(%rdi), %xmm36601; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]6602; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6603; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6604; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6605; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload6606; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]6607; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]6608; AVX2-FCP-NEXT:    vmovaps 1424(%rdi), %xmm26609; AVX2-FCP-NEXT:    vmovaps 1376(%rdi), %xmm16610; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]6611; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]6612; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6613; AVX2-FCP-NEXT:    vbroadcastsd 136(%rdi), %ymm06614; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm9[1],ymm0[3],ymm9[3]6615; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload6616; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload6617; AVX2-FCP-NEXT:    # xmm9 = xmm9[1],mem[1]6618; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]6619; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6620; AVX2-FCP-NEXT:    vbroadcastsd 328(%rdi), %ymm06621; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6622; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6623; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload6624; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload6625; AVX2-FCP-NEXT:    # xmm9 = xmm9[1],mem[1]6626; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm15 = ymm9[0,1,2,3],ymm0[4,5,6,7]6627; AVX2-FCP-NEXT:    vbroadcastsd 520(%rdi), %ymm06628; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6629; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6630; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm13[1],xmm14[1]6631; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm9[0,1,2,3],ymm0[4,5,6,7]6632; AVX2-FCP-NEXT:    vbroadcastsd 712(%rdi), %ymm06633; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6634; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6635; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm11[1],xmm12[1]6636; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]6637; AVX2-FCP-NEXT:    vbroadcastsd 904(%rdi), %ymm06638; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6639; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6640; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm10[1]6641; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]6642; AVX2-FCP-NEXT:    vbroadcastsd 1096(%rdi), %ymm06643; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6644; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6645; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]6646; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]6647; AVX2-FCP-NEXT:    vbroadcastsd 1288(%rdi), %ymm06648; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6649; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]6650; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm6[1]6651; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]6652; AVX2-FCP-NEXT:    vbroadcastsd 1480(%rdi), %ymm06653; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm4[1],ymm0[3],ymm4[3]6654; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]6655; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6656; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6657; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rsi)6658; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6659; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rsi)6660; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6661; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rsi)6662; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6663; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rsi)6664; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6665; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rsi)6666; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6667; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rsi)6668; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6669; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rsi)6670; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6671; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rsi)6672; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6673; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rdx)6674; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6675; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rdx)6676; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6677; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rdx)6678; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6679; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rdx)6680; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6681; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rdx)6682; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6683; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rdx)6684; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6685; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rdx)6686; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6687; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rdx)6688; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6689; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rcx)6690; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6691; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rcx)6692; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6693; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rcx)6694; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6695; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rcx)6696; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6697; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rcx)6698; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6699; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rcx)6700; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6701; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rcx)6702; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6703; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rcx)6704; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6705; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r8)6706; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6707; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r8)6708; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6709; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r8)6710; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6711; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r8)6712; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6713; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r8)6714; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6715; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r8)6716; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload6717; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r8)6718; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6719; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r8)6720; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6721; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r9)6722; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6723; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r9)6724; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6725; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r9)6726; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6727; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r9)6728; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6729; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r9)6730; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6731; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r9)6732; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6733; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r9)6734; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6735; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r9)6736; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax6737; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)6738; AVX2-FCP-NEXT:    vmovaps %ymm3, 192(%rax)6739; AVX2-FCP-NEXT:    vmovaps %ymm5, 160(%rax)6740; AVX2-FCP-NEXT:    vmovaps %ymm7, 128(%rax)6741; AVX2-FCP-NEXT:    vmovaps %ymm9, 96(%rax)6742; AVX2-FCP-NEXT:    vmovaps %ymm13, 64(%rax)6743; AVX2-FCP-NEXT:    vmovaps %ymm15, 32(%rax)6744; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6745; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)6746; AVX2-FCP-NEXT:    addq $1496, %rsp # imm = 0x5D86747; AVX2-FCP-NEXT:    vzeroupper6748; AVX2-FCP-NEXT:    retq6749;6750; AVX512-LABEL: load_i64_stride6_vf32:6751; AVX512:       # %bb.0:6752; AVX512-NEXT:    subq $2632, %rsp # imm = 0xA486753; AVX512-NEXT:    vmovdqa64 1280(%rdi), %zmm26754; AVX512-NEXT:    vmovdqa64 1344(%rdi), %zmm216755; AVX512-NEXT:    vmovdqa64 896(%rdi), %zmm16756; AVX512-NEXT:    vmovdqa64 960(%rdi), %zmm196757; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm186758; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm36759; AVX512-NEXT:    vmovdqa64 512(%rdi), %zmm46760; AVX512-NEXT:    vmovdqa64 576(%rdi), %zmm06761; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm56762; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm256763; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]6764; AVX512-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]6765; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm76766; AVX512-NEXT:    vpermt2q %zmm5, %zmm6, %zmm76767; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6768; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]6769; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm86770; AVX512-NEXT:    vpermt2q %zmm4, %zmm6, %zmm86771; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6772; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm86773; AVX512-NEXT:    vpermt2q %zmm18, %zmm7, %zmm86774; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6775; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm86776; AVX512-NEXT:    vpermt2q %zmm1, %zmm6, %zmm86777; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6778; AVX512-NEXT:    vpermi2q %zmm2, %zmm21, %zmm66779; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6780; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]6781; AVX512-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]6782; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm86783; AVX512-NEXT:    vpermt2q %zmm1, %zmm6, %zmm86784; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6785; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm86786; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm66787; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]6788; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm16789; AVX512-NEXT:    vpermt2q %zmm4, %zmm8, %zmm16790; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6791; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm16792; AVX512-NEXT:    vpermt2q %zmm5, %zmm8, %zmm16793; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6794; AVX512-NEXT:    vpermi2q %zmm2, %zmm21, %zmm86795; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6796; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]6797; AVX512-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]6798; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm86799; AVX512-NEXT:    vpermt2q %zmm19, %zmm1, %zmm86800; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6801; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm86802; AVX512-NEXT:    vpermt2q %zmm0, %zmm1, %zmm86803; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6804; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm86805; AVX512-NEXT:    vpermt2q %zmm25, %zmm1, %zmm86806; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6807; AVX512-NEXT:    vpermi2q %zmm21, %zmm2, %zmm16808; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6809; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]6810; AVX512-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]6811; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm86812; AVX512-NEXT:    vpermt2q %zmm19, %zmm1, %zmm86813; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6814; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm86815; AVX512-NEXT:    vpermt2q %zmm0, %zmm1, %zmm86816; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6817; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm86818; AVX512-NEXT:    vpermt2q %zmm25, %zmm1, %zmm86819; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6820; AVX512-NEXT:    vpermi2q %zmm21, %zmm2, %zmm16821; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6822; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]6823; AVX512-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]6824; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm16825; AVX512-NEXT:    vpermt2q %zmm0, %zmm8, %zmm16826; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6827; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]6828; AVX512-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]6829; AVX512-NEXT:    vpermt2q %zmm0, %zmm1, %zmm46830; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6831; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm06832; AVX512-NEXT:    vpermt2q %zmm25, %zmm8, %zmm06833; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6834; AVX512-NEXT:    vpermt2q %zmm25, %zmm1, %zmm56835; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6836; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm06837; AVX512-NEXT:    vpermt2q %zmm19, %zmm8, %zmm06838; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6839; AVX512-NEXT:    vpermi2q %zmm21, %zmm2, %zmm86840; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6841; AVX512-NEXT:    vpermt2q %zmm21, %zmm1, %zmm26842; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6843; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm06844; AVX512-NEXT:    vpermt2q %zmm18, %zmm10, %zmm06845; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6846; AVX512-NEXT:    vpermt2q %zmm19, %zmm1, %zmm66847; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6848; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]6849; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm06850; AVX512-NEXT:    vpermt2q %zmm3, %zmm11, %zmm06851; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6852; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]6853; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm06854; AVX512-NEXT:    vpermt2q %zmm3, %zmm31, %zmm06855; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6856; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]6857; AVX512-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]6858; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm06859; AVX512-NEXT:    vpermt2q %zmm18, %zmm5, %zmm06860; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6861; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]6862; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]6863; AVX512-NEXT:    vpermt2q %zmm18, %zmm0, %zmm36864; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6865; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm256866; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm16867; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm296868; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm26869; AVX512-NEXT:    vpermt2q %zmm1, %zmm7, %zmm296870; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm36871; AVX512-NEXT:    vpermt2q %zmm1, %zmm10, %zmm36872; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6873; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm36874; AVX512-NEXT:    vpermt2q %zmm25, %zmm11, %zmm36875; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6876; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm36877; AVX512-NEXT:    vpermt2q %zmm25, %zmm31, %zmm36878; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6879; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm36880; AVX512-NEXT:    vpermt2q %zmm1, %zmm5, %zmm36881; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6882; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm256883; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm96884; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm276885; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm16886; AVX512-NEXT:    vpermt2q %zmm9, %zmm5, %zmm16887; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6888; AVX512-NEXT:    vmovdqa64 704(%rdi), %zmm306889; AVX512-NEXT:    vmovdqa64 640(%rdi), %zmm266890; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm16891; AVX512-NEXT:    vpermt2q %zmm30, %zmm5, %zmm16892; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6893; AVX512-NEXT:    vmovdqa64 1088(%rdi), %zmm76894; AVX512-NEXT:    vmovdqa64 1024(%rdi), %zmm206895; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm16896; AVX512-NEXT:    vpermt2q %zmm7, %zmm5, %zmm16897; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6898; AVX512-NEXT:    vmovdqa64 832(%rdi), %zmm16899; AVX512-NEXT:    vmovdqa64 768(%rdi), %zmm86900; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm156901; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm156902; AVX512-NEXT:    vmovdqa64 1472(%rdi), %zmm66903; AVX512-NEXT:    vmovdqa64 1408(%rdi), %zmm216904; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm36905; AVX512-NEXT:    vpermt2q %zmm6, %zmm5, %zmm36906; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6907; AVX512-NEXT:    vmovdqa64 1216(%rdi), %zmm196908; AVX512-NEXT:    vmovdqa64 1152(%rdi), %zmm46909; AVX512-NEXT:    vpermi2q %zmm19, %zmm4, %zmm26910; AVX512-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill6911; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm126912; AVX512-NEXT:    vpermt2q %zmm1, %zmm10, %zmm126913; AVX512-NEXT:    vpermi2q %zmm19, %zmm4, %zmm106914; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6915; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm136916; AVX512-NEXT:    vpermt2q %zmm8, %zmm11, %zmm136917; AVX512-NEXT:    vpermi2q %zmm4, %zmm19, %zmm116918; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm146919; AVX512-NEXT:    vpermt2q %zmm8, %zmm31, %zmm146920; AVX512-NEXT:    vpermi2q %zmm4, %zmm19, %zmm316921; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm26922; AVX512-NEXT:    vpermt2q %zmm1, %zmm5, %zmm26923; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6924; AVX512-NEXT:    vpermi2q %zmm19, %zmm4, %zmm56925; AVX512-NEXT:    vpermt2q %zmm19, %zmm0, %zmm46926; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm86927; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm196928; AVX512-NEXT:    vpermt2q %zmm7, %zmm0, %zmm196929; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm226930; AVX512-NEXT:    vpermt2q %zmm30, %zmm0, %zmm226931; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm176932; AVX512-NEXT:    vpermt2q %zmm9, %zmm0, %zmm176933; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm186934; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm186935; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]6936; AVX512-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]6937; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm246938; AVX512-NEXT:    vpermt2q %zmm30, %zmm1, %zmm246939; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]6940; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]6941; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm06942; AVX512-NEXT:    vpermt2q %zmm30, %zmm2, %zmm06943; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6944; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]6945; AVX512-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]6946; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm06947; AVX512-NEXT:    vpermt2q %zmm26, %zmm3, %zmm06948; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6949; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]6950; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]6951; AVX512-NEXT:    vpermt2q %zmm26, %zmm0, %zmm306952; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm266953; AVX512-NEXT:    vpermt2q %zmm9, %zmm1, %zmm266954; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm236955; AVX512-NEXT:    vpermt2q %zmm9, %zmm2, %zmm236956; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm286957; AVX512-NEXT:    vpermt2q %zmm27, %zmm3, %zmm286958; AVX512-NEXT:    vpermt2q %zmm27, %zmm0, %zmm96959; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm276960; AVX512-NEXT:    vpermt2q %zmm7, %zmm1, %zmm276961; AVX512-NEXT:    vpermi2q %zmm6, %zmm21, %zmm16962; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm106963; AVX512-NEXT:    vpermt2q %zmm7, %zmm2, %zmm106964; AVX512-NEXT:    vpermi2q %zmm6, %zmm21, %zmm26965; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm166966; AVX512-NEXT:    vpermt2q %zmm20, %zmm3, %zmm166967; AVX512-NEXT:    vpermi2q %zmm21, %zmm6, %zmm36968; AVX512-NEXT:    vpermt2q %zmm21, %zmm0, %zmm66969; AVX512-NEXT:    vpermt2q %zmm20, %zmm0, %zmm76970; AVX512-NEXT:    movb $56, %al6971; AVX512-NEXT:    kmovw %eax, %k16972; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6973; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}6974; AVX512-NEXT:    movb $-64, %al6975; AVX512-NEXT:    kmovw %eax, %k26976; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6977; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}6978; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6979; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload6980; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}6981; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload6982; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}6983; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6984; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6985; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}6986; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6987; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}6988; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload6989; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload6990; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}6991; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload6992; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}6993; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm206994; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6995; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}6996; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}6997; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6998; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload6999; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}7000; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}7001; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7002; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7003; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7004; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}7005; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7006; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload7007; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}7008; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}7009; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm177010; AVX512-NEXT:    movb $24, %al7011; AVX512-NEXT:    kmovw %eax, %k27012; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7013; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}7014; AVX512-NEXT:    movb $-32, %al7015; AVX512-NEXT:    kmovw %eax, %k17016; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}7017; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7018; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload7019; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}7020; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}7021; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm227022; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7023; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload7024; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}7025; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}7026; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm247027; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7028; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}7029; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}7030; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7031; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}7032; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}7033; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7034; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7035; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}7036; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7037; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}7038; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm187039; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7040; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7041; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}7042; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}7043; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm237044; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7045; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}7046; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}7047; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7048; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload7049; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7050; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7051; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload7052; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload7053; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}7054; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload7055; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload7056; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}7057; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload7058; AVX512-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm57059; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}7060; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload7061; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload7062; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}7063; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload7064; AVX512-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm107065; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}7066; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload7067; AVX512-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm47068; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}7069; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload7070; AVX512-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm67071; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}7072; AVX512-NEXT:    vmovdqa64 %zmm20, 192(%rsi)7073; AVX512-NEXT:    vmovdqa64 %zmm15, 128(%rsi)7074; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload7075; AVX512-NEXT:    vmovaps %zmm7, 64(%rsi)7076; AVX512-NEXT:    vmovdqa64 %zmm29, (%rsi)7077; AVX512-NEXT:    vmovdqa64 %zmm17, 192(%rdx)7078; AVX512-NEXT:    vmovdqa64 %zmm19, (%rdx)7079; AVX512-NEXT:    vmovdqa64 %zmm21, 64(%rdx)7080; AVX512-NEXT:    vmovdqa64 %zmm12, 128(%rdx)7081; AVX512-NEXT:    vmovdqa64 %zmm11, 192(%rcx)7082; AVX512-NEXT:    vmovdqa64 %zmm24, (%rcx)7083; AVX512-NEXT:    vmovdqa64 %zmm22, 64(%rcx)7084; AVX512-NEXT:    vmovdqa64 %zmm13, 128(%rcx)7085; AVX512-NEXT:    vmovdqa64 %zmm31, 192(%r8)7086; AVX512-NEXT:    vmovdqa64 %zmm23, (%r8)7087; AVX512-NEXT:    vmovdqa64 %zmm18, 64(%r8)7088; AVX512-NEXT:    vmovdqa64 %zmm14, 128(%r8)7089; AVX512-NEXT:    vmovdqa64 %zmm5, 192(%r9)7090; AVX512-NEXT:    vmovdqa64 %zmm2, (%r9)7091; AVX512-NEXT:    vmovdqa64 %zmm1, 64(%r9)7092; AVX512-NEXT:    vmovdqa64 %zmm0, 128(%r9)7093; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax7094; AVX512-NEXT:    vmovdqa64 %zmm6, 128(%rax)7095; AVX512-NEXT:    vmovdqa64 %zmm4, 192(%rax)7096; AVX512-NEXT:    vmovdqa64 %zmm10, (%rax)7097; AVX512-NEXT:    vmovdqa64 %zmm3, 64(%rax)7098; AVX512-NEXT:    addq $2632, %rsp # imm = 0xA487099; AVX512-NEXT:    vzeroupper7100; AVX512-NEXT:    retq7101;7102; AVX512-FCP-LABEL: load_i64_stride6_vf32:7103; AVX512-FCP:       # %bb.0:7104; AVX512-FCP-NEXT:    subq $2632, %rsp # imm = 0xA487105; AVX512-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm27106; AVX512-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm217107; AVX512-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm17108; AVX512-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm197109; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm187110; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm37111; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm47112; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm07113; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm57114; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm257115; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]7116; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]7117; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm77118; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm6, %zmm77119; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7120; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]7121; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm87122; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm6, %zmm87123; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7124; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm87125; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm7, %zmm87126; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7127; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm87128; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm87129; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7130; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm67131; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7132; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]7133; AVX512-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]7134; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm87135; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm87136; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7137; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm87138; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm67139; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]7140; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm17141; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm17142; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7143; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm17144; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm8, %zmm17145; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7146; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm87147; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7148; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]7149; AVX512-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7150; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm87151; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm87152; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7153; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm87154; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm87155; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7156; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm87157; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm87158; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7159; AVX512-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm17160; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7161; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]7162; AVX512-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7163; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm87164; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm87165; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7166; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm87167; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm87168; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7169; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm87170; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm87171; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7172; AVX512-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm17173; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7174; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]7175; AVX512-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]7176; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm17177; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm8, %zmm17178; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7179; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]7180; AVX512-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]7181; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm47182; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7183; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm07184; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm8, %zmm07185; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7186; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm57187; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7188; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm07189; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm8, %zmm07190; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7191; AVX512-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm87192; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7193; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm1, %zmm27194; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7195; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm07196; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm10, %zmm07197; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7198; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm67199; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7200; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]7201; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm07202; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm07203; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7204; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]7205; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm07206; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm31, %zmm07207; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7208; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]7209; AVX512-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7210; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm07211; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm5, %zmm07212; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7213; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]7214; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7215; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm37216; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7217; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm257218; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm17219; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm297220; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm27221; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm297222; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm37223; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm37224; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7225; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm37226; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm11, %zmm37227; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7228; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm37229; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm31, %zmm37230; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7231; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm37232; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm37233; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7234; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm257235; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm97236; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm277237; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm17238; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm17239; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7240; AVX512-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm307241; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm267242; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm17243; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm5, %zmm17244; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7245; AVX512-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm77246; AVX512-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm207247; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm17248; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm5, %zmm17249; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7250; AVX512-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm17251; AVX512-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm87252; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm157253; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm157254; AVX512-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm67255; AVX512-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm217256; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm37257; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm37258; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7259; AVX512-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm197260; AVX512-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm47261; AVX512-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm27262; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill7263; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm127264; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm127265; AVX512-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm107266; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7267; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm137268; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm11, %zmm137269; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm117270; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm147271; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm31, %zmm147272; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm317273; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm27274; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm27275; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7276; AVX512-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm57277; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm47278; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm87279; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm197280; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm197281; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm227282; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm227283; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm177284; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm0, %zmm177285; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm187286; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm187287; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]7288; AVX512-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]7289; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm247290; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm1, %zmm247291; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]7292; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]7293; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm07294; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm2, %zmm07295; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7296; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]7297; AVX512-FCP-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]7298; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm07299; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm3, %zmm07300; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7301; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]7302; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]7303; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm307304; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm267305; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm1, %zmm267306; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm237307; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm237308; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm287309; AVX512-FCP-NEXT:    vpermt2q %zmm27, %zmm3, %zmm287310; AVX512-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm97311; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm277312; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm1, %zmm277313; AVX512-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm17314; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm107315; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm107316; AVX512-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm27317; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm167318; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm3, %zmm167319; AVX512-FCP-NEXT:    vpermi2q %zmm21, %zmm6, %zmm37320; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm67321; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm77322; AVX512-FCP-NEXT:    movb $56, %al7323; AVX512-FCP-NEXT:    kmovw %eax, %k17324; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7325; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}7326; AVX512-FCP-NEXT:    movb $-64, %al7327; AVX512-FCP-NEXT:    kmovw %eax, %k27328; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7329; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}7330; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7331; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7332; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}7333; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7334; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}7335; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7336; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7337; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}7338; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7339; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}7340; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload7341; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7342; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}7343; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7344; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}7345; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm207346; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7347; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}7348; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}7349; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7350; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload7351; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}7352; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}7353; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7354; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7355; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7356; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}7357; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7358; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload7359; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}7360; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}7361; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm177362; AVX512-FCP-NEXT:    movb $24, %al7363; AVX512-FCP-NEXT:    kmovw %eax, %k27364; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7365; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}7366; AVX512-FCP-NEXT:    movb $-32, %al7367; AVX512-FCP-NEXT:    kmovw %eax, %k17368; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}7369; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7370; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload7371; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}7372; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}7373; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm227374; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7375; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload7376; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}7377; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}7378; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm247379; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7380; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}7381; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}7382; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7383; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}7384; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}7385; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7386; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7387; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}7388; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7389; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}7390; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm187391; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7392; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7393; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}7394; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}7395; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm237396; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7397; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}7398; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}7399; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7400; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload7401; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7402; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7403; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload7404; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload7405; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}7406; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload7407; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload7408; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}7409; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload7410; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm57411; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}7412; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload7413; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload7414; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}7415; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload7416; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm107417; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}7418; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload7419; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm47420; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}7421; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload7422; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm67423; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}7424; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rsi)7425; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 128(%rsi)7426; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload7427; AVX512-FCP-NEXT:    vmovaps %zmm7, 64(%rsi)7428; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, (%rsi)7429; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, 192(%rdx)7430; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, (%rdx)7431; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, 64(%rdx)7432; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 128(%rdx)7433; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rcx)7434; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, (%rcx)7435; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, 64(%rcx)7436; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rcx)7437; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, 192(%r8)7438; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, (%r8)7439; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 64(%r8)7440; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 128(%r8)7441; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r9)7442; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%r9)7443; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r9)7444; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 128(%r9)7445; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax7446; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 128(%rax)7447; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 192(%rax)7448; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)7449; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)7450; AVX512-FCP-NEXT:    addq $2632, %rsp # imm = 0xA487451; AVX512-FCP-NEXT:    vzeroupper7452; AVX512-FCP-NEXT:    retq7453;7454; AVX512DQ-LABEL: load_i64_stride6_vf32:7455; AVX512DQ:       # %bb.0:7456; AVX512DQ-NEXT:    subq $2632, %rsp # imm = 0xA487457; AVX512DQ-NEXT:    vmovdqa64 1280(%rdi), %zmm27458; AVX512DQ-NEXT:    vmovdqa64 1344(%rdi), %zmm217459; AVX512DQ-NEXT:    vmovdqa64 896(%rdi), %zmm17460; AVX512DQ-NEXT:    vmovdqa64 960(%rdi), %zmm197461; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm187462; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm37463; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %zmm47464; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %zmm07465; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm57466; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm257467; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]7468; AVX512DQ-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]7469; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm77470; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm6, %zmm77471; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7472; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]7473; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm87474; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm6, %zmm87475; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7476; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm87477; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm7, %zmm87478; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7479; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm87480; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm6, %zmm87481; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7482; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm21, %zmm67483; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7484; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]7485; AVX512DQ-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]7486; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm87487; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm6, %zmm87488; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7489; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm87490; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm67491; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]7492; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm17493; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm8, %zmm17494; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7495; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm17496; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm8, %zmm17497; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7498; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm21, %zmm87499; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7500; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]7501; AVX512DQ-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7502; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm87503; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm1, %zmm87504; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7505; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm87506; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm1, %zmm87507; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7508; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm87509; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm1, %zmm87510; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7511; AVX512DQ-NEXT:    vpermi2q %zmm21, %zmm2, %zmm17512; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7513; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]7514; AVX512DQ-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7515; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm87516; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm1, %zmm87517; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7518; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm87519; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm1, %zmm87520; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7521; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm87522; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm1, %zmm87523; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7524; AVX512DQ-NEXT:    vpermi2q %zmm21, %zmm2, %zmm17525; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7526; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]7527; AVX512DQ-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]7528; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm17529; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm8, %zmm17530; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7531; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]7532; AVX512DQ-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]7533; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm1, %zmm47534; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7535; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm07536; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm8, %zmm07537; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7538; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm1, %zmm57539; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7540; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm07541; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm8, %zmm07542; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7543; AVX512DQ-NEXT:    vpermi2q %zmm21, %zmm2, %zmm87544; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7545; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm1, %zmm27546; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7547; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm07548; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm10, %zmm07549; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7550; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm1, %zmm67551; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7552; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]7553; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm07554; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm11, %zmm07555; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7556; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]7557; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm07558; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm31, %zmm07559; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7560; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]7561; AVX512DQ-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7562; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm07563; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm5, %zmm07564; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7565; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]7566; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7567; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm0, %zmm37568; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7569; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm257570; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm17571; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm297572; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm27573; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm7, %zmm297574; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm37575; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm10, %zmm37576; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7577; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm37578; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm11, %zmm37579; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7580; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm37581; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm31, %zmm37582; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7583; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm37584; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm5, %zmm37585; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7586; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm257587; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm97588; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm277589; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm17590; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm5, %zmm17591; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7592; AVX512DQ-NEXT:    vmovdqa64 704(%rdi), %zmm307593; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %zmm267594; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm17595; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm5, %zmm17596; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7597; AVX512DQ-NEXT:    vmovdqa64 1088(%rdi), %zmm77598; AVX512DQ-NEXT:    vmovdqa64 1024(%rdi), %zmm207599; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm17600; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm5, %zmm17601; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7602; AVX512DQ-NEXT:    vmovdqa64 832(%rdi), %zmm17603; AVX512DQ-NEXT:    vmovdqa64 768(%rdi), %zmm87604; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm157605; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm157606; AVX512DQ-NEXT:    vmovdqa64 1472(%rdi), %zmm67607; AVX512DQ-NEXT:    vmovdqa64 1408(%rdi), %zmm217608; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm37609; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm5, %zmm37610; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7611; AVX512DQ-NEXT:    vmovdqa64 1216(%rdi), %zmm197612; AVX512DQ-NEXT:    vmovdqa64 1152(%rdi), %zmm47613; AVX512DQ-NEXT:    vpermi2q %zmm19, %zmm4, %zmm27614; AVX512DQ-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill7615; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm127616; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm10, %zmm127617; AVX512DQ-NEXT:    vpermi2q %zmm19, %zmm4, %zmm107618; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7619; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm137620; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm11, %zmm137621; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm19, %zmm117622; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm147623; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm31, %zmm147624; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm19, %zmm317625; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm27626; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm5, %zmm27627; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7628; AVX512DQ-NEXT:    vpermi2q %zmm19, %zmm4, %zmm57629; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm0, %zmm47630; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm87631; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm197632; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm0, %zmm197633; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm227634; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm0, %zmm227635; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm177636; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm0, %zmm177637; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm187638; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm187639; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]7640; AVX512DQ-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]7641; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm247642; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm1, %zmm247643; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]7644; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]7645; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm07646; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm2, %zmm07647; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7648; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]7649; AVX512DQ-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]7650; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm07651; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm3, %zmm07652; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7653; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]7654; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]7655; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm0, %zmm307656; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm267657; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm1, %zmm267658; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm237659; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm2, %zmm237660; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm287661; AVX512DQ-NEXT:    vpermt2q %zmm27, %zmm3, %zmm287662; AVX512DQ-NEXT:    vpermt2q %zmm27, %zmm0, %zmm97663; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm277664; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm1, %zmm277665; AVX512DQ-NEXT:    vpermi2q %zmm6, %zmm21, %zmm17666; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm107667; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm2, %zmm107668; AVX512DQ-NEXT:    vpermi2q %zmm6, %zmm21, %zmm27669; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm167670; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm3, %zmm167671; AVX512DQ-NEXT:    vpermi2q %zmm21, %zmm6, %zmm37672; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm0, %zmm67673; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm0, %zmm77674; AVX512DQ-NEXT:    movb $56, %al7675; AVX512DQ-NEXT:    kmovw %eax, %k17676; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7677; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}7678; AVX512DQ-NEXT:    movb $-64, %al7679; AVX512DQ-NEXT:    kmovw %eax, %k27680; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7681; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}7682; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7683; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7684; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}7685; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7686; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}7687; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7688; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7689; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}7690; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7691; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}7692; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload7693; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7694; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}7695; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload7696; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}7697; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm207698; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7699; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}7700; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}7701; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7702; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload7703; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}7704; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}7705; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7706; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7707; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7708; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}7709; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7710; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload7711; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}7712; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}7713; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm177714; AVX512DQ-NEXT:    movb $24, %al7715; AVX512DQ-NEXT:    kmovw %eax, %k27716; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7717; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}7718; AVX512DQ-NEXT:    movb $-32, %al7719; AVX512DQ-NEXT:    kmovw %eax, %k17720; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}7721; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7722; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload7723; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}7724; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}7725; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm227726; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7727; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload7728; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}7729; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}7730; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm247731; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7732; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}7733; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}7734; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7735; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}7736; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}7737; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7738; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7739; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}7740; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7741; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}7742; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm187743; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7744; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7745; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}7746; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}7747; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm237748; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7749; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}7750; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}7751; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7752; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload7753; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7754; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload7755; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload7756; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload7757; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}7758; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload7759; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload7760; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}7761; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload7762; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm57763; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}7764; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload7765; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload7766; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}7767; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload7768; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm107769; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}7770; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload7771; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm47772; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}7773; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload7774; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm67775; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}7776; AVX512DQ-NEXT:    vmovdqa64 %zmm20, 192(%rsi)7777; AVX512DQ-NEXT:    vmovdqa64 %zmm15, 128(%rsi)7778; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload7779; AVX512DQ-NEXT:    vmovaps %zmm7, 64(%rsi)7780; AVX512DQ-NEXT:    vmovdqa64 %zmm29, (%rsi)7781; AVX512DQ-NEXT:    vmovdqa64 %zmm17, 192(%rdx)7782; AVX512DQ-NEXT:    vmovdqa64 %zmm19, (%rdx)7783; AVX512DQ-NEXT:    vmovdqa64 %zmm21, 64(%rdx)7784; AVX512DQ-NEXT:    vmovdqa64 %zmm12, 128(%rdx)7785; AVX512DQ-NEXT:    vmovdqa64 %zmm11, 192(%rcx)7786; AVX512DQ-NEXT:    vmovdqa64 %zmm24, (%rcx)7787; AVX512DQ-NEXT:    vmovdqa64 %zmm22, 64(%rcx)7788; AVX512DQ-NEXT:    vmovdqa64 %zmm13, 128(%rcx)7789; AVX512DQ-NEXT:    vmovdqa64 %zmm31, 192(%r8)7790; AVX512DQ-NEXT:    vmovdqa64 %zmm23, (%r8)7791; AVX512DQ-NEXT:    vmovdqa64 %zmm18, 64(%r8)7792; AVX512DQ-NEXT:    vmovdqa64 %zmm14, 128(%r8)7793; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 192(%r9)7794; AVX512DQ-NEXT:    vmovdqa64 %zmm2, (%r9)7795; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 64(%r9)7796; AVX512DQ-NEXT:    vmovdqa64 %zmm0, 128(%r9)7797; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax7798; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 128(%rax)7799; AVX512DQ-NEXT:    vmovdqa64 %zmm4, 192(%rax)7800; AVX512DQ-NEXT:    vmovdqa64 %zmm10, (%rax)7801; AVX512DQ-NEXT:    vmovdqa64 %zmm3, 64(%rax)7802; AVX512DQ-NEXT:    addq $2632, %rsp # imm = 0xA487803; AVX512DQ-NEXT:    vzeroupper7804; AVX512DQ-NEXT:    retq7805;7806; AVX512DQ-FCP-LABEL: load_i64_stride6_vf32:7807; AVX512DQ-FCP:       # %bb.0:7808; AVX512DQ-FCP-NEXT:    subq $2632, %rsp # imm = 0xA487809; AVX512DQ-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm27810; AVX512DQ-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm217811; AVX512DQ-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm17812; AVX512DQ-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm197813; AVX512DQ-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm187814; AVX512DQ-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm37815; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm47816; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm07817; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm57818; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm257819; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]7820; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]7821; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm77822; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm6, %zmm77823; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7824; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]7825; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm87826; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm6, %zmm87827; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7828; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm87829; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm7, %zmm87830; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7831; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm87832; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm87833; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7834; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm67835; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7836; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]7837; AVX512DQ-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]7838; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm87839; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm87840; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7841; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm87842; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm67843; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]7844; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm17845; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm17846; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7847; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm17848; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm8, %zmm17849; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7850; AVX512DQ-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm87851; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7852; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]7853; AVX512DQ-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7854; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm87855; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm87856; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7857; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm87858; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm87859; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7860; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm87861; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm87862; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7863; AVX512DQ-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm17864; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7865; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]7866; AVX512DQ-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7867; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm87868; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm87869; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7870; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm87871; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm87872; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7873; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm87874; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm87875; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7876; AVX512DQ-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm17877; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7878; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]7879; AVX512DQ-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]7880; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm17881; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm8, %zmm17882; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7883; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]7884; AVX512DQ-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]7885; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm47886; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7887; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm07888; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm8, %zmm07889; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7890; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm57891; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7892; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm07893; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm8, %zmm07894; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7895; AVX512DQ-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm87896; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7897; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm1, %zmm27898; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7899; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm07900; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm10, %zmm07901; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7902; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm67903; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7904; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]7905; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm07906; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm07907; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7908; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]7909; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm07910; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm31, %zmm07911; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7912; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]7913; AVX512DQ-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7914; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm07915; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm5, %zmm07916; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7917; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]7918; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7919; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm37920; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7921; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm257922; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm17923; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm297924; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm27925; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm297926; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm37927; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm37928; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7929; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm37930; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm11, %zmm37931; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7932; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm37933; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm31, %zmm37934; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7935; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm37936; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm37937; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7938; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm257939; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm97940; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm277941; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm17942; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm17943; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7944; AVX512DQ-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm307945; AVX512DQ-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm267946; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm17947; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm5, %zmm17948; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7949; AVX512DQ-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm77950; AVX512DQ-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm207951; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm17952; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm5, %zmm17953; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7954; AVX512DQ-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm17955; AVX512DQ-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm87956; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm157957; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm157958; AVX512DQ-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm67959; AVX512DQ-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm217960; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm37961; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm37962; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7963; AVX512DQ-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm197964; AVX512DQ-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm47965; AVX512DQ-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm27966; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill7967; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm127968; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm127969; AVX512DQ-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm107970; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7971; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm137972; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm11, %zmm137973; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm117974; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm147975; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm31, %zmm147976; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm317977; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm27978; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm27979; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7980; AVX512DQ-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm57981; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm47982; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm87983; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm197984; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm197985; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm227986; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm227987; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm177988; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm0, %zmm177989; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm187990; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm187991; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]7992; AVX512DQ-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]7993; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm247994; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm1, %zmm247995; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]7996; AVX512DQ-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]7997; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm07998; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm2, %zmm07999; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8000; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]8001; AVX512DQ-FCP-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]8002; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm08003; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm3, %zmm08004; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8005; AVX512DQ-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]8006; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]8007; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm308008; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm268009; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm1, %zmm268010; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm238011; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm238012; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm288013; AVX512DQ-FCP-NEXT:    vpermt2q %zmm27, %zmm3, %zmm288014; AVX512DQ-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm98015; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm278016; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm1, %zmm278017; AVX512DQ-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm18018; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm108019; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm108020; AVX512DQ-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm28021; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm168022; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm3, %zmm168023; AVX512DQ-FCP-NEXT:    vpermi2q %zmm21, %zmm6, %zmm38024; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm68025; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm78026; AVX512DQ-FCP-NEXT:    movb $56, %al8027; AVX512DQ-FCP-NEXT:    kmovw %eax, %k18028; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8029; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}8030; AVX512DQ-FCP-NEXT:    movb $-64, %al8031; AVX512DQ-FCP-NEXT:    kmovw %eax, %k28032; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8033; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}8034; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8035; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8036; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}8037; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8038; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}8039; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8040; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8041; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}8042; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8043; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}8044; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload8045; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8046; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}8047; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8048; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}8049; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm208050; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8051; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}8052; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}8053; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8054; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload8055; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}8056; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}8057; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8058; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8059; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8060; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}8061; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8062; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload8063; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}8064; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}8065; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm178066; AVX512DQ-FCP-NEXT:    movb $24, %al8067; AVX512DQ-FCP-NEXT:    kmovw %eax, %k28068; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8069; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}8070; AVX512DQ-FCP-NEXT:    movb $-32, %al8071; AVX512DQ-FCP-NEXT:    kmovw %eax, %k18072; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}8073; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8074; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload8075; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}8076; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}8077; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm228078; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8079; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload8080; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}8081; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}8082; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm248083; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8084; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}8085; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}8086; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8087; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}8088; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}8089; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8090; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8091; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}8092; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8093; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}8094; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm188095; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8096; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8097; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}8098; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}8099; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm238100; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8101; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}8102; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}8103; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8104; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload8105; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8106; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8107; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload8108; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8109; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}8110; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8111; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload8112; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}8113; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload8114; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm58115; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}8116; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8117; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload8118; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}8119; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload8120; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm108121; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}8122; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload8123; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm48124; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}8125; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload8126; AVX512DQ-FCP-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm68127; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}8128; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rsi)8129; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 128(%rsi)8130; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8131; AVX512DQ-FCP-NEXT:    vmovaps %zmm7, 64(%rsi)8132; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, (%rsi)8133; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, 192(%rdx)8134; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, (%rdx)8135; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, 64(%rdx)8136; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 128(%rdx)8137; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rcx)8138; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, (%rcx)8139; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, 64(%rcx)8140; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rcx)8141; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, 192(%r8)8142; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, (%r8)8143; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 64(%r8)8144; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 128(%r8)8145; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r9)8146; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%r9)8147; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r9)8148; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, 128(%r9)8149; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax8150; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 128(%rax)8151; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 192(%rax)8152; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)8153; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)8154; AVX512DQ-FCP-NEXT:    addq $2632, %rsp # imm = 0xA488155; AVX512DQ-FCP-NEXT:    vzeroupper8156; AVX512DQ-FCP-NEXT:    retq8157;8158; AVX512BW-LABEL: load_i64_stride6_vf32:8159; AVX512BW:       # %bb.0:8160; AVX512BW-NEXT:    subq $2632, %rsp # imm = 0xA488161; AVX512BW-NEXT:    vmovdqa64 1280(%rdi), %zmm28162; AVX512BW-NEXT:    vmovdqa64 1344(%rdi), %zmm218163; AVX512BW-NEXT:    vmovdqa64 896(%rdi), %zmm18164; AVX512BW-NEXT:    vmovdqa64 960(%rdi), %zmm198165; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm188166; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm38167; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm48168; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm08169; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm58170; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm258171; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]8172; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]8173; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm78174; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm6, %zmm78175; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8176; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]8177; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm88178; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm6, %zmm88179; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8180; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm88181; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm7, %zmm88182; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8183; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm88184; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm6, %zmm88185; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8186; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm21, %zmm68187; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8188; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]8189; AVX512BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]8190; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm88191; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm6, %zmm88192; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8193; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm88194; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm68195; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]8196; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm18197; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm18198; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8199; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm18200; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm8, %zmm18201; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8202; AVX512BW-NEXT:    vpermi2q %zmm2, %zmm21, %zmm88203; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8204; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]8205; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8206; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm88207; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm88208; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8209; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm88210; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm88211; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8212; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm88213; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm88214; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8215; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm18216; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8217; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]8218; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8219; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm88220; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm88221; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8222; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm88223; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm88224; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8225; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm88226; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm88227; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8228; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm18229; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8230; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]8231; AVX512BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]8232; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm18233; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm8, %zmm18234; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8235; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]8236; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]8237; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm48238; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8239; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm08240; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm8, %zmm08241; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8242; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm58243; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8244; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm08245; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm8, %zmm08246; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8247; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm88248; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8249; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm1, %zmm28250; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8251; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm08252; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm10, %zmm08253; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8254; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm68255; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8256; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]8257; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm08258; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm11, %zmm08259; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8260; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]8261; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm08262; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm31, %zmm08263; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8264; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]8265; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8266; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm08267; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm5, %zmm08268; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8269; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]8270; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8271; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm38272; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8273; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm258274; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm18275; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm298276; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm28277; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm298278; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm38279; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm38280; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8281; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm38282; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm11, %zmm38283; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8284; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm38285; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm31, %zmm38286; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8287; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm38288; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm5, %zmm38289; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8290; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm258291; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm98292; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm278293; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm18294; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm5, %zmm18295; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8296; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm308297; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm268298; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm18299; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm5, %zmm18300; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8301; AVX512BW-NEXT:    vmovdqa64 1088(%rdi), %zmm78302; AVX512BW-NEXT:    vmovdqa64 1024(%rdi), %zmm208303; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm18304; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm5, %zmm18305; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8306; AVX512BW-NEXT:    vmovdqa64 832(%rdi), %zmm18307; AVX512BW-NEXT:    vmovdqa64 768(%rdi), %zmm88308; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm158309; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm158310; AVX512BW-NEXT:    vmovdqa64 1472(%rdi), %zmm68311; AVX512BW-NEXT:    vmovdqa64 1408(%rdi), %zmm218312; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm38313; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm5, %zmm38314; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8315; AVX512BW-NEXT:    vmovdqa64 1216(%rdi), %zmm198316; AVX512BW-NEXT:    vmovdqa64 1152(%rdi), %zmm48317; AVX512BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm28318; AVX512BW-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill8319; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm128320; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm128321; AVX512BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm108322; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8323; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm138324; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm11, %zmm138325; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm19, %zmm118326; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm148327; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm31, %zmm148328; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm19, %zmm318329; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm28330; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm5, %zmm28331; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8332; AVX512BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm58333; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm48334; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm88335; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm198336; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm198337; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm228338; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm228339; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm178340; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm0, %zmm178341; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm188342; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm188343; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]8344; AVX512BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]8345; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm248346; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm1, %zmm248347; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]8348; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]8349; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm08350; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm2, %zmm08351; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8352; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]8353; AVX512BW-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]8354; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm08355; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm3, %zmm08356; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8357; AVX512BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]8358; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]8359; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm308360; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm268361; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm1, %zmm268362; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm238363; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm238364; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm288365; AVX512BW-NEXT:    vpermt2q %zmm27, %zmm3, %zmm288366; AVX512BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm98367; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm278368; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm1, %zmm278369; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm21, %zmm18370; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm108371; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm108372; AVX512BW-NEXT:    vpermi2q %zmm6, %zmm21, %zmm28373; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm168374; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm3, %zmm168375; AVX512BW-NEXT:    vpermi2q %zmm21, %zmm6, %zmm38376; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm68377; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm78378; AVX512BW-NEXT:    movb $56, %al8379; AVX512BW-NEXT:    kmovd %eax, %k18380; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8381; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}8382; AVX512BW-NEXT:    movb $-64, %al8383; AVX512BW-NEXT:    kmovd %eax, %k28384; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8385; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}8386; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8387; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8388; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}8389; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8390; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}8391; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8392; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8393; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}8394; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8395; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}8396; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload8397; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8398; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}8399; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8400; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}8401; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm208402; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8403; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}8404; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}8405; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8406; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload8407; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}8408; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}8409; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8410; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8411; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8412; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}8413; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8414; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload8415; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}8416; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}8417; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm178418; AVX512BW-NEXT:    movb $24, %al8419; AVX512BW-NEXT:    kmovd %eax, %k28420; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8421; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}8422; AVX512BW-NEXT:    movb $-32, %al8423; AVX512BW-NEXT:    kmovd %eax, %k18424; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}8425; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8426; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload8427; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}8428; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}8429; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm228430; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8431; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload8432; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}8433; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}8434; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm248435; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8436; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}8437; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}8438; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8439; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}8440; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}8441; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8442; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8443; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}8444; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8445; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}8446; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm188447; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8448; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8449; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}8450; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}8451; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm238452; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8453; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}8454; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}8455; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8456; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload8457; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8458; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8459; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload8460; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8461; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}8462; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8463; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload8464; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}8465; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload8466; AVX512BW-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm58467; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}8468; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8469; AVX512BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload8470; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}8471; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload8472; AVX512BW-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm108473; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}8474; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload8475; AVX512BW-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm48476; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}8477; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload8478; AVX512BW-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm68479; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}8480; AVX512BW-NEXT:    vmovdqa64 %zmm20, 192(%rsi)8481; AVX512BW-NEXT:    vmovdqa64 %zmm15, 128(%rsi)8482; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8483; AVX512BW-NEXT:    vmovaps %zmm7, 64(%rsi)8484; AVX512BW-NEXT:    vmovdqa64 %zmm29, (%rsi)8485; AVX512BW-NEXT:    vmovdqa64 %zmm17, 192(%rdx)8486; AVX512BW-NEXT:    vmovdqa64 %zmm19, (%rdx)8487; AVX512BW-NEXT:    vmovdqa64 %zmm21, 64(%rdx)8488; AVX512BW-NEXT:    vmovdqa64 %zmm12, 128(%rdx)8489; AVX512BW-NEXT:    vmovdqa64 %zmm11, 192(%rcx)8490; AVX512BW-NEXT:    vmovdqa64 %zmm24, (%rcx)8491; AVX512BW-NEXT:    vmovdqa64 %zmm22, 64(%rcx)8492; AVX512BW-NEXT:    vmovdqa64 %zmm13, 128(%rcx)8493; AVX512BW-NEXT:    vmovdqa64 %zmm31, 192(%r8)8494; AVX512BW-NEXT:    vmovdqa64 %zmm23, (%r8)8495; AVX512BW-NEXT:    vmovdqa64 %zmm18, 64(%r8)8496; AVX512BW-NEXT:    vmovdqa64 %zmm14, 128(%r8)8497; AVX512BW-NEXT:    vmovdqa64 %zmm5, 192(%r9)8498; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%r9)8499; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%r9)8500; AVX512BW-NEXT:    vmovdqa64 %zmm0, 128(%r9)8501; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax8502; AVX512BW-NEXT:    vmovdqa64 %zmm6, 128(%rax)8503; AVX512BW-NEXT:    vmovdqa64 %zmm4, 192(%rax)8504; AVX512BW-NEXT:    vmovdqa64 %zmm10, (%rax)8505; AVX512BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)8506; AVX512BW-NEXT:    addq $2632, %rsp # imm = 0xA488507; AVX512BW-NEXT:    vzeroupper8508; AVX512BW-NEXT:    retq8509;8510; AVX512BW-FCP-LABEL: load_i64_stride6_vf32:8511; AVX512BW-FCP:       # %bb.0:8512; AVX512BW-FCP-NEXT:    subq $2632, %rsp # imm = 0xA488513; AVX512BW-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm28514; AVX512BW-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm218515; AVX512BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm18516; AVX512BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm198517; AVX512BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm188518; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm38519; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm48520; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm08521; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm58522; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm258523; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]8524; AVX512BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]8525; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm78526; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm6, %zmm78527; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8528; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]8529; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm88530; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm6, %zmm88531; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8532; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm88533; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm7, %zmm88534; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8535; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm88536; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm88537; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8538; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm68539; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8540; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]8541; AVX512BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]8542; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm88543; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm88544; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8545; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm88546; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm68547; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]8548; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18549; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm18550; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8551; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm18552; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm8, %zmm18553; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8554; AVX512BW-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm88555; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8556; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]8557; AVX512BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8558; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm88559; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm88560; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8561; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm88562; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm88563; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8564; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm88565; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm88566; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8567; AVX512BW-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm18568; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8569; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]8570; AVX512BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8571; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm88572; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm88573; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8574; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm88575; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm88576; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8577; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm88578; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm88579; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8580; AVX512BW-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm18581; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8582; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]8583; AVX512BW-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]8584; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm18585; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm8, %zmm18586; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8587; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]8588; AVX512BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]8589; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm48590; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8591; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm08592; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm8, %zmm08593; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8594; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm58595; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8596; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm08597; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm8, %zmm08598; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8599; AVX512BW-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm88600; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8601; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm1, %zmm28602; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8603; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm08604; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm10, %zmm08605; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8606; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm68607; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8608; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]8609; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm08610; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm08611; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8612; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]8613; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm08614; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm31, %zmm08615; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8616; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]8617; AVX512BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8618; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm08619; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm5, %zmm08620; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8621; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]8622; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8623; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm38624; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8625; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm258626; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm18627; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm298628; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm28629; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm298630; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm38631; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm38632; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8633; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm38634; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm11, %zmm38635; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8636; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm38637; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm31, %zmm38638; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8639; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm38640; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm38641; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8642; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm258643; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm98644; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm278645; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm18646; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm18647; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8648; AVX512BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm308649; AVX512BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm268650; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm18651; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm5, %zmm18652; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8653; AVX512BW-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm78654; AVX512BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm208655; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18656; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm5, %zmm18657; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8658; AVX512BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm18659; AVX512BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm88660; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm158661; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm158662; AVX512BW-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm68663; AVX512BW-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm218664; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm38665; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm38666; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8667; AVX512BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm198668; AVX512BW-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm48669; AVX512BW-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm28670; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill8671; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm128672; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm128673; AVX512BW-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm108674; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8675; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm138676; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm11, %zmm138677; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm118678; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm148679; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm31, %zmm148680; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm318681; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm28682; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm28683; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8684; AVX512BW-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm58685; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm48686; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm88687; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm198688; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm198689; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm228690; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm228691; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm178692; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm0, %zmm178693; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm188694; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm188695; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]8696; AVX512BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]8697; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm248698; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm1, %zmm248699; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]8700; AVX512BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]8701; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm08702; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm2, %zmm08703; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8704; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]8705; AVX512BW-FCP-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]8706; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm08707; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm3, %zmm08708; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8709; AVX512BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]8710; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]8711; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm308712; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm268713; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm1, %zmm268714; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm238715; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm238716; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm288717; AVX512BW-FCP-NEXT:    vpermt2q %zmm27, %zmm3, %zmm288718; AVX512BW-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm98719; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm278720; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm1, %zmm278721; AVX512BW-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm18722; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm108723; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm108724; AVX512BW-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm28725; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm168726; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm3, %zmm168727; AVX512BW-FCP-NEXT:    vpermi2q %zmm21, %zmm6, %zmm38728; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm68729; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm78730; AVX512BW-FCP-NEXT:    movb $56, %al8731; AVX512BW-FCP-NEXT:    kmovd %eax, %k18732; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8733; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}8734; AVX512BW-FCP-NEXT:    movb $-64, %al8735; AVX512BW-FCP-NEXT:    kmovd %eax, %k28736; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8737; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}8738; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8739; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8740; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}8741; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8742; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}8743; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8744; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8745; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}8746; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8747; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}8748; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload8749; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8750; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}8751; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload8752; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}8753; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm208754; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8755; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}8756; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}8757; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8758; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload8759; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}8760; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}8761; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8762; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8763; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8764; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}8765; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8766; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload8767; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}8768; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}8769; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm178770; AVX512BW-FCP-NEXT:    movb $24, %al8771; AVX512BW-FCP-NEXT:    kmovd %eax, %k28772; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8773; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}8774; AVX512BW-FCP-NEXT:    movb $-32, %al8775; AVX512BW-FCP-NEXT:    kmovd %eax, %k18776; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}8777; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8778; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload8779; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}8780; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}8781; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm228782; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8783; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload8784; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}8785; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}8786; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm248787; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8788; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}8789; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}8790; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8791; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}8792; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}8793; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8794; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8795; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}8796; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8797; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}8798; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm188799; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8800; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8801; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}8802; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}8803; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm238804; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8805; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}8806; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}8807; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8808; AVX512BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload8809; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8810; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload8811; AVX512BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload8812; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8813; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}8814; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8815; AVX512BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload8816; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}8817; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload8818; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm58819; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}8820; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8821; AVX512BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload8822; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}8823; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload8824; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm108825; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}8826; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload8827; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm48828; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}8829; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload8830; AVX512BW-FCP-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm68831; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}8832; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rsi)8833; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, 128(%rsi)8834; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8835; AVX512BW-FCP-NEXT:    vmovaps %zmm7, 64(%rsi)8836; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, (%rsi)8837; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, 192(%rdx)8838; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, (%rdx)8839; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, 64(%rdx)8840; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, 128(%rdx)8841; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rcx)8842; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, (%rcx)8843; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, 64(%rcx)8844; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rcx)8845; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, 192(%r8)8846; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, (%r8)8847; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, 64(%r8)8848; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, 128(%r8)8849; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r9)8850; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%r9)8851; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r9)8852; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, 128(%r9)8853; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax8854; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 128(%rax)8855; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, 192(%rax)8856; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)8857; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)8858; AVX512BW-FCP-NEXT:    addq $2632, %rsp # imm = 0xA488859; AVX512BW-FCP-NEXT:    vzeroupper8860; AVX512BW-FCP-NEXT:    retq8861;8862; AVX512DQ-BW-LABEL: load_i64_stride6_vf32:8863; AVX512DQ-BW:       # %bb.0:8864; AVX512DQ-BW-NEXT:    subq $2632, %rsp # imm = 0xA488865; AVX512DQ-BW-NEXT:    vmovdqa64 1280(%rdi), %zmm28866; AVX512DQ-BW-NEXT:    vmovdqa64 1344(%rdi), %zmm218867; AVX512DQ-BW-NEXT:    vmovdqa64 896(%rdi), %zmm18868; AVX512DQ-BW-NEXT:    vmovdqa64 960(%rdi), %zmm198869; AVX512DQ-BW-NEXT:    vmovdqa64 448(%rdi), %zmm188870; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm38871; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %zmm48872; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %zmm08873; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm58874; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm258875; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]8876; AVX512DQ-BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]8877; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm78878; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm6, %zmm78879; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8880; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]8881; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm88882; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm6, %zmm88883; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8884; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm88885; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm7, %zmm88886; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8887; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm88888; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm6, %zmm88889; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8890; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm21, %zmm68891; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8892; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]8893; AVX512DQ-BW-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]8894; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm88895; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm6, %zmm88896; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8897; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm88898; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm68899; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]8900; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm18901; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm18902; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8903; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm18904; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm8, %zmm18905; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8906; AVX512DQ-BW-NEXT:    vpermi2q %zmm2, %zmm21, %zmm88907; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8908; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]8909; AVX512DQ-BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8910; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm88911; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm88912; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8913; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm88914; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm88915; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8916; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm88917; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm88918; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8919; AVX512DQ-BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm18920; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8921; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]8922; AVX512DQ-BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8923; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm88924; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm88925; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8926; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm88927; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm88928; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8929; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm88930; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm88931; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8932; AVX512DQ-BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm18933; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8934; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]8935; AVX512DQ-BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]8936; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm18937; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm8, %zmm18938; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8939; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]8940; AVX512DQ-BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]8941; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm1, %zmm48942; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8943; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm08944; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm8, %zmm08945; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8946; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm1, %zmm58947; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8948; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm08949; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm8, %zmm08950; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8951; AVX512DQ-BW-NEXT:    vpermi2q %zmm21, %zmm2, %zmm88952; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8953; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm1, %zmm28954; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8955; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm08956; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm10, %zmm08957; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8958; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm1, %zmm68959; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8960; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]8961; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm08962; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm11, %zmm08963; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8964; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]8965; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm08966; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm31, %zmm08967; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8968; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]8969; AVX512DQ-BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8970; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm08971; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm5, %zmm08972; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8973; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]8974; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8975; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm38976; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8977; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm258978; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm18979; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm298980; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm28981; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm7, %zmm298982; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm38983; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm38984; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8985; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm38986; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm11, %zmm38987; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8988; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm38989; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm31, %zmm38990; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8991; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm38992; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm5, %zmm38993; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8994; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm258995; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm98996; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm278997; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm18998; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm5, %zmm18999; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9000; AVX512DQ-BW-NEXT:    vmovdqa64 704(%rdi), %zmm309001; AVX512DQ-BW-NEXT:    vmovdqa64 640(%rdi), %zmm269002; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm19003; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm5, %zmm19004; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9005; AVX512DQ-BW-NEXT:    vmovdqa64 1088(%rdi), %zmm79006; AVX512DQ-BW-NEXT:    vmovdqa64 1024(%rdi), %zmm209007; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm19008; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm5, %zmm19009; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9010; AVX512DQ-BW-NEXT:    vmovdqa64 832(%rdi), %zmm19011; AVX512DQ-BW-NEXT:    vmovdqa64 768(%rdi), %zmm89012; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm159013; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm159014; AVX512DQ-BW-NEXT:    vmovdqa64 1472(%rdi), %zmm69015; AVX512DQ-BW-NEXT:    vmovdqa64 1408(%rdi), %zmm219016; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm39017; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm5, %zmm39018; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9019; AVX512DQ-BW-NEXT:    vmovdqa64 1216(%rdi), %zmm199020; AVX512DQ-BW-NEXT:    vmovdqa64 1152(%rdi), %zmm49021; AVX512DQ-BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm29022; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill9023; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm129024; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm10, %zmm129025; AVX512DQ-BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm109026; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9027; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm139028; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm11, %zmm139029; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm19, %zmm119030; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm149031; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm31, %zmm149032; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm19, %zmm319033; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm29034; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm5, %zmm29035; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9036; AVX512DQ-BW-NEXT:    vpermi2q %zmm19, %zmm4, %zmm59037; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm49038; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm0, %zmm89039; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm199040; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm199041; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm229042; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm229043; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm179044; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm0, %zmm179045; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm189046; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm189047; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]9048; AVX512DQ-BW-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]9049; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm249050; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm1, %zmm249051; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]9052; AVX512DQ-BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]9053; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm09054; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm2, %zmm09055; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9056; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]9057; AVX512DQ-BW-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]9058; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm30, %zmm09059; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm3, %zmm09060; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9061; AVX512DQ-BW-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]9062; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]9063; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm309064; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm269065; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm1, %zmm269066; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm239067; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm239068; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm289069; AVX512DQ-BW-NEXT:    vpermt2q %zmm27, %zmm3, %zmm289070; AVX512DQ-BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm99071; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm279072; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm1, %zmm279073; AVX512DQ-BW-NEXT:    vpermi2q %zmm6, %zmm21, %zmm19074; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm109075; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm109076; AVX512DQ-BW-NEXT:    vpermi2q %zmm6, %zmm21, %zmm29077; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm169078; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm3, %zmm169079; AVX512DQ-BW-NEXT:    vpermi2q %zmm21, %zmm6, %zmm39080; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm69081; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm79082; AVX512DQ-BW-NEXT:    movb $56, %al9083; AVX512DQ-BW-NEXT:    kmovd %eax, %k19084; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9085; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}9086; AVX512DQ-BW-NEXT:    movb $-64, %al9087; AVX512DQ-BW-NEXT:    kmovd %eax, %k29088; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9089; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}9090; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9091; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9092; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}9093; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9094; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}9095; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9096; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9097; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}9098; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9099; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}9100; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload9101; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9102; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}9103; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9104; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}9105; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm209106; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9107; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}9108; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}9109; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9110; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload9111; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}9112; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}9113; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9114; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload9115; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}9116; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}9117; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9118; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload9119; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}9120; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}9121; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm179122; AVX512DQ-BW-NEXT:    movb $24, %al9123; AVX512DQ-BW-NEXT:    kmovd %eax, %k29124; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9125; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}9126; AVX512DQ-BW-NEXT:    movb $-32, %al9127; AVX512DQ-BW-NEXT:    kmovd %eax, %k19128; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}9129; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9130; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload9131; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}9132; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}9133; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm229134; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9135; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload9136; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}9137; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}9138; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm249139; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9140; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}9141; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}9142; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9143; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}9144; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}9145; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9146; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload9147; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}9148; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9149; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}9150; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm189151; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9152; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload9153; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}9154; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}9155; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm239156; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9157; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}9158; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}9159; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9160; AVX512DQ-BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload9161; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}9162; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload9163; AVX512DQ-BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload9164; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9165; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}9166; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9167; AVX512DQ-BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload9168; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}9169; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload9170; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm59171; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}9172; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9173; AVX512DQ-BW-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload9174; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}9175; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload9176; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm109177; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}9178; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload9179; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm49180; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}9181; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload9182; AVX512DQ-BW-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm69183; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}9184; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, 192(%rsi)9185; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, 128(%rsi)9186; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload9187; AVX512DQ-BW-NEXT:    vmovaps %zmm7, 64(%rsi)9188; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, (%rsi)9189; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, 192(%rdx)9190; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, (%rdx)9191; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, 64(%rdx)9192; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, 128(%rdx)9193; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, 192(%rcx)9194; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, (%rcx)9195; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, 64(%rcx)9196; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, 128(%rcx)9197; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, 192(%r8)9198; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, (%r8)9199; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, 64(%r8)9200; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, 128(%r8)9201; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, 192(%r9)9202; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%r9)9203; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 64(%r9)9204; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, 128(%r9)9205; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax9206; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, 128(%rax)9207; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, 192(%rax)9208; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, (%rax)9209; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)9210; AVX512DQ-BW-NEXT:    addq $2632, %rsp # imm = 0xA489211; AVX512DQ-BW-NEXT:    vzeroupper9212; AVX512DQ-BW-NEXT:    retq9213;9214; AVX512DQ-BW-FCP-LABEL: load_i64_stride6_vf32:9215; AVX512DQ-BW-FCP:       # %bb.0:9216; AVX512DQ-BW-FCP-NEXT:    subq $2632, %rsp # imm = 0xA489217; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm29218; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm219219; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm19220; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm199221; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm189222; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm39223; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm49224; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm09225; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm59226; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm259227; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [0,6,0,10,0,6,0,10]9228; AVX512DQ-BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]9229; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm79230; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm6, %zmm79231; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9232; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm7 = [0,6,12,0]9233; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm89234; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm6, %zmm89235; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9236; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm89237; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm7, %zmm89238; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9239; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm89240; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm89241; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9242; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm69243; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9244; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm6 = [1,7,0,11,1,7,0,11]9245; AVX512DQ-BW-FCP-NEXT:    # zmm6 = mem[0,1,2,3,0,1,2,3]9246; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm89247; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm6, %zmm89248; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9249; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm89250; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm69251; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [1,7,13,0]9252; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19253; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm19254; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9255; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm19256; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm8, %zmm19257; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9258; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm2, %zmm21, %zmm89259; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9260; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [10,4,10,4,10,4,10,4]9261; AVX512DQ-BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9262; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm89263; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm89264; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9265; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm89266; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm89267; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9268; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm89269; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm89270; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9271; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm19272; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9273; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm1 = [11,5,11,5,11,5,11,5]9274; AVX512DQ-BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9275; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm89276; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm89277; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9278; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm89279; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm89280; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9281; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm89282; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm89283; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9284; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm19285; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9286; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [12,0,0,6,12,0,0,6]9287; AVX512DQ-BW-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]9288; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm19289; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm8, %zmm19290; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9291; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [13,0,1,7,13,0,1,7]9292; AVX512DQ-BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]9293; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm1, %zmm49294; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9295; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm09296; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm8, %zmm09297; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9298; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm59299; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9300; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm09301; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm8, %zmm09302; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9303; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm21, %zmm2, %zmm89304; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9305; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm1, %zmm29306; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9307; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm09308; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm10, %zmm09309; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9310; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm1, %zmm69311; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9312; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [10,0,6,0]9313; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm09314; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm11, %zmm09315; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9316; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm31 = [11,1,7,0]9317; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm09318; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm31, %zmm09319; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9320; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [4,10,4,10,4,10,4,10]9321; AVX512DQ-BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9322; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm09323; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm5, %zmm09324; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9325; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]9326; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9327; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm39328; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9329; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm259330; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm19331; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm299332; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm29333; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm299334; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm39335; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm39336; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9337; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm39338; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm11, %zmm39339; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9340; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm39341; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm31, %zmm39342; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9343; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm39344; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm39345; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9346; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm259347; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm99348; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm279349; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm19350; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm19351; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9352; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm309353; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm269354; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm19355; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm5, %zmm19356; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9357; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm79358; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm209359; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm19360; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm5, %zmm19361; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9362; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm19363; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm89364; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm159365; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm159366; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm69367; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm219368; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm39369; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm5, %zmm39370; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9371; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm199372; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm49373; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm29374; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill9375; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm129376; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm129377; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm109378; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9379; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm139380; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm11, %zmm139381; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm119382; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm149383; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm31, %zmm149384; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm19, %zmm319385; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm29386; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm5, %zmm29387; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9388; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm19, %zmm4, %zmm59389; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm49390; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm89391; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm199392; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm199393; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm229394; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm229395; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm179396; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm0, %zmm179397; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm189398; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm189399; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]9400; AVX512DQ-BW-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]9401; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm249402; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm1, %zmm249403; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]9404; AVX512DQ-BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]9405; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm09406; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm2, %zmm09407; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9408; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]9409; AVX512DQ-BW-FCP-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]9410; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm09411; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm3, %zmm09412; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9413; AVX512DQ-BW-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]9414; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]9415; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm309416; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm269417; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm1, %zmm269418; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm239419; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm239420; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm289421; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm27, %zmm3, %zmm289422; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm99423; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm279424; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm1, %zmm279425; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm19426; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm109427; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm109428; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm6, %zmm21, %zmm29429; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm169430; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm3, %zmm169431; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm21, %zmm6, %zmm39432; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm69433; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm79434; AVX512DQ-BW-FCP-NEXT:    movb $56, %al9435; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k19436; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9437; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}9438; AVX512DQ-BW-FCP-NEXT:    movb $-64, %al9439; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k29440; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9441; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}9442; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9443; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9444; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}9445; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9446; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}9447; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9448; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9449; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k1}9450; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9451; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm15 {%k2}9452; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload9453; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9454; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k1}9455; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload9456; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm0 {%k2}9457; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm209458; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9459; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k1}9460; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm12 {%k2}9461; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9462; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload9463; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k1}9464; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm21 {%k2}9465; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9466; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload9467; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}9468; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm19 {%k2}9469; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9470; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload9471; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0 {%k1}9472; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm0 {%k2}9473; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm179474; AVX512DQ-BW-FCP-NEXT:    movb $24, %al9475; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k29476; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9477; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm13 {%k2}9478; AVX512DQ-BW-FCP-NEXT:    movb $-32, %al9479; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k19480; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm13 {%k1}9481; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9482; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload9483; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}9484; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm18 {%k1}9485; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm229486; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9487; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload9488; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}9489; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm18 {%k1}9490; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm249491; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9492; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}9493; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm11 {%k1}9494; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9495; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}9496; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm14 {%k1}9497; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9498; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload9499; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}9500; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9501; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}9502; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm189503; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9504; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload9505; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}9506; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1 {%k1}9507; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm239508; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9509; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}9510; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm31 {%k1}9511; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9512; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload9513; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}9514; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload9515; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload9516; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9517; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}9518; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9519; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload9520; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k1}9521; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload9522; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm5, %zmm10, %zmm59523; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}9524; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9525; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload9526; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm3 {%k1}9527; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload9528; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm25, %zmm10, %zmm109529; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}9530; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload9531; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm4, %zmm9, %zmm49532; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k1}9533; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload9534; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $0, %xmm8, %zmm6, %zmm69535; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm6 {%k1}9536; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, 192(%rsi)9537; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, 128(%rsi)9538; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload9539; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm7, 64(%rsi)9540; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, (%rsi)9541; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, 192(%rdx)9542; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, (%rdx)9543; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, 64(%rdx)9544; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, 128(%rdx)9545; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rcx)9546; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, (%rcx)9547; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, 64(%rcx)9548; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rcx)9549; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, 192(%r8)9550; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, (%r8)9551; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, 64(%r8)9552; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, 128(%r8)9553; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, 192(%r9)9554; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%r9)9555; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r9)9556; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, 128(%r9)9557; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax9558; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 128(%rax)9559; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, 192(%rax)9560; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)9561; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)9562; AVX512DQ-BW-FCP-NEXT:    addq $2632, %rsp # imm = 0xA489563; AVX512DQ-BW-FCP-NEXT:    vzeroupper9564; AVX512DQ-BW-FCP-NEXT:    retq9565  %wide.vec = load <192 x i64>, ptr %in.vec, align 649566  %strided.vec0 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90, i32 96, i32 102, i32 108, i32 114, i32 120, i32 126, i32 132, i32 138, i32 144, i32 150, i32 156, i32 162, i32 168, i32 174, i32 180, i32 186>9567  %strided.vec1 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91, i32 97, i32 103, i32 109, i32 115, i32 121, i32 127, i32 133, i32 139, i32 145, i32 151, i32 157, i32 163, i32 169, i32 175, i32 181, i32 187>9568  %strided.vec2 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92, i32 98, i32 104, i32 110, i32 116, i32 122, i32 128, i32 134, i32 140, i32 146, i32 152, i32 158, i32 164, i32 170, i32 176, i32 182, i32 188>9569  %strided.vec3 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93, i32 99, i32 105, i32 111, i32 117, i32 123, i32 129, i32 135, i32 141, i32 147, i32 153, i32 159, i32 165, i32 171, i32 177, i32 183, i32 189>9570  %strided.vec4 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94, i32 100, i32 106, i32 112, i32 118, i32 124, i32 130, i32 136, i32 142, i32 148, i32 154, i32 160, i32 166, i32 172, i32 178, i32 184, i32 190>9571  %strided.vec5 = shufflevector <192 x i64> %wide.vec, <192 x i64> poison, <32 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95, i32 101, i32 107, i32 113, i32 119, i32 125, i32 131, i32 137, i32 143, i32 149, i32 155, i32 161, i32 167, i32 173, i32 179, i32 185, i32 191>9572  store <32 x i64> %strided.vec0, ptr %out.vec0, align 649573  store <32 x i64> %strided.vec1, ptr %out.vec1, align 649574  store <32 x i64> %strided.vec2, ptr %out.vec2, align 649575  store <32 x i64> %strided.vec3, ptr %out.vec3, align 649576  store <32 x i64> %strided.vec4, ptr %out.vec4, align 649577  store <32 x i64> %strided.vec5, ptr %out.vec5, align 649578  ret void9579}9580 9581define void @load_i64_stride6_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {9582; SSE-LABEL: load_i64_stride6_vf64:9583; SSE:       # %bb.0:9584; SSE-NEXT:    subq $2712, %rsp # imm = 0xA989585; SSE-NEXT:    movaps 816(%rdi), %xmm09586; SSE-NEXT:    movaps 720(%rdi), %xmm19587; SSE-NEXT:    movaps 672(%rdi), %xmm89588; SSE-NEXT:    movaps 624(%rdi), %xmm29589; SSE-NEXT:    movaps 576(%rdi), %xmm99590; SSE-NEXT:    movaps 528(%rdi), %xmm39591; SSE-NEXT:    movaps 480(%rdi), %xmm109592; SSE-NEXT:    movaps 432(%rdi), %xmm49593; SSE-NEXT:    movaps 384(%rdi), %xmm119594; SSE-NEXT:    movaps 336(%rdi), %xmm59595; SSE-NEXT:    movaps 288(%rdi), %xmm129596; SSE-NEXT:    movaps 240(%rdi), %xmm69597; SSE-NEXT:    movaps 192(%rdi), %xmm139598; SSE-NEXT:    movaps 144(%rdi), %xmm79599; SSE-NEXT:    movaps 96(%rdi), %xmm149600; SSE-NEXT:    movaps %xmm14, %xmm159601; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]9602; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9603; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm7[1]9604; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9605; SSE-NEXT:    movaps %xmm13, %xmm79606; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]9607; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9608; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]9609; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9610; SSE-NEXT:    movaps %xmm12, %xmm69611; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]9612; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9613; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]9614; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9615; SSE-NEXT:    movaps %xmm11, %xmm59616; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]9617; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9618; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm4[1]9619; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9620; SSE-NEXT:    movaps %xmm10, %xmm49621; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]9622; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9623; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm3[1]9624; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9625; SSE-NEXT:    movaps %xmm9, %xmm39626; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]9627; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9628; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm2[1]9629; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9630; SSE-NEXT:    movaps %xmm8, %xmm29631; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]9632; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9633; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]9634; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9635; SSE-NEXT:    movaps 768(%rdi), %xmm19636; SSE-NEXT:    movaps %xmm1, %xmm29637; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9638; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9639; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9640; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9641; SSE-NEXT:    movaps 912(%rdi), %xmm09642; SSE-NEXT:    movaps 864(%rdi), %xmm19643; SSE-NEXT:    movaps %xmm1, %xmm29644; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9645; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9646; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9647; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9648; SSE-NEXT:    movaps 1008(%rdi), %xmm09649; SSE-NEXT:    movaps 960(%rdi), %xmm19650; SSE-NEXT:    movaps %xmm1, %xmm29651; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9652; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9653; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9654; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9655; SSE-NEXT:    movaps 1104(%rdi), %xmm09656; SSE-NEXT:    movaps 1056(%rdi), %xmm19657; SSE-NEXT:    movaps %xmm1, %xmm29658; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9659; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9660; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9661; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9662; SSE-NEXT:    movaps 1200(%rdi), %xmm09663; SSE-NEXT:    movaps 1152(%rdi), %xmm19664; SSE-NEXT:    movaps %xmm1, %xmm29665; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9666; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9667; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9668; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9669; SSE-NEXT:    movaps 1296(%rdi), %xmm09670; SSE-NEXT:    movaps 1248(%rdi), %xmm19671; SSE-NEXT:    movaps %xmm1, %xmm29672; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9673; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9674; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9675; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9676; SSE-NEXT:    movaps 1392(%rdi), %xmm09677; SSE-NEXT:    movaps 1344(%rdi), %xmm19678; SSE-NEXT:    movaps %xmm1, %xmm29679; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9680; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9681; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9682; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9683; SSE-NEXT:    movaps 1488(%rdi), %xmm09684; SSE-NEXT:    movaps 1440(%rdi), %xmm19685; SSE-NEXT:    movaps %xmm1, %xmm29686; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9687; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9688; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9689; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9690; SSE-NEXT:    movaps 1584(%rdi), %xmm09691; SSE-NEXT:    movaps 1536(%rdi), %xmm19692; SSE-NEXT:    movaps %xmm1, %xmm29693; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9694; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9695; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9696; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9697; SSE-NEXT:    movaps 1680(%rdi), %xmm09698; SSE-NEXT:    movaps 1632(%rdi), %xmm19699; SSE-NEXT:    movaps %xmm1, %xmm29700; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9701; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9702; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9703; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9704; SSE-NEXT:    movaps 1776(%rdi), %xmm09705; SSE-NEXT:    movaps 1728(%rdi), %xmm19706; SSE-NEXT:    movaps %xmm1, %xmm29707; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9708; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9709; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9710; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9711; SSE-NEXT:    movaps 1872(%rdi), %xmm09712; SSE-NEXT:    movaps 1824(%rdi), %xmm19713; SSE-NEXT:    movaps %xmm1, %xmm29714; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9715; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9716; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9717; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9718; SSE-NEXT:    movaps 1968(%rdi), %xmm09719; SSE-NEXT:    movaps 1920(%rdi), %xmm19720; SSE-NEXT:    movaps %xmm1, %xmm29721; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9722; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9723; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9724; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9725; SSE-NEXT:    movaps 2064(%rdi), %xmm09726; SSE-NEXT:    movaps 2016(%rdi), %xmm19727; SSE-NEXT:    movaps %xmm1, %xmm29728; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9729; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9730; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9731; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9732; SSE-NEXT:    movaps 2160(%rdi), %xmm09733; SSE-NEXT:    movaps 2112(%rdi), %xmm19734; SSE-NEXT:    movaps %xmm1, %xmm29735; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9736; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9737; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9738; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9739; SSE-NEXT:    movaps 2256(%rdi), %xmm09740; SSE-NEXT:    movaps 2208(%rdi), %xmm19741; SSE-NEXT:    movaps %xmm1, %xmm29742; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9743; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9744; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9745; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9746; SSE-NEXT:    movaps 2352(%rdi), %xmm09747; SSE-NEXT:    movaps 2304(%rdi), %xmm19748; SSE-NEXT:    movaps %xmm1, %xmm29749; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9750; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9751; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9752; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9753; SSE-NEXT:    movaps 2448(%rdi), %xmm09754; SSE-NEXT:    movaps 2400(%rdi), %xmm19755; SSE-NEXT:    movaps %xmm1, %xmm29756; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9757; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9758; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9759; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9760; SSE-NEXT:    movaps 2544(%rdi), %xmm09761; SSE-NEXT:    movaps 2496(%rdi), %xmm19762; SSE-NEXT:    movaps %xmm1, %xmm29763; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9764; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9765; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9766; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9767; SSE-NEXT:    movaps 2640(%rdi), %xmm09768; SSE-NEXT:    movaps 2592(%rdi), %xmm19769; SSE-NEXT:    movaps %xmm1, %xmm29770; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9771; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9772; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9773; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9774; SSE-NEXT:    movaps 2736(%rdi), %xmm09775; SSE-NEXT:    movaps 2688(%rdi), %xmm19776; SSE-NEXT:    movaps %xmm1, %xmm29777; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9778; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9779; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9780; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9781; SSE-NEXT:    movaps 2832(%rdi), %xmm09782; SSE-NEXT:    movaps 2784(%rdi), %xmm19783; SSE-NEXT:    movaps %xmm1, %xmm29784; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9785; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9786; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9787; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9788; SSE-NEXT:    movaps 2928(%rdi), %xmm09789; SSE-NEXT:    movaps 2880(%rdi), %xmm19790; SSE-NEXT:    movaps %xmm1, %xmm29791; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9792; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9793; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9794; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9795; SSE-NEXT:    movaps 3024(%rdi), %xmm09796; SSE-NEXT:    movaps 2976(%rdi), %xmm19797; SSE-NEXT:    movaps %xmm1, %xmm29798; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9799; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9800; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9801; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9802; SSE-NEXT:    movaps (%rdi), %xmm19803; SSE-NEXT:    movaps 48(%rdi), %xmm09804; SSE-NEXT:    movaps %xmm1, %xmm29805; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9806; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9807; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9808; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9809; SSE-NEXT:    movaps 64(%rdi), %xmm09810; SSE-NEXT:    movaps 16(%rdi), %xmm19811; SSE-NEXT:    movaps %xmm1, %xmm29812; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9813; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9814; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9815; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9816; SSE-NEXT:    movaps 160(%rdi), %xmm09817; SSE-NEXT:    movaps 112(%rdi), %xmm19818; SSE-NEXT:    movaps %xmm1, %xmm29819; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9820; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9821; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9822; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9823; SSE-NEXT:    movaps 256(%rdi), %xmm09824; SSE-NEXT:    movaps 208(%rdi), %xmm19825; SSE-NEXT:    movaps %xmm1, %xmm29826; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9827; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9828; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9829; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9830; SSE-NEXT:    movaps 352(%rdi), %xmm09831; SSE-NEXT:    movaps 304(%rdi), %xmm19832; SSE-NEXT:    movaps %xmm1, %xmm29833; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9834; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9835; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9836; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9837; SSE-NEXT:    movaps 448(%rdi), %xmm09838; SSE-NEXT:    movaps 400(%rdi), %xmm19839; SSE-NEXT:    movaps %xmm1, %xmm29840; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9841; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9842; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9843; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9844; SSE-NEXT:    movaps 544(%rdi), %xmm09845; SSE-NEXT:    movaps 496(%rdi), %xmm19846; SSE-NEXT:    movaps %xmm1, %xmm29847; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9848; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9849; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9850; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9851; SSE-NEXT:    movaps 640(%rdi), %xmm09852; SSE-NEXT:    movaps 592(%rdi), %xmm19853; SSE-NEXT:    movaps %xmm1, %xmm29854; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9855; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9856; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9857; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9858; SSE-NEXT:    movaps 736(%rdi), %xmm09859; SSE-NEXT:    movaps 688(%rdi), %xmm19860; SSE-NEXT:    movaps %xmm1, %xmm29861; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9862; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9863; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9864; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9865; SSE-NEXT:    movaps 832(%rdi), %xmm09866; SSE-NEXT:    movaps 784(%rdi), %xmm19867; SSE-NEXT:    movaps %xmm1, %xmm29868; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9869; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9870; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9871; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9872; SSE-NEXT:    movaps 928(%rdi), %xmm09873; SSE-NEXT:    movaps 880(%rdi), %xmm19874; SSE-NEXT:    movaps %xmm1, %xmm29875; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9876; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9877; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9878; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9879; SSE-NEXT:    movaps 1024(%rdi), %xmm09880; SSE-NEXT:    movaps 976(%rdi), %xmm19881; SSE-NEXT:    movaps %xmm1, %xmm29882; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9883; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9884; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9885; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9886; SSE-NEXT:    movaps 1120(%rdi), %xmm09887; SSE-NEXT:    movaps 1072(%rdi), %xmm19888; SSE-NEXT:    movaps %xmm1, %xmm29889; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9890; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9891; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9892; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9893; SSE-NEXT:    movaps 1216(%rdi), %xmm09894; SSE-NEXT:    movaps 1168(%rdi), %xmm19895; SSE-NEXT:    movaps %xmm1, %xmm29896; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9897; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9898; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9899; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9900; SSE-NEXT:    movaps 1312(%rdi), %xmm09901; SSE-NEXT:    movaps 1264(%rdi), %xmm19902; SSE-NEXT:    movaps %xmm1, %xmm29903; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9904; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9905; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9906; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9907; SSE-NEXT:    movaps 1408(%rdi), %xmm09908; SSE-NEXT:    movaps 1360(%rdi), %xmm19909; SSE-NEXT:    movaps %xmm1, %xmm29910; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9911; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9912; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9913; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9914; SSE-NEXT:    movaps 1504(%rdi), %xmm09915; SSE-NEXT:    movaps 1456(%rdi), %xmm19916; SSE-NEXT:    movaps %xmm1, %xmm29917; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9918; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9919; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9920; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9921; SSE-NEXT:    movaps 1600(%rdi), %xmm09922; SSE-NEXT:    movaps 1552(%rdi), %xmm19923; SSE-NEXT:    movaps %xmm1, %xmm29924; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9925; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9926; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9927; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9928; SSE-NEXT:    movaps 1696(%rdi), %xmm09929; SSE-NEXT:    movaps 1648(%rdi), %xmm19930; SSE-NEXT:    movaps %xmm1, %xmm29931; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9932; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9933; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9934; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9935; SSE-NEXT:    movaps 1792(%rdi), %xmm09936; SSE-NEXT:    movaps 1744(%rdi), %xmm19937; SSE-NEXT:    movaps %xmm1, %xmm29938; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9939; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9940; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9941; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9942; SSE-NEXT:    movaps 1888(%rdi), %xmm09943; SSE-NEXT:    movaps 1840(%rdi), %xmm19944; SSE-NEXT:    movaps %xmm1, %xmm29945; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9946; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9947; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9948; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9949; SSE-NEXT:    movaps 1984(%rdi), %xmm09950; SSE-NEXT:    movaps 1936(%rdi), %xmm19951; SSE-NEXT:    movaps %xmm1, %xmm29952; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9953; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9954; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9955; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9956; SSE-NEXT:    movaps 2080(%rdi), %xmm09957; SSE-NEXT:    movaps 2032(%rdi), %xmm19958; SSE-NEXT:    movaps %xmm1, %xmm29959; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9960; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9961; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9962; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9963; SSE-NEXT:    movaps 2176(%rdi), %xmm09964; SSE-NEXT:    movaps 2128(%rdi), %xmm19965; SSE-NEXT:    movaps %xmm1, %xmm29966; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9967; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9968; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9969; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9970; SSE-NEXT:    movaps 2272(%rdi), %xmm09971; SSE-NEXT:    movaps 2224(%rdi), %xmm19972; SSE-NEXT:    movaps %xmm1, %xmm29973; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9974; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9975; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9976; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9977; SSE-NEXT:    movaps 2368(%rdi), %xmm09978; SSE-NEXT:    movaps 2320(%rdi), %xmm19979; SSE-NEXT:    movaps %xmm1, %xmm29980; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9981; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9982; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9983; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9984; SSE-NEXT:    movaps 2464(%rdi), %xmm09985; SSE-NEXT:    movaps 2416(%rdi), %xmm19986; SSE-NEXT:    movaps %xmm1, %xmm29987; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9988; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9989; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9990; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9991; SSE-NEXT:    movaps 2560(%rdi), %xmm09992; SSE-NEXT:    movaps 2512(%rdi), %xmm19993; SSE-NEXT:    movaps %xmm1, %xmm29994; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9995; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9996; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9997; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9998; SSE-NEXT:    movaps 2656(%rdi), %xmm09999; SSE-NEXT:    movaps 2608(%rdi), %xmm110000; SSE-NEXT:    movaps %xmm1, %xmm210001; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10002; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10003; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10004; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10005; SSE-NEXT:    movaps 2752(%rdi), %xmm010006; SSE-NEXT:    movaps 2704(%rdi), %xmm110007; SSE-NEXT:    movaps %xmm1, %xmm210008; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10009; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10010; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10011; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10012; SSE-NEXT:    movaps 2848(%rdi), %xmm010013; SSE-NEXT:    movaps 2800(%rdi), %xmm110014; SSE-NEXT:    movaps %xmm1, %xmm210015; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10016; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10017; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10018; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10019; SSE-NEXT:    movaps 2944(%rdi), %xmm010020; SSE-NEXT:    movaps 2896(%rdi), %xmm110021; SSE-NEXT:    movaps %xmm1, %xmm210022; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10023; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10024; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10025; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10026; SSE-NEXT:    movaps 3040(%rdi), %xmm010027; SSE-NEXT:    movaps 2992(%rdi), %xmm110028; SSE-NEXT:    movaps %xmm1, %xmm210029; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10030; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10031; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10032; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10033; SSE-NEXT:    movaps 80(%rdi), %xmm010034; SSE-NEXT:    movaps 32(%rdi), %xmm110035; SSE-NEXT:    movaps %xmm1, %xmm210036; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10037; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10038; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10039; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10040; SSE-NEXT:    movaps 176(%rdi), %xmm010041; SSE-NEXT:    movaps 128(%rdi), %xmm110042; SSE-NEXT:    movaps %xmm1, %xmm210043; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10044; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10045; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10046; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10047; SSE-NEXT:    movaps 272(%rdi), %xmm010048; SSE-NEXT:    movaps 224(%rdi), %xmm110049; SSE-NEXT:    movaps %xmm1, %xmm210050; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10051; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10052; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10053; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10054; SSE-NEXT:    movaps 368(%rdi), %xmm010055; SSE-NEXT:    movaps 320(%rdi), %xmm110056; SSE-NEXT:    movaps %xmm1, %xmm210057; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10058; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10059; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10060; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10061; SSE-NEXT:    movaps 464(%rdi), %xmm010062; SSE-NEXT:    movaps 416(%rdi), %xmm110063; SSE-NEXT:    movaps %xmm1, %xmm210064; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10065; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10066; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10067; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10068; SSE-NEXT:    movaps 560(%rdi), %xmm010069; SSE-NEXT:    movaps 512(%rdi), %xmm110070; SSE-NEXT:    movaps %xmm1, %xmm210071; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10072; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10073; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10074; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10075; SSE-NEXT:    movaps 656(%rdi), %xmm010076; SSE-NEXT:    movaps 608(%rdi), %xmm110077; SSE-NEXT:    movaps %xmm1, %xmm210078; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10079; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10080; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10081; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10082; SSE-NEXT:    movaps 752(%rdi), %xmm010083; SSE-NEXT:    movaps 704(%rdi), %xmm110084; SSE-NEXT:    movaps %xmm1, %xmm210085; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10086; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10087; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10088; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10089; SSE-NEXT:    movaps 848(%rdi), %xmm010090; SSE-NEXT:    movaps 800(%rdi), %xmm110091; SSE-NEXT:    movaps %xmm1, %xmm210092; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10093; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10094; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10095; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10096; SSE-NEXT:    movaps 944(%rdi), %xmm010097; SSE-NEXT:    movaps 896(%rdi), %xmm110098; SSE-NEXT:    movaps %xmm1, %xmm210099; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10100; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10101; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10102; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10103; SSE-NEXT:    movaps 1040(%rdi), %xmm010104; SSE-NEXT:    movaps 992(%rdi), %xmm110105; SSE-NEXT:    movaps %xmm1, %xmm210106; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10107; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10108; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10109; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10110; SSE-NEXT:    movaps 1136(%rdi), %xmm010111; SSE-NEXT:    movaps 1088(%rdi), %xmm110112; SSE-NEXT:    movaps %xmm1, %xmm210113; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10114; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10115; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10116; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10117; SSE-NEXT:    movaps 1232(%rdi), %xmm010118; SSE-NEXT:    movaps 1184(%rdi), %xmm110119; SSE-NEXT:    movaps %xmm1, %xmm210120; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10121; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10122; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10123; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10124; SSE-NEXT:    movaps 1328(%rdi), %xmm010125; SSE-NEXT:    movaps 1280(%rdi), %xmm110126; SSE-NEXT:    movaps %xmm1, %xmm210127; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10128; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10129; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10130; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10131; SSE-NEXT:    movaps 1424(%rdi), %xmm010132; SSE-NEXT:    movaps 1376(%rdi), %xmm110133; SSE-NEXT:    movaps %xmm1, %xmm210134; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10135; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10136; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10137; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10138; SSE-NEXT:    movaps 1520(%rdi), %xmm010139; SSE-NEXT:    movaps 1472(%rdi), %xmm110140; SSE-NEXT:    movaps %xmm1, %xmm210141; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10142; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10143; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10144; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10145; SSE-NEXT:    movaps 1616(%rdi), %xmm010146; SSE-NEXT:    movaps 1568(%rdi), %xmm110147; SSE-NEXT:    movaps %xmm1, %xmm210148; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10149; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10150; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10151; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10152; SSE-NEXT:    movaps 1712(%rdi), %xmm010153; SSE-NEXT:    movaps 1664(%rdi), %xmm110154; SSE-NEXT:    movaps %xmm1, %xmm210155; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10156; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10157; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10158; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10159; SSE-NEXT:    movaps 1808(%rdi), %xmm010160; SSE-NEXT:    movaps 1760(%rdi), %xmm110161; SSE-NEXT:    movaps %xmm1, %xmm210162; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10163; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10164; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10165; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10166; SSE-NEXT:    movaps 1904(%rdi), %xmm010167; SSE-NEXT:    movaps 1856(%rdi), %xmm110168; SSE-NEXT:    movaps %xmm1, %xmm210169; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10170; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10171; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10172; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10173; SSE-NEXT:    movaps 2000(%rdi), %xmm010174; SSE-NEXT:    movaps 1952(%rdi), %xmm110175; SSE-NEXT:    movaps %xmm1, %xmm210176; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10177; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10178; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10179; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10180; SSE-NEXT:    movaps 2096(%rdi), %xmm010181; SSE-NEXT:    movaps 2048(%rdi), %xmm110182; SSE-NEXT:    movaps %xmm1, %xmm210183; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10184; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10185; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10186; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10187; SSE-NEXT:    movaps 2192(%rdi), %xmm010188; SSE-NEXT:    movaps 2144(%rdi), %xmm110189; SSE-NEXT:    movaps %xmm1, %xmm210190; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10191; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10192; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10193; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill10194; SSE-NEXT:    movaps 2288(%rdi), %xmm010195; SSE-NEXT:    movaps 2240(%rdi), %xmm1510196; SSE-NEXT:    movaps %xmm15, %xmm110197; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]10198; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10199; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]10200; SSE-NEXT:    movaps 2384(%rdi), %xmm010201; SSE-NEXT:    movaps 2336(%rdi), %xmm1210202; SSE-NEXT:    movaps %xmm12, %xmm110203; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]10204; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10205; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]10206; SSE-NEXT:    movaps 2480(%rdi), %xmm010207; SSE-NEXT:    movaps 2432(%rdi), %xmm1410208; SSE-NEXT:    movaps %xmm14, %xmm110209; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]10210; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10211; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]10212; SSE-NEXT:    movaps 2576(%rdi), %xmm010213; SSE-NEXT:    movaps 2528(%rdi), %xmm1010214; SSE-NEXT:    movaps %xmm10, %xmm1310215; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]10216; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]10217; SSE-NEXT:    movaps 2672(%rdi), %xmm010218; SSE-NEXT:    movaps 2624(%rdi), %xmm910219; SSE-NEXT:    movaps %xmm9, %xmm1110220; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]10221; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]10222; SSE-NEXT:    movaps 2768(%rdi), %xmm010223; SSE-NEXT:    movaps 2720(%rdi), %xmm510224; SSE-NEXT:    movaps %xmm5, %xmm810225; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]10226; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]10227; SSE-NEXT:    movaps 2864(%rdi), %xmm010228; SSE-NEXT:    movaps 2816(%rdi), %xmm610229; SSE-NEXT:    movaps %xmm6, %xmm710230; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]10231; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]10232; SSE-NEXT:    movaps 2960(%rdi), %xmm010233; SSE-NEXT:    movaps 2912(%rdi), %xmm310234; SSE-NEXT:    movaps %xmm3, %xmm410235; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]10236; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]10237; SSE-NEXT:    movaps 3056(%rdi), %xmm010238; SSE-NEXT:    movaps 3008(%rdi), %xmm110239; SSE-NEXT:    movaps %xmm1, %xmm210240; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]10241; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]10242; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10243; SSE-NEXT:    movaps %xmm0, 496(%rsi)10244; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10245; SSE-NEXT:    movaps %xmm0, 480(%rsi)10246; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10247; SSE-NEXT:    movaps %xmm0, 464(%rsi)10248; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10249; SSE-NEXT:    movaps %xmm0, 448(%rsi)10250; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10251; SSE-NEXT:    movaps %xmm0, 432(%rsi)10252; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10253; SSE-NEXT:    movaps %xmm0, 416(%rsi)10254; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10255; SSE-NEXT:    movaps %xmm0, 400(%rsi)10256; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10257; SSE-NEXT:    movaps %xmm0, 384(%rsi)10258; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10259; SSE-NEXT:    movaps %xmm0, 368(%rsi)10260; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10261; SSE-NEXT:    movaps %xmm0, 352(%rsi)10262; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10263; SSE-NEXT:    movaps %xmm0, 336(%rsi)10264; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10265; SSE-NEXT:    movaps %xmm0, 320(%rsi)10266; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10267; SSE-NEXT:    movaps %xmm0, 304(%rsi)10268; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10269; SSE-NEXT:    movaps %xmm0, 288(%rsi)10270; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10271; SSE-NEXT:    movaps %xmm0, 272(%rsi)10272; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10273; SSE-NEXT:    movaps %xmm0, 256(%rsi)10274; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10275; SSE-NEXT:    movaps %xmm0, 240(%rsi)10276; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10277; SSE-NEXT:    movaps %xmm0, 224(%rsi)10278; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10279; SSE-NEXT:    movaps %xmm0, 208(%rsi)10280; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10281; SSE-NEXT:    movaps %xmm0, 192(%rsi)10282; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10283; SSE-NEXT:    movaps %xmm0, 176(%rsi)10284; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10285; SSE-NEXT:    movaps %xmm0, 160(%rsi)10286; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10287; SSE-NEXT:    movaps %xmm0, 144(%rsi)10288; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10289; SSE-NEXT:    movaps %xmm0, 128(%rsi)10290; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10291; SSE-NEXT:    movaps %xmm0, 112(%rsi)10292; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10293; SSE-NEXT:    movaps %xmm0, 96(%rsi)10294; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10295; SSE-NEXT:    movaps %xmm0, 80(%rsi)10296; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10297; SSE-NEXT:    movaps %xmm0, 64(%rsi)10298; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10299; SSE-NEXT:    movaps %xmm0, 48(%rsi)10300; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10301; SSE-NEXT:    movaps %xmm0, 32(%rsi)10302; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10303; SSE-NEXT:    movaps %xmm0, 16(%rsi)10304; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10305; SSE-NEXT:    movaps %xmm0, (%rsi)10306; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10307; SSE-NEXT:    movaps %xmm0, 496(%rdx)10308; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10309; SSE-NEXT:    movaps %xmm0, 480(%rdx)10310; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10311; SSE-NEXT:    movaps %xmm0, 464(%rdx)10312; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10313; SSE-NEXT:    movaps %xmm0, 448(%rdx)10314; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10315; SSE-NEXT:    movaps %xmm0, 432(%rdx)10316; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10317; SSE-NEXT:    movaps %xmm0, 416(%rdx)10318; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10319; SSE-NEXT:    movaps %xmm0, 400(%rdx)10320; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10321; SSE-NEXT:    movaps %xmm0, 384(%rdx)10322; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10323; SSE-NEXT:    movaps %xmm0, 368(%rdx)10324; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10325; SSE-NEXT:    movaps %xmm0, 352(%rdx)10326; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10327; SSE-NEXT:    movaps %xmm0, 336(%rdx)10328; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10329; SSE-NEXT:    movaps %xmm0, 320(%rdx)10330; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10331; SSE-NEXT:    movaps %xmm0, 304(%rdx)10332; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10333; SSE-NEXT:    movaps %xmm0, 288(%rdx)10334; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10335; SSE-NEXT:    movaps %xmm0, 272(%rdx)10336; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10337; SSE-NEXT:    movaps %xmm0, 256(%rdx)10338; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10339; SSE-NEXT:    movaps %xmm0, 240(%rdx)10340; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10341; SSE-NEXT:    movaps %xmm0, 224(%rdx)10342; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10343; SSE-NEXT:    movaps %xmm0, 208(%rdx)10344; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10345; SSE-NEXT:    movaps %xmm0, 192(%rdx)10346; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10347; SSE-NEXT:    movaps %xmm0, 176(%rdx)10348; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10349; SSE-NEXT:    movaps %xmm0, 160(%rdx)10350; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10351; SSE-NEXT:    movaps %xmm0, 144(%rdx)10352; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10353; SSE-NEXT:    movaps %xmm0, 128(%rdx)10354; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10355; SSE-NEXT:    movaps %xmm0, 112(%rdx)10356; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10357; SSE-NEXT:    movaps %xmm0, 96(%rdx)10358; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10359; SSE-NEXT:    movaps %xmm0, 80(%rdx)10360; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10361; SSE-NEXT:    movaps %xmm0, 64(%rdx)10362; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10363; SSE-NEXT:    movaps %xmm0, 48(%rdx)10364; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10365; SSE-NEXT:    movaps %xmm0, 32(%rdx)10366; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10367; SSE-NEXT:    movaps %xmm0, 16(%rdx)10368; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10369; SSE-NEXT:    movaps %xmm0, (%rdx)10370; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10371; SSE-NEXT:    movaps %xmm0, 496(%rcx)10372; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10373; SSE-NEXT:    movaps %xmm0, 480(%rcx)10374; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10375; SSE-NEXT:    movaps %xmm0, 464(%rcx)10376; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10377; SSE-NEXT:    movaps %xmm0, 448(%rcx)10378; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10379; SSE-NEXT:    movaps %xmm0, 432(%rcx)10380; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10381; SSE-NEXT:    movaps %xmm0, 416(%rcx)10382; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10383; SSE-NEXT:    movaps %xmm0, 400(%rcx)10384; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10385; SSE-NEXT:    movaps %xmm0, 384(%rcx)10386; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10387; SSE-NEXT:    movaps %xmm0, 368(%rcx)10388; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10389; SSE-NEXT:    movaps %xmm0, 352(%rcx)10390; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10391; SSE-NEXT:    movaps %xmm0, 336(%rcx)10392; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10393; SSE-NEXT:    movaps %xmm0, 320(%rcx)10394; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10395; SSE-NEXT:    movaps %xmm0, 304(%rcx)10396; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10397; SSE-NEXT:    movaps %xmm0, 288(%rcx)10398; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10399; SSE-NEXT:    movaps %xmm0, 272(%rcx)10400; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10401; SSE-NEXT:    movaps %xmm0, 256(%rcx)10402; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10403; SSE-NEXT:    movaps %xmm0, 240(%rcx)10404; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10405; SSE-NEXT:    movaps %xmm0, 224(%rcx)10406; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10407; SSE-NEXT:    movaps %xmm0, 208(%rcx)10408; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10409; SSE-NEXT:    movaps %xmm0, 192(%rcx)10410; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10411; SSE-NEXT:    movaps %xmm0, 176(%rcx)10412; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10413; SSE-NEXT:    movaps %xmm0, 160(%rcx)10414; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10415; SSE-NEXT:    movaps %xmm0, 144(%rcx)10416; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10417; SSE-NEXT:    movaps %xmm0, 128(%rcx)10418; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10419; SSE-NEXT:    movaps %xmm0, 112(%rcx)10420; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10421; SSE-NEXT:    movaps %xmm0, 96(%rcx)10422; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10423; SSE-NEXT:    movaps %xmm0, 80(%rcx)10424; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10425; SSE-NEXT:    movaps %xmm0, 64(%rcx)10426; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10427; SSE-NEXT:    movaps %xmm0, 48(%rcx)10428; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10429; SSE-NEXT:    movaps %xmm0, 32(%rcx)10430; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10431; SSE-NEXT:    movaps %xmm0, 16(%rcx)10432; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10433; SSE-NEXT:    movaps %xmm0, (%rcx)10434; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10435; SSE-NEXT:    movaps %xmm0, 496(%r8)10436; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10437; SSE-NEXT:    movaps %xmm0, 480(%r8)10438; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10439; SSE-NEXT:    movaps %xmm0, 464(%r8)10440; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10441; SSE-NEXT:    movaps %xmm0, 448(%r8)10442; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10443; SSE-NEXT:    movaps %xmm0, 432(%r8)10444; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10445; SSE-NEXT:    movaps %xmm0, 416(%r8)10446; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10447; SSE-NEXT:    movaps %xmm0, 400(%r8)10448; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10449; SSE-NEXT:    movaps %xmm0, 384(%r8)10450; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10451; SSE-NEXT:    movaps %xmm0, 368(%r8)10452; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10453; SSE-NEXT:    movaps %xmm0, 352(%r8)10454; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10455; SSE-NEXT:    movaps %xmm0, 336(%r8)10456; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10457; SSE-NEXT:    movaps %xmm0, 320(%r8)10458; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10459; SSE-NEXT:    movaps %xmm0, 304(%r8)10460; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10461; SSE-NEXT:    movaps %xmm0, 288(%r8)10462; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10463; SSE-NEXT:    movaps %xmm0, 272(%r8)10464; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10465; SSE-NEXT:    movaps %xmm0, 256(%r8)10466; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10467; SSE-NEXT:    movaps %xmm0, 240(%r8)10468; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10469; SSE-NEXT:    movaps %xmm0, 224(%r8)10470; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10471; SSE-NEXT:    movaps %xmm0, 208(%r8)10472; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10473; SSE-NEXT:    movaps %xmm0, 192(%r8)10474; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10475; SSE-NEXT:    movaps %xmm0, 176(%r8)10476; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10477; SSE-NEXT:    movaps %xmm0, 160(%r8)10478; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10479; SSE-NEXT:    movaps %xmm0, 144(%r8)10480; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10481; SSE-NEXT:    movaps %xmm0, 128(%r8)10482; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10483; SSE-NEXT:    movaps %xmm0, 112(%r8)10484; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10485; SSE-NEXT:    movaps %xmm0, 96(%r8)10486; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10487; SSE-NEXT:    movaps %xmm0, 80(%r8)10488; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10489; SSE-NEXT:    movaps %xmm0, 64(%r8)10490; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10491; SSE-NEXT:    movaps %xmm0, 48(%r8)10492; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10493; SSE-NEXT:    movaps %xmm0, 32(%r8)10494; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10495; SSE-NEXT:    movaps %xmm0, 16(%r8)10496; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10497; SSE-NEXT:    movaps %xmm0, (%r8)10498; SSE-NEXT:    movaps %xmm2, 496(%r9)10499; SSE-NEXT:    movaps %xmm4, 480(%r9)10500; SSE-NEXT:    movaps %xmm7, 464(%r9)10501; SSE-NEXT:    movaps %xmm8, 448(%r9)10502; SSE-NEXT:    movaps %xmm11, 432(%r9)10503; SSE-NEXT:    movaps %xmm13, 416(%r9)10504; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10505; SSE-NEXT:    movaps %xmm0, 400(%r9)10506; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10507; SSE-NEXT:    movaps %xmm0, 384(%r9)10508; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10509; SSE-NEXT:    movaps %xmm0, 368(%r9)10510; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10511; SSE-NEXT:    movaps %xmm0, 352(%r9)10512; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10513; SSE-NEXT:    movaps %xmm0, 336(%r9)10514; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10515; SSE-NEXT:    movaps %xmm0, 320(%r9)10516; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10517; SSE-NEXT:    movaps %xmm0, 304(%r9)10518; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10519; SSE-NEXT:    movaps %xmm0, 288(%r9)10520; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10521; SSE-NEXT:    movaps %xmm0, 272(%r9)10522; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10523; SSE-NEXT:    movaps %xmm0, 256(%r9)10524; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10525; SSE-NEXT:    movaps %xmm0, 240(%r9)10526; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10527; SSE-NEXT:    movaps %xmm0, 224(%r9)10528; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10529; SSE-NEXT:    movaps %xmm0, 208(%r9)10530; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10531; SSE-NEXT:    movaps %xmm0, 192(%r9)10532; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10533; SSE-NEXT:    movaps %xmm0, 176(%r9)10534; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10535; SSE-NEXT:    movaps %xmm0, 160(%r9)10536; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10537; SSE-NEXT:    movaps %xmm0, 144(%r9)10538; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10539; SSE-NEXT:    movaps %xmm0, 128(%r9)10540; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10541; SSE-NEXT:    movaps %xmm0, 112(%r9)10542; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10543; SSE-NEXT:    movaps %xmm0, 96(%r9)10544; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10545; SSE-NEXT:    movaps %xmm0, 80(%r9)10546; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10547; SSE-NEXT:    movaps %xmm0, 64(%r9)10548; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10549; SSE-NEXT:    movaps %xmm0, 48(%r9)10550; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10551; SSE-NEXT:    movaps %xmm0, 32(%r9)10552; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10553; SSE-NEXT:    movaps %xmm0, 16(%r9)10554; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10555; SSE-NEXT:    movaps %xmm0, (%r9)10556; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax10557; SSE-NEXT:    movaps %xmm1, 496(%rax)10558; SSE-NEXT:    movaps %xmm3, 480(%rax)10559; SSE-NEXT:    movaps %xmm6, 464(%rax)10560; SSE-NEXT:    movaps %xmm5, 448(%rax)10561; SSE-NEXT:    movaps %xmm9, 432(%rax)10562; SSE-NEXT:    movaps %xmm10, 416(%rax)10563; SSE-NEXT:    movaps %xmm14, 400(%rax)10564; SSE-NEXT:    movaps %xmm12, 384(%rax)10565; SSE-NEXT:    movaps %xmm15, 368(%rax)10566; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload10567; SSE-NEXT:    movaps %xmm0, 352(%rax)10568; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10569; SSE-NEXT:    movaps %xmm0, 336(%rax)10570; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10571; SSE-NEXT:    movaps %xmm0, 320(%rax)10572; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10573; SSE-NEXT:    movaps %xmm0, 304(%rax)10574; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10575; SSE-NEXT:    movaps %xmm0, 288(%rax)10576; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10577; SSE-NEXT:    movaps %xmm0, 272(%rax)10578; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10579; SSE-NEXT:    movaps %xmm0, 256(%rax)10580; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10581; SSE-NEXT:    movaps %xmm0, 240(%rax)10582; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10583; SSE-NEXT:    movaps %xmm0, 224(%rax)10584; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10585; SSE-NEXT:    movaps %xmm0, 208(%rax)10586; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10587; SSE-NEXT:    movaps %xmm0, 192(%rax)10588; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10589; SSE-NEXT:    movaps %xmm0, 176(%rax)10590; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10591; SSE-NEXT:    movaps %xmm0, 160(%rax)10592; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10593; SSE-NEXT:    movaps %xmm0, 144(%rax)10594; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10595; SSE-NEXT:    movaps %xmm0, 128(%rax)10596; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10597; SSE-NEXT:    movaps %xmm0, 112(%rax)10598; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10599; SSE-NEXT:    movaps %xmm0, 96(%rax)10600; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10601; SSE-NEXT:    movaps %xmm0, 80(%rax)10602; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10603; SSE-NEXT:    movaps %xmm0, 64(%rax)10604; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10605; SSE-NEXT:    movaps %xmm0, 48(%rax)10606; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10607; SSE-NEXT:    movaps %xmm0, 32(%rax)10608; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10609; SSE-NEXT:    movaps %xmm0, 16(%rax)10610; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10611; SSE-NEXT:    movaps %xmm0, (%rax)10612; SSE-NEXT:    addq $2712, %rsp # imm = 0xA9810613; SSE-NEXT:    retq10614;10615; AVX-LABEL: load_i64_stride6_vf64:10616; AVX:       # %bb.0:10617; AVX-NEXT:    subq $3768, %rsp # imm = 0xEB810618; AVX-NEXT:    vmovaps 1088(%rdi), %ymm210619; AVX-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10620; AVX-NEXT:    vmovaps 704(%rdi), %ymm310621; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10622; AVX-NEXT:    vmovaps 320(%rdi), %ymm110623; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10624; AVX-NEXT:    vinsertf128 $1, 288(%rdi), %ymm0, %ymm010625; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10626; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10627; AVX-NEXT:    vmovaps 240(%rdi), %xmm410628; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10629; AVX-NEXT:    vmovaps 192(%rdi), %xmm110630; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10631; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]10632; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10633; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10634; AVX-NEXT:    vinsertf128 $1, 672(%rdi), %ymm0, %ymm010635; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10636; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]10637; AVX-NEXT:    vmovaps 624(%rdi), %xmm310638; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10639; AVX-NEXT:    vmovaps 576(%rdi), %xmm110640; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10641; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]10642; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10643; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10644; AVX-NEXT:    vinsertf128 $1, 1056(%rdi), %ymm0, %ymm010645; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10646; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]10647; AVX-NEXT:    vmovaps 1008(%rdi), %xmm210648; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10649; AVX-NEXT:    vmovaps 960(%rdi), %xmm110650; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10651; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10652; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10653; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10654; AVX-NEXT:    vmovaps 1472(%rdi), %ymm110655; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10656; AVX-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm0, %ymm010657; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10658; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10659; AVX-NEXT:    vmovaps 1392(%rdi), %xmm210660; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10661; AVX-NEXT:    vmovaps 1344(%rdi), %xmm110662; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10663; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10664; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10665; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10666; AVX-NEXT:    vmovaps 1856(%rdi), %ymm110667; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10668; AVX-NEXT:    vinsertf128 $1, 1824(%rdi), %ymm0, %ymm010669; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10670; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10671; AVX-NEXT:    vmovaps 1776(%rdi), %xmm210672; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10673; AVX-NEXT:    vmovaps 1728(%rdi), %xmm110674; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10675; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10676; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10677; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10678; AVX-NEXT:    vmovaps 2240(%rdi), %ymm110679; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10680; AVX-NEXT:    vinsertf128 $1, 2208(%rdi), %ymm0, %ymm010681; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10682; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10683; AVX-NEXT:    vmovaps 2160(%rdi), %xmm210684; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10685; AVX-NEXT:    vmovaps 2112(%rdi), %xmm110686; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10687; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10688; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10689; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10690; AVX-NEXT:    vmovaps 2624(%rdi), %ymm110691; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10692; AVX-NEXT:    vinsertf128 $1, 2592(%rdi), %ymm0, %ymm010693; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10694; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10695; AVX-NEXT:    vmovaps 2544(%rdi), %xmm210696; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10697; AVX-NEXT:    vmovaps 2496(%rdi), %xmm110698; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10699; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10700; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10701; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10702; AVX-NEXT:    vmovaps 3008(%rdi), %ymm110703; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10704; AVX-NEXT:    vinsertf128 $1, 2976(%rdi), %ymm0, %ymm010705; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10706; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10707; AVX-NEXT:    vmovaps 2928(%rdi), %xmm210708; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10709; AVX-NEXT:    vmovaps 2880(%rdi), %xmm110710; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10711; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10712; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10713; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10714; AVX-NEXT:    vmovaps 128(%rdi), %ymm110715; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10716; AVX-NEXT:    vinsertf128 $1, 96(%rdi), %ymm0, %ymm010717; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10718; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10719; AVX-NEXT:    vmovaps (%rdi), %xmm210720; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10721; AVX-NEXT:    vmovaps 48(%rdi), %xmm110722; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10723; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]10724; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10725; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10726; AVX-NEXT:    vmovaps 512(%rdi), %ymm110727; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10728; AVX-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm010729; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10730; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10731; AVX-NEXT:    vmovaps 432(%rdi), %xmm210732; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10733; AVX-NEXT:    vmovaps 384(%rdi), %xmm110734; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10735; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10736; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10737; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10738; AVX-NEXT:    vmovaps 896(%rdi), %ymm110739; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10740; AVX-NEXT:    vinsertf128 $1, 864(%rdi), %ymm0, %ymm010741; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10742; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10743; AVX-NEXT:    vmovaps 816(%rdi), %xmm210744; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10745; AVX-NEXT:    vmovaps 768(%rdi), %xmm110746; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10747; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10748; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10749; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10750; AVX-NEXT:    vmovaps 1280(%rdi), %ymm110751; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10752; AVX-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm010753; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10754; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10755; AVX-NEXT:    vmovaps 1200(%rdi), %xmm210756; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10757; AVX-NEXT:    vmovaps 1152(%rdi), %xmm110758; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10759; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10760; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10761; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10762; AVX-NEXT:    vmovaps 1664(%rdi), %ymm110763; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10764; AVX-NEXT:    vinsertf128 $1, 1632(%rdi), %ymm0, %ymm010765; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10766; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10767; AVX-NEXT:    vmovaps 1584(%rdi), %xmm1410768; AVX-NEXT:    vmovaps 1536(%rdi), %xmm1310769; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]10770; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10771; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10772; AVX-NEXT:    vmovaps 2048(%rdi), %ymm1210773; AVX-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm1110774; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]10775; AVX-NEXT:    vmovaps 1968(%rdi), %xmm1010776; AVX-NEXT:    vmovaps 1920(%rdi), %xmm910777; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]10778; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10779; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10780; AVX-NEXT:    vmovaps 2432(%rdi), %ymm810781; AVX-NEXT:    vinsertf128 $1, 2400(%rdi), %ymm0, %ymm710782; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]10783; AVX-NEXT:    vmovaps 2352(%rdi), %xmm610784; AVX-NEXT:    vmovaps 2304(%rdi), %xmm510785; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]10786; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10787; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10788; AVX-NEXT:    vmovaps 2816(%rdi), %ymm410789; AVX-NEXT:    vinsertf128 $1, 2784(%rdi), %ymm0, %ymm310790; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]10791; AVX-NEXT:    vmovaps 2736(%rdi), %xmm210792; AVX-NEXT:    vmovaps 2688(%rdi), %xmm110793; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]10794; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10795; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10796; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10797; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10798; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10799; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10800; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10801; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10802; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10803; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10804; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10805; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10806; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10807; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10808; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10809; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10810; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10811; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10812; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10813; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10814; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10815; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10816; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10817; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10818; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10819; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10820; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10821; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10822; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10823; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10824; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10825; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10826; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10827; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10828; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10829; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10830; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10831; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10832; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10833; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10834; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10835; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10836; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10837; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10838; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10839; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10840; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10841; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10842; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10843; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10844; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10845; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10846; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10847; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10848; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10849; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10850; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10851; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10852; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10853; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10854; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10855; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10856; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10857; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10858; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10859; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10860; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10861; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10862; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10863; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10864; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10865; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10866; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10867; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10868; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10869; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10870; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10871; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10872; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10873; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10874; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10875; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10876; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10877; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10878; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10879; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10880; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10881; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10882; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10883; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10884; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10885; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10886; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10887; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10888; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10889; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]10890; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10891; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10892; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10893; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10894; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]10895; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]10896; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]10897; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10898; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]10899; AVX-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]10900; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]10901; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10902; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]10903; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]10904; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]10905; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10906; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]10907; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]10908; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10909; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10910; AVX-NEXT:    vmovaps 288(%rdi), %ymm110911; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10912; AVX-NEXT:    vinsertf128 $1, 352(%rdi), %ymm0, %ymm010913; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10914; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10915; AVX-NEXT:    vmovaps 256(%rdi), %xmm210916; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10917; AVX-NEXT:    vmovaps 208(%rdi), %xmm110918; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10919; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10920; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10921; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10922; AVX-NEXT:    vmovaps 672(%rdi), %ymm110923; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10924; AVX-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm010925; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10926; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10927; AVX-NEXT:    vmovaps 640(%rdi), %xmm210928; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10929; AVX-NEXT:    vmovaps 592(%rdi), %xmm110930; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10931; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10932; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10933; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10934; AVX-NEXT:    vmovaps 1056(%rdi), %ymm110935; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10936; AVX-NEXT:    vinsertf128 $1, 1120(%rdi), %ymm0, %ymm010937; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10938; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10939; AVX-NEXT:    vmovaps 1024(%rdi), %xmm210940; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10941; AVX-NEXT:    vmovaps 976(%rdi), %xmm110942; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10943; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10944; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10945; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10946; AVX-NEXT:    vmovaps 1440(%rdi), %ymm110947; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10948; AVX-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm010949; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10950; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10951; AVX-NEXT:    vmovaps 1408(%rdi), %xmm210952; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10953; AVX-NEXT:    vmovaps 1360(%rdi), %xmm110954; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10955; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10956; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10957; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10958; AVX-NEXT:    vmovaps 1824(%rdi), %ymm110959; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10960; AVX-NEXT:    vinsertf128 $1, 1888(%rdi), %ymm0, %ymm010961; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10962; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10963; AVX-NEXT:    vmovaps 1792(%rdi), %xmm210964; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10965; AVX-NEXT:    vmovaps 1744(%rdi), %xmm110966; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10967; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10968; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10969; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10970; AVX-NEXT:    vmovaps 2208(%rdi), %ymm110971; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10972; AVX-NEXT:    vinsertf128 $1, 2272(%rdi), %ymm0, %ymm010973; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10974; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10975; AVX-NEXT:    vmovaps 2176(%rdi), %xmm210976; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10977; AVX-NEXT:    vmovaps 2128(%rdi), %xmm110978; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10979; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10980; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10981; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10982; AVX-NEXT:    vmovaps 2592(%rdi), %ymm110983; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10984; AVX-NEXT:    vinsertf128 $1, 2656(%rdi), %ymm0, %ymm010985; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10986; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10987; AVX-NEXT:    vmovaps 2560(%rdi), %xmm210988; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10989; AVX-NEXT:    vmovaps 2512(%rdi), %xmm110990; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10991; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10992; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10993; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10994; AVX-NEXT:    vmovaps 2976(%rdi), %ymm110995; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10996; AVX-NEXT:    vinsertf128 $1, 3040(%rdi), %ymm0, %ymm010997; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10998; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]10999; AVX-NEXT:    vmovaps 2944(%rdi), %xmm211000; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11001; AVX-NEXT:    vmovaps 2896(%rdi), %xmm111002; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11003; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11004; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11005; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11006; AVX-NEXT:    vmovaps 2784(%rdi), %ymm111007; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11008; AVX-NEXT:    vinsertf128 $1, 2848(%rdi), %ymm0, %ymm011009; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11010; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11011; AVX-NEXT:    vmovaps 2752(%rdi), %xmm211012; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11013; AVX-NEXT:    vmovaps 2704(%rdi), %xmm111014; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11015; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11016; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11017; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11018; AVX-NEXT:    vmovaps 2400(%rdi), %ymm111019; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11020; AVX-NEXT:    vinsertf128 $1, 2464(%rdi), %ymm0, %ymm011021; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11022; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11023; AVX-NEXT:    vmovaps 2368(%rdi), %xmm211024; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11025; AVX-NEXT:    vmovaps 2320(%rdi), %xmm111026; AVX-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill11027; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11028; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11029; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11030; AVX-NEXT:    vmovaps 2016(%rdi), %ymm111031; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11032; AVX-NEXT:    vinsertf128 $1, 2080(%rdi), %ymm0, %ymm011033; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11034; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11035; AVX-NEXT:    vmovaps 1984(%rdi), %xmm211036; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11037; AVX-NEXT:    vmovaps 1936(%rdi), %xmm111038; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11039; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11040; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11041; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11042; AVX-NEXT:    vmovaps 1632(%rdi), %ymm111043; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11044; AVX-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm0, %ymm011045; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11046; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11047; AVX-NEXT:    vmovaps 1600(%rdi), %xmm211048; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11049; AVX-NEXT:    vmovaps 1552(%rdi), %xmm111050; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11051; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11052; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11053; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11054; AVX-NEXT:    vmovaps 1248(%rdi), %ymm111055; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11056; AVX-NEXT:    vinsertf128 $1, 1312(%rdi), %ymm0, %ymm011057; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11058; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11059; AVX-NEXT:    vmovaps 1216(%rdi), %xmm1211060; AVX-NEXT:    vmovaps 1168(%rdi), %xmm1111061; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]11062; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11063; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11064; AVX-NEXT:    vmovaps 864(%rdi), %ymm1011065; AVX-NEXT:    vinsertf128 $1, 928(%rdi), %ymm0, %ymm911066; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]11067; AVX-NEXT:    vmovaps 832(%rdi), %xmm811068; AVX-NEXT:    vmovaps 784(%rdi), %xmm711069; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm7[0],xmm8[0]11070; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11071; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11072; AVX-NEXT:    vmovaps 480(%rdi), %ymm611073; AVX-NEXT:    vinsertf128 $1, 544(%rdi), %ymm0, %ymm1511074; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm15[0],ymm6[2],ymm15[2]11075; AVX-NEXT:    vmovaps 448(%rdi), %xmm511076; AVX-NEXT:    vmovaps 400(%rdi), %xmm411077; AVX-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm4[0],xmm5[0]11078; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm0[4,5,6,7]11079; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11080; AVX-NEXT:    vmovaps 96(%rdi), %ymm211081; AVX-NEXT:    vmovaps 16(%rdi), %xmm1411082; AVX-NEXT:    vinsertf128 $1, 160(%rdi), %ymm0, %ymm111083; AVX-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]11084; AVX-NEXT:    vmovaps 64(%rdi), %xmm011085; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm14[0],xmm0[0]11086; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm13[0,1,2,3],ymm3[4,5,6,7]11087; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11088; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]11089; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm14[1],xmm0[1]11090; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11091; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11092; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11093; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11094; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11095; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11096; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11097; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11098; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11099; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11100; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm15[1],ymm6[3],ymm15[3]11101; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm5[1]11102; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11103; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11104; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11105; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11106; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11107; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11108; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11109; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11110; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11111; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11112; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]11113; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm8[1]11114; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11115; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11116; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11117; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11118; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11119; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11120; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11121; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11122; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11123; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11124; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11125; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11126; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11127; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm11[1],xmm12[1]11128; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11129; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11130; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11131; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11132; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11133; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11134; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11135; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11136; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11137; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11138; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11139; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11140; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11141; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11142; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11143; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11144; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11145; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11146; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11147; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11148; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11149; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11150; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11151; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11152; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11153; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11154; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11155; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11156; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11157; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11158; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11159; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11160; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11161; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11162; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11163; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11164; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11165; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11166; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11167; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11168; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11169; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11170; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11171; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11172; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11173; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload11174; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11175; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11176; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11177; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11178; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11179; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11180; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11181; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11182; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11183; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11184; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11185; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11186; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11187; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11188; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11189; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11190; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11191; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11192; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11193; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11194; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11195; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11196; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11197; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11198; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11199; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]11200; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11201; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11202; AVX-NEXT:    vmovaps 160(%rdi), %ymm111203; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11204; AVX-NEXT:    vmovaps 32(%rdi), %xmm211205; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11206; AVX-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm011207; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11208; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11209; AVX-NEXT:    vmovaps 80(%rdi), %xmm111210; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11211; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]11212; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11213; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11214; AVX-NEXT:    vmovaps 352(%rdi), %ymm111215; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11216; AVX-NEXT:    vinsertf128 $1, 320(%rdi), %ymm0, %ymm011217; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11218; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11219; AVX-NEXT:    vmovaps 272(%rdi), %xmm211220; AVX-NEXT:    vmovaps %xmm2, (%rsp) # 16-byte Spill11221; AVX-NEXT:    vmovaps 224(%rdi), %xmm111222; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11223; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11224; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11225; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11226; AVX-NEXT:    vmovaps 544(%rdi), %ymm111227; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11228; AVX-NEXT:    vinsertf128 $1, 512(%rdi), %ymm0, %ymm011229; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11230; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11231; AVX-NEXT:    vmovaps 464(%rdi), %xmm211232; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11233; AVX-NEXT:    vmovaps 416(%rdi), %xmm111234; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11235; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11236; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11237; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11238; AVX-NEXT:    vmovaps 736(%rdi), %ymm111239; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11240; AVX-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm011241; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11242; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11243; AVX-NEXT:    vmovaps 656(%rdi), %xmm211244; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11245; AVX-NEXT:    vmovaps 608(%rdi), %xmm111246; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11247; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11248; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11249; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11250; AVX-NEXT:    vmovaps 928(%rdi), %ymm111251; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11252; AVX-NEXT:    vinsertf128 $1, 896(%rdi), %ymm0, %ymm011253; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11254; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11255; AVX-NEXT:    vmovaps 848(%rdi), %xmm211256; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11257; AVX-NEXT:    vmovaps 800(%rdi), %xmm111258; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11259; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11260; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11261; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11262; AVX-NEXT:    vmovaps 1120(%rdi), %ymm111263; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11264; AVX-NEXT:    vinsertf128 $1, 1088(%rdi), %ymm0, %ymm011265; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11266; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11267; AVX-NEXT:    vmovaps 1040(%rdi), %xmm211268; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11269; AVX-NEXT:    vmovaps 992(%rdi), %xmm111270; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11271; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11272; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11273; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11274; AVX-NEXT:    vmovaps 1312(%rdi), %ymm111275; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11276; AVX-NEXT:    vinsertf128 $1, 1280(%rdi), %ymm0, %ymm011277; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11278; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11279; AVX-NEXT:    vmovaps 1232(%rdi), %xmm211280; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11281; AVX-NEXT:    vmovaps 1184(%rdi), %xmm111282; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11283; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11284; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11285; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11286; AVX-NEXT:    vmovaps 1504(%rdi), %ymm111287; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11288; AVX-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm0, %ymm011289; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11290; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11291; AVX-NEXT:    vmovaps 1424(%rdi), %xmm211292; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11293; AVX-NEXT:    vmovaps 1376(%rdi), %xmm111294; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11295; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11296; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11297; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11298; AVX-NEXT:    vmovaps 1696(%rdi), %ymm111299; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11300; AVX-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm0, %ymm011301; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11302; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11303; AVX-NEXT:    vmovaps 1616(%rdi), %xmm211304; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11305; AVX-NEXT:    vmovaps 1568(%rdi), %xmm111306; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11307; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11308; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11309; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11310; AVX-NEXT:    vmovaps 1888(%rdi), %ymm111311; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11312; AVX-NEXT:    vinsertf128 $1, 1856(%rdi), %ymm0, %ymm011313; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11314; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11315; AVX-NEXT:    vmovaps 1808(%rdi), %xmm211316; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11317; AVX-NEXT:    vmovaps 1760(%rdi), %xmm111318; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11319; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11320; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11321; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11322; AVX-NEXT:    vmovaps 2080(%rdi), %ymm111323; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11324; AVX-NEXT:    vinsertf128 $1, 2048(%rdi), %ymm0, %ymm011325; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11326; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11327; AVX-NEXT:    vmovaps 2000(%rdi), %xmm211328; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11329; AVX-NEXT:    vmovaps 1952(%rdi), %xmm111330; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11331; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11332; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11333; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11334; AVX-NEXT:    vmovaps 2272(%rdi), %ymm111335; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11336; AVX-NEXT:    vinsertf128 $1, 2240(%rdi), %ymm0, %ymm011337; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11338; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11339; AVX-NEXT:    vmovaps 2192(%rdi), %xmm211340; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11341; AVX-NEXT:    vmovaps 2144(%rdi), %xmm111342; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11343; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11344; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11345; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11346; AVX-NEXT:    vmovaps 2464(%rdi), %ymm111347; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11348; AVX-NEXT:    vinsertf128 $1, 2432(%rdi), %ymm0, %ymm011349; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11350; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11351; AVX-NEXT:    vmovaps 2384(%rdi), %xmm1411352; AVX-NEXT:    vmovaps 2336(%rdi), %xmm1311353; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]11354; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11355; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11356; AVX-NEXT:    vmovaps 2656(%rdi), %ymm1211357; AVX-NEXT:    vinsertf128 $1, 2624(%rdi), %ymm0, %ymm1111358; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]11359; AVX-NEXT:    vmovaps 2576(%rdi), %xmm1011360; AVX-NEXT:    vmovaps 2528(%rdi), %xmm911361; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]11362; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11363; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11364; AVX-NEXT:    vmovaps 2848(%rdi), %ymm811365; AVX-NEXT:    vinsertf128 $1, 2816(%rdi), %ymm0, %ymm711366; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]11367; AVX-NEXT:    vmovaps 2768(%rdi), %xmm611368; AVX-NEXT:    vmovaps 2720(%rdi), %xmm511369; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]11370; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11371; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11372; AVX-NEXT:    vmovaps 3040(%rdi), %ymm411373; AVX-NEXT:    vinsertf128 $1, 3008(%rdi), %ymm0, %ymm311374; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]11375; AVX-NEXT:    vmovaps 2960(%rdi), %xmm211376; AVX-NEXT:    vmovaps 2912(%rdi), %xmm111377; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]11378; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11379; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11380; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11381; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11382; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11383; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11384; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11385; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11386; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11387; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11388; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11389; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11390; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11391; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11392; AVX-NEXT:    vunpckhpd (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload11393; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11394; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11395; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11396; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11397; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11398; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11399; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11400; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11401; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11402; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11403; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11404; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11405; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11406; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11407; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11408; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11409; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11410; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11411; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11412; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11413; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11414; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11415; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11416; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11417; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11418; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11419; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11420; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11421; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11422; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11423; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11424; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11425; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11426; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11427; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11428; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11429; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11430; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11431; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11432; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11433; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11434; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11435; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11436; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11437; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11438; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11439; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11440; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11441; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11442; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11443; AVX-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill11444; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11445; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11446; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11447; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11448; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11449; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11450; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11451; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11452; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11453; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11454; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11455; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11456; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11457; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11458; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11459; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11460; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11461; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11462; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11463; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11464; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11465; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11466; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11467; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11468; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11469; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11470; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11471; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11472; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload11473; AVX-NEXT:    # xmm15 = xmm15[1],mem[1]11474; AVX-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]11475; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11476; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11477; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11478; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]11479; AVX-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]11480; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]11481; AVX-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]11482; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]11483; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]11484; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]11485; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]11486; AVX-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]11487; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]11488; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm3[4,5,6,7]11489; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11490; AVX-NEXT:    vmovaps %ymm1, 448(%rsi)11491; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11492; AVX-NEXT:    vmovaps %ymm1, 384(%rsi)11493; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11494; AVX-NEXT:    vmovaps %ymm1, 320(%rsi)11495; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11496; AVX-NEXT:    vmovaps %ymm1, 256(%rsi)11497; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11498; AVX-NEXT:    vmovaps %ymm1, 192(%rsi)11499; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11500; AVX-NEXT:    vmovaps %ymm1, 128(%rsi)11501; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11502; AVX-NEXT:    vmovaps %ymm1, 64(%rsi)11503; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11504; AVX-NEXT:    vmovaps %ymm1, (%rsi)11505; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11506; AVX-NEXT:    vmovaps %ymm1, 480(%rsi)11507; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11508; AVX-NEXT:    vmovaps %ymm1, 416(%rsi)11509; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11510; AVX-NEXT:    vmovaps %ymm1, 352(%rsi)11511; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11512; AVX-NEXT:    vmovaps %ymm1, 288(%rsi)11513; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11514; AVX-NEXT:    vmovaps %ymm1, 224(%rsi)11515; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11516; AVX-NEXT:    vmovaps %ymm1, 160(%rsi)11517; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11518; AVX-NEXT:    vmovaps %ymm1, 96(%rsi)11519; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11520; AVX-NEXT:    vmovaps %ymm1, 32(%rsi)11521; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11522; AVX-NEXT:    vmovaps %ymm1, 448(%rdx)11523; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11524; AVX-NEXT:    vmovaps %ymm1, 384(%rdx)11525; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11526; AVX-NEXT:    vmovaps %ymm1, 320(%rdx)11527; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11528; AVX-NEXT:    vmovaps %ymm1, 256(%rdx)11529; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11530; AVX-NEXT:    vmovaps %ymm1, 192(%rdx)11531; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11532; AVX-NEXT:    vmovaps %ymm1, 128(%rdx)11533; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11534; AVX-NEXT:    vmovaps %ymm1, 64(%rdx)11535; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11536; AVX-NEXT:    vmovaps %ymm1, (%rdx)11537; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11538; AVX-NEXT:    vmovaps %ymm1, 480(%rdx)11539; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11540; AVX-NEXT:    vmovaps %ymm1, 416(%rdx)11541; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11542; AVX-NEXT:    vmovaps %ymm1, 352(%rdx)11543; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11544; AVX-NEXT:    vmovaps %ymm1, 288(%rdx)11545; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11546; AVX-NEXT:    vmovaps %ymm1, 224(%rdx)11547; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11548; AVX-NEXT:    vmovaps %ymm1, 160(%rdx)11549; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11550; AVX-NEXT:    vmovaps %ymm1, 96(%rdx)11551; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11552; AVX-NEXT:    vmovaps %ymm1, 32(%rdx)11553; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11554; AVX-NEXT:    vmovaps %ymm1, (%rcx)11555; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11556; AVX-NEXT:    vmovaps %ymm1, 64(%rcx)11557; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11558; AVX-NEXT:    vmovaps %ymm1, 128(%rcx)11559; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11560; AVX-NEXT:    vmovaps %ymm1, 192(%rcx)11561; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11562; AVX-NEXT:    vmovaps %ymm1, 256(%rcx)11563; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11564; AVX-NEXT:    vmovaps %ymm1, 320(%rcx)11565; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11566; AVX-NEXT:    vmovaps %ymm1, 384(%rcx)11567; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11568; AVX-NEXT:    vmovaps %ymm1, 448(%rcx)11569; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11570; AVX-NEXT:    vmovaps %ymm1, 480(%rcx)11571; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11572; AVX-NEXT:    vmovaps %ymm1, 416(%rcx)11573; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11574; AVX-NEXT:    vmovaps %ymm1, 352(%rcx)11575; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11576; AVX-NEXT:    vmovaps %ymm1, 288(%rcx)11577; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11578; AVX-NEXT:    vmovaps %ymm1, 224(%rcx)11579; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11580; AVX-NEXT:    vmovaps %ymm1, 160(%rcx)11581; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11582; AVX-NEXT:    vmovaps %ymm1, 96(%rcx)11583; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11584; AVX-NEXT:    vmovaps %ymm1, 32(%rcx)11585; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11586; AVX-NEXT:    vmovaps %ymm1, 480(%r8)11587; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11588; AVX-NEXT:    vmovaps %ymm1, 448(%r8)11589; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11590; AVX-NEXT:    vmovaps %ymm1, 416(%r8)11591; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11592; AVX-NEXT:    vmovaps %ymm1, 384(%r8)11593; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11594; AVX-NEXT:    vmovaps %ymm1, 352(%r8)11595; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11596; AVX-NEXT:    vmovaps %ymm1, 320(%r8)11597; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11598; AVX-NEXT:    vmovaps %ymm1, 288(%r8)11599; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11600; AVX-NEXT:    vmovaps %ymm1, 256(%r8)11601; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11602; AVX-NEXT:    vmovaps %ymm1, 224(%r8)11603; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11604; AVX-NEXT:    vmovaps %ymm1, 192(%r8)11605; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11606; AVX-NEXT:    vmovaps %ymm1, 160(%r8)11607; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11608; AVX-NEXT:    vmovaps %ymm1, 128(%r8)11609; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11610; AVX-NEXT:    vmovaps %ymm1, 96(%r8)11611; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11612; AVX-NEXT:    vmovaps %ymm1, 64(%r8)11613; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11614; AVX-NEXT:    vmovaps %ymm1, 32(%r8)11615; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11616; AVX-NEXT:    vmovaps %ymm1, (%r8)11617; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11618; AVX-NEXT:    vmovaps %ymm1, 480(%r9)11619; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11620; AVX-NEXT:    vmovaps %ymm1, 448(%r9)11621; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11622; AVX-NEXT:    vmovaps %ymm1, 416(%r9)11623; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11624; AVX-NEXT:    vmovaps %ymm1, 384(%r9)11625; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11626; AVX-NEXT:    vmovaps %ymm1, 352(%r9)11627; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11628; AVX-NEXT:    vmovaps %ymm1, 320(%r9)11629; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11630; AVX-NEXT:    vmovaps %ymm1, 288(%r9)11631; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11632; AVX-NEXT:    vmovaps %ymm1, 256(%r9)11633; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11634; AVX-NEXT:    vmovaps %ymm1, 224(%r9)11635; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11636; AVX-NEXT:    vmovaps %ymm1, 192(%r9)11637; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11638; AVX-NEXT:    vmovaps %ymm1, 160(%r9)11639; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11640; AVX-NEXT:    vmovaps %ymm1, 128(%r9)11641; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11642; AVX-NEXT:    vmovaps %ymm1, 96(%r9)11643; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11644; AVX-NEXT:    vmovaps %ymm1, 64(%r9)11645; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11646; AVX-NEXT:    vmovaps %ymm1, 32(%r9)11647; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11648; AVX-NEXT:    vmovaps %ymm1, (%r9)11649; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax11650; AVX-NEXT:    vmovaps %ymm4, 480(%rax)11651; AVX-NEXT:    vmovaps %ymm5, 448(%rax)11652; AVX-NEXT:    vmovaps %ymm0, 416(%rax)11653; AVX-NEXT:    vmovaps %ymm13, 384(%rax)11654; AVX-NEXT:    vmovaps %ymm15, 352(%rax)11655; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11656; AVX-NEXT:    vmovaps %ymm0, 320(%rax)11657; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11658; AVX-NEXT:    vmovaps %ymm0, 288(%rax)11659; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11660; AVX-NEXT:    vmovaps %ymm0, 256(%rax)11661; AVX-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload11662; AVX-NEXT:    vmovaps %ymm0, 224(%rax)11663; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11664; AVX-NEXT:    vmovaps %ymm0, 192(%rax)11665; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11666; AVX-NEXT:    vmovaps %ymm0, 160(%rax)11667; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11668; AVX-NEXT:    vmovaps %ymm0, 128(%rax)11669; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11670; AVX-NEXT:    vmovaps %ymm0, 96(%rax)11671; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11672; AVX-NEXT:    vmovaps %ymm0, 64(%rax)11673; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11674; AVX-NEXT:    vmovaps %ymm0, 32(%rax)11675; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11676; AVX-NEXT:    vmovaps %ymm0, (%rax)11677; AVX-NEXT:    addq $3768, %rsp # imm = 0xEB811678; AVX-NEXT:    vzeroupper11679; AVX-NEXT:    retq11680;11681; AVX2-LABEL: load_i64_stride6_vf64:11682; AVX2:       # %bb.0:11683; AVX2-NEXT:    subq $3432, %rsp # imm = 0xD6811684; AVX2-NEXT:    vmovaps 1088(%rdi), %ymm211685; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11686; AVX2-NEXT:    vmovaps 1056(%rdi), %ymm411687; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11688; AVX2-NEXT:    vmovaps 704(%rdi), %ymm311689; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11690; AVX2-NEXT:    vmovaps 672(%rdi), %ymm511691; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11692; AVX2-NEXT:    vmovaps 320(%rdi), %ymm111693; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11694; AVX2-NEXT:    vmovaps 288(%rdi), %ymm611695; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11696; AVX2-NEXT:    vmovaps 240(%rdi), %xmm711697; AVX2-NEXT:    vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11698; AVX2-NEXT:    vmovaps 192(%rdi), %xmm011699; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11700; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm7[0]11701; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]11702; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]11703; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11704; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11705; AVX2-NEXT:    vmovaps 624(%rdi), %xmm111706; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11707; AVX2-NEXT:    vmovaps 576(%rdi), %xmm011708; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11709; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]11710; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]11711; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]11712; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11713; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11714; AVX2-NEXT:    vmovaps 1008(%rdi), %xmm111715; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11716; AVX2-NEXT:    vmovaps 960(%rdi), %xmm011717; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11718; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]11719; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]11720; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]11721; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11722; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11723; AVX2-NEXT:    vmovaps 1472(%rdi), %ymm011724; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11725; AVX2-NEXT:    vmovaps 1440(%rdi), %ymm111726; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11727; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11728; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11729; AVX2-NEXT:    vmovaps 1392(%rdi), %xmm211730; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11731; AVX2-NEXT:    vmovaps 1344(%rdi), %xmm111732; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11733; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11734; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11735; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11736; AVX2-NEXT:    vmovaps 1856(%rdi), %ymm011737; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11738; AVX2-NEXT:    vmovaps 1824(%rdi), %ymm111739; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11740; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11741; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11742; AVX2-NEXT:    vmovaps 1776(%rdi), %xmm211743; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11744; AVX2-NEXT:    vmovaps 1728(%rdi), %xmm111745; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11746; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11747; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11748; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11749; AVX2-NEXT:    vmovaps 2240(%rdi), %ymm011750; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11751; AVX2-NEXT:    vmovaps 2208(%rdi), %ymm111752; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11753; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11754; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11755; AVX2-NEXT:    vmovaps 2160(%rdi), %xmm211756; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11757; AVX2-NEXT:    vmovaps 2112(%rdi), %xmm111758; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11759; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11760; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11761; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11762; AVX2-NEXT:    vmovaps 2624(%rdi), %ymm011763; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11764; AVX2-NEXT:    vmovaps 2592(%rdi), %ymm111765; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11766; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11767; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11768; AVX2-NEXT:    vmovaps 2544(%rdi), %xmm211769; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11770; AVX2-NEXT:    vmovaps 2496(%rdi), %xmm111771; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11772; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11773; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11774; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11775; AVX2-NEXT:    vmovaps 3008(%rdi), %ymm011776; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11777; AVX2-NEXT:    vmovaps 2976(%rdi), %ymm111778; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11779; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11780; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11781; AVX2-NEXT:    vmovaps 2928(%rdi), %xmm211782; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11783; AVX2-NEXT:    vmovaps 2880(%rdi), %xmm111784; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11785; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]11786; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11787; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11788; AVX2-NEXT:    vmovaps 128(%rdi), %ymm011789; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11790; AVX2-NEXT:    vmovaps 96(%rdi), %ymm111791; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11792; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11793; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11794; AVX2-NEXT:    vmovaps (%rdi), %xmm211795; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11796; AVX2-NEXT:    vmovaps 48(%rdi), %xmm111797; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11798; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]11799; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11800; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11801; AVX2-NEXT:    vmovaps 512(%rdi), %ymm011802; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11803; AVX2-NEXT:    vmovaps 480(%rdi), %ymm111804; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11805; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11806; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11807; AVX2-NEXT:    vmovaps 432(%rdi), %xmm111808; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11809; AVX2-NEXT:    vmovaps 384(%rdi), %xmm1311810; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]11811; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11812; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11813; AVX2-NEXT:    vmovaps 896(%rdi), %ymm011814; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11815; AVX2-NEXT:    vmovaps 864(%rdi), %ymm111816; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11817; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11818; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11819; AVX2-NEXT:    vmovaps 816(%rdi), %xmm111820; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11821; AVX2-NEXT:    vmovaps 768(%rdi), %xmm1111822; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]11823; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11824; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11825; AVX2-NEXT:    vmovaps 1280(%rdi), %ymm011826; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11827; AVX2-NEXT:    vmovaps 1248(%rdi), %ymm111828; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11829; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11830; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11831; AVX2-NEXT:    vmovaps 1200(%rdi), %xmm111832; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11833; AVX2-NEXT:    vmovaps 1152(%rdi), %xmm911834; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]11835; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11836; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11837; AVX2-NEXT:    vmovaps 1664(%rdi), %ymm011838; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11839; AVX2-NEXT:    vmovaps 1632(%rdi), %ymm111840; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11841; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11842; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11843; AVX2-NEXT:    vmovaps 1584(%rdi), %xmm1411844; AVX2-NEXT:    vmovaps 1536(%rdi), %xmm711845; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm14[0]11846; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11847; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11848; AVX2-NEXT:    vmovaps 2048(%rdi), %ymm1211849; AVX2-NEXT:    vmovaps 2016(%rdi), %ymm011850; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11851; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]11852; AVX2-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11853; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11854; AVX2-NEXT:    vmovaps 1968(%rdi), %xmm1011855; AVX2-NEXT:    vmovaps 1920(%rdi), %xmm511856; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm10[0]11857; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11858; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11859; AVX2-NEXT:    vmovaps 2432(%rdi), %ymm811860; AVX2-NEXT:    vmovaps 2400(%rdi), %ymm011861; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11862; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]11863; AVX2-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11864; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11865; AVX2-NEXT:    vmovaps 2352(%rdi), %xmm611866; AVX2-NEXT:    vmovaps 2304(%rdi), %xmm311867; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]11868; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11869; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11870; AVX2-NEXT:    vmovaps 2816(%rdi), %ymm411871; AVX2-NEXT:    vmovaps 2784(%rdi), %ymm011872; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11873; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]11874; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11875; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]11876; AVX2-NEXT:    vmovaps 2736(%rdi), %xmm211877; AVX2-NEXT:    vmovaps 2688(%rdi), %xmm111878; AVX2-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]11879; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]11880; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11881; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11882; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11883; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11884; AVX2-NEXT:    vbroadcastsd 296(%rdi), %ymm1511885; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11886; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11887; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11888; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11889; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11890; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11891; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11892; AVX2-NEXT:    vbroadcastsd 680(%rdi), %ymm1511893; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11894; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11895; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11896; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11897; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11898; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11899; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11900; AVX2-NEXT:    vbroadcastsd 1064(%rdi), %ymm1511901; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11902; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11903; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11904; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11905; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11906; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11907; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11908; AVX2-NEXT:    vbroadcastsd 1448(%rdi), %ymm1511909; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11910; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11911; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11912; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11913; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11914; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11915; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11916; AVX2-NEXT:    vbroadcastsd 1832(%rdi), %ymm1511917; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11918; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11919; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11920; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11921; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11922; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11923; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11924; AVX2-NEXT:    vbroadcastsd 2216(%rdi), %ymm1511925; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11926; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11927; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11928; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11929; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11930; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11931; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11932; AVX2-NEXT:    vbroadcastsd 2600(%rdi), %ymm1511933; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11934; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11935; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11936; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11937; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11938; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11939; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11940; AVX2-NEXT:    vbroadcastsd 2984(%rdi), %ymm1511941; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11942; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11943; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11944; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11945; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload11946; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11947; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]11948; AVX2-NEXT:    vbroadcastsd 104(%rdi), %ymm1511949; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11950; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]11951; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]11952; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11953; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm0 # 16-byte Folded Reload11954; AVX2-NEXT:    # xmm0 = xmm13[1],mem[1]11955; AVX2-NEXT:    vbroadcastsd 488(%rdi), %ymm1311956; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload11957; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]11958; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]11959; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11960; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm0 # 16-byte Folded Reload11961; AVX2-NEXT:    # xmm0 = xmm11[1],mem[1]11962; AVX2-NEXT:    vbroadcastsd 872(%rdi), %ymm1111963; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload11964; AVX2-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]11965; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]11966; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11967; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm0 # 16-byte Folded Reload11968; AVX2-NEXT:    # xmm0 = xmm9[1],mem[1]11969; AVX2-NEXT:    vbroadcastsd 1256(%rdi), %ymm911970; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload11971; AVX2-NEXT:    # ymm9 = ymm9[1],mem[1],ymm9[3],mem[3]11972; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]11973; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11974; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm14[1]11975; AVX2-NEXT:    vbroadcastsd 1640(%rdi), %ymm711976; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload11977; AVX2-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]11978; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]11979; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11980; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm10[1]11981; AVX2-NEXT:    vbroadcastsd 2024(%rdi), %ymm511982; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm12[1],ymm5[3],ymm12[3]11983; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]11984; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11985; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm6[1]11986; AVX2-NEXT:    vbroadcastsd 2408(%rdi), %ymm311987; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm8[1],ymm3[3],ymm8[3]11988; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]11989; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11990; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]11991; AVX2-NEXT:    vbroadcastsd 2792(%rdi), %ymm111992; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]11993; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11994; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11995; AVX2-NEXT:    vbroadcastsd 352(%rdi), %ymm011996; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11997; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11998; AVX2-NEXT:    vmovaps 256(%rdi), %xmm211999; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12000; AVX2-NEXT:    vmovaps 208(%rdi), %xmm112001; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12002; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12003; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12004; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12005; AVX2-NEXT:    vbroadcastsd 736(%rdi), %ymm012006; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12007; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12008; AVX2-NEXT:    vmovaps 640(%rdi), %xmm212009; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12010; AVX2-NEXT:    vmovaps 592(%rdi), %xmm112011; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12012; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12013; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12014; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12015; AVX2-NEXT:    vbroadcastsd 1120(%rdi), %ymm012016; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12017; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12018; AVX2-NEXT:    vmovaps 1024(%rdi), %xmm212019; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12020; AVX2-NEXT:    vmovaps 976(%rdi), %xmm112021; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12022; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12023; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12024; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12025; AVX2-NEXT:    vbroadcastsd 1504(%rdi), %ymm012026; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12027; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12028; AVX2-NEXT:    vmovaps 1408(%rdi), %xmm212029; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12030; AVX2-NEXT:    vmovaps 1360(%rdi), %xmm112031; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12032; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12033; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12034; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12035; AVX2-NEXT:    vbroadcastsd 1888(%rdi), %ymm012036; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12037; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12038; AVX2-NEXT:    vmovaps 1792(%rdi), %xmm212039; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12040; AVX2-NEXT:    vmovaps 1744(%rdi), %xmm112041; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12042; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12043; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12044; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12045; AVX2-NEXT:    vbroadcastsd 2272(%rdi), %ymm012046; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12047; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12048; AVX2-NEXT:    vmovaps 2176(%rdi), %xmm212049; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12050; AVX2-NEXT:    vmovaps 2128(%rdi), %xmm112051; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12052; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12053; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12054; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12055; AVX2-NEXT:    vbroadcastsd 2656(%rdi), %ymm012056; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12057; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12058; AVX2-NEXT:    vmovaps 2560(%rdi), %xmm212059; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12060; AVX2-NEXT:    vmovaps 2512(%rdi), %xmm112061; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12062; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12063; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12064; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12065; AVX2-NEXT:    vbroadcastsd 3040(%rdi), %ymm012066; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12067; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12068; AVX2-NEXT:    vmovaps 2944(%rdi), %xmm212069; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12070; AVX2-NEXT:    vmovaps 2896(%rdi), %xmm112071; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12072; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12073; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12074; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12075; AVX2-NEXT:    vbroadcastsd 2848(%rdi), %ymm012076; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12077; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12078; AVX2-NEXT:    vmovaps 2752(%rdi), %xmm212079; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12080; AVX2-NEXT:    vmovaps 2704(%rdi), %xmm112081; AVX2-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill12082; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12083; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12084; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12085; AVX2-NEXT:    vbroadcastsd 2464(%rdi), %ymm012086; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12087; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12088; AVX2-NEXT:    vmovaps 2368(%rdi), %xmm212089; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12090; AVX2-NEXT:    vmovaps 2320(%rdi), %xmm112091; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12092; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12093; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12094; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12095; AVX2-NEXT:    vbroadcastsd 2080(%rdi), %ymm012096; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12097; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12098; AVX2-NEXT:    vmovaps 1984(%rdi), %xmm212099; AVX2-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12100; AVX2-NEXT:    vmovaps 1936(%rdi), %xmm112101; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12102; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12103; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12104; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12105; AVX2-NEXT:    vbroadcastsd 1696(%rdi), %ymm012106; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12107; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12108; AVX2-NEXT:    vmovaps 1600(%rdi), %xmm1212109; AVX2-NEXT:    vmovaps 1552(%rdi), %xmm1112110; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]12111; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12112; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12113; AVX2-NEXT:    vbroadcastsd 1312(%rdi), %ymm012114; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload12115; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]12116; AVX2-NEXT:    vmovaps 1216(%rdi), %xmm912117; AVX2-NEXT:    vmovaps 1168(%rdi), %xmm812118; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]12119; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12120; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12121; AVX2-NEXT:    vbroadcastsd 928(%rdi), %ymm012122; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload12123; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[2],ymm0[2]12124; AVX2-NEXT:    vmovaps 832(%rdi), %xmm612125; AVX2-NEXT:    vmovaps 784(%rdi), %xmm512126; AVX2-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm5[0],xmm6[0]12127; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]12128; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12129; AVX2-NEXT:    vbroadcastsd 544(%rdi), %ymm012130; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload12131; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[2],ymm0[2]12132; AVX2-NEXT:    vmovaps 448(%rdi), %xmm1312133; AVX2-NEXT:    vmovaps 400(%rdi), %xmm312134; AVX2-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm3[0],xmm13[0]12135; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12136; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12137; AVX2-NEXT:    vbroadcastsd 160(%rdi), %ymm012138; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12139; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]12140; AVX2-NEXT:    vmovaps 16(%rdi), %xmm1512141; AVX2-NEXT:    vmovaps 64(%rdi), %xmm012142; AVX2-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm15[0],xmm0[0]12143; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3],ymm1[4,5,6,7]12144; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12145; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm15[1],xmm0[1]12146; AVX2-NEXT:    vmovaps 160(%rdi), %ymm1412147; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm14[1],ymm2[3],ymm14[3]12148; AVX2-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12149; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12150; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12151; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12152; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12153; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12154; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12155; AVX2-NEXT:    vmovaps 352(%rdi), %ymm1512156; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12157; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm15[1],ymm1[3],ymm15[3]12158; AVX2-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12159; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12160; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12161; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12162; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm13[1]12163; AVX2-NEXT:    vmovaps 544(%rdi), %ymm312164; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm3[1],ymm4[3],ymm3[3]12165; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12166; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12167; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12168; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12169; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12170; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12171; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12172; AVX2-NEXT:    vmovaps 736(%rdi), %ymm412173; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12174; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]12175; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12176; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12177; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12178; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12179; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm6[1]12180; AVX2-NEXT:    vmovaps 928(%rdi), %ymm512181; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm5[1],ymm7[3],ymm5[3]12182; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12183; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12184; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12185; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12186; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12187; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12188; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12189; AVX2-NEXT:    vmovaps 1120(%rdi), %ymm612190; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12191; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm6[1],ymm1[3],ymm6[3]12192; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12193; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12194; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12195; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12196; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm9[1]12197; AVX2-NEXT:    vmovaps 1312(%rdi), %ymm712198; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm10[1],ymm7[1],ymm10[3],ymm7[3]12199; AVX2-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12200; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12201; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12202; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12203; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12204; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12205; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12206; AVX2-NEXT:    vmovaps 1504(%rdi), %ymm812207; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12208; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm8[1],ymm1[3],ymm8[3]12209; AVX2-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12210; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12211; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12212; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12213; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm11[1],xmm12[1]12214; AVX2-NEXT:    vmovaps 1696(%rdi), %ymm912215; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12216; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]12217; AVX2-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12218; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12219; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12220; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12221; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12222; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12223; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12224; AVX2-NEXT:    vmovaps 1888(%rdi), %ymm1012225; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12226; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]12227; AVX2-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12228; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12229; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12230; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12231; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12232; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12233; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12234; AVX2-NEXT:    vmovaps 2080(%rdi), %ymm1112235; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12236; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]12237; AVX2-NEXT:    vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12238; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12239; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12240; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12241; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12242; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12243; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12244; AVX2-NEXT:    vmovaps 2272(%rdi), %ymm212245; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12246; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12247; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]12248; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12249; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12250; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12251; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12252; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12253; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12254; AVX2-NEXT:    vmovaps 2464(%rdi), %ymm212255; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12256; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12257; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]12258; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12259; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12260; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12261; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12262; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12263; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12264; AVX2-NEXT:    vmovaps 2656(%rdi), %ymm212265; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12266; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12267; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]12268; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12269; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12270; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12271; AVX2-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload12272; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12273; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12274; AVX2-NEXT:    vmovaps 2848(%rdi), %ymm212275; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12276; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12277; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]12278; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12279; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12280; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12281; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12282; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12283; AVX2-NEXT:    # xmm0 = xmm0[1],mem[1]12284; AVX2-NEXT:    vmovaps 3040(%rdi), %ymm212285; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12286; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]12287; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12288; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]12289; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12290; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12291; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12292; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]12293; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12294; AVX2-NEXT:    vmovaps 32(%rdi), %xmm1212295; AVX2-NEXT:    vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12296; AVX2-NEXT:    vmovaps 80(%rdi), %xmm112297; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12298; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm1[0]12299; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12300; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12301; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12302; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]12303; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12304; AVX2-NEXT:    vmovaps 272(%rdi), %xmm1212305; AVX2-NEXT:    vmovaps %xmm12, (%rsp) # 16-byte Spill12306; AVX2-NEXT:    vmovaps 224(%rdi), %xmm112307; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12308; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm12[0]12309; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12310; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12311; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12312; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]12313; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12314; AVX2-NEXT:    vmovaps 464(%rdi), %xmm312315; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12316; AVX2-NEXT:    vmovaps 416(%rdi), %xmm112317; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12318; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12319; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12320; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12321; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12322; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]12323; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12324; AVX2-NEXT:    vmovaps 656(%rdi), %xmm312325; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12326; AVX2-NEXT:    vmovaps 608(%rdi), %xmm112327; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12328; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12329; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12330; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12331; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12332; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]12333; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12334; AVX2-NEXT:    vmovaps 848(%rdi), %xmm312335; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12336; AVX2-NEXT:    vmovaps 800(%rdi), %xmm112337; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12338; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12339; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12340; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12341; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12342; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]12343; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12344; AVX2-NEXT:    vmovaps 1040(%rdi), %xmm312345; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12346; AVX2-NEXT:    vmovaps 992(%rdi), %xmm112347; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12348; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12349; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12350; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12351; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12352; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]12353; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12354; AVX2-NEXT:    vmovaps 1232(%rdi), %xmm312355; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12356; AVX2-NEXT:    vmovaps 1184(%rdi), %xmm112357; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12358; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12359; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12360; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12361; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12362; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]12363; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12364; AVX2-NEXT:    vmovaps 1424(%rdi), %xmm312365; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12366; AVX2-NEXT:    vmovaps 1376(%rdi), %xmm112367; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12368; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12369; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12370; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12371; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12372; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]12373; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12374; AVX2-NEXT:    vmovaps 1616(%rdi), %xmm312375; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12376; AVX2-NEXT:    vmovaps 1568(%rdi), %xmm112377; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12378; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]12379; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12380; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12381; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12382; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]12383; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12384; AVX2-NEXT:    vmovaps 1808(%rdi), %xmm112385; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12386; AVX2-NEXT:    vmovaps 1760(%rdi), %xmm1312387; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]12388; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12389; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12390; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12391; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm11[0],ymm0[2],ymm11[2]12392; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12393; AVX2-NEXT:    vmovaps 2000(%rdi), %xmm112394; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12395; AVX2-NEXT:    vmovaps 1952(%rdi), %xmm1112396; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]12397; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12398; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12399; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12400; AVX2-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12401; AVX2-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]12402; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12403; AVX2-NEXT:    vmovaps 2192(%rdi), %xmm112404; AVX2-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12405; AVX2-NEXT:    vmovaps 2144(%rdi), %xmm912406; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]12407; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12408; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12409; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12410; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload12411; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]12412; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12413; AVX2-NEXT:    vmovaps 2384(%rdi), %xmm1212414; AVX2-NEXT:    vmovaps 2336(%rdi), %xmm712415; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm12[0]12416; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12417; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12418; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12419; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload12420; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]12421; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12422; AVX2-NEXT:    vmovaps 2576(%rdi), %xmm812423; AVX2-NEXT:    vmovaps 2528(%rdi), %xmm512424; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]12425; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12426; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12427; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12428; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload12429; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]12430; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12431; AVX2-NEXT:    vmovaps 2768(%rdi), %xmm412432; AVX2-NEXT:    vmovaps 2720(%rdi), %xmm312433; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm4[0]12434; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12435; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12436; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12437; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]12438; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12439; AVX2-NEXT:    vmovaps 2960(%rdi), %xmm212440; AVX2-NEXT:    vmovaps 2912(%rdi), %xmm112441; AVX2-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]12442; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12443; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12444; AVX2-NEXT:    vbroadcastsd 136(%rdi), %ymm012445; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12446; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12447; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12448; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12449; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12450; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12451; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12452; AVX2-NEXT:    vbroadcastsd 328(%rdi), %ymm012453; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12454; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12455; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12456; AVX2-NEXT:    vunpckhpd (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload12457; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12458; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12459; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12460; AVX2-NEXT:    vbroadcastsd 520(%rdi), %ymm012461; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12462; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12463; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12464; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12465; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12466; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12467; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12468; AVX2-NEXT:    vbroadcastsd 712(%rdi), %ymm012469; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12470; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12471; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12472; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12473; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12474; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12475; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12476; AVX2-NEXT:    vbroadcastsd 904(%rdi), %ymm012477; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12478; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12479; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12480; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12481; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12482; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12483; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12484; AVX2-NEXT:    vbroadcastsd 1096(%rdi), %ymm012485; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12486; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12487; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12488; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12489; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12490; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12491; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12492; AVX2-NEXT:    vbroadcastsd 1288(%rdi), %ymm012493; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12494; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12495; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12496; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12497; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12498; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12499; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12500; AVX2-NEXT:    vbroadcastsd 1480(%rdi), %ymm012501; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12502; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12503; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12504; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12505; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12506; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12507; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12508; AVX2-NEXT:    vbroadcastsd 1672(%rdi), %ymm012509; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12510; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12511; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload12512; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload12513; AVX2-NEXT:    # xmm15 = xmm15[1],mem[1]12514; AVX2-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]12515; AVX2-NEXT:    vbroadcastsd 1864(%rdi), %ymm012516; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12517; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12518; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload12519; AVX2-NEXT:    # xmm13 = xmm13[1],mem[1]12520; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]12521; AVX2-NEXT:    vbroadcastsd 2056(%rdi), %ymm012522; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12523; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12524; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload12525; AVX2-NEXT:    # xmm11 = xmm11[1],mem[1]12526; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm0[4,5,6,7]12527; AVX2-NEXT:    vbroadcastsd 2248(%rdi), %ymm012528; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12529; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12530; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload12531; AVX2-NEXT:    # xmm9 = xmm9[1],mem[1]12532; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]12533; AVX2-NEXT:    vbroadcastsd 2440(%rdi), %ymm012534; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm14[1],ymm0[3],ymm14[3]12535; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm12[1]12536; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]12537; AVX2-NEXT:    vbroadcastsd 2632(%rdi), %ymm012538; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]12539; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]12540; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]12541; AVX2-NEXT:    vbroadcastsd 2824(%rdi), %ymm012542; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm6[1],ymm0[3],ymm6[3]12543; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]12544; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]12545; AVX2-NEXT:    vbroadcastsd 3016(%rdi), %ymm012546; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12547; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]12548; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]12549; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12550; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12551; AVX2-NEXT:    vmovaps %ymm1, 448(%rsi)12552; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12553; AVX2-NEXT:    vmovaps %ymm1, 384(%rsi)12554; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12555; AVX2-NEXT:    vmovaps %ymm1, 320(%rsi)12556; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12557; AVX2-NEXT:    vmovaps %ymm1, 256(%rsi)12558; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12559; AVX2-NEXT:    vmovaps %ymm1, 192(%rsi)12560; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12561; AVX2-NEXT:    vmovaps %ymm1, 128(%rsi)12562; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12563; AVX2-NEXT:    vmovaps %ymm1, 64(%rsi)12564; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12565; AVX2-NEXT:    vmovaps %ymm1, (%rsi)12566; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12567; AVX2-NEXT:    vmovaps %ymm1, 480(%rsi)12568; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12569; AVX2-NEXT:    vmovaps %ymm1, 416(%rsi)12570; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12571; AVX2-NEXT:    vmovaps %ymm1, 352(%rsi)12572; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12573; AVX2-NEXT:    vmovaps %ymm1, 288(%rsi)12574; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12575; AVX2-NEXT:    vmovaps %ymm1, 224(%rsi)12576; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12577; AVX2-NEXT:    vmovaps %ymm1, 160(%rsi)12578; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12579; AVX2-NEXT:    vmovaps %ymm1, 96(%rsi)12580; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12581; AVX2-NEXT:    vmovaps %ymm1, 32(%rsi)12582; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12583; AVX2-NEXT:    vmovaps %ymm1, 448(%rdx)12584; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12585; AVX2-NEXT:    vmovaps %ymm1, 384(%rdx)12586; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12587; AVX2-NEXT:    vmovaps %ymm1, 320(%rdx)12588; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12589; AVX2-NEXT:    vmovaps %ymm1, 256(%rdx)12590; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12591; AVX2-NEXT:    vmovaps %ymm1, 192(%rdx)12592; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12593; AVX2-NEXT:    vmovaps %ymm1, 128(%rdx)12594; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12595; AVX2-NEXT:    vmovaps %ymm1, 64(%rdx)12596; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12597; AVX2-NEXT:    vmovaps %ymm1, (%rdx)12598; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12599; AVX2-NEXT:    vmovaps %ymm1, 480(%rdx)12600; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12601; AVX2-NEXT:    vmovaps %ymm1, 416(%rdx)12602; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12603; AVX2-NEXT:    vmovaps %ymm1, 352(%rdx)12604; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12605; AVX2-NEXT:    vmovaps %ymm1, 288(%rdx)12606; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12607; AVX2-NEXT:    vmovaps %ymm1, 224(%rdx)12608; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12609; AVX2-NEXT:    vmovaps %ymm1, 160(%rdx)12610; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12611; AVX2-NEXT:    vmovaps %ymm1, 96(%rdx)12612; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12613; AVX2-NEXT:    vmovaps %ymm1, 32(%rdx)12614; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12615; AVX2-NEXT:    vmovaps %ymm1, (%rcx)12616; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12617; AVX2-NEXT:    vmovaps %ymm1, 64(%rcx)12618; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12619; AVX2-NEXT:    vmovaps %ymm1, 128(%rcx)12620; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12621; AVX2-NEXT:    vmovaps %ymm1, 192(%rcx)12622; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12623; AVX2-NEXT:    vmovaps %ymm1, 256(%rcx)12624; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12625; AVX2-NEXT:    vmovaps %ymm1, 320(%rcx)12626; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12627; AVX2-NEXT:    vmovaps %ymm1, 384(%rcx)12628; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12629; AVX2-NEXT:    vmovaps %ymm1, 448(%rcx)12630; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12631; AVX2-NEXT:    vmovaps %ymm1, 480(%rcx)12632; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12633; AVX2-NEXT:    vmovaps %ymm1, 416(%rcx)12634; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12635; AVX2-NEXT:    vmovaps %ymm1, 352(%rcx)12636; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12637; AVX2-NEXT:    vmovaps %ymm1, 288(%rcx)12638; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12639; AVX2-NEXT:    vmovaps %ymm1, 224(%rcx)12640; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12641; AVX2-NEXT:    vmovaps %ymm1, 160(%rcx)12642; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12643; AVX2-NEXT:    vmovaps %ymm1, 96(%rcx)12644; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12645; AVX2-NEXT:    vmovaps %ymm1, 32(%rcx)12646; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12647; AVX2-NEXT:    vmovaps %ymm1, 480(%r8)12648; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12649; AVX2-NEXT:    vmovaps %ymm1, 448(%r8)12650; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12651; AVX2-NEXT:    vmovaps %ymm1, 416(%r8)12652; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12653; AVX2-NEXT:    vmovaps %ymm1, 384(%r8)12654; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12655; AVX2-NEXT:    vmovaps %ymm1, 352(%r8)12656; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12657; AVX2-NEXT:    vmovaps %ymm1, 320(%r8)12658; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12659; AVX2-NEXT:    vmovaps %ymm1, 288(%r8)12660; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12661; AVX2-NEXT:    vmovaps %ymm1, 256(%r8)12662; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12663; AVX2-NEXT:    vmovaps %ymm1, 224(%r8)12664; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12665; AVX2-NEXT:    vmovaps %ymm1, 192(%r8)12666; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12667; AVX2-NEXT:    vmovaps %ymm1, 160(%r8)12668; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12669; AVX2-NEXT:    vmovaps %ymm1, 128(%r8)12670; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12671; AVX2-NEXT:    vmovaps %ymm1, 96(%r8)12672; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12673; AVX2-NEXT:    vmovaps %ymm1, 64(%r8)12674; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12675; AVX2-NEXT:    vmovaps %ymm1, 32(%r8)12676; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12677; AVX2-NEXT:    vmovaps %ymm1, (%r8)12678; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12679; AVX2-NEXT:    vmovaps %ymm1, 480(%r9)12680; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12681; AVX2-NEXT:    vmovaps %ymm1, 448(%r9)12682; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12683; AVX2-NEXT:    vmovaps %ymm1, 416(%r9)12684; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12685; AVX2-NEXT:    vmovaps %ymm1, 384(%r9)12686; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12687; AVX2-NEXT:    vmovaps %ymm1, 352(%r9)12688; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12689; AVX2-NEXT:    vmovaps %ymm1, 320(%r9)12690; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12691; AVX2-NEXT:    vmovaps %ymm1, 288(%r9)12692; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12693; AVX2-NEXT:    vmovaps %ymm1, 256(%r9)12694; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12695; AVX2-NEXT:    vmovaps %ymm1, 224(%r9)12696; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12697; AVX2-NEXT:    vmovaps %ymm1, 192(%r9)12698; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12699; AVX2-NEXT:    vmovaps %ymm1, 160(%r9)12700; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12701; AVX2-NEXT:    vmovaps %ymm1, 128(%r9)12702; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12703; AVX2-NEXT:    vmovaps %ymm1, 96(%r9)12704; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12705; AVX2-NEXT:    vmovaps %ymm1, 64(%r9)12706; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12707; AVX2-NEXT:    vmovaps %ymm1, 32(%r9)12708; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12709; AVX2-NEXT:    vmovaps %ymm1, (%r9)12710; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax12711; AVX2-NEXT:    vmovaps %ymm0, 480(%rax)12712; AVX2-NEXT:    vmovaps %ymm3, 448(%rax)12713; AVX2-NEXT:    vmovaps %ymm5, 416(%rax)12714; AVX2-NEXT:    vmovaps %ymm7, 384(%rax)12715; AVX2-NEXT:    vmovaps %ymm9, 352(%rax)12716; AVX2-NEXT:    vmovaps %ymm11, 320(%rax)12717; AVX2-NEXT:    vmovaps %ymm13, 288(%rax)12718; AVX2-NEXT:    vmovaps %ymm15, 256(%rax)12719; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12720; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)12721; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12722; AVX2-NEXT:    vmovaps %ymm0, 192(%rax)12723; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12724; AVX2-NEXT:    vmovaps %ymm0, 160(%rax)12725; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12726; AVX2-NEXT:    vmovaps %ymm0, 128(%rax)12727; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12728; AVX2-NEXT:    vmovaps %ymm0, 96(%rax)12729; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12730; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)12731; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12732; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)12733; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12734; AVX2-NEXT:    vmovaps %ymm0, (%rax)12735; AVX2-NEXT:    addq $3432, %rsp # imm = 0xD6812736; AVX2-NEXT:    vzeroupper12737; AVX2-NEXT:    retq12738;12739; AVX2-FP-LABEL: load_i64_stride6_vf64:12740; AVX2-FP:       # %bb.0:12741; AVX2-FP-NEXT:    subq $3432, %rsp # imm = 0xD6812742; AVX2-FP-NEXT:    vmovaps 1088(%rdi), %ymm212743; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12744; AVX2-FP-NEXT:    vmovaps 1056(%rdi), %ymm412745; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12746; AVX2-FP-NEXT:    vmovaps 704(%rdi), %ymm312747; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12748; AVX2-FP-NEXT:    vmovaps 672(%rdi), %ymm512749; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12750; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm112751; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12752; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm612753; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12754; AVX2-FP-NEXT:    vmovaps 240(%rdi), %xmm712755; AVX2-FP-NEXT:    vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12756; AVX2-FP-NEXT:    vmovaps 192(%rdi), %xmm012757; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12758; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm7[0]12759; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]12760; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]12761; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12762; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12763; AVX2-FP-NEXT:    vmovaps 624(%rdi), %xmm112764; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12765; AVX2-FP-NEXT:    vmovaps 576(%rdi), %xmm012766; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12767; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]12768; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]12769; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]12770; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12771; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12772; AVX2-FP-NEXT:    vmovaps 1008(%rdi), %xmm112773; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12774; AVX2-FP-NEXT:    vmovaps 960(%rdi), %xmm012775; AVX2-FP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12776; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]12777; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]12778; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]12779; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12780; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12781; AVX2-FP-NEXT:    vmovaps 1472(%rdi), %ymm012782; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12783; AVX2-FP-NEXT:    vmovaps 1440(%rdi), %ymm112784; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12785; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12786; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12787; AVX2-FP-NEXT:    vmovaps 1392(%rdi), %xmm212788; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12789; AVX2-FP-NEXT:    vmovaps 1344(%rdi), %xmm112790; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12791; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12792; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12793; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12794; AVX2-FP-NEXT:    vmovaps 1856(%rdi), %ymm012795; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12796; AVX2-FP-NEXT:    vmovaps 1824(%rdi), %ymm112797; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12798; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12799; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12800; AVX2-FP-NEXT:    vmovaps 1776(%rdi), %xmm212801; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12802; AVX2-FP-NEXT:    vmovaps 1728(%rdi), %xmm112803; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12804; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12805; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12806; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12807; AVX2-FP-NEXT:    vmovaps 2240(%rdi), %ymm012808; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12809; AVX2-FP-NEXT:    vmovaps 2208(%rdi), %ymm112810; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12811; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12812; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12813; AVX2-FP-NEXT:    vmovaps 2160(%rdi), %xmm212814; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12815; AVX2-FP-NEXT:    vmovaps 2112(%rdi), %xmm112816; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12817; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12818; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12819; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12820; AVX2-FP-NEXT:    vmovaps 2624(%rdi), %ymm012821; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12822; AVX2-FP-NEXT:    vmovaps 2592(%rdi), %ymm112823; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12824; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12825; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12826; AVX2-FP-NEXT:    vmovaps 2544(%rdi), %xmm212827; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12828; AVX2-FP-NEXT:    vmovaps 2496(%rdi), %xmm112829; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12830; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12831; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12832; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12833; AVX2-FP-NEXT:    vmovaps 3008(%rdi), %ymm012834; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12835; AVX2-FP-NEXT:    vmovaps 2976(%rdi), %ymm112836; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12837; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12838; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12839; AVX2-FP-NEXT:    vmovaps 2928(%rdi), %xmm212840; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12841; AVX2-FP-NEXT:    vmovaps 2880(%rdi), %xmm112842; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12843; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]12844; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12845; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12846; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm012847; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12848; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm112849; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12850; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12851; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12852; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm212853; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12854; AVX2-FP-NEXT:    vmovaps 48(%rdi), %xmm112855; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12856; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]12857; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12858; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12859; AVX2-FP-NEXT:    vmovaps 512(%rdi), %ymm012860; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12861; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm112862; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12863; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12864; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12865; AVX2-FP-NEXT:    vmovaps 432(%rdi), %xmm112866; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12867; AVX2-FP-NEXT:    vmovaps 384(%rdi), %xmm1312868; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]12869; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12870; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12871; AVX2-FP-NEXT:    vmovaps 896(%rdi), %ymm012872; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12873; AVX2-FP-NEXT:    vmovaps 864(%rdi), %ymm112874; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12875; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12876; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12877; AVX2-FP-NEXT:    vmovaps 816(%rdi), %xmm112878; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12879; AVX2-FP-NEXT:    vmovaps 768(%rdi), %xmm1112880; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]12881; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12882; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12883; AVX2-FP-NEXT:    vmovaps 1280(%rdi), %ymm012884; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12885; AVX2-FP-NEXT:    vmovaps 1248(%rdi), %ymm112886; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12887; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12888; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12889; AVX2-FP-NEXT:    vmovaps 1200(%rdi), %xmm112890; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12891; AVX2-FP-NEXT:    vmovaps 1152(%rdi), %xmm912892; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]12893; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12894; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12895; AVX2-FP-NEXT:    vmovaps 1664(%rdi), %ymm012896; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12897; AVX2-FP-NEXT:    vmovaps 1632(%rdi), %ymm112898; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12899; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]12900; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12901; AVX2-FP-NEXT:    vmovaps 1584(%rdi), %xmm1412902; AVX2-FP-NEXT:    vmovaps 1536(%rdi), %xmm712903; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm14[0]12904; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12905; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12906; AVX2-FP-NEXT:    vmovaps 2048(%rdi), %ymm1212907; AVX2-FP-NEXT:    vmovaps 2016(%rdi), %ymm012908; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12909; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]12910; AVX2-FP-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12911; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12912; AVX2-FP-NEXT:    vmovaps 1968(%rdi), %xmm1012913; AVX2-FP-NEXT:    vmovaps 1920(%rdi), %xmm512914; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm10[0]12915; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12916; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12917; AVX2-FP-NEXT:    vmovaps 2432(%rdi), %ymm812918; AVX2-FP-NEXT:    vmovaps 2400(%rdi), %ymm012919; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12920; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]12921; AVX2-FP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12922; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12923; AVX2-FP-NEXT:    vmovaps 2352(%rdi), %xmm612924; AVX2-FP-NEXT:    vmovaps 2304(%rdi), %xmm312925; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]12926; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12927; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12928; AVX2-FP-NEXT:    vmovaps 2816(%rdi), %ymm412929; AVX2-FP-NEXT:    vmovaps 2784(%rdi), %ymm012930; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12931; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]12932; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12933; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]12934; AVX2-FP-NEXT:    vmovaps 2736(%rdi), %xmm212935; AVX2-FP-NEXT:    vmovaps 2688(%rdi), %xmm112936; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]12937; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12938; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12939; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12940; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12941; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12942; AVX2-FP-NEXT:    vbroadcastsd 296(%rdi), %ymm1512943; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12944; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12945; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12946; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12947; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12948; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12949; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12950; AVX2-FP-NEXT:    vbroadcastsd 680(%rdi), %ymm1512951; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12952; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12953; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12954; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12955; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12956; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12957; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12958; AVX2-FP-NEXT:    vbroadcastsd 1064(%rdi), %ymm1512959; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12960; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12961; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12962; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12963; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12964; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12965; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12966; AVX2-FP-NEXT:    vbroadcastsd 1448(%rdi), %ymm1512967; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12968; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12969; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12970; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12971; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12972; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12973; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12974; AVX2-FP-NEXT:    vbroadcastsd 1832(%rdi), %ymm1512975; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12976; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12977; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12978; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12979; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12980; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12981; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12982; AVX2-FP-NEXT:    vbroadcastsd 2216(%rdi), %ymm1512983; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12984; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12985; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12986; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12987; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12988; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12989; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12990; AVX2-FP-NEXT:    vbroadcastsd 2600(%rdi), %ymm1512991; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload12992; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]12993; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]12994; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12995; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12996; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12997; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]12998; AVX2-FP-NEXT:    vbroadcastsd 2984(%rdi), %ymm1512999; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload13000; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]13001; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]13002; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13003; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13004; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13005; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13006; AVX2-FP-NEXT:    vbroadcastsd 104(%rdi), %ymm1513007; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload13008; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]13009; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]13010; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13011; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm0 # 16-byte Folded Reload13012; AVX2-FP-NEXT:    # xmm0 = xmm13[1],mem[1]13013; AVX2-FP-NEXT:    vbroadcastsd 488(%rdi), %ymm1313014; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload13015; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]13016; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]13017; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13018; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm0 # 16-byte Folded Reload13019; AVX2-FP-NEXT:    # xmm0 = xmm11[1],mem[1]13020; AVX2-FP-NEXT:    vbroadcastsd 872(%rdi), %ymm1113021; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload13022; AVX2-FP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]13023; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]13024; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13025; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm0 # 16-byte Folded Reload13026; AVX2-FP-NEXT:    # xmm0 = xmm9[1],mem[1]13027; AVX2-FP-NEXT:    vbroadcastsd 1256(%rdi), %ymm913028; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload13029; AVX2-FP-NEXT:    # ymm9 = ymm9[1],mem[1],ymm9[3],mem[3]13030; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]13031; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13032; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm14[1]13033; AVX2-FP-NEXT:    vbroadcastsd 1640(%rdi), %ymm713034; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload13035; AVX2-FP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]13036; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]13037; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13038; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm10[1]13039; AVX2-FP-NEXT:    vbroadcastsd 2024(%rdi), %ymm513040; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm12[1],ymm5[3],ymm12[3]13041; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]13042; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13043; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm6[1]13044; AVX2-FP-NEXT:    vbroadcastsd 2408(%rdi), %ymm313045; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm8[1],ymm3[3],ymm8[3]13046; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]13047; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13048; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]13049; AVX2-FP-NEXT:    vbroadcastsd 2792(%rdi), %ymm113050; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]13051; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13052; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13053; AVX2-FP-NEXT:    vbroadcastsd 352(%rdi), %ymm013054; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13055; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13056; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm213057; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13058; AVX2-FP-NEXT:    vmovaps 208(%rdi), %xmm113059; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13060; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13061; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13062; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13063; AVX2-FP-NEXT:    vbroadcastsd 736(%rdi), %ymm013064; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13065; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13066; AVX2-FP-NEXT:    vmovaps 640(%rdi), %xmm213067; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13068; AVX2-FP-NEXT:    vmovaps 592(%rdi), %xmm113069; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13070; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13071; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13072; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13073; AVX2-FP-NEXT:    vbroadcastsd 1120(%rdi), %ymm013074; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13075; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13076; AVX2-FP-NEXT:    vmovaps 1024(%rdi), %xmm213077; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13078; AVX2-FP-NEXT:    vmovaps 976(%rdi), %xmm113079; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13080; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13081; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13082; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13083; AVX2-FP-NEXT:    vbroadcastsd 1504(%rdi), %ymm013084; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13085; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13086; AVX2-FP-NEXT:    vmovaps 1408(%rdi), %xmm213087; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13088; AVX2-FP-NEXT:    vmovaps 1360(%rdi), %xmm113089; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13090; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13091; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13092; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13093; AVX2-FP-NEXT:    vbroadcastsd 1888(%rdi), %ymm013094; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13095; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13096; AVX2-FP-NEXT:    vmovaps 1792(%rdi), %xmm213097; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13098; AVX2-FP-NEXT:    vmovaps 1744(%rdi), %xmm113099; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13100; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13101; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13102; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13103; AVX2-FP-NEXT:    vbroadcastsd 2272(%rdi), %ymm013104; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13105; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13106; AVX2-FP-NEXT:    vmovaps 2176(%rdi), %xmm213107; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13108; AVX2-FP-NEXT:    vmovaps 2128(%rdi), %xmm113109; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13110; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13111; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13112; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13113; AVX2-FP-NEXT:    vbroadcastsd 2656(%rdi), %ymm013114; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13115; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13116; AVX2-FP-NEXT:    vmovaps 2560(%rdi), %xmm213117; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13118; AVX2-FP-NEXT:    vmovaps 2512(%rdi), %xmm113119; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13120; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13121; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13122; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13123; AVX2-FP-NEXT:    vbroadcastsd 3040(%rdi), %ymm013124; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13125; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13126; AVX2-FP-NEXT:    vmovaps 2944(%rdi), %xmm213127; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13128; AVX2-FP-NEXT:    vmovaps 2896(%rdi), %xmm113129; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13130; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13131; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13132; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13133; AVX2-FP-NEXT:    vbroadcastsd 2848(%rdi), %ymm013134; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13135; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13136; AVX2-FP-NEXT:    vmovaps 2752(%rdi), %xmm213137; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13138; AVX2-FP-NEXT:    vmovaps 2704(%rdi), %xmm113139; AVX2-FP-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill13140; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13141; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13142; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13143; AVX2-FP-NEXT:    vbroadcastsd 2464(%rdi), %ymm013144; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13145; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13146; AVX2-FP-NEXT:    vmovaps 2368(%rdi), %xmm213147; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13148; AVX2-FP-NEXT:    vmovaps 2320(%rdi), %xmm113149; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13150; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13151; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13152; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13153; AVX2-FP-NEXT:    vbroadcastsd 2080(%rdi), %ymm013154; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13155; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13156; AVX2-FP-NEXT:    vmovaps 1984(%rdi), %xmm213157; AVX2-FP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13158; AVX2-FP-NEXT:    vmovaps 1936(%rdi), %xmm113159; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13160; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13161; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13162; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13163; AVX2-FP-NEXT:    vbroadcastsd 1696(%rdi), %ymm013164; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13165; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13166; AVX2-FP-NEXT:    vmovaps 1600(%rdi), %xmm1213167; AVX2-FP-NEXT:    vmovaps 1552(%rdi), %xmm1113168; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]13169; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13170; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13171; AVX2-FP-NEXT:    vbroadcastsd 1312(%rdi), %ymm013172; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload13173; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]13174; AVX2-FP-NEXT:    vmovaps 1216(%rdi), %xmm913175; AVX2-FP-NEXT:    vmovaps 1168(%rdi), %xmm813176; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]13177; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13178; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13179; AVX2-FP-NEXT:    vbroadcastsd 928(%rdi), %ymm013180; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload13181; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[2],ymm0[2]13182; AVX2-FP-NEXT:    vmovaps 832(%rdi), %xmm613183; AVX2-FP-NEXT:    vmovaps 784(%rdi), %xmm513184; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm5[0],xmm6[0]13185; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]13186; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13187; AVX2-FP-NEXT:    vbroadcastsd 544(%rdi), %ymm013188; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload13189; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[2],ymm0[2]13190; AVX2-FP-NEXT:    vmovaps 448(%rdi), %xmm1313191; AVX2-FP-NEXT:    vmovaps 400(%rdi), %xmm313192; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm3[0],xmm13[0]13193; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13194; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13195; AVX2-FP-NEXT:    vbroadcastsd 160(%rdi), %ymm013196; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload13197; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]13198; AVX2-FP-NEXT:    vmovaps 16(%rdi), %xmm1513199; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm013200; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm15[0],xmm0[0]13201; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3],ymm1[4,5,6,7]13202; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13203; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm15[1],xmm0[1]13204; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm1413205; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm14[1],ymm2[3],ymm14[3]13206; AVX2-FP-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13207; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13208; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13209; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13210; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13211; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13212; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13213; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm1513214; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13215; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm15[1],ymm1[3],ymm15[3]13216; AVX2-FP-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13217; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13218; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13219; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13220; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm13[1]13221; AVX2-FP-NEXT:    vmovaps 544(%rdi), %ymm313222; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm3[1],ymm4[3],ymm3[3]13223; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13224; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13225; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13226; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13227; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13228; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13229; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13230; AVX2-FP-NEXT:    vmovaps 736(%rdi), %ymm413231; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13232; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]13233; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13234; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13235; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13236; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13237; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm6[1]13238; AVX2-FP-NEXT:    vmovaps 928(%rdi), %ymm513239; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm5[1],ymm7[3],ymm5[3]13240; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13241; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13242; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13243; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13244; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13245; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13246; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13247; AVX2-FP-NEXT:    vmovaps 1120(%rdi), %ymm613248; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13249; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm6[1],ymm1[3],ymm6[3]13250; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13251; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13252; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13253; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13254; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm9[1]13255; AVX2-FP-NEXT:    vmovaps 1312(%rdi), %ymm713256; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm10[1],ymm7[1],ymm10[3],ymm7[3]13257; AVX2-FP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13258; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13259; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13260; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13261; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13262; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13263; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13264; AVX2-FP-NEXT:    vmovaps 1504(%rdi), %ymm813265; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13266; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm8[1],ymm1[3],ymm8[3]13267; AVX2-FP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13268; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13269; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13270; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13271; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm11[1],xmm12[1]13272; AVX2-FP-NEXT:    vmovaps 1696(%rdi), %ymm913273; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13274; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]13275; AVX2-FP-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13276; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13277; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13278; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13279; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13280; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13281; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13282; AVX2-FP-NEXT:    vmovaps 1888(%rdi), %ymm1013283; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13284; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]13285; AVX2-FP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13286; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13287; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13288; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13289; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13290; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13291; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13292; AVX2-FP-NEXT:    vmovaps 2080(%rdi), %ymm1113293; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13294; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]13295; AVX2-FP-NEXT:    vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13296; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13297; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13298; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13299; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13300; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13301; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13302; AVX2-FP-NEXT:    vmovaps 2272(%rdi), %ymm213303; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13304; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13305; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]13306; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13307; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13308; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13309; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13310; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13311; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13312; AVX2-FP-NEXT:    vmovaps 2464(%rdi), %ymm213313; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13314; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13315; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]13316; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13317; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13318; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13319; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13320; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13321; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13322; AVX2-FP-NEXT:    vmovaps 2656(%rdi), %ymm213323; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13324; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13325; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]13326; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13327; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13328; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13329; AVX2-FP-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload13330; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13331; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13332; AVX2-FP-NEXT:    vmovaps 2848(%rdi), %ymm213333; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13334; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13335; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]13336; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13337; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13338; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13339; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13340; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13341; AVX2-FP-NEXT:    # xmm0 = xmm0[1],mem[1]13342; AVX2-FP-NEXT:    vmovaps 3040(%rdi), %ymm213343; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13344; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]13345; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13346; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]13347; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13348; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13349; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13350; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]13351; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13352; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm1213353; AVX2-FP-NEXT:    vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13354; AVX2-FP-NEXT:    vmovaps 80(%rdi), %xmm113355; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13356; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm1[0]13357; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13358; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13359; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13360; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]13361; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13362; AVX2-FP-NEXT:    vmovaps 272(%rdi), %xmm1213363; AVX2-FP-NEXT:    vmovaps %xmm12, (%rsp) # 16-byte Spill13364; AVX2-FP-NEXT:    vmovaps 224(%rdi), %xmm113365; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13366; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm12[0]13367; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13368; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13369; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13370; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]13371; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13372; AVX2-FP-NEXT:    vmovaps 464(%rdi), %xmm313373; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13374; AVX2-FP-NEXT:    vmovaps 416(%rdi), %xmm113375; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13376; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13377; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13378; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13379; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13380; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]13381; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13382; AVX2-FP-NEXT:    vmovaps 656(%rdi), %xmm313383; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13384; AVX2-FP-NEXT:    vmovaps 608(%rdi), %xmm113385; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13386; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13387; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13388; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13389; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13390; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]13391; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13392; AVX2-FP-NEXT:    vmovaps 848(%rdi), %xmm313393; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13394; AVX2-FP-NEXT:    vmovaps 800(%rdi), %xmm113395; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13396; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13397; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13398; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13399; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13400; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]13401; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13402; AVX2-FP-NEXT:    vmovaps 1040(%rdi), %xmm313403; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13404; AVX2-FP-NEXT:    vmovaps 992(%rdi), %xmm113405; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13406; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13407; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13408; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13409; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13410; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]13411; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13412; AVX2-FP-NEXT:    vmovaps 1232(%rdi), %xmm313413; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13414; AVX2-FP-NEXT:    vmovaps 1184(%rdi), %xmm113415; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13416; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13417; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13418; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13419; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13420; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]13421; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13422; AVX2-FP-NEXT:    vmovaps 1424(%rdi), %xmm313423; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13424; AVX2-FP-NEXT:    vmovaps 1376(%rdi), %xmm113425; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13426; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13427; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13428; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13429; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13430; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]13431; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13432; AVX2-FP-NEXT:    vmovaps 1616(%rdi), %xmm313433; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13434; AVX2-FP-NEXT:    vmovaps 1568(%rdi), %xmm113435; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13436; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]13437; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13438; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13439; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13440; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]13441; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13442; AVX2-FP-NEXT:    vmovaps 1808(%rdi), %xmm113443; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13444; AVX2-FP-NEXT:    vmovaps 1760(%rdi), %xmm1313445; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]13446; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13447; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13448; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13449; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm11[0],ymm0[2],ymm11[2]13450; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13451; AVX2-FP-NEXT:    vmovaps 2000(%rdi), %xmm113452; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13453; AVX2-FP-NEXT:    vmovaps 1952(%rdi), %xmm1113454; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]13455; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13456; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13457; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13458; AVX2-FP-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13459; AVX2-FP-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]13460; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13461; AVX2-FP-NEXT:    vmovaps 2192(%rdi), %xmm113462; AVX2-FP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13463; AVX2-FP-NEXT:    vmovaps 2144(%rdi), %xmm913464; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]13465; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13466; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13467; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13468; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload13469; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]13470; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13471; AVX2-FP-NEXT:    vmovaps 2384(%rdi), %xmm1213472; AVX2-FP-NEXT:    vmovaps 2336(%rdi), %xmm713473; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm12[0]13474; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13475; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13476; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13477; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload13478; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]13479; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13480; AVX2-FP-NEXT:    vmovaps 2576(%rdi), %xmm813481; AVX2-FP-NEXT:    vmovaps 2528(%rdi), %xmm513482; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]13483; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13484; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13485; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13486; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload13487; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]13488; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13489; AVX2-FP-NEXT:    vmovaps 2768(%rdi), %xmm413490; AVX2-FP-NEXT:    vmovaps 2720(%rdi), %xmm313491; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm4[0]13492; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13493; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13494; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13495; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]13496; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13497; AVX2-FP-NEXT:    vmovaps 2960(%rdi), %xmm213498; AVX2-FP-NEXT:    vmovaps 2912(%rdi), %xmm113499; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]13500; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13501; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13502; AVX2-FP-NEXT:    vbroadcastsd 136(%rdi), %ymm013503; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13504; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13505; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13506; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13507; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13508; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13509; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13510; AVX2-FP-NEXT:    vbroadcastsd 328(%rdi), %ymm013511; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13512; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13513; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13514; AVX2-FP-NEXT:    vunpckhpd (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload13515; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13516; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13517; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13518; AVX2-FP-NEXT:    vbroadcastsd 520(%rdi), %ymm013519; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13520; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13521; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13522; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13523; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13524; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13525; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13526; AVX2-FP-NEXT:    vbroadcastsd 712(%rdi), %ymm013527; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13528; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13529; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13530; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13531; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13532; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13533; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13534; AVX2-FP-NEXT:    vbroadcastsd 904(%rdi), %ymm013535; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13536; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13537; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13538; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13539; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13540; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13541; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13542; AVX2-FP-NEXT:    vbroadcastsd 1096(%rdi), %ymm013543; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13544; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13545; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13546; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13547; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13548; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13549; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13550; AVX2-FP-NEXT:    vbroadcastsd 1288(%rdi), %ymm013551; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13552; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13553; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13554; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13555; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13556; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13557; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13558; AVX2-FP-NEXT:    vbroadcastsd 1480(%rdi), %ymm013559; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13560; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13561; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13562; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13563; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13564; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13565; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13566; AVX2-FP-NEXT:    vbroadcastsd 1672(%rdi), %ymm013567; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13568; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13569; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13570; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload13571; AVX2-FP-NEXT:    # xmm15 = xmm15[1],mem[1]13572; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]13573; AVX2-FP-NEXT:    vbroadcastsd 1864(%rdi), %ymm013574; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13575; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13576; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload13577; AVX2-FP-NEXT:    # xmm13 = xmm13[1],mem[1]13578; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]13579; AVX2-FP-NEXT:    vbroadcastsd 2056(%rdi), %ymm013580; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13581; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13582; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload13583; AVX2-FP-NEXT:    # xmm11 = xmm11[1],mem[1]13584; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm0[4,5,6,7]13585; AVX2-FP-NEXT:    vbroadcastsd 2248(%rdi), %ymm013586; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13587; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13588; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload13589; AVX2-FP-NEXT:    # xmm9 = xmm9[1],mem[1]13590; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]13591; AVX2-FP-NEXT:    vbroadcastsd 2440(%rdi), %ymm013592; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm14[1],ymm0[3],ymm14[3]13593; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm12[1]13594; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]13595; AVX2-FP-NEXT:    vbroadcastsd 2632(%rdi), %ymm013596; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]13597; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]13598; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]13599; AVX2-FP-NEXT:    vbroadcastsd 2824(%rdi), %ymm013600; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm6[1],ymm0[3],ymm6[3]13601; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]13602; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]13603; AVX2-FP-NEXT:    vbroadcastsd 3016(%rdi), %ymm013604; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13605; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]13606; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]13607; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13608; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13609; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rsi)13610; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13611; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rsi)13612; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13613; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rsi)13614; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13615; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rsi)13616; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13617; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rsi)13618; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13619; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rsi)13620; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13621; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rsi)13622; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13623; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)13624; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13625; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rsi)13626; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13627; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rsi)13628; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13629; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rsi)13630; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13631; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rsi)13632; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13633; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rsi)13634; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13635; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rsi)13636; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13637; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rsi)13638; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13639; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rsi)13640; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13641; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rdx)13642; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13643; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rdx)13644; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13645; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rdx)13646; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13647; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rdx)13648; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13649; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rdx)13650; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13651; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rdx)13652; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13653; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rdx)13654; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13655; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)13656; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13657; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rdx)13658; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13659; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rdx)13660; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13661; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rdx)13662; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13663; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rdx)13664; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13665; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rdx)13666; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13667; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rdx)13668; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13669; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rdx)13670; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13671; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rdx)13672; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13673; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)13674; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13675; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rcx)13676; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13677; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rcx)13678; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13679; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rcx)13680; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13681; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rcx)13682; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13683; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rcx)13684; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13685; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rcx)13686; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13687; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rcx)13688; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13689; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rcx)13690; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13691; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rcx)13692; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13693; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rcx)13694; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13695; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rcx)13696; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13697; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rcx)13698; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13699; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rcx)13700; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13701; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rcx)13702; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13703; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rcx)13704; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13705; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%r8)13706; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13707; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%r8)13708; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13709; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%r8)13710; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13711; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%r8)13712; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13713; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%r8)13714; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13715; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%r8)13716; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13717; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%r8)13718; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13719; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%r8)13720; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13721; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r8)13722; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13723; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r8)13724; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13725; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r8)13726; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13727; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r8)13728; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13729; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r8)13730; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13731; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r8)13732; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13733; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r8)13734; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13735; AVX2-FP-NEXT:    vmovaps %ymm1, (%r8)13736; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13737; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%r9)13738; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13739; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%r9)13740; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13741; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%r9)13742; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13743; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%r9)13744; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13745; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%r9)13746; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13747; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%r9)13748; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13749; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%r9)13750; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13751; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%r9)13752; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13753; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r9)13754; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13755; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r9)13756; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13757; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r9)13758; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13759; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r9)13760; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13761; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r9)13762; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13763; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r9)13764; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13765; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r9)13766; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13767; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)13768; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax13769; AVX2-FP-NEXT:    vmovaps %ymm0, 480(%rax)13770; AVX2-FP-NEXT:    vmovaps %ymm3, 448(%rax)13771; AVX2-FP-NEXT:    vmovaps %ymm5, 416(%rax)13772; AVX2-FP-NEXT:    vmovaps %ymm7, 384(%rax)13773; AVX2-FP-NEXT:    vmovaps %ymm9, 352(%rax)13774; AVX2-FP-NEXT:    vmovaps %ymm11, 320(%rax)13775; AVX2-FP-NEXT:    vmovaps %ymm13, 288(%rax)13776; AVX2-FP-NEXT:    vmovaps %ymm15, 256(%rax)13777; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13778; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)13779; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13780; AVX2-FP-NEXT:    vmovaps %ymm0, 192(%rax)13781; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13782; AVX2-FP-NEXT:    vmovaps %ymm0, 160(%rax)13783; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13784; AVX2-FP-NEXT:    vmovaps %ymm0, 128(%rax)13785; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13786; AVX2-FP-NEXT:    vmovaps %ymm0, 96(%rax)13787; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13788; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)13789; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13790; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)13791; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13792; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)13793; AVX2-FP-NEXT:    addq $3432, %rsp # imm = 0xD6813794; AVX2-FP-NEXT:    vzeroupper13795; AVX2-FP-NEXT:    retq13796;13797; AVX2-FCP-LABEL: load_i64_stride6_vf64:13798; AVX2-FCP:       # %bb.0:13799; AVX2-FCP-NEXT:    subq $3432, %rsp # imm = 0xD6813800; AVX2-FCP-NEXT:    vmovaps 1088(%rdi), %ymm213801; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13802; AVX2-FCP-NEXT:    vmovaps 1056(%rdi), %ymm413803; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13804; AVX2-FCP-NEXT:    vmovaps 704(%rdi), %ymm313805; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13806; AVX2-FCP-NEXT:    vmovaps 672(%rdi), %ymm513807; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13808; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm113809; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13810; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm613811; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13812; AVX2-FCP-NEXT:    vmovaps 240(%rdi), %xmm713813; AVX2-FCP-NEXT:    vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13814; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %xmm013815; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13816; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm7[0]13817; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]13818; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]13819; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13820; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13821; AVX2-FCP-NEXT:    vmovaps 624(%rdi), %xmm113822; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13823; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %xmm013824; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13825; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]13826; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]13827; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]13828; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13829; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13830; AVX2-FCP-NEXT:    vmovaps 1008(%rdi), %xmm113831; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13832; AVX2-FCP-NEXT:    vmovaps 960(%rdi), %xmm013833; AVX2-FCP-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13834; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]13835; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]13836; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]13837; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13838; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13839; AVX2-FCP-NEXT:    vmovaps 1472(%rdi), %ymm013840; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13841; AVX2-FCP-NEXT:    vmovaps 1440(%rdi), %ymm113842; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13843; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13844; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13845; AVX2-FCP-NEXT:    vmovaps 1392(%rdi), %xmm213846; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13847; AVX2-FCP-NEXT:    vmovaps 1344(%rdi), %xmm113848; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13849; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13850; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13851; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13852; AVX2-FCP-NEXT:    vmovaps 1856(%rdi), %ymm013853; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13854; AVX2-FCP-NEXT:    vmovaps 1824(%rdi), %ymm113855; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13856; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13857; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13858; AVX2-FCP-NEXT:    vmovaps 1776(%rdi), %xmm213859; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13860; AVX2-FCP-NEXT:    vmovaps 1728(%rdi), %xmm113861; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13862; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13863; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13864; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13865; AVX2-FCP-NEXT:    vmovaps 2240(%rdi), %ymm013866; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13867; AVX2-FCP-NEXT:    vmovaps 2208(%rdi), %ymm113868; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13869; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13870; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13871; AVX2-FCP-NEXT:    vmovaps 2160(%rdi), %xmm213872; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13873; AVX2-FCP-NEXT:    vmovaps 2112(%rdi), %xmm113874; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13875; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13876; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13877; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13878; AVX2-FCP-NEXT:    vmovaps 2624(%rdi), %ymm013879; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13880; AVX2-FCP-NEXT:    vmovaps 2592(%rdi), %ymm113881; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13882; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13883; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13884; AVX2-FCP-NEXT:    vmovaps 2544(%rdi), %xmm213885; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13886; AVX2-FCP-NEXT:    vmovaps 2496(%rdi), %xmm113887; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13888; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13889; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13890; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13891; AVX2-FCP-NEXT:    vmovaps 3008(%rdi), %ymm013892; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13893; AVX2-FCP-NEXT:    vmovaps 2976(%rdi), %ymm113894; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13895; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13896; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13897; AVX2-FCP-NEXT:    vmovaps 2928(%rdi), %xmm213898; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13899; AVX2-FCP-NEXT:    vmovaps 2880(%rdi), %xmm113900; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13901; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]13902; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13903; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13904; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm013905; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13906; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm113907; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13908; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13909; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13910; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm213911; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13912; AVX2-FCP-NEXT:    vmovaps 48(%rdi), %xmm113913; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13914; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]13915; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13916; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13917; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %ymm013918; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13919; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm113920; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13921; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13922; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13923; AVX2-FCP-NEXT:    vmovaps 432(%rdi), %xmm113924; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13925; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %xmm1313926; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]13927; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13928; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13929; AVX2-FCP-NEXT:    vmovaps 896(%rdi), %ymm013930; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13931; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %ymm113932; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13933; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13934; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13935; AVX2-FCP-NEXT:    vmovaps 816(%rdi), %xmm113936; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13937; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %xmm1113938; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]13939; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13940; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13941; AVX2-FCP-NEXT:    vmovaps 1280(%rdi), %ymm013942; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13943; AVX2-FCP-NEXT:    vmovaps 1248(%rdi), %ymm113944; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13945; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13946; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13947; AVX2-FCP-NEXT:    vmovaps 1200(%rdi), %xmm113948; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13949; AVX2-FCP-NEXT:    vmovaps 1152(%rdi), %xmm913950; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]13951; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13952; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13953; AVX2-FCP-NEXT:    vmovaps 1664(%rdi), %ymm013954; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13955; AVX2-FCP-NEXT:    vmovaps 1632(%rdi), %ymm113956; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13957; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]13958; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13959; AVX2-FCP-NEXT:    vmovaps 1584(%rdi), %xmm1413960; AVX2-FCP-NEXT:    vmovaps 1536(%rdi), %xmm713961; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm14[0]13962; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13963; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13964; AVX2-FCP-NEXT:    vmovaps 2048(%rdi), %ymm1213965; AVX2-FCP-NEXT:    vmovaps 2016(%rdi), %ymm013966; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13967; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm12[0],ymm0[2],ymm12[2]13968; AVX2-FCP-NEXT:    vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13969; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13970; AVX2-FCP-NEXT:    vmovaps 1968(%rdi), %xmm1013971; AVX2-FCP-NEXT:    vmovaps 1920(%rdi), %xmm513972; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm10[0]13973; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13974; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13975; AVX2-FCP-NEXT:    vmovaps 2432(%rdi), %ymm813976; AVX2-FCP-NEXT:    vmovaps 2400(%rdi), %ymm013977; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13978; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]13979; AVX2-FCP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13980; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13981; AVX2-FCP-NEXT:    vmovaps 2352(%rdi), %xmm613982; AVX2-FCP-NEXT:    vmovaps 2304(%rdi), %xmm313983; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm6[0]13984; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13985; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13986; AVX2-FCP-NEXT:    vmovaps 2816(%rdi), %ymm413987; AVX2-FCP-NEXT:    vmovaps 2784(%rdi), %ymm013988; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13989; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]13990; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13991; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]13992; AVX2-FCP-NEXT:    vmovaps 2736(%rdi), %xmm213993; AVX2-FCP-NEXT:    vmovaps 2688(%rdi), %xmm113994; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]13995; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13996; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13997; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13998; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13999; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14000; AVX2-FCP-NEXT:    vbroadcastsd 296(%rdi), %ymm1514001; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14002; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14003; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14004; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14005; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14006; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14007; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14008; AVX2-FCP-NEXT:    vbroadcastsd 680(%rdi), %ymm1514009; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14010; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14011; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14012; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14013; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14014; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14015; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14016; AVX2-FCP-NEXT:    vbroadcastsd 1064(%rdi), %ymm1514017; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14018; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14019; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14020; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14021; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14022; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14023; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14024; AVX2-FCP-NEXT:    vbroadcastsd 1448(%rdi), %ymm1514025; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14026; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14027; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14028; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14029; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14030; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14031; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14032; AVX2-FCP-NEXT:    vbroadcastsd 1832(%rdi), %ymm1514033; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14034; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14035; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14036; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14037; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14038; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14039; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14040; AVX2-FCP-NEXT:    vbroadcastsd 2216(%rdi), %ymm1514041; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14042; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14043; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14044; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14045; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14046; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14047; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14048; AVX2-FCP-NEXT:    vbroadcastsd 2600(%rdi), %ymm1514049; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14050; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14051; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14052; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14053; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14054; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14055; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14056; AVX2-FCP-NEXT:    vbroadcastsd 2984(%rdi), %ymm1514057; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14058; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14059; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14060; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14061; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14062; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14063; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14064; AVX2-FCP-NEXT:    vbroadcastsd 104(%rdi), %ymm1514065; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload14066; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]14067; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]14068; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14069; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm0 # 16-byte Folded Reload14070; AVX2-FCP-NEXT:    # xmm0 = xmm13[1],mem[1]14071; AVX2-FCP-NEXT:    vbroadcastsd 488(%rdi), %ymm1314072; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload14073; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]14074; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]14075; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14076; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm0 # 16-byte Folded Reload14077; AVX2-FCP-NEXT:    # xmm0 = xmm11[1],mem[1]14078; AVX2-FCP-NEXT:    vbroadcastsd 872(%rdi), %ymm1114079; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload14080; AVX2-FCP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]14081; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]14082; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14083; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm0 # 16-byte Folded Reload14084; AVX2-FCP-NEXT:    # xmm0 = xmm9[1],mem[1]14085; AVX2-FCP-NEXT:    vbroadcastsd 1256(%rdi), %ymm914086; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload14087; AVX2-FCP-NEXT:    # ymm9 = ymm9[1],mem[1],ymm9[3],mem[3]14088; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]14089; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14090; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm14[1]14091; AVX2-FCP-NEXT:    vbroadcastsd 1640(%rdi), %ymm714092; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload14093; AVX2-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]14094; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]14095; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14096; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm10[1]14097; AVX2-FCP-NEXT:    vbroadcastsd 2024(%rdi), %ymm514098; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm12[1],ymm5[3],ymm12[3]14099; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]14100; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14101; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm6[1]14102; AVX2-FCP-NEXT:    vbroadcastsd 2408(%rdi), %ymm314103; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm8[1],ymm3[3],ymm8[3]14104; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]14105; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14106; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]14107; AVX2-FCP-NEXT:    vbroadcastsd 2792(%rdi), %ymm114108; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]14109; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14110; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14111; AVX2-FCP-NEXT:    vbroadcastsd 352(%rdi), %ymm014112; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14113; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14114; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm214115; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14116; AVX2-FCP-NEXT:    vmovaps 208(%rdi), %xmm114117; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14118; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14119; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14120; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14121; AVX2-FCP-NEXT:    vbroadcastsd 736(%rdi), %ymm014122; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14123; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14124; AVX2-FCP-NEXT:    vmovaps 640(%rdi), %xmm214125; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14126; AVX2-FCP-NEXT:    vmovaps 592(%rdi), %xmm114127; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14128; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14129; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14130; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14131; AVX2-FCP-NEXT:    vbroadcastsd 1120(%rdi), %ymm014132; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14133; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14134; AVX2-FCP-NEXT:    vmovaps 1024(%rdi), %xmm214135; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14136; AVX2-FCP-NEXT:    vmovaps 976(%rdi), %xmm114137; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14138; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14139; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14140; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14141; AVX2-FCP-NEXT:    vbroadcastsd 1504(%rdi), %ymm014142; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14143; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14144; AVX2-FCP-NEXT:    vmovaps 1408(%rdi), %xmm214145; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14146; AVX2-FCP-NEXT:    vmovaps 1360(%rdi), %xmm114147; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14148; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14149; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14150; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14151; AVX2-FCP-NEXT:    vbroadcastsd 1888(%rdi), %ymm014152; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14153; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14154; AVX2-FCP-NEXT:    vmovaps 1792(%rdi), %xmm214155; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14156; AVX2-FCP-NEXT:    vmovaps 1744(%rdi), %xmm114157; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14158; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14159; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14160; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14161; AVX2-FCP-NEXT:    vbroadcastsd 2272(%rdi), %ymm014162; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14163; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14164; AVX2-FCP-NEXT:    vmovaps 2176(%rdi), %xmm214165; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14166; AVX2-FCP-NEXT:    vmovaps 2128(%rdi), %xmm114167; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14168; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14169; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14170; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14171; AVX2-FCP-NEXT:    vbroadcastsd 2656(%rdi), %ymm014172; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14173; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14174; AVX2-FCP-NEXT:    vmovaps 2560(%rdi), %xmm214175; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14176; AVX2-FCP-NEXT:    vmovaps 2512(%rdi), %xmm114177; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14178; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14179; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14180; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14181; AVX2-FCP-NEXT:    vbroadcastsd 3040(%rdi), %ymm014182; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14183; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14184; AVX2-FCP-NEXT:    vmovaps 2944(%rdi), %xmm214185; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14186; AVX2-FCP-NEXT:    vmovaps 2896(%rdi), %xmm114187; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14188; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14189; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14190; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14191; AVX2-FCP-NEXT:    vbroadcastsd 2848(%rdi), %ymm014192; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14193; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14194; AVX2-FCP-NEXT:    vmovaps 2752(%rdi), %xmm214195; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14196; AVX2-FCP-NEXT:    vmovaps 2704(%rdi), %xmm114197; AVX2-FCP-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill14198; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14199; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14200; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14201; AVX2-FCP-NEXT:    vbroadcastsd 2464(%rdi), %ymm014202; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14203; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14204; AVX2-FCP-NEXT:    vmovaps 2368(%rdi), %xmm214205; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14206; AVX2-FCP-NEXT:    vmovaps 2320(%rdi), %xmm114207; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14208; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14209; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14210; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14211; AVX2-FCP-NEXT:    vbroadcastsd 2080(%rdi), %ymm014212; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14213; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14214; AVX2-FCP-NEXT:    vmovaps 1984(%rdi), %xmm214215; AVX2-FCP-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14216; AVX2-FCP-NEXT:    vmovaps 1936(%rdi), %xmm114217; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14218; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]14219; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14220; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14221; AVX2-FCP-NEXT:    vbroadcastsd 1696(%rdi), %ymm014222; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14223; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]14224; AVX2-FCP-NEXT:    vmovaps 1600(%rdi), %xmm1214225; AVX2-FCP-NEXT:    vmovaps 1552(%rdi), %xmm1114226; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0]14227; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14228; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14229; AVX2-FCP-NEXT:    vbroadcastsd 1312(%rdi), %ymm014230; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload14231; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[2],ymm0[2]14232; AVX2-FCP-NEXT:    vmovaps 1216(%rdi), %xmm914233; AVX2-FCP-NEXT:    vmovaps 1168(%rdi), %xmm814234; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]14235; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14236; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14237; AVX2-FCP-NEXT:    vbroadcastsd 928(%rdi), %ymm014238; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload14239; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[2],ymm0[2]14240; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %xmm614241; AVX2-FCP-NEXT:    vmovaps 784(%rdi), %xmm514242; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm5[0],xmm6[0]14243; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]14244; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14245; AVX2-FCP-NEXT:    vbroadcastsd 544(%rdi), %ymm014246; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload14247; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[2],ymm0[2]14248; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %xmm1314249; AVX2-FCP-NEXT:    vmovaps 400(%rdi), %xmm314250; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm3[0],xmm13[0]14251; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14252; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14253; AVX2-FCP-NEXT:    vbroadcastsd 160(%rdi), %ymm014254; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14255; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]14256; AVX2-FCP-NEXT:    vmovaps 16(%rdi), %xmm1514257; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm014258; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm15[0],xmm0[0]14259; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3],ymm1[4,5,6,7]14260; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14261; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm15[1],xmm0[1]14262; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm1414263; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm14[1],ymm2[3],ymm14[3]14264; AVX2-FCP-NEXT:    vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14265; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14266; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14267; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14268; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14269; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14270; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14271; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm1514272; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14273; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm15[1],ymm1[3],ymm15[3]14274; AVX2-FCP-NEXT:    vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14275; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14276; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14277; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14278; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm13[1]14279; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %ymm314280; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm3[1],ymm4[3],ymm3[3]14281; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14282; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14283; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14284; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14285; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14286; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14287; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14288; AVX2-FCP-NEXT:    vmovaps 736(%rdi), %ymm414289; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14290; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]14291; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14292; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14293; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14294; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14295; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm6[1]14296; AVX2-FCP-NEXT:    vmovaps 928(%rdi), %ymm514297; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm5[1],ymm7[3],ymm5[3]14298; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14299; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14300; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14301; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14302; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14303; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14304; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14305; AVX2-FCP-NEXT:    vmovaps 1120(%rdi), %ymm614306; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14307; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm6[1],ymm1[3],ymm6[3]14308; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14309; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14310; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14311; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14312; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm8[1],xmm9[1]14313; AVX2-FCP-NEXT:    vmovaps 1312(%rdi), %ymm714314; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm10[1],ymm7[1],ymm10[3],ymm7[3]14315; AVX2-FCP-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14316; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14317; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14318; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14319; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14320; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14321; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14322; AVX2-FCP-NEXT:    vmovaps 1504(%rdi), %ymm814323; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14324; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm8[1],ymm1[3],ymm8[3]14325; AVX2-FCP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14326; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14327; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14328; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14329; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm11[1],xmm12[1]14330; AVX2-FCP-NEXT:    vmovaps 1696(%rdi), %ymm914331; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14332; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm9[1],ymm1[3],ymm9[3]14333; AVX2-FCP-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14334; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14335; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14336; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14337; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14338; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14339; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14340; AVX2-FCP-NEXT:    vmovaps 1888(%rdi), %ymm1014341; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14342; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3]14343; AVX2-FCP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14344; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14345; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14346; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14347; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14348; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14349; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14350; AVX2-FCP-NEXT:    vmovaps 2080(%rdi), %ymm1114351; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14352; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]14353; AVX2-FCP-NEXT:    vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14354; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14355; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14356; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14357; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14358; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14359; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14360; AVX2-FCP-NEXT:    vmovaps 2272(%rdi), %ymm214361; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14362; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14363; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]14364; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14365; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14366; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14367; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14368; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14369; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14370; AVX2-FCP-NEXT:    vmovaps 2464(%rdi), %ymm214371; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14372; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14373; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]14374; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14375; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14376; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14377; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14378; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14379; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14380; AVX2-FCP-NEXT:    vmovaps 2656(%rdi), %ymm214381; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14382; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14383; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]14384; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14385; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14386; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14387; AVX2-FCP-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload14388; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14389; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14390; AVX2-FCP-NEXT:    vmovaps 2848(%rdi), %ymm214391; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14392; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14393; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]14394; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14395; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14396; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14397; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14398; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14399; AVX2-FCP-NEXT:    # xmm0 = xmm0[1],mem[1]14400; AVX2-FCP-NEXT:    vmovaps 3040(%rdi), %ymm214401; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14402; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]14403; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14404; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]14405; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14406; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14407; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14408; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]14409; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14410; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm1214411; AVX2-FCP-NEXT:    vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14412; AVX2-FCP-NEXT:    vmovaps 80(%rdi), %xmm114413; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14414; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm1[0]14415; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14416; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14417; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14418; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm15[0],ymm0[2],ymm15[2]14419; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14420; AVX2-FCP-NEXT:    vmovaps 272(%rdi), %xmm1214421; AVX2-FCP-NEXT:    vmovaps %xmm12, (%rsp) # 16-byte Spill14422; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %xmm114423; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14424; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm12[0]14425; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14426; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14427; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14428; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]14429; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14430; AVX2-FCP-NEXT:    vmovaps 464(%rdi), %xmm314431; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14432; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %xmm114433; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14434; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14435; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14436; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14437; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14438; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]14439; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14440; AVX2-FCP-NEXT:    vmovaps 656(%rdi), %xmm314441; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14442; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %xmm114443; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14444; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14445; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14446; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14447; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14448; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]14449; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14450; AVX2-FCP-NEXT:    vmovaps 848(%rdi), %xmm314451; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14452; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %xmm114453; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14454; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14455; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14456; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14457; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14458; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]14459; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14460; AVX2-FCP-NEXT:    vmovaps 1040(%rdi), %xmm314461; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14462; AVX2-FCP-NEXT:    vmovaps 992(%rdi), %xmm114463; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14464; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14465; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14466; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14467; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14468; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]14469; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14470; AVX2-FCP-NEXT:    vmovaps 1232(%rdi), %xmm314471; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14472; AVX2-FCP-NEXT:    vmovaps 1184(%rdi), %xmm114473; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14474; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14475; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14476; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14477; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14478; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]14479; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14480; AVX2-FCP-NEXT:    vmovaps 1424(%rdi), %xmm314481; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14482; AVX2-FCP-NEXT:    vmovaps 1376(%rdi), %xmm114483; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14484; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14485; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14486; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14487; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14488; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm9[0],ymm0[2],ymm9[2]14489; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14490; AVX2-FCP-NEXT:    vmovaps 1616(%rdi), %xmm314491; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14492; AVX2-FCP-NEXT:    vmovaps 1568(%rdi), %xmm114493; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14494; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]14495; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14496; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14497; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14498; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]14499; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14500; AVX2-FCP-NEXT:    vmovaps 1808(%rdi), %xmm114501; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14502; AVX2-FCP-NEXT:    vmovaps 1760(%rdi), %xmm1314503; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm1[0]14504; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14505; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14506; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14507; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm11[0],ymm0[2],ymm11[2]14508; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14509; AVX2-FCP-NEXT:    vmovaps 2000(%rdi), %xmm114510; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14511; AVX2-FCP-NEXT:    vmovaps 1952(%rdi), %xmm1114512; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm1[0]14513; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14514; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14515; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14516; AVX2-FCP-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14517; AVX2-FCP-NEXT:    # ymm0 = ymm0[0],mem[0],ymm0[2],mem[2]14518; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14519; AVX2-FCP-NEXT:    vmovaps 2192(%rdi), %xmm114520; AVX2-FCP-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14521; AVX2-FCP-NEXT:    vmovaps 2144(%rdi), %xmm914522; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm1[0]14523; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14524; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14525; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14526; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload14527; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm14[0],ymm0[2],ymm14[2]14528; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14529; AVX2-FCP-NEXT:    vmovaps 2384(%rdi), %xmm1214530; AVX2-FCP-NEXT:    vmovaps 2336(%rdi), %xmm714531; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm7[0],xmm12[0]14532; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14533; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14534; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14535; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload14536; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]14537; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14538; AVX2-FCP-NEXT:    vmovaps 2576(%rdi), %xmm814539; AVX2-FCP-NEXT:    vmovaps 2528(%rdi), %xmm514540; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm8[0]14541; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14542; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14543; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14544; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload14545; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[2],ymm6[2]14546; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14547; AVX2-FCP-NEXT:    vmovaps 2768(%rdi), %xmm414548; AVX2-FCP-NEXT:    vmovaps 2720(%rdi), %xmm314549; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm4[0]14550; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14551; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14552; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14553; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]14554; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]14555; AVX2-FCP-NEXT:    vmovaps 2960(%rdi), %xmm214556; AVX2-FCP-NEXT:    vmovaps 2912(%rdi), %xmm114557; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]14558; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14559; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14560; AVX2-FCP-NEXT:    vbroadcastsd 136(%rdi), %ymm014561; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14562; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14563; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14564; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14565; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14566; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14567; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14568; AVX2-FCP-NEXT:    vbroadcastsd 328(%rdi), %ymm014569; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14570; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14571; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14572; AVX2-FCP-NEXT:    vunpckhpd (%rsp), %xmm15, %xmm15 # 16-byte Folded Reload14573; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14574; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14575; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14576; AVX2-FCP-NEXT:    vbroadcastsd 520(%rdi), %ymm014577; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14578; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14579; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14580; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14581; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14582; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14583; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14584; AVX2-FCP-NEXT:    vbroadcastsd 712(%rdi), %ymm014585; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14586; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14587; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14588; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14589; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14590; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14591; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14592; AVX2-FCP-NEXT:    vbroadcastsd 904(%rdi), %ymm014593; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14594; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14595; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14596; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14597; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14598; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14599; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14600; AVX2-FCP-NEXT:    vbroadcastsd 1096(%rdi), %ymm014601; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14602; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14603; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14604; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14605; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14606; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14607; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14608; AVX2-FCP-NEXT:    vbroadcastsd 1288(%rdi), %ymm014609; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14610; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14611; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14612; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14613; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14614; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14615; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14616; AVX2-FCP-NEXT:    vbroadcastsd 1480(%rdi), %ymm014617; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14618; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14619; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14620; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14621; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14622; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14623; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14624; AVX2-FCP-NEXT:    vbroadcastsd 1672(%rdi), %ymm014625; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14626; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14627; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14628; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload14629; AVX2-FCP-NEXT:    # xmm15 = xmm15[1],mem[1]14630; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]14631; AVX2-FCP-NEXT:    vbroadcastsd 1864(%rdi), %ymm014632; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14633; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14634; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload14635; AVX2-FCP-NEXT:    # xmm13 = xmm13[1],mem[1]14636; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]14637; AVX2-FCP-NEXT:    vbroadcastsd 2056(%rdi), %ymm014638; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14639; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14640; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload14641; AVX2-FCP-NEXT:    # xmm11 = xmm11[1],mem[1]14642; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm0[4,5,6,7]14643; AVX2-FCP-NEXT:    vbroadcastsd 2248(%rdi), %ymm014644; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14645; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14646; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload14647; AVX2-FCP-NEXT:    # xmm9 = xmm9[1],mem[1]14648; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]14649; AVX2-FCP-NEXT:    vbroadcastsd 2440(%rdi), %ymm014650; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm14[1],ymm0[3],ymm14[3]14651; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm7[1],xmm12[1]14652; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4,5,6,7]14653; AVX2-FCP-NEXT:    vbroadcastsd 2632(%rdi), %ymm014654; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]14655; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]14656; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]14657; AVX2-FCP-NEXT:    vbroadcastsd 2824(%rdi), %ymm014658; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm6[1],ymm0[3],ymm6[3]14659; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]14660; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm0[4,5,6,7]14661; AVX2-FCP-NEXT:    vbroadcastsd 3016(%rdi), %ymm014662; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14663; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]14664; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]14665; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14666; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14667; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rsi)14668; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14669; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rsi)14670; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14671; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rsi)14672; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14673; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rsi)14674; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14675; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rsi)14676; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14677; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rsi)14678; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14679; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rsi)14680; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14681; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rsi)14682; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14683; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rsi)14684; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14685; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rsi)14686; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14687; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rsi)14688; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14689; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rsi)14690; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14691; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rsi)14692; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14693; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rsi)14694; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14695; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rsi)14696; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14697; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rsi)14698; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14699; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rdx)14700; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14701; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rdx)14702; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14703; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rdx)14704; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14705; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rdx)14706; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14707; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rdx)14708; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14709; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rdx)14710; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14711; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rdx)14712; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14713; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rdx)14714; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14715; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rdx)14716; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14717; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rdx)14718; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14719; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rdx)14720; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14721; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rdx)14722; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14723; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rdx)14724; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14725; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rdx)14726; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14727; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rdx)14728; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14729; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rdx)14730; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14731; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rcx)14732; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14733; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rcx)14734; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14735; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rcx)14736; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14737; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rcx)14738; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14739; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rcx)14740; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14741; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rcx)14742; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14743; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rcx)14744; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14745; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rcx)14746; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14747; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rcx)14748; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14749; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rcx)14750; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14751; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rcx)14752; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14753; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rcx)14754; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14755; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rcx)14756; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14757; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rcx)14758; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14759; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rcx)14760; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14761; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rcx)14762; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14763; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%r8)14764; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14765; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%r8)14766; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14767; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%r8)14768; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14769; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%r8)14770; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14771; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%r8)14772; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14773; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%r8)14774; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14775; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%r8)14776; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14777; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%r8)14778; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14779; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r8)14780; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14781; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r8)14782; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14783; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r8)14784; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14785; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r8)14786; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14787; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r8)14788; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14789; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r8)14790; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14791; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r8)14792; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14793; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r8)14794; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14795; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%r9)14796; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14797; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%r9)14798; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14799; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%r9)14800; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14801; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%r9)14802; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14803; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%r9)14804; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14805; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%r9)14806; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14807; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%r9)14808; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14809; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%r9)14810; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14811; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r9)14812; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14813; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r9)14814; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14815; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r9)14816; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14817; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r9)14818; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14819; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r9)14820; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14821; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r9)14822; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14823; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r9)14824; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14825; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r9)14826; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax14827; AVX2-FCP-NEXT:    vmovaps %ymm0, 480(%rax)14828; AVX2-FCP-NEXT:    vmovaps %ymm3, 448(%rax)14829; AVX2-FCP-NEXT:    vmovaps %ymm5, 416(%rax)14830; AVX2-FCP-NEXT:    vmovaps %ymm7, 384(%rax)14831; AVX2-FCP-NEXT:    vmovaps %ymm9, 352(%rax)14832; AVX2-FCP-NEXT:    vmovaps %ymm11, 320(%rax)14833; AVX2-FCP-NEXT:    vmovaps %ymm13, 288(%rax)14834; AVX2-FCP-NEXT:    vmovaps %ymm15, 256(%rax)14835; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14836; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)14837; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14838; AVX2-FCP-NEXT:    vmovaps %ymm0, 192(%rax)14839; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14840; AVX2-FCP-NEXT:    vmovaps %ymm0, 160(%rax)14841; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14842; AVX2-FCP-NEXT:    vmovaps %ymm0, 128(%rax)14843; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14844; AVX2-FCP-NEXT:    vmovaps %ymm0, 96(%rax)14845; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14846; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)14847; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14848; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)14849; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14850; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)14851; AVX2-FCP-NEXT:    addq $3432, %rsp # imm = 0xD6814852; AVX2-FCP-NEXT:    vzeroupper14853; AVX2-FCP-NEXT:    retq14854;14855; AVX512-LABEL: load_i64_stride6_vf64:14856; AVX512:       # %bb.0:14857; AVX512-NEXT:    subq $7240, %rsp # imm = 0x1C4814858; AVX512-NEXT:    vmovdqa64 2048(%rdi), %zmm314859; AVX512-NEXT:    vmovdqa64 1280(%rdi), %zmm414860; AVX512-NEXT:    vmovdqa64 1344(%rdi), %zmm014861; AVX512-NEXT:    vmovdqa64 896(%rdi), %zmm514862; AVX512-NEXT:    vmovdqa64 960(%rdi), %zmm2614863; AVX512-NEXT:    vmovdqa64 512(%rdi), %zmm214864; AVX512-NEXT:    vmovdqa64 576(%rdi), %zmm114865; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm614866; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm2914867; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,0,10,0,6,0,10]14868; AVX512-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]14869; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm814870; AVX512-NEXT:    vpermt2q %zmm6, %zmm7, %zmm814871; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14872; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm814873; AVX512-NEXT:    vpermt2q %zmm2, %zmm7, %zmm814874; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14875; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm814876; AVX512-NEXT:    vpermt2q %zmm5, %zmm7, %zmm814877; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14878; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm814879; AVX512-NEXT:    vpermt2q %zmm4, %zmm7, %zmm814880; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14881; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,0,11,1,7,0,11]14882; AVX512-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]14883; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm914884; AVX512-NEXT:    vpermt2q %zmm2, %zmm8, %zmm914885; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14886; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm914887; AVX512-NEXT:    vpermt2q %zmm6, %zmm8, %zmm914888; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14889; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm914890; AVX512-NEXT:    vpermt2q %zmm4, %zmm8, %zmm914891; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14892; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm914893; AVX512-NEXT:    vpermt2q %zmm5, %zmm8, %zmm914894; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14895; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm9 = [10,4,10,4,10,4,10,4]14896; AVX512-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]14897; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm1014898; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm1014899; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14900; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1014901; AVX512-NEXT:    vpermt2q %zmm29, %zmm9, %zmm1014902; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14903; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm1014904; AVX512-NEXT:    vpermt2q %zmm0, %zmm9, %zmm1014905; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14906; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm1014907; AVX512-NEXT:    vpermt2q %zmm26, %zmm9, %zmm1014908; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14909; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [11,5,11,5,11,5,11,5]14910; AVX512-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]14911; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm1114912; AVX512-NEXT:    vpermt2q %zmm1, %zmm10, %zmm1114913; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14914; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1114915; AVX512-NEXT:    vpermt2q %zmm29, %zmm10, %zmm1114916; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14917; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm1114918; AVX512-NEXT:    vpermt2q %zmm0, %zmm10, %zmm1114919; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14920; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm1114921; AVX512-NEXT:    vpermt2q %zmm26, %zmm10, %zmm1114922; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14923; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [12,0,0,6,12,0,0,6]14924; AVX512-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]14925; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm1314926; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm1214927; AVX512-NEXT:    vpermt2q %zmm1, %zmm11, %zmm1314928; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14929; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [13,0,1,7,13,0,1,7]14930; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]14931; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm1214932; AVX512-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14933; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm114934; AVX512-NEXT:    vpermt2q %zmm29, %zmm11, %zmm114935; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14936; AVX512-NEXT:    vpermt2q %zmm29, %zmm2, %zmm614937; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14938; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm114939; AVX512-NEXT:    vpermt2q %zmm0, %zmm11, %zmm114940; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14941; AVX512-NEXT:    vpermt2q %zmm0, %zmm2, %zmm414942; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14943; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm014944; AVX512-NEXT:    vpermt2q %zmm26, %zmm11, %zmm014945; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14946; AVX512-NEXT:    vpermt2q %zmm26, %zmm2, %zmm514947; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14948; AVX512-NEXT:    vmovdqa64 2112(%rdi), %zmm014949; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm114950; AVX512-NEXT:    vpermt2q %zmm3, %zmm7, %zmm114951; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14952; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm114953; AVX512-NEXT:    vpermt2q %zmm3, %zmm8, %zmm114954; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14955; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm114956; AVX512-NEXT:    vpermt2q %zmm0, %zmm9, %zmm114957; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14958; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm114959; AVX512-NEXT:    vpermt2q %zmm0, %zmm10, %zmm114960; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14961; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm114962; AVX512-NEXT:    vpermt2q %zmm0, %zmm11, %zmm114963; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14964; AVX512-NEXT:    vpermt2q %zmm0, %zmm2, %zmm314965; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14966; AVX512-NEXT:    vmovdqa64 1664(%rdi), %zmm114967; AVX512-NEXT:    vmovdqa64 1728(%rdi), %zmm014968; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm314969; AVX512-NEXT:    vpermt2q %zmm1, %zmm7, %zmm314970; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14971; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm314972; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm314973; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14974; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm314975; AVX512-NEXT:    vpermt2q %zmm0, %zmm9, %zmm314976; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14977; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm314978; AVX512-NEXT:    vpermt2q %zmm0, %zmm10, %zmm314979; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14980; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm314981; AVX512-NEXT:    vpermt2q %zmm0, %zmm11, %zmm314982; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14983; AVX512-NEXT:    vpermt2q %zmm0, %zmm2, %zmm114984; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14985; AVX512-NEXT:    vmovdqa64 2432(%rdi), %zmm314986; AVX512-NEXT:    vmovdqa64 2496(%rdi), %zmm014987; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm114988; AVX512-NEXT:    vpermt2q %zmm3, %zmm7, %zmm114989; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14990; AVX512-NEXT:    vmovdqa64 2816(%rdi), %zmm414991; AVX512-NEXT:    vmovdqa64 2880(%rdi), %zmm114992; AVX512-NEXT:    vpermi2q %zmm4, %zmm1, %zmm714993; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14994; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm514995; AVX512-NEXT:    vpermt2q %zmm3, %zmm8, %zmm514996; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14997; AVX512-NEXT:    vpermi2q %zmm4, %zmm1, %zmm814998; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14999; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm515000; AVX512-NEXT:    vpermt2q %zmm0, %zmm9, %zmm515001; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15002; AVX512-NEXT:    vpermi2q %zmm1, %zmm4, %zmm915003; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15004; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm515005; AVX512-NEXT:    vpermt2q %zmm0, %zmm10, %zmm515006; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15007; AVX512-NEXT:    vpermi2q %zmm1, %zmm4, %zmm1015008; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15009; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm515010; AVX512-NEXT:    vpermt2q %zmm0, %zmm11, %zmm515011; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15012; AVX512-NEXT:    vpermi2q %zmm1, %zmm4, %zmm1115013; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15014; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm415015; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15016; AVX512-NEXT:    vpermt2q %zmm0, %zmm2, %zmm315017; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15018; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm115019; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm215020; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [0,6,12,0]15021; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm015022; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm015023; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15024; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [1,7,13,0]15025; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm015026; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm015027; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15028; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [10,0,6,0]15029; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm015030; AVX512-NEXT:    vpermt2q %zmm2, %zmm10, %zmm015031; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15032; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [11,1,7,0]15033; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm015034; AVX512-NEXT:    vpermt2q %zmm2, %zmm11, %zmm015035; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15036; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm4 = [4,10,4,10,4,10,4,10]15037; AVX512-NEXT:    # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15038; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm015039; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm015040; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15041; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]15042; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15043; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215044; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15045; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm215046; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm115047; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315048; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315049; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15050; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315051; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315052; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15053; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315054; AVX512-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315055; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15056; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315057; AVX512-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315058; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15059; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315060; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315061; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15062; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215063; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15064; AVX512-NEXT:    vmovdqa64 1216(%rdi), %zmm115065; AVX512-NEXT:    vmovdqa64 1152(%rdi), %zmm215066; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315067; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315068; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15069; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315070; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315071; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15072; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315073; AVX512-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315074; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15075; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315076; AVX512-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315077; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15078; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315079; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315080; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15081; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215082; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15083; AVX512-NEXT:    vmovdqa64 832(%rdi), %zmm115084; AVX512-NEXT:    vmovdqa64 768(%rdi), %zmm215085; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315086; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315087; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15088; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315089; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315090; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15091; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315092; AVX512-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315093; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15094; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315095; AVX512-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315096; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15097; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315098; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315099; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15100; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215101; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15102; AVX512-NEXT:    vmovdqa64 1984(%rdi), %zmm115103; AVX512-NEXT:    vmovdqa64 1920(%rdi), %zmm215104; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315105; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315106; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15107; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315108; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315109; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15110; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315111; AVX512-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315112; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15113; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315114; AVX512-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315115; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15116; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm315117; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315118; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15119; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215120; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15121; AVX512-NEXT:    vmovdqa64 1600(%rdi), %zmm115122; AVX512-NEXT:    vmovdqa64 1536(%rdi), %zmm3015123; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm215124; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm215125; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15126; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm215127; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm215128; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15129; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm215130; AVX512-NEXT:    vpermt2q %zmm30, %zmm10, %zmm215131; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15132; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm215133; AVX512-NEXT:    vpermt2q %zmm30, %zmm11, %zmm215134; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15135; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm215136; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm215137; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15138; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm3015139; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm2615140; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm1815141; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm115142; AVX512-NEXT:    vpermt2q %zmm26, %zmm4, %zmm115143; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15144; AVX512-NEXT:    vmovdqa64 704(%rdi), %zmm2115145; AVX512-NEXT:    vmovdqa64 640(%rdi), %zmm1315146; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm115147; AVX512-NEXT:    vpermt2q %zmm21, %zmm4, %zmm115148; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15149; AVX512-NEXT:    vmovdqa64 1088(%rdi), %zmm2415150; AVX512-NEXT:    vmovdqa64 1024(%rdi), %zmm1915151; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm115152; AVX512-NEXT:    vpermt2q %zmm24, %zmm4, %zmm115153; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15154; AVX512-NEXT:    vmovdqa64 1472(%rdi), %zmm2815155; AVX512-NEXT:    vmovdqa64 1408(%rdi), %zmm1615156; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm115157; AVX512-NEXT:    vpermt2q %zmm28, %zmm4, %zmm115158; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15159; AVX512-NEXT:    vmovdqa64 1856(%rdi), %zmm2515160; AVX512-NEXT:    vmovdqa64 1792(%rdi), %zmm715161; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm115162; AVX512-NEXT:    vpermt2q %zmm25, %zmm4, %zmm115163; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15164; AVX512-NEXT:    vmovdqa64 2240(%rdi), %zmm2315165; AVX512-NEXT:    vmovdqa64 2176(%rdi), %zmm1715166; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm115167; AVX512-NEXT:    vpermt2q %zmm23, %zmm4, %zmm115168; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15169; AVX512-NEXT:    vmovdqa64 2624(%rdi), %zmm2015170; AVX512-NEXT:    vmovdqa64 2560(%rdi), %zmm515171; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm115172; AVX512-NEXT:    vpermt2q %zmm20, %zmm4, %zmm115173; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15174; AVX512-NEXT:    vmovdqa64 2368(%rdi), %zmm115175; AVX512-NEXT:    vmovdqa64 2304(%rdi), %zmm2215176; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm2915177; AVX512-NEXT:    vpermt2q %zmm1, %zmm8, %zmm2915178; AVX512-NEXT:    vmovdqa64 3008(%rdi), %zmm1415179; AVX512-NEXT:    vmovdqa64 2944(%rdi), %zmm3115180; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm215181; AVX512-NEXT:    vpermt2q %zmm14, %zmm4, %zmm215182; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15183; AVX512-NEXT:    vmovdqa64 2752(%rdi), %zmm215184; AVX512-NEXT:    vmovdqa64 2688(%rdi), %zmm1215185; AVX512-NEXT:    vpermi2q %zmm2, %zmm12, %zmm815186; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm315187; AVX512-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315188; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15189; AVX512-NEXT:    vpermi2q %zmm2, %zmm12, %zmm915190; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315191; AVX512-NEXT:    vpermt2q %zmm22, %zmm10, %zmm315192; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15193; AVX512-NEXT:    vpermi2q %zmm12, %zmm2, %zmm1015194; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm315195; AVX512-NEXT:    vpermt2q %zmm22, %zmm11, %zmm315196; AVX512-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill15197; AVX512-NEXT:    vpermi2q %zmm12, %zmm2, %zmm1115198; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm315199; AVX512-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315200; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15201; AVX512-NEXT:    vpermi2q %zmm2, %zmm12, %zmm415202; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1215203; AVX512-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2215204; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm115205; AVX512-NEXT:    vpermt2q %zmm20, %zmm0, %zmm115206; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15207; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm115208; AVX512-NEXT:    vpermt2q %zmm21, %zmm0, %zmm115209; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15210; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm115211; AVX512-NEXT:    vpermt2q %zmm26, %zmm0, %zmm115212; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15213; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm115214; AVX512-NEXT:    vpermt2q %zmm28, %zmm0, %zmm115215; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15216; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm115217; AVX512-NEXT:    vpermt2q %zmm24, %zmm0, %zmm115218; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15219; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm115220; AVX512-NEXT:    vpermt2q %zmm23, %zmm0, %zmm115221; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15222; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm115223; AVX512-NEXT:    vpermt2q %zmm25, %zmm0, %zmm115224; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15225; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm115226; AVX512-NEXT:    vpermt2q %zmm14, %zmm0, %zmm115227; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15228; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]15229; AVX512-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]15230; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm015231; AVX512-NEXT:    vpermt2q %zmm21, %zmm1, %zmm015232; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15233; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]15234; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]15235; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm015236; AVX512-NEXT:    vpermt2q %zmm21, %zmm2, %zmm015237; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15238; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]15239; AVX512-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]15240; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm015241; AVX512-NEXT:    vpermt2q %zmm13, %zmm3, %zmm015242; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15243; AVX512-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]15244; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]15245; AVX512-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2115246; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm615247; AVX512-NEXT:    vpermt2q %zmm26, %zmm1, %zmm615248; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15249; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm615250; AVX512-NEXT:    vpermt2q %zmm26, %zmm2, %zmm615251; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15252; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm615253; AVX512-NEXT:    vpermt2q %zmm18, %zmm3, %zmm615254; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15255; AVX512-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2615256; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm2715257; AVX512-NEXT:    vpermt2q %zmm28, %zmm1, %zmm2715258; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm615259; AVX512-NEXT:    vpermt2q %zmm28, %zmm2, %zmm615260; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15261; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm615262; AVX512-NEXT:    vpermt2q %zmm16, %zmm3, %zmm615263; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15264; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2815265; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm1815266; AVX512-NEXT:    vpermt2q %zmm24, %zmm1, %zmm1815267; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm615268; AVX512-NEXT:    vpermt2q %zmm24, %zmm2, %zmm615269; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15270; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm615271; AVX512-NEXT:    vpermt2q %zmm19, %zmm3, %zmm615272; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15273; AVX512-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2415274; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm1515275; AVX512-NEXT:    vpermt2q %zmm23, %zmm1, %zmm1515276; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm1915277; AVX512-NEXT:    vpermt2q %zmm23, %zmm2, %zmm1915278; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm615279; AVX512-NEXT:    vpermt2q %zmm17, %zmm3, %zmm615280; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15281; AVX512-NEXT:    vpermt2q %zmm17, %zmm0, %zmm2315282; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1315283; AVX512-NEXT:    vpermt2q %zmm25, %zmm1, %zmm1315284; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1715285; AVX512-NEXT:    vpermt2q %zmm25, %zmm2, %zmm1715286; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm615287; AVX512-NEXT:    vpermt2q %zmm7, %zmm3, %zmm615288; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15289; AVX512-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2515290; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm615291; AVX512-NEXT:    vpermt2q %zmm20, %zmm1, %zmm615292; AVX512-NEXT:    vpermi2q %zmm14, %zmm31, %zmm115293; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm715294; AVX512-NEXT:    vpermt2q %zmm20, %zmm2, %zmm715295; AVX512-NEXT:    vpermi2q %zmm14, %zmm31, %zmm215296; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm1615297; AVX512-NEXT:    vpermt2q %zmm5, %zmm3, %zmm1615298; AVX512-NEXT:    vpermi2q %zmm31, %zmm14, %zmm315299; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1415300; AVX512-NEXT:    vpermt2q %zmm5, %zmm0, %zmm2015301; AVX512-NEXT:    movb $56, %al15302; AVX512-NEXT:    kmovw %eax, %k115303; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15304; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15305; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15306; AVX512-NEXT:    movb $-64, %al15307; AVX512-NEXT:    kmovw %eax, %k215308; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15309; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15310; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15311; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15312; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15313; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15314; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15315; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15316; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15317; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15318; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15319; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15320; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15321; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15322; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15323; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15324; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15325; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15326; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15327; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15328; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15329; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15330; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15331; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15332; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15333; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15334; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15335; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15336; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15337; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15338; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15339; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15340; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15341; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15342; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}15343; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15344; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}15345; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15346; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15347; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15348; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k2}15349; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15350; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15351; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15352; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15353; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15354; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15355; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15356; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15357; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15358; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15359; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15360; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15361; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15362; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15363; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15364; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15365; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15366; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15367; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15368; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15369; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15370; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15371; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15372; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15373; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15374; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15375; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15376; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15377; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15378; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15379; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15380; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15381; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15382; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15383; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15384; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15385; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15386; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15387; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15388; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15389; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15390; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15391; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15392; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k1}15393; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15394; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}15395; AVX512-NEXT:    movb $24, %al15396; AVX512-NEXT:    kmovw %eax, %k215397; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15398; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15399; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15400; AVX512-NEXT:    movb $-32, %al15401; AVX512-NEXT:    kmovw %eax, %k115402; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm5 {%k1}15403; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15404; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15405; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15406; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15407; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15408; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15409; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15410; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15411; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15412; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15413; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15414; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}15415; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15416; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15417; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15418; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15419; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm5 {%k1}15420; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15421; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15422; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15423; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15424; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm5 {%k1}15425; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15426; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15427; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15428; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15429; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm5 {%k1}15430; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15431; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15432; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15433; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}15434; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm5 {%k1}15435; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15436; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15437; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}15438; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}15439; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15440; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload15441; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15442; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1 {%k1}15443; AVX512-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill15444; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15445; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15446; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15447; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15448; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15449; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15450; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15451; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15452; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15453; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15454; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15455; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15456; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15457; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15458; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15459; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15460; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15461; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15462; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15463; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15464; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15465; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15466; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15467; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15468; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15469; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15470; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}15471; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm31 {%k1}15472; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15473; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload15474; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k2}15475; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm27 {%k1}15476; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15477; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}15478; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm11 {%k1}15479; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15480; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload15481; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}15482; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15483; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload15484; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15485; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}15486; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15487; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 16-byte Folded Reload15488; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15489; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm5 {%k1}15490; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15491; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 16-byte Folded Reload15492; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15493; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm6 {%k1}15494; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15495; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm7 # 16-byte Folded Reload15496; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15497; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm7 {%k1}15498; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15499; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm13 # 16-byte Folded Reload15500; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15501; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm13 {%k1}15502; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15503; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 16-byte Folded Reload15504; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15505; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm15 {%k1}15506; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15507; AVX512-NEXT:    vinserti32x4 $0, %xmm4, %zmm2, %zmm415508; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}15509; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15510; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload15511; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm2 {%k1}15512; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload15513; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload15514; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm3 {%k1}15515; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload15516; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm16 # 16-byte Folded Reload15517; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm16 {%k1}15518; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload15519; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 16-byte Folded Reload15520; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm17 {%k1}15521; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload15522; AVX512-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm18 # 16-byte Folded Reload15523; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}15524; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload15525; AVX512-NEXT:    vinserti32x4 $0, %xmm30, %zmm19, %zmm1915526; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm19 {%k1}15527; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload15528; AVX512-NEXT:    vinserti32x4 $0, %xmm12, %zmm21, %zmm1215529; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm12 {%k1}15530; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload15531; AVX512-NEXT:    vinserti32x4 $0, %xmm22, %zmm14, %zmm1415532; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm14 {%k1}15533; AVX512-NEXT:    vmovdqa64 %zmm8, 448(%rsi)15534; AVX512-NEXT:    vmovdqa64 %zmm29, 384(%rsi)15535; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15536; AVX512-NEXT:    vmovaps %zmm8, 320(%rsi)15537; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15538; AVX512-NEXT:    vmovaps %zmm8, 256(%rsi)15539; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15540; AVX512-NEXT:    vmovaps %zmm8, 192(%rsi)15541; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15542; AVX512-NEXT:    vmovaps %zmm8, 128(%rsi)15543; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15544; AVX512-NEXT:    vmovaps %zmm8, 64(%rsi)15545; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15546; AVX512-NEXT:    vmovaps %zmm8, (%rsi)15547; AVX512-NEXT:    vmovdqa64 %zmm9, 448(%rdx)15548; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15549; AVX512-NEXT:    vmovaps %zmm8, 256(%rdx)15550; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15551; AVX512-NEXT:    vmovaps %zmm8, 320(%rdx)15552; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15553; AVX512-NEXT:    vmovaps %zmm8, 128(%rdx)15554; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15555; AVX512-NEXT:    vmovaps %zmm8, 192(%rdx)15556; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15557; AVX512-NEXT:    vmovaps %zmm8, (%rdx)15558; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15559; AVX512-NEXT:    vmovaps %zmm8, 64(%rdx)15560; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15561; AVX512-NEXT:    vmovaps %zmm8, 384(%rdx)15562; AVX512-NEXT:    vmovdqa64 %zmm10, 448(%rcx)15563; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15564; AVX512-NEXT:    vmovaps %zmm8, 256(%rcx)15565; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15566; AVX512-NEXT:    vmovaps %zmm8, 320(%rcx)15567; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15568; AVX512-NEXT:    vmovaps %zmm8, 128(%rcx)15569; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15570; AVX512-NEXT:    vmovaps %zmm8, 192(%rcx)15571; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15572; AVX512-NEXT:    vmovaps %zmm8, (%rcx)15573; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15574; AVX512-NEXT:    vmovaps %zmm8, 64(%rcx)15575; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15576; AVX512-NEXT:    vmovaps %zmm8, 384(%rcx)15577; AVX512-NEXT:    vmovdqa64 %zmm11, 448(%r8)15578; AVX512-NEXT:    vmovdqa64 %zmm27, 256(%r8)15579; AVX512-NEXT:    vmovdqa64 %zmm31, 320(%r8)15580; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15581; AVX512-NEXT:    vmovaps %zmm8, 128(%r8)15582; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15583; AVX512-NEXT:    vmovaps %zmm8, 192(%r8)15584; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15585; AVX512-NEXT:    vmovaps %zmm8, (%r8)15586; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15587; AVX512-NEXT:    vmovaps %zmm8, 64(%r8)15588; AVX512-NEXT:    vmovups (%rsp), %zmm8 # 64-byte Reload15589; AVX512-NEXT:    vmovaps %zmm8, 384(%r8)15590; AVX512-NEXT:    vmovdqa64 %zmm4, 448(%r9)15591; AVX512-NEXT:    vmovdqa64 %zmm15, 256(%r9)15592; AVX512-NEXT:    vmovdqa64 %zmm13, 320(%r9)15593; AVX512-NEXT:    vmovdqa64 %zmm7, 128(%r9)15594; AVX512-NEXT:    vmovdqa64 %zmm6, 192(%r9)15595; AVX512-NEXT:    vmovdqa64 %zmm5, (%r9)15596; AVX512-NEXT:    vmovdqa64 %zmm1, 64(%r9)15597; AVX512-NEXT:    vmovdqa64 %zmm0, 384(%r9)15598; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax15599; AVX512-NEXT:    vmovdqa64 %zmm14, 384(%rax)15600; AVX512-NEXT:    vmovdqa64 %zmm12, 448(%rax)15601; AVX512-NEXT:    vmovdqa64 %zmm19, 256(%rax)15602; AVX512-NEXT:    vmovdqa64 %zmm18, 320(%rax)15603; AVX512-NEXT:    vmovdqa64 %zmm17, 128(%rax)15604; AVX512-NEXT:    vmovdqa64 %zmm16, 192(%rax)15605; AVX512-NEXT:    vmovdqa64 %zmm3, (%rax)15606; AVX512-NEXT:    vmovdqa64 %zmm2, 64(%rax)15607; AVX512-NEXT:    addq $7240, %rsp # imm = 0x1C4815608; AVX512-NEXT:    vzeroupper15609; AVX512-NEXT:    retq15610;15611; AVX512-FCP-LABEL: load_i64_stride6_vf64:15612; AVX512-FCP:       # %bb.0:15613; AVX512-FCP-NEXT:    subq $7240, %rsp # imm = 0x1C4815614; AVX512-FCP-NEXT:    vmovdqa64 2048(%rdi), %zmm315615; AVX512-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm415616; AVX512-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm015617; AVX512-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm515618; AVX512-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm2615619; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm215620; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm115621; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm615622; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2915623; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,0,10,0,6,0,10]15624; AVX512-FCP-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]15625; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm815626; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm7, %zmm815627; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15628; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm815629; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm7, %zmm815630; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15631; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm815632; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm7, %zmm815633; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15634; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm815635; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm7, %zmm815636; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15637; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,0,11,1,7,0,11]15638; AVX512-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]15639; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm915640; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm8, %zmm915641; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15642; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm915643; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm8, %zmm915644; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15645; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm915646; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm915647; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15648; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm915649; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm8, %zmm915650; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15651; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm9 = [10,4,10,4,10,4,10,4]15652; AVX512-FCP-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15653; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1015654; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm1015655; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15656; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1015657; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm9, %zmm1015658; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15659; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1015660; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm9, %zmm1015661; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15662; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1015663; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm9, %zmm1015664; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15665; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [11,5,11,5,11,5,11,5]15666; AVX512-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15667; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1115668; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm10, %zmm1115669; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15670; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115671; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm10, %zmm1115672; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15673; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1115674; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm10, %zmm1115675; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15676; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1115677; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm10, %zmm1115678; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15679; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [12,0,0,6,12,0,0,6]15680; AVX512-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]15681; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1315682; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1215683; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm11, %zmm1315684; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15685; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [13,0,1,7,13,0,1,7]15686; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]15687; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm1215688; AVX512-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15689; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm115690; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm11, %zmm115691; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15692; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm2, %zmm615693; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15694; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115695; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm11, %zmm115696; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15697; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm415698; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15699; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm015700; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm11, %zmm015701; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15702; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm2, %zmm515703; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15704; AVX512-FCP-NEXT:    vmovdqa64 2112(%rdi), %zmm015705; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm115706; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm7, %zmm115707; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15708; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm115709; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm8, %zmm115710; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15711; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm115712; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm9, %zmm115713; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15714; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm115715; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm10, %zmm115716; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15717; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm115718; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm11, %zmm115719; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15720; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm315721; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15722; AVX512-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm115723; AVX512-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm015724; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm315725; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm7, %zmm315726; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15727; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm315728; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315729; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15730; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315731; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm9, %zmm315732; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15733; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315734; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm10, %zmm315735; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15736; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315737; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm11, %zmm315738; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15739; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm115740; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15741; AVX512-FCP-NEXT:    vmovdqa64 2432(%rdi), %zmm315742; AVX512-FCP-NEXT:    vmovdqa64 2496(%rdi), %zmm015743; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm115744; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm7, %zmm115745; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15746; AVX512-FCP-NEXT:    vmovdqa64 2816(%rdi), %zmm415747; AVX512-FCP-NEXT:    vmovdqa64 2880(%rdi), %zmm115748; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm1, %zmm715749; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15750; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm515751; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm8, %zmm515752; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15753; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm1, %zmm815754; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15755; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm515756; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm9, %zmm515757; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15758; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm915759; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15760; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm515761; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm10, %zmm515762; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15763; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm1015764; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15765; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm515766; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm11, %zmm515767; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15768; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm1115769; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15770; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm415771; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15772; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm315773; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15774; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm115775; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm215776; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [0,6,12,0]15777; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015778; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm015779; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15780; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [1,7,13,0]15781; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015782; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm015783; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15784; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [10,0,6,0]15785; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm015786; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm10, %zmm015787; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15788; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [11,1,7,0]15789; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm015790; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm015791; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15792; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm4 = [4,10,4,10,4,10,4,10]15793; AVX512-FCP-NEXT:    # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15794; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015795; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm015796; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15797; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]15798; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15799; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215800; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15801; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm215802; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm115803; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315804; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315805; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15806; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315807; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315808; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15809; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315810; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315811; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15812; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315813; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315814; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15815; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315816; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315817; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15818; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215819; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15820; AVX512-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm115821; AVX512-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm215822; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315823; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315824; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15825; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315826; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315827; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15828; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315829; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315830; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15831; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315832; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315833; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15834; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315835; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315836; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15837; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215838; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15839; AVX512-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm115840; AVX512-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm215841; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315842; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315843; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15844; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315845; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315846; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15847; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315848; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315849; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15850; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315851; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315852; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15853; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315854; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315855; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15856; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215857; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15858; AVX512-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm115859; AVX512-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm215860; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315861; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm315862; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15863; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315864; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315865; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15866; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315867; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm10, %zmm315868; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15869; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315870; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm315871; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15872; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm315873; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315874; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15875; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm215876; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15877; AVX512-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm115878; AVX512-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm3015879; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm215880; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm215881; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15882; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm215883; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm215884; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15885; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm215886; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm10, %zmm215887; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15888; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm215889; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm11, %zmm215890; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15891; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm215892; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm215893; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15894; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm3015895; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm2615896; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm1815897; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm115898; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm4, %zmm115899; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15900; AVX512-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm2115901; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1315902; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm115903; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm4, %zmm115904; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15905; AVX512-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm2415906; AVX512-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm1915907; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm115908; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm4, %zmm115909; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15910; AVX512-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm2815911; AVX512-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm1615912; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm115913; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm4, %zmm115914; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15915; AVX512-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm2515916; AVX512-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm715917; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm115918; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm4, %zmm115919; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15920; AVX512-FCP-NEXT:    vmovdqa64 2240(%rdi), %zmm2315921; AVX512-FCP-NEXT:    vmovdqa64 2176(%rdi), %zmm1715922; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm115923; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm4, %zmm115924; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15925; AVX512-FCP-NEXT:    vmovdqa64 2624(%rdi), %zmm2015926; AVX512-FCP-NEXT:    vmovdqa64 2560(%rdi), %zmm515927; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115928; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm4, %zmm115929; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15930; AVX512-FCP-NEXT:    vmovdqa64 2368(%rdi), %zmm115931; AVX512-FCP-NEXT:    vmovdqa64 2304(%rdi), %zmm2215932; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm2915933; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm8, %zmm2915934; AVX512-FCP-NEXT:    vmovdqa64 3008(%rdi), %zmm1415935; AVX512-FCP-NEXT:    vmovdqa64 2944(%rdi), %zmm3115936; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm215937; AVX512-FCP-NEXT:    vpermt2q %zmm14, %zmm4, %zmm215938; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15939; AVX512-FCP-NEXT:    vmovdqa64 2752(%rdi), %zmm215940; AVX512-FCP-NEXT:    vmovdqa64 2688(%rdi), %zmm1215941; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm12, %zmm815942; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm315943; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm9, %zmm315944; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15945; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm12, %zmm915946; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315947; AVX512-FCP-NEXT:    vpermt2q %zmm22, %zmm10, %zmm315948; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15949; AVX512-FCP-NEXT:    vpermi2q %zmm12, %zmm2, %zmm1015950; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315951; AVX512-FCP-NEXT:    vpermt2q %zmm22, %zmm11, %zmm315952; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill15953; AVX512-FCP-NEXT:    vpermi2q %zmm12, %zmm2, %zmm1115954; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm315955; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm4, %zmm315956; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15957; AVX512-FCP-NEXT:    vpermi2q %zmm2, %zmm12, %zmm415958; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1215959; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2215960; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115961; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm115962; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15963; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm115964; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm115965; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15966; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm115967; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm115968; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15969; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm115970; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm0, %zmm115971; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15972; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm115973; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm115974; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15975; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm115976; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm115977; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15978; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm115979; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm115980; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15981; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm115982; AVX512-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm115983; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15984; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]15985; AVX512-FCP-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]15986; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm015987; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm1, %zmm015988; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15989; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]15990; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]15991; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm015992; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm2, %zmm015993; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15994; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]15995; AVX512-FCP-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]15996; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm015997; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm3, %zmm015998; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15999; AVX512-FCP-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]16000; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]16001; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2116002; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm616003; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm1, %zmm616004; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16005; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm616006; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm2, %zmm616007; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16008; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm616009; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm3, %zmm616010; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16011; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2616012; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm2716013; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm1, %zmm2716014; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm616015; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm2, %zmm616016; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16017; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm616018; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm3, %zmm616019; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16020; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2816021; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm1816022; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm1, %zmm1816023; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm616024; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm616025; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16026; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm616027; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm3, %zmm616028; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16029; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2416030; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm1516031; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm1, %zmm1516032; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm1916033; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm2, %zmm1916034; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm616035; AVX512-FCP-NEXT:    vpermt2q %zmm17, %zmm3, %zmm616036; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16037; AVX512-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm2316038; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1316039; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm1, %zmm1316040; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1716041; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm2, %zmm1716042; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm616043; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm3, %zmm616044; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16045; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2516046; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm616047; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm1, %zmm616048; AVX512-FCP-NEXT:    vpermi2q %zmm14, %zmm31, %zmm116049; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm716050; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm2, %zmm716051; AVX512-FCP-NEXT:    vpermi2q %zmm14, %zmm31, %zmm216052; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm1616053; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm3, %zmm1616054; AVX512-FCP-NEXT:    vpermi2q %zmm31, %zmm14, %zmm316055; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1416056; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm2016057; AVX512-FCP-NEXT:    movb $56, %al16058; AVX512-FCP-NEXT:    kmovw %eax, %k116059; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16060; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16061; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16062; AVX512-FCP-NEXT:    movb $-64, %al16063; AVX512-FCP-NEXT:    kmovw %eax, %k216064; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16065; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16066; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16067; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16068; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16069; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16070; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16071; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16072; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16073; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16074; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16075; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16076; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16077; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16078; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16079; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16080; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16081; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16082; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16083; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16084; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16085; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16086; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16087; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16088; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16089; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16090; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16091; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16092; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16093; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16094; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16095; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16096; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16097; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16098; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}16099; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16100; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}16101; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16102; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}16103; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16104; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k2}16105; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16106; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16107; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16108; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16109; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16110; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16111; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16112; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16113; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16114; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16115; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16116; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16117; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16118; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16119; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16120; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16121; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16122; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16123; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16124; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16125; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16126; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16127; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16128; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16129; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16130; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16131; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16132; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16133; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16134; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16135; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16136; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16137; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16138; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16139; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16140; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16141; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16142; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16143; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16144; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16145; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16146; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16147; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16148; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k1}16149; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16150; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}16151; AVX512-FCP-NEXT:    movb $24, %al16152; AVX512-FCP-NEXT:    kmovw %eax, %k216153; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16154; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16155; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16156; AVX512-FCP-NEXT:    movb $-32, %al16157; AVX512-FCP-NEXT:    kmovw %eax, %k116158; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm5 {%k1}16159; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16160; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16161; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16162; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16163; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16164; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16165; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16166; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16167; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16168; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16169; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16170; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16171; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16172; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16173; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16174; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16175; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm5 {%k1}16176; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16177; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16178; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16179; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16180; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm5 {%k1}16181; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16182; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16183; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16184; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16185; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm5 {%k1}16186; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16187; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16188; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16189; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16190; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm5 {%k1}16191; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16192; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16193; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}16194; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}16195; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16196; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload16197; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16198; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1 {%k1}16199; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill16200; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16201; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16202; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16203; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16204; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16205; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16206; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16207; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16208; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16209; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16210; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16211; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16212; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16213; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16214; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16215; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16216; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16217; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16218; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16219; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16220; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16221; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16222; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16223; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16224; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16225; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload16226; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}16227; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm31 {%k1}16228; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16229; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload16230; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k2}16231; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm27 {%k1}16232; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16233; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}16234; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm11 {%k1}16235; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16236; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload16237; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}16238; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16239; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload16240; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16241; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}16242; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16243; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 16-byte Folded Reload16244; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16245; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm5 {%k1}16246; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16247; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 16-byte Folded Reload16248; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16249; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm6 {%k1}16250; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16251; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm7 # 16-byte Folded Reload16252; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16253; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm7 {%k1}16254; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16255; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm13 # 16-byte Folded Reload16256; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16257; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm13 {%k1}16258; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16259; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 16-byte Folded Reload16260; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16261; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm15 {%k1}16262; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16263; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm4, %zmm2, %zmm416264; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}16265; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16266; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload16267; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm2 {%k1}16268; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload16269; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload16270; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm3 {%k1}16271; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload16272; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm16 # 16-byte Folded Reload16273; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm16 {%k1}16274; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload16275; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 16-byte Folded Reload16276; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm17 {%k1}16277; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload16278; AVX512-FCP-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm18 # 16-byte Folded Reload16279; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}16280; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload16281; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm30, %zmm19, %zmm1916282; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm19 {%k1}16283; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload16284; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm12, %zmm21, %zmm1216285; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm12 {%k1}16286; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload16287; AVX512-FCP-NEXT:    vinserti32x4 $0, %xmm22, %zmm14, %zmm1416288; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm14 {%k1}16289; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 448(%rsi)16290; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, 384(%rsi)16291; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16292; AVX512-FCP-NEXT:    vmovaps %zmm8, 320(%rsi)16293; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16294; AVX512-FCP-NEXT:    vmovaps %zmm8, 256(%rsi)16295; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16296; AVX512-FCP-NEXT:    vmovaps %zmm8, 192(%rsi)16297; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16298; AVX512-FCP-NEXT:    vmovaps %zmm8, 128(%rsi)16299; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16300; AVX512-FCP-NEXT:    vmovaps %zmm8, 64(%rsi)16301; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16302; AVX512-FCP-NEXT:    vmovaps %zmm8, (%rsi)16303; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rdx)16304; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16305; AVX512-FCP-NEXT:    vmovaps %zmm8, 256(%rdx)16306; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16307; AVX512-FCP-NEXT:    vmovaps %zmm8, 320(%rdx)16308; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16309; AVX512-FCP-NEXT:    vmovaps %zmm8, 128(%rdx)16310; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16311; AVX512-FCP-NEXT:    vmovaps %zmm8, 192(%rdx)16312; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16313; AVX512-FCP-NEXT:    vmovaps %zmm8, (%rdx)16314; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16315; AVX512-FCP-NEXT:    vmovaps %zmm8, 64(%rdx)16316; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16317; AVX512-FCP-NEXT:    vmovaps %zmm8, 384(%rdx)16318; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 448(%rcx)16319; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16320; AVX512-FCP-NEXT:    vmovaps %zmm8, 256(%rcx)16321; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16322; AVX512-FCP-NEXT:    vmovaps %zmm8, 320(%rcx)16323; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16324; AVX512-FCP-NEXT:    vmovaps %zmm8, 128(%rcx)16325; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16326; AVX512-FCP-NEXT:    vmovaps %zmm8, 192(%rcx)16327; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16328; AVX512-FCP-NEXT:    vmovaps %zmm8, (%rcx)16329; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16330; AVX512-FCP-NEXT:    vmovaps %zmm8, 64(%rcx)16331; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16332; AVX512-FCP-NEXT:    vmovaps %zmm8, 384(%rcx)16333; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 448(%r8)16334; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, 256(%r8)16335; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, 320(%r8)16336; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16337; AVX512-FCP-NEXT:    vmovaps %zmm8, 128(%r8)16338; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16339; AVX512-FCP-NEXT:    vmovaps %zmm8, 192(%r8)16340; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16341; AVX512-FCP-NEXT:    vmovaps %zmm8, (%r8)16342; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16343; AVX512-FCP-NEXT:    vmovaps %zmm8, 64(%r8)16344; AVX512-FCP-NEXT:    vmovups (%rsp), %zmm8 # 64-byte Reload16345; AVX512-FCP-NEXT:    vmovaps %zmm8, 384(%r8)16346; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 448(%r9)16347; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 256(%r9)16348; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 320(%r9)16349; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 128(%r9)16350; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 192(%r9)16351; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, (%r9)16352; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 64(%r9)16353; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, 384(%r9)16354; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax16355; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 384(%rax)16356; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 448(%rax)16357; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, 256(%rax)16358; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rax)16359; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, 128(%rax)16360; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, 192(%rax)16361; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, (%rax)16362; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)16363; AVX512-FCP-NEXT:    addq $7240, %rsp # imm = 0x1C4816364; AVX512-FCP-NEXT:    vzeroupper16365; AVX512-FCP-NEXT:    retq16366;16367; AVX512DQ-LABEL: load_i64_stride6_vf64:16368; AVX512DQ:       # %bb.0:16369; AVX512DQ-NEXT:    subq $7240, %rsp # imm = 0x1C4816370; AVX512DQ-NEXT:    vmovdqa64 2048(%rdi), %zmm316371; AVX512DQ-NEXT:    vmovdqa64 1280(%rdi), %zmm416372; AVX512DQ-NEXT:    vmovdqa64 1344(%rdi), %zmm016373; AVX512DQ-NEXT:    vmovdqa64 896(%rdi), %zmm516374; AVX512DQ-NEXT:    vmovdqa64 960(%rdi), %zmm2616375; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %zmm216376; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %zmm116377; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm616378; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm2916379; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,0,10,0,6,0,10]16380; AVX512DQ-NEXT:    # zmm7 = mem[0,1,2,3,0,1,2,3]16381; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm816382; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm7, %zmm816383; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16384; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm816385; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm7, %zmm816386; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16387; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm816388; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm7, %zmm816389; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16390; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm816391; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm7, %zmm816392; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16393; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,0,11,1,7,0,11]16394; AVX512DQ-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3]16395; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm916396; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm8, %zmm916397; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16398; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm916399; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm8, %zmm916400; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16401; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm916402; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm8, %zmm916403; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16404; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm916405; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm8, %zmm916406; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16407; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm9 = [10,4,10,4,10,4,10,4]16408; AVX512DQ-NEXT:    # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16409; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm1016410; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm1016411; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16412; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1016413; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm9, %zmm1016414; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16415; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm1016416; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm9, %zmm1016417; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16418; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm1016419; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm9, %zmm1016420; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16421; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [11,5,11,5,11,5,11,5]16422; AVX512DQ-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16423; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm1116424; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm10, %zmm1116425; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16426; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1116427; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm10, %zmm1116428; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16429; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm1116430; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm10, %zmm1116431; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16432; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm1116433; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm10, %zmm1116434; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16435; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm11 = [12,0,0,6,12,0,0,6]16436; AVX512DQ-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3]16437; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm1316438; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm1216439; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm11, %zmm1316440; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16441; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [13,0,1,7,13,0,1,7]16442; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]16443; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm1216444; AVX512DQ-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16445; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm116446; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm11, %zmm116447; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16448; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm2, %zmm616449; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16450; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm116451; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm11, %zmm116452; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16453; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm2, %zmm416454; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16455; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm016456; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm11, %zmm016457; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16458; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm2, %zmm516459; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16460; AVX512DQ-NEXT:    vmovdqa64 2112(%rdi), %zmm016461; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm116462; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm7, %zmm116463; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16464; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm116465; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm8, %zmm116466; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16467; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm116468; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm9, %zmm116469; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16470; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm116471; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm10, %zmm116472; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16473; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm116474; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm11, %zmm116475; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16476; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm2, %zmm316477; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16478; AVX512DQ-NEXT:    vmovdqa64 1664(%rdi), %zmm116479; AVX512DQ-NEXT:    vmovdqa64 1728(%rdi), %zmm016480; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm316481; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm7, %zmm316482; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16483; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm316484; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm316485; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16486; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316487; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm9, %zmm316488; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16489; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316490; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm10, %zmm316491; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16492; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316493; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm11, %zmm316494; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16495; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm2, %zmm116496; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16497; AVX512DQ-NEXT:    vmovdqa64 2432(%rdi), %zmm316498; AVX512DQ-NEXT:    vmovdqa64 2496(%rdi), %zmm016499; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm116500; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm7, %zmm116501; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16502; AVX512DQ-NEXT:    vmovdqa64 2816(%rdi), %zmm416503; AVX512DQ-NEXT:    vmovdqa64 2880(%rdi), %zmm116504; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm1, %zmm716505; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16506; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm516507; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm8, %zmm516508; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16509; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm1, %zmm816510; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16511; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm516512; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm9, %zmm516513; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16514; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm4, %zmm916515; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16516; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm516517; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm10, %zmm516518; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16519; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm4, %zmm1016520; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16521; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm516522; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm11, %zmm516523; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16524; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm4, %zmm1116525; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16526; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm416527; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16528; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm2, %zmm316529; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16530; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm116531; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm216532; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm8 = [0,6,12,0]16533; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm016534; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm016535; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16536; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [1,7,13,0]16537; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm016538; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm016539; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16540; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [10,0,6,0]16541; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm016542; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm10, %zmm016543; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16544; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [11,1,7,0]16545; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm016546; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm11, %zmm016547; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16548; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm4 = [4,10,4,10,4,10,4,10]16549; AVX512DQ-NEXT:    # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16550; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm016551; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm016552; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16553; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,11,5,11,5,11,5,11]16554; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16555; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm216556; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16557; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm216558; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm116559; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316560; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm316561; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16562; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316563; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm316564; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16565; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316566; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm10, %zmm316567; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16568; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316569; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm11, %zmm316570; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16571; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316572; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm316573; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16574; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm216575; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16576; AVX512DQ-NEXT:    vmovdqa64 1216(%rdi), %zmm116577; AVX512DQ-NEXT:    vmovdqa64 1152(%rdi), %zmm216578; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316579; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm316580; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16581; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316582; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm316583; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16584; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316585; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm10, %zmm316586; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16587; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316588; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm11, %zmm316589; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16590; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316591; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm316592; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16593; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm216594; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16595; AVX512DQ-NEXT:    vmovdqa64 832(%rdi), %zmm116596; AVX512DQ-NEXT:    vmovdqa64 768(%rdi), %zmm216597; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316598; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm316599; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16600; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316601; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm316602; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16603; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316604; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm10, %zmm316605; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16606; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316607; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm11, %zmm316608; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16609; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316610; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm316611; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16612; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm216613; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16614; AVX512DQ-NEXT:    vmovdqa64 1984(%rdi), %zmm116615; AVX512DQ-NEXT:    vmovdqa64 1920(%rdi), %zmm216616; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316617; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm316618; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16619; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316620; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm316621; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16622; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316623; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm10, %zmm316624; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16625; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316626; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm11, %zmm316627; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16628; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm316629; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm316630; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16631; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm216632; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16633; AVX512DQ-NEXT:    vmovdqa64 1600(%rdi), %zmm116634; AVX512DQ-NEXT:    vmovdqa64 1536(%rdi), %zmm3016635; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm216636; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm216637; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16638; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm216639; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm216640; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16641; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm216642; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm10, %zmm216643; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16644; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm216645; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm11, %zmm216646; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16647; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm216648; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm216649; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16650; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm3016651; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm2616652; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm1816653; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm116654; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm4, %zmm116655; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16656; AVX512DQ-NEXT:    vmovdqa64 704(%rdi), %zmm2116657; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %zmm1316658; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm116659; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm4, %zmm116660; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16661; AVX512DQ-NEXT:    vmovdqa64 1088(%rdi), %zmm2416662; AVX512DQ-NEXT:    vmovdqa64 1024(%rdi), %zmm1916663; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm116664; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm4, %zmm116665; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16666; AVX512DQ-NEXT:    vmovdqa64 1472(%rdi), %zmm2816667; AVX512DQ-NEXT:    vmovdqa64 1408(%rdi), %zmm1616668; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm116669; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm4, %zmm116670; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16671; AVX512DQ-NEXT:    vmovdqa64 1856(%rdi), %zmm2516672; AVX512DQ-NEXT:    vmovdqa64 1792(%rdi), %zmm716673; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm116674; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm4, %zmm116675; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16676; AVX512DQ-NEXT:    vmovdqa64 2240(%rdi), %zmm2316677; AVX512DQ-NEXT:    vmovdqa64 2176(%rdi), %zmm1716678; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm116679; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm4, %zmm116680; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16681; AVX512DQ-NEXT:    vmovdqa64 2624(%rdi), %zmm2016682; AVX512DQ-NEXT:    vmovdqa64 2560(%rdi), %zmm516683; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm116684; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm4, %zmm116685; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16686; AVX512DQ-NEXT:    vmovdqa64 2368(%rdi), %zmm116687; AVX512DQ-NEXT:    vmovdqa64 2304(%rdi), %zmm2216688; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm2916689; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm8, %zmm2916690; AVX512DQ-NEXT:    vmovdqa64 3008(%rdi), %zmm1416691; AVX512DQ-NEXT:    vmovdqa64 2944(%rdi), %zmm3116692; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm216693; AVX512DQ-NEXT:    vpermt2q %zmm14, %zmm4, %zmm216694; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16695; AVX512DQ-NEXT:    vmovdqa64 2752(%rdi), %zmm216696; AVX512DQ-NEXT:    vmovdqa64 2688(%rdi), %zmm1216697; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm12, %zmm816698; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm316699; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm9, %zmm316700; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16701; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm12, %zmm916702; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316703; AVX512DQ-NEXT:    vpermt2q %zmm22, %zmm10, %zmm316704; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16705; AVX512DQ-NEXT:    vpermi2q %zmm12, %zmm2, %zmm1016706; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm316707; AVX512DQ-NEXT:    vpermt2q %zmm22, %zmm11, %zmm316708; AVX512DQ-NEXT:    vmovdqu64 %zmm3, (%rsp) # 64-byte Spill16709; AVX512DQ-NEXT:    vpermi2q %zmm12, %zmm2, %zmm1116710; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm316711; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm4, %zmm316712; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16713; AVX512DQ-NEXT:    vpermi2q %zmm2, %zmm12, %zmm416714; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1216715; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm0, %zmm2216716; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm116717; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm0, %zmm116718; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16719; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm116720; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm0, %zmm116721; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16722; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm116723; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm0, %zmm116724; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16725; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm116726; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm0, %zmm116727; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16728; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm116729; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm0, %zmm116730; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16731; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm116732; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm0, %zmm116733; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16734; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm116735; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm0, %zmm116736; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16737; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm116738; AVX512DQ-NEXT:    vpermt2q %zmm14, %zmm0, %zmm116739; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16740; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm1 = [0,0,6,12,0,0,6,12]16741; AVX512DQ-NEXT:    # zmm1 = mem[0,1,2,3,0,1,2,3]16742; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm016743; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm1, %zmm016744; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16745; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,7,13,0,1,7,13]16746; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3]16747; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm016748; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm2, %zmm016749; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16750; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm3 = [0,10,0,6,0,10,0,6]16751; AVX512DQ-NEXT:    # zmm3 = mem[0,1,2,3,0,1,2,3]16752; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm016753; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm3, %zmm016754; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16755; AVX512DQ-NEXT:    vbroadcasti64x4 {{.*#+}} zmm0 = [0,11,1,7,0,11,1,7]16756; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3]16757; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2116758; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm616759; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm1, %zmm616760; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16761; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm616762; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm2, %zmm616763; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16764; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm616765; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm3, %zmm616766; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16767; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2616768; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm2716769; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm1, %zmm2716770; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm616771; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm2, %zmm616772; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16773; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm616774; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm3, %zmm616775; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16776; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2816777; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm1816778; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm1, %zmm1816779; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm616780; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm2, %zmm616781; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16782; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm616783; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm3, %zmm616784; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16785; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2416786; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm1516787; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm1, %zmm1516788; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm1916789; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm2, %zmm1916790; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm616791; AVX512DQ-NEXT:    vpermt2q %zmm17, %zmm3, %zmm616792; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16793; AVX512DQ-NEXT:    vpermt2q %zmm17, %zmm0, %zmm2316794; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1316795; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm1, %zmm1316796; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1716797; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm2, %zmm1716798; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm616799; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm3, %zmm616800; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16801; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2516802; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm616803; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm1, %zmm616804; AVX512DQ-NEXT:    vpermi2q %zmm14, %zmm31, %zmm116805; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm716806; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm2, %zmm716807; AVX512DQ-NEXT:    vpermi2q %zmm14, %zmm31, %zmm216808; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm1616809; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm3, %zmm1616810; AVX512DQ-NEXT:    vpermi2q %zmm31, %zmm14, %zmm316811; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1416812; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm0, %zmm2016813; AVX512DQ-NEXT:    movb $56, %al16814; AVX512DQ-NEXT:    kmovw %eax, %k116815; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16816; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16817; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16818; AVX512DQ-NEXT:    movb $-64, %al16819; AVX512DQ-NEXT:    kmovw %eax, %k216820; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16821; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16822; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16823; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16824; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16825; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16826; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16827; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16828; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16829; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16830; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16831; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16832; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16833; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16834; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16835; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16836; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16837; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16838; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16839; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16840; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16841; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16842; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16843; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16844; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16845; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16846; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16847; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16848; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16849; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16850; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16851; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16852; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16853; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16854; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k1}16855; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16856; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}16857; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16858; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}16859; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16860; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k2}16861; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16862; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16863; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16864; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16865; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16866; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16867; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16868; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16869; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16870; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16871; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16872; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16873; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16874; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16875; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16876; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16877; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16878; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16879; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16880; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16881; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16882; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16883; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16884; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16885; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16886; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16887; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16888; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16889; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16890; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16891; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16892; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16893; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16894; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16895; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16896; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16897; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16898; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16899; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16900; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16901; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16902; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16903; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16904; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k1}16905; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16906; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}16907; AVX512DQ-NEXT:    movb $24, %al16908; AVX512DQ-NEXT:    kmovw %eax, %k216909; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16910; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16911; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16912; AVX512DQ-NEXT:    movb $-32, %al16913; AVX512DQ-NEXT:    kmovw %eax, %k116914; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm5 {%k1}16915; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16916; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16917; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16918; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16919; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16920; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16921; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16922; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16923; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16924; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16925; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16926; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k1}16927; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16928; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16929; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16930; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16931; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm5 {%k1}16932; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16933; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16934; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16935; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16936; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm5 {%k1}16937; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16938; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16939; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16940; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16941; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm5 {%k1}16942; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16943; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16944; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload16945; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm5 {%k2}16946; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm5 {%k1}16947; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16948; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16949; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}16950; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm10 {%k1}16951; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16952; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload16953; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16954; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1 {%k1}16955; AVX512DQ-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill16956; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16957; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16958; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16959; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16960; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16961; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16962; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16963; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16964; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16965; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16966; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16967; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16968; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16969; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16970; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16971; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16972; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16973; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16974; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16975; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16976; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}16977; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16978; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16979; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16980; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16981; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload16982; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}16983; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm31 {%k1}16984; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16985; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload16986; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k2}16987; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm27 {%k1}16988; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16989; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}16990; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm11 {%k1}16991; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16992; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 16-byte Folded Reload16993; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}16994; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16995; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload16996; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16997; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm1 {%k1}16998; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16999; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 16-byte Folded Reload17000; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17001; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm5 {%k1}17002; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17003; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 16-byte Folded Reload17004; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17005; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm6 {%k1}17006; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17007; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm7 # 16-byte Folded Reload17008; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17009; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm7 {%k1}17010; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17011; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm13 # 16-byte Folded Reload17012; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17013; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm13 {%k1}17014; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17015; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm15 # 16-byte Folded Reload17016; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17017; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm15 {%k1}17018; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17019; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm4, %zmm2, %zmm417020; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm4 {%k1}17021; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload17022; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload17023; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm2 {%k1}17024; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload17025; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 16-byte Folded Reload17026; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm3 {%k1}17027; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload17028; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm16 # 16-byte Folded Reload17029; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm16 {%k1}17030; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload17031; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 16-byte Folded Reload17032; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm17 {%k1}17033; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload17034; AVX512DQ-NEXT:    vinserti32x4 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm18 # 16-byte Folded Reload17035; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm18 {%k1}17036; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload17037; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm30, %zmm19, %zmm1917038; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm19 {%k1}17039; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload17040; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm12, %zmm21, %zmm1217041; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm12 {%k1}17042; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload17043; AVX512DQ-NEXT:    vinserti32x4 $0, %xmm22, %zmm14, %zmm1417044; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm14 {%k1}17045; AVX512DQ-NEXT:    vmovdqa64 %zmm8, 448(%rsi)17046; AVX512DQ-NEXT:    vmovdqa64 %zmm29, 384(%rsi)17047; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17048; AVX512DQ-NEXT:    vmovaps %zmm8, 320(%rsi)17049; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17050; AVX512DQ-NEXT:    vmovaps %zmm8, 256(%rsi)17051; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17052; AVX512DQ-NEXT:    vmovaps %zmm8, 192(%rsi)17053; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17054; AVX512DQ-NEXT:    vmovaps %zmm8, 128(%rsi)17055; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17056; AVX512DQ-NEXT:    vmovaps %zmm8, 64(%rsi)17057; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17058; AVX512DQ-NEXT:    vmovaps %zmm8, (%rsi)17059; AVX512DQ-NEXT:    vmovdqa64 %zmm9, 448(%rdx)17060; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17061; AVX512DQ-NEXT:    vmovaps %zmm8, 256(%rdx)17062; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17063; AVX512DQ-NEXT:    vmovaps %zmm8, 320(%rdx)17064; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload17065; AVX512DQ-NEXT:    vmovaps %zmm8, 128(%rdx)17066; AVX512DQ-NE

File truncated to the first 1 MiB.