16314 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_i32_stride8_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {19; SSE-LABEL: load_i32_stride8_vf2:20; SSE: # %bb.0:21; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax22; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r1023; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r1124; SSE-NEXT: movdqa (%rdi), %xmm025; SSE-NEXT: movdqa 16(%rdi), %xmm126; SSE-NEXT: movdqa 32(%rdi), %xmm227; SSE-NEXT: movdqa 48(%rdi), %xmm328; SSE-NEXT: movdqa %xmm0, %xmm429; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]30; SSE-NEXT: movq %xmm4, (%rsi)31; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,3,2,3]32; SSE-NEXT: movq %xmm4, (%rdx)33; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]34; SSE-NEXT: movq %xmm0, (%rcx)35; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]36; SSE-NEXT: movq %xmm0, (%r8)37; SSE-NEXT: movdqa %xmm1, %xmm038; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]39; SSE-NEXT: movq %xmm0, (%r9)40; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]41; SSE-NEXT: movq %xmm0, (%r11)42; SSE-NEXT: punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]43; SSE-NEXT: movq %xmm1, (%r10)44; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]45; SSE-NEXT: movq %xmm0, (%rax)46; SSE-NEXT: retq47;48; AVX-LABEL: load_i32_stride8_vf2:49; AVX: # %bb.0:50; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax51; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r1052; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r1153; AVX-NEXT: vmovaps (%rdi), %ymm054; AVX-NEXT: vmovaps 32(%rdi), %ymm155; AVX-NEXT: vmovdqa 32(%rdi), %xmm256; AVX-NEXT: vmovdqa (%rdi), %xmm357; AVX-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]58; AVX-NEXT: vmovq %xmm4, (%rsi)59; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,1,1,1]60; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm2[2,3],xmm4[4,5,6,7]61; AVX-NEXT: vmovq %xmm4, (%rdx)62; AVX-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]63; AVX-NEXT: vmovq %xmm2, (%rcx)64; AVX-NEXT: vpextrq $1, %xmm2, (%r8)65; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]66; AVX-NEXT: vextractf128 $1, %ymm2, %xmm267; AVX-NEXT: vmovlps %xmm2, (%r9)68; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]69; AVX-NEXT: vextractf128 $1, %ymm2, %xmm270; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[2,0,2,3]71; AVX-NEXT: vmovlps %xmm2, (%r11)72; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]73; AVX-NEXT: vextractf128 $1, %ymm2, %xmm274; AVX-NEXT: vmovlps %xmm2, (%r10)75; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]76; AVX-NEXT: vextractf128 $1, %ymm0, %xmm077; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0,2,3]78; AVX-NEXT: vmovlps %xmm0, (%rax)79; AVX-NEXT: vzeroupper80; AVX-NEXT: retq81;82; AVX2-LABEL: load_i32_stride8_vf2:83; AVX2: # %bb.0:84; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax85; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r1086; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r1187; AVX2-NEXT: vmovaps (%rdi), %ymm088; AVX2-NEXT: vmovaps 32(%rdi), %ymm189; AVX2-NEXT: vmovdqa (%rdi), %xmm290; AVX2-NEXT: vmovdqa 32(%rdi), %xmm391; AVX2-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]92; AVX2-NEXT: vmovq %xmm4, (%rsi)93; AVX2-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[1,1,1,1]94; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0],xmm3[1],xmm4[2,3]95; AVX2-NEXT: vmovq %xmm4, (%rdx)96; AVX2-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]97; AVX2-NEXT: vmovq %xmm2, (%rcx)98; AVX2-NEXT: vpextrq $1, %xmm2, (%r8)99; AVX2-NEXT: vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]100; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm2101; AVX2-NEXT: vmovlps %xmm2, (%r9)102; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm0[1,1,1,1,5,5,5,5]103; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7]104; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm2105; AVX2-NEXT: vmovlps %xmm2, (%r11)106; AVX2-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]107; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm1108; AVX2-NEXT: vmovlps %xmm1, (%r10)109; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]110; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm0111; AVX2-NEXT: vmovlps %xmm0, (%rax)112; AVX2-NEXT: vzeroupper113; AVX2-NEXT: retq114;115; AVX2-FP-LABEL: load_i32_stride8_vf2:116; AVX2-FP: # %bb.0:117; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax118; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10119; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11120; AVX2-FP-NEXT: vmovaps (%rdi), %ymm0121; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm1122; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm2123; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm3124; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]125; AVX2-FP-NEXT: vmovq %xmm4, (%rsi)126; AVX2-FP-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[1,1,1,1]127; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0],xmm3[1],xmm4[2,3]128; AVX2-FP-NEXT: vmovq %xmm4, (%rdx)129; AVX2-FP-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]130; AVX2-FP-NEXT: vmovq %xmm2, (%rcx)131; AVX2-FP-NEXT: vpextrq $1, %xmm2, (%r8)132; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]133; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm2134; AVX2-FP-NEXT: vmovlps %xmm2, (%r9)135; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm0[1,1,1,1,5,5,5,5]136; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7]137; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm2138; AVX2-FP-NEXT: vmovlps %xmm2, (%r11)139; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]140; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm1141; AVX2-FP-NEXT: vmovlps %xmm1, (%r10)142; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]143; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm0144; AVX2-FP-NEXT: vmovlps %xmm0, (%rax)145; AVX2-FP-NEXT: vzeroupper146; AVX2-FP-NEXT: retq147;148; AVX2-FCP-LABEL: load_i32_stride8_vf2:149; AVX2-FCP: # %bb.0:150; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax151; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10152; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11153; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm0154; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm1155; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm2156; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm3157; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]158; AVX2-FCP-NEXT: vmovq %xmm4, (%rsi)159; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[1,1,1,1]160; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0],xmm3[1],xmm4[2,3]161; AVX2-FCP-NEXT: vmovq %xmm4, (%rdx)162; AVX2-FCP-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]163; AVX2-FCP-NEXT: vmovq %xmm2, (%rcx)164; AVX2-FCP-NEXT: vpextrq $1, %xmm2, (%r8)165; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]166; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm2167; AVX2-FCP-NEXT: vmovlps %xmm2, (%r9)168; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm0[1,1,1,1,5,5,5,5]169; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7]170; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm2171; AVX2-FCP-NEXT: vmovlps %xmm2, (%r11)172; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]173; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm1174; AVX2-FCP-NEXT: vmovlps %xmm1, (%r10)175; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]176; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0177; AVX2-FCP-NEXT: vmovlps %xmm0, (%rax)178; AVX2-FCP-NEXT: vzeroupper179; AVX2-FCP-NEXT: retq180;181; AVX512-LABEL: load_i32_stride8_vf2:182; AVX512: # %bb.0:183; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax184; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10185; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11186; AVX512-NEXT: vmovdqa (%rdi), %xmm0187; AVX512-NEXT: vmovdqa 32(%rdi), %xmm1188; AVX512-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]189; AVX512-NEXT: vmovaps 32(%rdi), %ymm3190; AVX512-NEXT: vmovaps (%rdi), %ymm4191; AVX512-NEXT: vmovq %xmm2, (%rsi)192; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]193; AVX512-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2,3]194; AVX512-NEXT: vmovq %xmm2, (%rdx)195; AVX512-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]196; AVX512-NEXT: vmovq %xmm0, (%rcx)197; AVX512-NEXT: vpextrq $1, %xmm0, (%r8)198; AVX512-NEXT: vunpcklps {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[4],ymm3[4],ymm4[5],ymm3[5]199; AVX512-NEXT: vextractf128 $1, %ymm0, %xmm0200; AVX512-NEXT: vmovlps %xmm0, (%r9)201; AVX512-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]202; AVX512-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]203; AVX512-NEXT: vextractf128 $1, %ymm0, %xmm0204; AVX512-NEXT: vmovlps %xmm0, (%r11)205; AVX512-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[6],ymm3[6],ymm4[7],ymm3[7]206; AVX512-NEXT: vextractf128 $1, %ymm0, %xmm1207; AVX512-NEXT: vmovlps %xmm1, (%r10)208; AVX512-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]209; AVX512-NEXT: vextractf128 $1, %ymm0, %xmm0210; AVX512-NEXT: vmovlps %xmm0, (%rax)211; AVX512-NEXT: vzeroupper212; AVX512-NEXT: retq213;214; AVX512-FCP-LABEL: load_i32_stride8_vf2:215; AVX512-FCP: # %bb.0:216; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax217; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10218; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11219; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0220; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm1221; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]222; AVX512-FCP-NEXT: vmovaps {{\.?LCPI[0-9]+_[0-9]+}}+16(%rip), %xmm3223; AVX512-FCP-NEXT: vpermps (%rdi), %zmm3, %zmm3224; AVX512-FCP-NEXT: vmovaps 32(%rdi), %ymm4225; AVX512-FCP-NEXT: vmovaps (%rdi), %ymm5226; AVX512-FCP-NEXT: vmovq %xmm2, (%rsi)227; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [1,5,0,0]228; AVX512-FCP-NEXT: vpermi2d %xmm1, %xmm0, %xmm2229; AVX512-FCP-NEXT: vmovq %xmm2, (%rdx)230; AVX512-FCP-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]231; AVX512-FCP-NEXT: vmovq %xmm0, (%rcx)232; AVX512-FCP-NEXT: vpextrq $1, %xmm0, (%r8)233; AVX512-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]234; AVX512-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0235; AVX512-FCP-NEXT: vmovlps %xmm0, (%r9)236; AVX512-FCP-NEXT: vmovlps %xmm3, (%r11)237; AVX512-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]238; AVX512-FCP-NEXT: vextractf128 $1, %ymm0, %xmm1239; AVX512-FCP-NEXT: vmovlps %xmm1, (%r10)240; AVX512-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]241; AVX512-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0242; AVX512-FCP-NEXT: vmovlps %xmm0, (%rax)243; AVX512-FCP-NEXT: vzeroupper244; AVX512-FCP-NEXT: retq245;246; AVX512DQ-LABEL: load_i32_stride8_vf2:247; AVX512DQ: # %bb.0:248; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax249; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10250; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11251; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0252; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm1253; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]254; AVX512DQ-NEXT: vmovaps 32(%rdi), %ymm3255; AVX512DQ-NEXT: vmovaps (%rdi), %ymm4256; AVX512DQ-NEXT: vmovq %xmm2, (%rsi)257; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]258; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2,3]259; AVX512DQ-NEXT: vmovq %xmm2, (%rdx)260; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]261; AVX512DQ-NEXT: vmovq %xmm0, (%rcx)262; AVX512DQ-NEXT: vpextrq $1, %xmm0, (%r8)263; AVX512DQ-NEXT: vunpcklps {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[4],ymm3[4],ymm4[5],ymm3[5]264; AVX512DQ-NEXT: vextractf128 $1, %ymm0, %xmm0265; AVX512DQ-NEXT: vmovlps %xmm0, (%r9)266; AVX512DQ-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]267; AVX512DQ-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]268; AVX512DQ-NEXT: vextractf128 $1, %ymm0, %xmm0269; AVX512DQ-NEXT: vmovlps %xmm0, (%r11)270; AVX512DQ-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[6],ymm3[6],ymm4[7],ymm3[7]271; AVX512DQ-NEXT: vextractf128 $1, %ymm0, %xmm1272; AVX512DQ-NEXT: vmovlps %xmm1, (%r10)273; AVX512DQ-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]274; AVX512DQ-NEXT: vextractf128 $1, %ymm0, %xmm0275; AVX512DQ-NEXT: vmovlps %xmm0, (%rax)276; AVX512DQ-NEXT: vzeroupper277; AVX512DQ-NEXT: retq278;279; AVX512DQ-FCP-LABEL: load_i32_stride8_vf2:280; AVX512DQ-FCP: # %bb.0:281; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax282; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10283; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11284; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0285; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm1286; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]287; AVX512DQ-FCP-NEXT: vmovaps {{\.?LCPI[0-9]+_[0-9]+}}+16(%rip), %xmm3288; AVX512DQ-FCP-NEXT: vpermps (%rdi), %zmm3, %zmm3289; AVX512DQ-FCP-NEXT: vmovaps 32(%rdi), %ymm4290; AVX512DQ-FCP-NEXT: vmovaps (%rdi), %ymm5291; AVX512DQ-FCP-NEXT: vmovq %xmm2, (%rsi)292; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [1,5,0,0]293; AVX512DQ-FCP-NEXT: vpermi2d %xmm1, %xmm0, %xmm2294; AVX512DQ-FCP-NEXT: vmovq %xmm2, (%rdx)295; AVX512DQ-FCP-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]296; AVX512DQ-FCP-NEXT: vmovq %xmm0, (%rcx)297; AVX512DQ-FCP-NEXT: vpextrq $1, %xmm0, (%r8)298; AVX512DQ-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]299; AVX512DQ-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0300; AVX512DQ-FCP-NEXT: vmovlps %xmm0, (%r9)301; AVX512DQ-FCP-NEXT: vmovlps %xmm3, (%r11)302; AVX512DQ-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]303; AVX512DQ-FCP-NEXT: vextractf128 $1, %ymm0, %xmm1304; AVX512DQ-FCP-NEXT: vmovlps %xmm1, (%r10)305; AVX512DQ-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]306; AVX512DQ-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0307; AVX512DQ-FCP-NEXT: vmovlps %xmm0, (%rax)308; AVX512DQ-FCP-NEXT: vzeroupper309; AVX512DQ-FCP-NEXT: retq310;311; AVX512BW-LABEL: load_i32_stride8_vf2:312; AVX512BW: # %bb.0:313; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax314; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10315; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11316; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0317; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm1318; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]319; AVX512BW-NEXT: vmovaps 32(%rdi), %ymm3320; AVX512BW-NEXT: vmovaps (%rdi), %ymm4321; AVX512BW-NEXT: vmovq %xmm2, (%rsi)322; AVX512BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]323; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2,3]324; AVX512BW-NEXT: vmovq %xmm2, (%rdx)325; AVX512BW-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]326; AVX512BW-NEXT: vmovq %xmm0, (%rcx)327; AVX512BW-NEXT: vpextrq $1, %xmm0, (%r8)328; AVX512BW-NEXT: vunpcklps {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[4],ymm3[4],ymm4[5],ymm3[5]329; AVX512BW-NEXT: vextractf128 $1, %ymm0, %xmm0330; AVX512BW-NEXT: vmovlps %xmm0, (%r9)331; AVX512BW-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]332; AVX512BW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]333; AVX512BW-NEXT: vextractf128 $1, %ymm0, %xmm0334; AVX512BW-NEXT: vmovlps %xmm0, (%r11)335; AVX512BW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[6],ymm3[6],ymm4[7],ymm3[7]336; AVX512BW-NEXT: vextractf128 $1, %ymm0, %xmm1337; AVX512BW-NEXT: vmovlps %xmm1, (%r10)338; AVX512BW-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]339; AVX512BW-NEXT: vextractf128 $1, %ymm0, %xmm0340; AVX512BW-NEXT: vmovlps %xmm0, (%rax)341; AVX512BW-NEXT: vzeroupper342; AVX512BW-NEXT: retq343;344; AVX512BW-FCP-LABEL: load_i32_stride8_vf2:345; AVX512BW-FCP: # %bb.0:346; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax347; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10348; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11349; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0350; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm1351; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]352; AVX512BW-FCP-NEXT: vmovaps {{\.?LCPI[0-9]+_[0-9]+}}+16(%rip), %xmm3353; AVX512BW-FCP-NEXT: vpermps (%rdi), %zmm3, %zmm3354; AVX512BW-FCP-NEXT: vmovaps 32(%rdi), %ymm4355; AVX512BW-FCP-NEXT: vmovaps (%rdi), %ymm5356; AVX512BW-FCP-NEXT: vmovq %xmm2, (%rsi)357; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [1,5,0,0]358; AVX512BW-FCP-NEXT: vpermi2d %xmm1, %xmm0, %xmm2359; AVX512BW-FCP-NEXT: vmovq %xmm2, (%rdx)360; AVX512BW-FCP-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]361; AVX512BW-FCP-NEXT: vmovq %xmm0, (%rcx)362; AVX512BW-FCP-NEXT: vpextrq $1, %xmm0, (%r8)363; AVX512BW-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]364; AVX512BW-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0365; AVX512BW-FCP-NEXT: vmovlps %xmm0, (%r9)366; AVX512BW-FCP-NEXT: vmovlps %xmm3, (%r11)367; AVX512BW-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]368; AVX512BW-FCP-NEXT: vextractf128 $1, %ymm0, %xmm1369; AVX512BW-FCP-NEXT: vmovlps %xmm1, (%r10)370; AVX512BW-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]371; AVX512BW-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0372; AVX512BW-FCP-NEXT: vmovlps %xmm0, (%rax)373; AVX512BW-FCP-NEXT: vzeroupper374; AVX512BW-FCP-NEXT: retq375;376; AVX512DQ-BW-LABEL: load_i32_stride8_vf2:377; AVX512DQ-BW: # %bb.0:378; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax379; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10380; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11381; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0382; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm1383; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]384; AVX512DQ-BW-NEXT: vmovaps 32(%rdi), %ymm3385; AVX512DQ-BW-NEXT: vmovaps (%rdi), %ymm4386; AVX512DQ-BW-NEXT: vmovq %xmm2, (%rsi)387; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]388; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2,3]389; AVX512DQ-BW-NEXT: vmovq %xmm2, (%rdx)390; AVX512DQ-BW-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]391; AVX512DQ-BW-NEXT: vmovq %xmm0, (%rcx)392; AVX512DQ-BW-NEXT: vpextrq $1, %xmm0, (%r8)393; AVX512DQ-BW-NEXT: vunpcklps {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[4],ymm3[4],ymm4[5],ymm3[5]394; AVX512DQ-BW-NEXT: vextractf128 $1, %ymm0, %xmm0395; AVX512DQ-BW-NEXT: vmovlps %xmm0, (%r9)396; AVX512DQ-BW-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]397; AVX512DQ-BW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]398; AVX512DQ-BW-NEXT: vextractf128 $1, %ymm0, %xmm0399; AVX512DQ-BW-NEXT: vmovlps %xmm0, (%r11)400; AVX512DQ-BW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[6],ymm3[6],ymm4[7],ymm3[7]401; AVX512DQ-BW-NEXT: vextractf128 $1, %ymm0, %xmm1402; AVX512DQ-BW-NEXT: vmovlps %xmm1, (%r10)403; AVX512DQ-BW-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]404; AVX512DQ-BW-NEXT: vextractf128 $1, %ymm0, %xmm0405; AVX512DQ-BW-NEXT: vmovlps %xmm0, (%rax)406; AVX512DQ-BW-NEXT: vzeroupper407; AVX512DQ-BW-NEXT: retq408;409; AVX512DQ-BW-FCP-LABEL: load_i32_stride8_vf2:410; AVX512DQ-BW-FCP: # %bb.0:411; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax412; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10413; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11414; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0415; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm1416; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]417; AVX512DQ-BW-FCP-NEXT: vmovaps {{\.?LCPI[0-9]+_[0-9]+}}+16(%rip), %xmm3418; AVX512DQ-BW-FCP-NEXT: vpermps (%rdi), %zmm3, %zmm3419; AVX512DQ-BW-FCP-NEXT: vmovaps 32(%rdi), %ymm4420; AVX512DQ-BW-FCP-NEXT: vmovaps (%rdi), %ymm5421; AVX512DQ-BW-FCP-NEXT: vmovq %xmm2, (%rsi)422; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [1,5,0,0]423; AVX512DQ-BW-FCP-NEXT: vpermi2d %xmm1, %xmm0, %xmm2424; AVX512DQ-BW-FCP-NEXT: vmovq %xmm2, (%rdx)425; AVX512DQ-BW-FCP-NEXT: vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]426; AVX512DQ-BW-FCP-NEXT: vmovq %xmm0, (%rcx)427; AVX512DQ-BW-FCP-NEXT: vpextrq $1, %xmm0, (%r8)428; AVX512DQ-BW-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]429; AVX512DQ-BW-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0430; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm0, (%r9)431; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm3, (%r11)432; AVX512DQ-BW-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]433; AVX512DQ-BW-FCP-NEXT: vextractf128 $1, %ymm0, %xmm1434; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm1, (%r10)435; AVX512DQ-BW-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]436; AVX512DQ-BW-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0437; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm0, (%rax)438; AVX512DQ-BW-FCP-NEXT: vzeroupper439; AVX512DQ-BW-FCP-NEXT: retq440 %wide.vec = load <16 x i32>, ptr %in.vec, align 64441 %strided.vec0 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 0, i32 8>442 %strided.vec1 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 1, i32 9>443 %strided.vec2 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 2, i32 10>444 %strided.vec3 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 3, i32 11>445 %strided.vec4 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 4, i32 12>446 %strided.vec5 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 5, i32 13>447 %strided.vec6 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 6, i32 14>448 %strided.vec7 = shufflevector <16 x i32> %wide.vec, <16 x i32> poison, <2 x i32> <i32 7, i32 15>449 store <2 x i32> %strided.vec0, ptr %out.vec0, align 64450 store <2 x i32> %strided.vec1, ptr %out.vec1, align 64451 store <2 x i32> %strided.vec2, ptr %out.vec2, align 64452 store <2 x i32> %strided.vec3, ptr %out.vec3, align 64453 store <2 x i32> %strided.vec4, ptr %out.vec4, align 64454 store <2 x i32> %strided.vec5, ptr %out.vec5, align 64455 store <2 x i32> %strided.vec6, ptr %out.vec6, align 64456 store <2 x i32> %strided.vec7, ptr %out.vec7, align 64457 ret void458}459 460define void @load_i32_stride8_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {461; SSE-LABEL: load_i32_stride8_vf4:462; SSE: # %bb.0:463; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax464; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10465; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r11466; SSE-NEXT: movaps 112(%rdi), %xmm3467; SSE-NEXT: movaps 80(%rdi), %xmm2468; SSE-NEXT: movaps (%rdi), %xmm1469; SSE-NEXT: movaps 16(%rdi), %xmm0470; SSE-NEXT: movaps 32(%rdi), %xmm4471; SSE-NEXT: movaps 48(%rdi), %xmm5472; SSE-NEXT: movaps 96(%rdi), %xmm6473; SSE-NEXT: movaps 64(%rdi), %xmm7474; SSE-NEXT: movaps %xmm7, %xmm8475; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1]476; SSE-NEXT: movaps %xmm1, %xmm9477; SSE-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm4[0],xmm9[1],xmm4[1]478; SSE-NEXT: movaps %xmm9, %xmm10479; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm8[0]480; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm8[1]481; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]482; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm4[2],xmm1[3],xmm4[3]483; SSE-NEXT: movaps %xmm1, %xmm4484; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm7[0]485; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm7[1]486; SSE-NEXT: movaps %xmm2, %xmm6487; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1]488; SSE-NEXT: movaps %xmm0, %xmm7489; SSE-NEXT: unpcklps {{.*#+}} xmm7 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]490; SSE-NEXT: movaps %xmm7, %xmm8491; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm6[0]492; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm6[1]493; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]494; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3]495; SSE-NEXT: movaps %xmm0, %xmm3496; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]497; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]498; SSE-NEXT: movaps %xmm10, (%rsi)499; SSE-NEXT: movaps %xmm9, (%rdx)500; SSE-NEXT: movaps %xmm4, (%rcx)501; SSE-NEXT: movaps %xmm1, (%r8)502; SSE-NEXT: movaps %xmm8, (%r9)503; SSE-NEXT: movaps %xmm7, (%r11)504; SSE-NEXT: movaps %xmm3, (%r10)505; SSE-NEXT: movaps %xmm0, (%rax)506; SSE-NEXT: retq507;508; AVX-LABEL: load_i32_stride8_vf4:509; AVX: # %bb.0:510; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax511; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10512; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11513; AVX-NEXT: vmovaps (%rdi), %ymm0514; AVX-NEXT: vmovaps 32(%rdi), %ymm1515; AVX-NEXT: vmovaps 64(%rdi), %ymm2516; AVX-NEXT: vmovaps 96(%rdi), %ymm3517; AVX-NEXT: vmovaps 32(%rdi), %xmm4518; AVX-NEXT: vmovaps (%rdi), %xmm5519; AVX-NEXT: vunpcklps {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]520; AVX-NEXT: vmovaps 96(%rdi), %xmm7521; AVX-NEXT: vmovaps 64(%rdi), %xmm8522; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]523; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm6[0],xmm9[0]524; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm5[1,1,1,1]525; AVX-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm4[1],xmm10[2,3]526; AVX-NEXT: vblendps {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]527; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm7[2,2,2,2]528; AVX-NEXT: vblendps {{.*#+}} xmm10 = xmm8[0,1,2],xmm10[3]529; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]530; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm4[0,1],xmm10[2,3]531; AVX-NEXT: vunpckhps {{.*#+}} xmm7 = xmm8[2],xmm7[2],xmm8[3],xmm7[3]532; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm7[1]533; AVX-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]534; AVX-NEXT: vextractf128 $1, %ymm7, %xmm7535; AVX-NEXT: vunpcklps {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]536; AVX-NEXT: vextractf128 $1, %ymm8, %xmm8537; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,0]538; AVX-NEXT: vunpcklps {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]539; AVX-NEXT: vextractf128 $1, %ymm8, %xmm8540; AVX-NEXT: vshufps {{.*#+}} ymm10 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]541; AVX-NEXT: vextractf128 $1, %ymm10, %xmm10542; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm10[2,0],xmm8[2,3]543; AVX-NEXT: vunpckhpd {{.*#+}} ymm10 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]544; AVX-NEXT: vextractf128 $1, %ymm10, %xmm10545; AVX-NEXT: vunpckhps {{.*#+}} ymm11 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]546; AVX-NEXT: vextractf128 $1, %ymm11, %xmm11547; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,0]548; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]549; AVX-NEXT: vextractf128 $1, %ymm2, %xmm2550; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]551; AVX-NEXT: vextractf128 $1, %ymm0, %xmm0552; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[2,3]553; AVX-NEXT: vmovaps %xmm6, (%rsi)554; AVX-NEXT: vmovaps %xmm9, (%rdx)555; AVX-NEXT: vmovaps %xmm5, (%rcx)556; AVX-NEXT: vmovaps %xmm4, (%r8)557; AVX-NEXT: vmovaps %xmm7, (%r9)558; AVX-NEXT: vmovaps %xmm8, (%r11)559; AVX-NEXT: vmovaps %xmm10, (%r10)560; AVX-NEXT: vmovaps %xmm0, (%rax)561; AVX-NEXT: vzeroupper562; AVX-NEXT: retq563;564; AVX2-LABEL: load_i32_stride8_vf4:565; AVX2: # %bb.0:566; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax567; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10568; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11569; AVX2-NEXT: vmovaps 96(%rdi), %ymm0570; AVX2-NEXT: vmovaps 64(%rdi), %ymm1571; AVX2-NEXT: vmovaps 32(%rdi), %ymm2572; AVX2-NEXT: vmovaps (%rdi), %ymm3573; AVX2-NEXT: vmovaps 96(%rdi), %xmm4574; AVX2-NEXT: vbroadcastss %xmm4, %xmm5575; AVX2-NEXT: vmovaps (%rdi), %xmm6576; AVX2-NEXT: vmovaps 32(%rdi), %xmm7577; AVX2-NEXT: vmovaps 64(%rdi), %xmm8578; AVX2-NEXT: vbroadcastss %xmm8, %xmm9579; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]580; AVX2-NEXT: vunpcklps {{.*#+}} xmm9 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]581; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm9[0,1],xmm5[2,3]582; AVX2-NEXT: vunpcklps {{.*#+}} xmm9 = xmm8[0],xmm4[0],xmm8[1],xmm4[1]583; AVX2-NEXT: vshufps {{.*#+}} xmm10 = xmm6[1,1,1,1]584; AVX2-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm7[1],xmm10[2,3]585; AVX2-NEXT: vblendps {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]586; AVX2-NEXT: vshufps {{.*#+}} xmm10 = xmm4[2,2,2,2]587; AVX2-NEXT: vblendps {{.*#+}} xmm10 = xmm8[0,1,2],xmm10[3]588; AVX2-NEXT: vunpckhps {{.*#+}} xmm6 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]589; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm10[2,3]590; AVX2-NEXT: vunpckhps {{.*#+}} xmm4 = xmm8[2],xmm4[2],xmm8[3],xmm4[3]591; AVX2-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm4[1]592; AVX2-NEXT: vunpcklps {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]593; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm6594; AVX2-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]595; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm8[2,2,2,2]596; AVX2-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm10[2,3]597; AVX2-NEXT: vextractf128 $1, %ymm8, %xmm8598; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]599; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm2[1],ymm10[2,3,4],ymm2[5],ymm10[6,7]600; AVX2-NEXT: vextractf128 $1, %ymm10, %xmm10601; AVX2-NEXT: vblendps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]602; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]603; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm3604; AVX2-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]605; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]606; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]607; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm0608; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]609; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm2610; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]611; AVX2-NEXT: vmovaps %xmm5, (%rsi)612; AVX2-NEXT: vmovaps %xmm9, (%rdx)613; AVX2-NEXT: vmovaps %xmm7, (%rcx)614; AVX2-NEXT: vmovaps %xmm4, (%r8)615; AVX2-NEXT: vmovaps %xmm6, (%r9)616; AVX2-NEXT: vmovaps %xmm8, (%r11)617; AVX2-NEXT: vmovaps %xmm1, (%r10)618; AVX2-NEXT: vmovaps %xmm0, (%rax)619; AVX2-NEXT: vzeroupper620; AVX2-NEXT: retq621;622; AVX2-FP-LABEL: load_i32_stride8_vf4:623; AVX2-FP: # %bb.0:624; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax625; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10626; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11627; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm0628; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm1629; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm2630; AVX2-FP-NEXT: vmovaps (%rdi), %ymm3631; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm4632; AVX2-FP-NEXT: vbroadcastss %xmm4, %xmm5633; AVX2-FP-NEXT: vmovaps (%rdi), %xmm6634; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm7635; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm8636; AVX2-FP-NEXT: vbroadcastss %xmm8, %xmm9637; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]638; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]639; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm9[0,1],xmm5[2,3]640; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm8[0],xmm4[0],xmm8[1],xmm4[1]641; AVX2-FP-NEXT: vshufps {{.*#+}} xmm10 = xmm6[1,1,1,1]642; AVX2-FP-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm7[1],xmm10[2,3]643; AVX2-FP-NEXT: vblendps {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]644; AVX2-FP-NEXT: vshufps {{.*#+}} xmm10 = xmm4[2,2,2,2]645; AVX2-FP-NEXT: vblendps {{.*#+}} xmm10 = xmm8[0,1,2],xmm10[3]646; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm6 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]647; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm10[2,3]648; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm4 = xmm8[2],xmm4[2],xmm8[3],xmm4[3]649; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm4[1]650; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]651; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm6652; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]653; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm8[2,2,2,2]654; AVX2-FP-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm10[2,3]655; AVX2-FP-NEXT: vextractf128 $1, %ymm8, %xmm8656; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]657; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm2[1],ymm10[2,3,4],ymm2[5],ymm10[6,7]658; AVX2-FP-NEXT: vextractf128 $1, %ymm10, %xmm10659; AVX2-FP-NEXT: vblendps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]660; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]661; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm3662; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]663; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]664; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]665; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm0666; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]667; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm2668; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]669; AVX2-FP-NEXT: vmovaps %xmm5, (%rsi)670; AVX2-FP-NEXT: vmovaps %xmm9, (%rdx)671; AVX2-FP-NEXT: vmovaps %xmm7, (%rcx)672; AVX2-FP-NEXT: vmovaps %xmm4, (%r8)673; AVX2-FP-NEXT: vmovaps %xmm6, (%r9)674; AVX2-FP-NEXT: vmovaps %xmm8, (%r11)675; AVX2-FP-NEXT: vmovaps %xmm1, (%r10)676; AVX2-FP-NEXT: vmovaps %xmm0, (%rax)677; AVX2-FP-NEXT: vzeroupper678; AVX2-FP-NEXT: retq679;680; AVX2-FCP-LABEL: load_i32_stride8_vf4:681; AVX2-FCP: # %bb.0:682; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax683; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10684; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11685; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm0686; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm1687; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm2688; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm3689; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm4690; AVX2-FCP-NEXT: vbroadcastss %xmm4, %xmm5691; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm6692; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm7693; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm8694; AVX2-FCP-NEXT: vbroadcastss %xmm8, %xmm9695; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]696; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]697; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm9[0,1],xmm5[2,3]698; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm8[0],xmm4[0],xmm8[1],xmm4[1]699; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm10 = xmm6[1,1,1,1]700; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm7[1],xmm10[2,3]701; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]702; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm10 = xmm4[2,2,2,2]703; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm10 = xmm8[0,1,2],xmm10[3]704; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm6 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]705; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm10[2,3]706; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm4 = xmm8[2],xmm4[2],xmm8[3],xmm4[3]707; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm4[1]708; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]709; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm6710; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]711; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm10 = ymm8[2,2,2,2]712; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm10[2,3]713; AVX2-FCP-NEXT: vextractf128 $1, %ymm8, %xmm8714; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]715; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm2[1],ymm10[2,3,4],ymm2[5],ymm10[6,7]716; AVX2-FCP-NEXT: vextractf128 $1, %ymm10, %xmm10717; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]718; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]719; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm3720; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]721; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]722; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]723; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm0724; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]725; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm2726; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]727; AVX2-FCP-NEXT: vmovaps %xmm5, (%rsi)728; AVX2-FCP-NEXT: vmovaps %xmm9, (%rdx)729; AVX2-FCP-NEXT: vmovaps %xmm7, (%rcx)730; AVX2-FCP-NEXT: vmovaps %xmm4, (%r8)731; AVX2-FCP-NEXT: vmovaps %xmm6, (%r9)732; AVX2-FCP-NEXT: vmovaps %xmm8, (%r11)733; AVX2-FCP-NEXT: vmovaps %xmm1, (%r10)734; AVX2-FCP-NEXT: vmovaps %xmm0, (%rax)735; AVX2-FCP-NEXT: vzeroupper736; AVX2-FCP-NEXT: retq737;738; AVX512-LABEL: load_i32_stride8_vf4:739; AVX512: # %bb.0:740; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax741; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10742; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11743; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]744; AVX512-NEXT: vmovdqa64 (%rdi), %zmm1745; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm2746; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm0747; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]748; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm3749; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]750; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm4751; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]752; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm5753; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]754; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm6755; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]756; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm7757; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]758; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm8759; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]760; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm9761; AVX512-NEXT: vmovdqa %xmm0, (%rsi)762; AVX512-NEXT: vmovdqa %xmm3, (%rdx)763; AVX512-NEXT: vmovdqa %xmm4, (%rcx)764; AVX512-NEXT: vmovdqa %xmm5, (%r8)765; AVX512-NEXT: vmovdqa %xmm6, (%r9)766; AVX512-NEXT: vmovdqa %xmm7, (%r11)767; AVX512-NEXT: vmovdqa %xmm8, (%r10)768; AVX512-NEXT: vmovdqa %xmm9, (%rax)769; AVX512-NEXT: vzeroupper770; AVX512-NEXT: retq771;772; AVX512-FCP-LABEL: load_i32_stride8_vf4:773; AVX512-FCP: # %bb.0:774; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax775; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10776; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11777; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]778; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm1779; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2780; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm0781; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]782; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm3783; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]784; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm4785; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]786; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm5787; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]788; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm6789; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]790; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm7791; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]792; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm8793; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]794; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm9795; AVX512-FCP-NEXT: vmovdqa %xmm0, (%rsi)796; AVX512-FCP-NEXT: vmovdqa %xmm3, (%rdx)797; AVX512-FCP-NEXT: vmovdqa %xmm4, (%rcx)798; AVX512-FCP-NEXT: vmovdqa %xmm5, (%r8)799; AVX512-FCP-NEXT: vmovdqa %xmm6, (%r9)800; AVX512-FCP-NEXT: vmovdqa %xmm7, (%r11)801; AVX512-FCP-NEXT: vmovdqa %xmm8, (%r10)802; AVX512-FCP-NEXT: vmovdqa %xmm9, (%rax)803; AVX512-FCP-NEXT: vzeroupper804; AVX512-FCP-NEXT: retq805;806; AVX512DQ-LABEL: load_i32_stride8_vf4:807; AVX512DQ: # %bb.0:808; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax809; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10810; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11811; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]812; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm1813; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm2814; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm0815; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]816; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm3817; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]818; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm4819; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]820; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm5821; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]822; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm6823; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]824; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm7825; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]826; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm8827; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]828; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm9829; AVX512DQ-NEXT: vmovdqa %xmm0, (%rsi)830; AVX512DQ-NEXT: vmovdqa %xmm3, (%rdx)831; AVX512DQ-NEXT: vmovdqa %xmm4, (%rcx)832; AVX512DQ-NEXT: vmovdqa %xmm5, (%r8)833; AVX512DQ-NEXT: vmovdqa %xmm6, (%r9)834; AVX512DQ-NEXT: vmovdqa %xmm7, (%r11)835; AVX512DQ-NEXT: vmovdqa %xmm8, (%r10)836; AVX512DQ-NEXT: vmovdqa %xmm9, (%rax)837; AVX512DQ-NEXT: vzeroupper838; AVX512DQ-NEXT: retq839;840; AVX512DQ-FCP-LABEL: load_i32_stride8_vf4:841; AVX512DQ-FCP: # %bb.0:842; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax843; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10844; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11845; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]846; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm1847; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2848; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm0849; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]850; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm3851; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]852; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm4853; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]854; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm5855; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]856; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm6857; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]858; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm7859; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]860; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm8861; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]862; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm9863; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, (%rsi)864; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, (%rdx)865; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, (%rcx)866; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, (%r8)867; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, (%r9)868; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, (%r11)869; AVX512DQ-FCP-NEXT: vmovdqa %xmm8, (%r10)870; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, (%rax)871; AVX512DQ-FCP-NEXT: vzeroupper872; AVX512DQ-FCP-NEXT: retq873;874; AVX512BW-LABEL: load_i32_stride8_vf4:875; AVX512BW: # %bb.0:876; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax877; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10878; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11879; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]880; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm1881; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm2882; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm0883; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]884; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm3885; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]886; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm4887; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]888; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm5889; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]890; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm6891; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]892; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm7893; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]894; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm8895; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]896; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm9897; AVX512BW-NEXT: vmovdqa %xmm0, (%rsi)898; AVX512BW-NEXT: vmovdqa %xmm3, (%rdx)899; AVX512BW-NEXT: vmovdqa %xmm4, (%rcx)900; AVX512BW-NEXT: vmovdqa %xmm5, (%r8)901; AVX512BW-NEXT: vmovdqa %xmm6, (%r9)902; AVX512BW-NEXT: vmovdqa %xmm7, (%r11)903; AVX512BW-NEXT: vmovdqa %xmm8, (%r10)904; AVX512BW-NEXT: vmovdqa %xmm9, (%rax)905; AVX512BW-NEXT: vzeroupper906; AVX512BW-NEXT: retq907;908; AVX512BW-FCP-LABEL: load_i32_stride8_vf4:909; AVX512BW-FCP: # %bb.0:910; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax911; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10912; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11913; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]914; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm1915; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2916; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm0917; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]918; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm3919; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]920; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm4921; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]922; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm5923; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]924; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm6925; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]926; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm7927; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]928; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm8929; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]930; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm9931; AVX512BW-FCP-NEXT: vmovdqa %xmm0, (%rsi)932; AVX512BW-FCP-NEXT: vmovdqa %xmm3, (%rdx)933; AVX512BW-FCP-NEXT: vmovdqa %xmm4, (%rcx)934; AVX512BW-FCP-NEXT: vmovdqa %xmm5, (%r8)935; AVX512BW-FCP-NEXT: vmovdqa %xmm6, (%r9)936; AVX512BW-FCP-NEXT: vmovdqa %xmm7, (%r11)937; AVX512BW-FCP-NEXT: vmovdqa %xmm8, (%r10)938; AVX512BW-FCP-NEXT: vmovdqa %xmm9, (%rax)939; AVX512BW-FCP-NEXT: vzeroupper940; AVX512BW-FCP-NEXT: retq941;942; AVX512DQ-BW-LABEL: load_i32_stride8_vf4:943; AVX512DQ-BW: # %bb.0:944; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax945; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10946; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11947; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]948; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm1949; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm2950; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm0951; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]952; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm3953; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]954; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm4955; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]956; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm5957; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]958; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm6959; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]960; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm7961; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]962; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm8963; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]964; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm9965; AVX512DQ-BW-NEXT: vmovdqa %xmm0, (%rsi)966; AVX512DQ-BW-NEXT: vmovdqa %xmm3, (%rdx)967; AVX512DQ-BW-NEXT: vmovdqa %xmm4, (%rcx)968; AVX512DQ-BW-NEXT: vmovdqa %xmm5, (%r8)969; AVX512DQ-BW-NEXT: vmovdqa %xmm6, (%r9)970; AVX512DQ-BW-NEXT: vmovdqa %xmm7, (%r11)971; AVX512DQ-BW-NEXT: vmovdqa %xmm8, (%r10)972; AVX512DQ-BW-NEXT: vmovdqa %xmm9, (%rax)973; AVX512DQ-BW-NEXT: vzeroupper974; AVX512DQ-BW-NEXT: retq975;976; AVX512DQ-BW-FCP-LABEL: load_i32_stride8_vf4:977; AVX512DQ-BW-FCP: # %bb.0:978; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax979; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10980; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11981; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,8,16,24]982; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm1983; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2984; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm0985; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,9,17,25]986; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm3987; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,10,18,26]988; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm4989; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,11,19,27]990; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm5991; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,12,20,28]992; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm6993; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,13,21,29]994; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm7995; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [6,14,22,30]996; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm8997; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [7,15,23,31]998; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm1, %zmm9999; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, (%rsi)1000; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm3, (%rdx)1001; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm4, (%rcx)1002; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm5, (%r8)1003; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm6, (%r9)1004; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm7, (%r11)1005; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm8, (%r10)1006; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm9, (%rax)1007; AVX512DQ-BW-FCP-NEXT: vzeroupper1008; AVX512DQ-BW-FCP-NEXT: retq1009 %wide.vec = load <32 x i32>, ptr %in.vec, align 641010 %strided.vec0 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 0, i32 8, i32 16, i32 24>1011 %strided.vec1 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 1, i32 9, i32 17, i32 25>1012 %strided.vec2 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 2, i32 10, i32 18, i32 26>1013 %strided.vec3 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 3, i32 11, i32 19, i32 27>1014 %strided.vec4 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 4, i32 12, i32 20, i32 28>1015 %strided.vec5 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 5, i32 13, i32 21, i32 29>1016 %strided.vec6 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 6, i32 14, i32 22, i32 30>1017 %strided.vec7 = shufflevector <32 x i32> %wide.vec, <32 x i32> poison, <4 x i32> <i32 7, i32 15, i32 23, i32 31>1018 store <4 x i32> %strided.vec0, ptr %out.vec0, align 641019 store <4 x i32> %strided.vec1, ptr %out.vec1, align 641020 store <4 x i32> %strided.vec2, ptr %out.vec2, align 641021 store <4 x i32> %strided.vec3, ptr %out.vec3, align 641022 store <4 x i32> %strided.vec4, ptr %out.vec4, align 641023 store <4 x i32> %strided.vec5, ptr %out.vec5, align 641024 store <4 x i32> %strided.vec6, ptr %out.vec6, align 641025 store <4 x i32> %strided.vec7, ptr %out.vec7, align 641026 ret void1027}1028 1029define void @load_i32_stride8_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {1030; SSE-LABEL: load_i32_stride8_vf8:1031; SSE: # %bb.0:1032; SSE-NEXT: movaps 112(%rdi), %xmm151033; SSE-NEXT: movaps 176(%rdi), %xmm41034; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1035; SSE-NEXT: movaps 144(%rdi), %xmm91036; SSE-NEXT: movaps (%rdi), %xmm101037; SSE-NEXT: movaps 32(%rdi), %xmm11038; SSE-NEXT: movaps 96(%rdi), %xmm131039; SSE-NEXT: movaps 64(%rdi), %xmm111040; SSE-NEXT: movaps 160(%rdi), %xmm21041; SSE-NEXT: movaps 128(%rdi), %xmm61042; SSE-NEXT: movaps 224(%rdi), %xmm121043; SSE-NEXT: movaps 192(%rdi), %xmm01044; SSE-NEXT: movaps %xmm0, %xmm81045; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1]1046; SSE-NEXT: movaps %xmm6, %xmm51047; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]1048; SSE-NEXT: movaps %xmm5, %xmm71049; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]1050; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1051; SSE-NEXT: movaps %xmm11, %xmm141052; SSE-NEXT: unpcklps {{.*#+}} xmm14 = xmm14[0],xmm13[0],xmm14[1],xmm13[1]1053; SSE-NEXT: movaps %xmm10, %xmm71054; SSE-NEXT: unpcklps {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]1055; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]1056; SSE-NEXT: movaps %xmm7, %xmm81057; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm14[0]1058; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1059; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm14[1]1060; SSE-NEXT: movaps 240(%rdi), %xmm141061; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1062; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm12[2],xmm0[3],xmm12[3]1063; SSE-NEXT: movaps 208(%rdi), %xmm121064; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm2[2],xmm6[3],xmm2[3]1065; SSE-NEXT: movaps %xmm6, %xmm21066; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]1067; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1068; SSE-NEXT: unpckhps {{.*#+}} xmm11 = xmm11[2],xmm13[2],xmm11[3],xmm13[3]1069; SSE-NEXT: unpckhps {{.*#+}} xmm10 = xmm10[2],xmm1[2],xmm10[3],xmm1[3]1070; SSE-NEXT: movaps %xmm10, %xmm81071; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm11[0]1072; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]1073; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm11[1]1074; SSE-NEXT: movaps %xmm12, %xmm01075; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1]1076; SSE-NEXT: movaps %xmm9, %xmm111077; SSE-NEXT: unpcklps {{.*#+}} xmm11 = xmm11[0],xmm4[0],xmm11[1],xmm4[1]1078; SSE-NEXT: movaps %xmm11, %xmm131079; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]1080; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]1081; SSE-NEXT: movaps 80(%rdi), %xmm21082; SSE-NEXT: movaps %xmm2, %xmm11083; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1]1084; SSE-NEXT: movaps 16(%rdi), %xmm01085; SSE-NEXT: movaps 48(%rdi), %xmm31086; SSE-NEXT: movaps %xmm0, %xmm141087; SSE-NEXT: unpcklps {{.*#+}} xmm14 = xmm14[0],xmm3[0],xmm14[1],xmm3[1]1088; SSE-NEXT: movaps %xmm14, %xmm41089; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm1[0]1090; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm1[1]1091; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload1092; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]1093; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload1094; SSE-NEXT: # xmm9 = xmm9[2],mem[2],xmm9[3],mem[3]1095; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm15[2],xmm2[3],xmm15[3]1096; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]1097; SSE-NEXT: movaps %xmm9, %xmm11098; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm12[0]1099; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm12[1]1100; SSE-NEXT: movaps %xmm0, %xmm31101; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]1102; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]1103; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1104; SSE-NEXT: movaps %xmm2, (%rsi)1105; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1106; SSE-NEXT: movaps %xmm2, 16(%rsi)1107; SSE-NEXT: movaps %xmm7, (%rdx)1108; SSE-NEXT: movaps %xmm5, 16(%rdx)1109; SSE-NEXT: movaps %xmm8, (%rcx)1110; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1111; SSE-NEXT: movaps %xmm2, 16(%rcx)1112; SSE-NEXT: movaps %xmm10, (%r8)1113; SSE-NEXT: movaps %xmm6, 16(%r8)1114; SSE-NEXT: movaps %xmm4, (%r9)1115; SSE-NEXT: movaps %xmm13, 16(%r9)1116; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax1117; SSE-NEXT: movaps %xmm14, (%rax)1118; SSE-NEXT: movaps %xmm11, 16(%rax)1119; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax1120; SSE-NEXT: movaps %xmm3, (%rax)1121; SSE-NEXT: movaps %xmm1, 16(%rax)1122; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax1123; SSE-NEXT: movaps %xmm9, 16(%rax)1124; SSE-NEXT: movaps %xmm0, (%rax)1125; SSE-NEXT: retq1126;1127; AVX-LABEL: load_i32_stride8_vf8:1128; AVX: # %bb.0:1129; AVX-NEXT: vmovaps (%rdi), %ymm01130; AVX-NEXT: vmovaps 32(%rdi), %ymm11131; AVX-NEXT: vmovaps 64(%rdi), %ymm21132; AVX-NEXT: vmovaps 96(%rdi), %ymm31133; AVX-NEXT: vmovaps 32(%rdi), %xmm81134; AVX-NEXT: vmovaps (%rdi), %xmm111135; AVX-NEXT: vunpcklps {{.*#+}} xmm4 = xmm11[0],xmm8[0],xmm11[1],xmm8[1]1136; AVX-NEXT: vmovaps 96(%rdi), %xmm91137; AVX-NEXT: vmovaps 64(%rdi), %xmm101138; AVX-NEXT: vmovaps 160(%rdi), %xmm141139; AVX-NEXT: vmovaps 128(%rdi), %xmm151140; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]1141; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm51142; AVX-NEXT: vmovaps 224(%rdi), %xmm121143; AVX-NEXT: vmovaps 192(%rdi), %xmm131144; AVX-NEXT: vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1145; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm6[0,1,0,1]1146; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm71147; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]1148; AVX-NEXT: vunpcklps {{.*#+}} xmm7 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1149; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm7[0]1150; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]1151; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1152; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm11[1,1,1,1]1153; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0],xmm8[1],xmm5[2,3]1154; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]1155; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm61156; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm15[1,1,1,1]1157; AVX-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm14[1],xmm7[2,3]1158; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm71159; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm6[6,7]1160; AVX-NEXT: vmovaps 160(%rdi), %ymm61161; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]1162; AVX-NEXT: vmovaps 128(%rdi), %ymm71163; AVX-NEXT: vunpckhps {{.*#+}} xmm8 = xmm11[2],xmm8[2],xmm11[3],xmm8[3]1164; AVX-NEXT: vunpckhps {{.*#+}} xmm15 = xmm15[2],xmm14[2],xmm15[3],xmm14[3]1165; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm111166; AVX-NEXT: vshufps {{.*#+}} xmm14 = xmm12[2,2,2,2]1167; AVX-NEXT: vblendps {{.*#+}} xmm14 = xmm13[0,1,2],xmm14[3]1168; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm141169; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],ymm14[6,7]1170; AVX-NEXT: vshufps {{.*#+}} xmm14 = xmm9[2,2,2,2]1171; AVX-NEXT: vblendps {{.*#+}} xmm14 = xmm10[0,1,2],xmm14[3]1172; AVX-NEXT: vblendps {{.*#+}} xmm14 = xmm8[0,1],xmm14[2,3]1173; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm11[4,5,6,7]1174; AVX-NEXT: vmovaps 192(%rdi), %ymm111175; AVX-NEXT: vunpckhps {{.*#+}} xmm9 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1176; AVX-NEXT: vmovaps 224(%rdi), %ymm101177; AVX-NEXT: vunpckhpd {{.*#+}} xmm8 = xmm8[1],xmm9[1]1178; AVX-NEXT: vunpckhps {{.*#+}} xmm9 = xmm13[2],xmm12[2],xmm13[3],xmm12[3]1179; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm91180; AVX-NEXT: vshufps {{.*#+}} xmm12 = xmm15[2,3,2,3]1181; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm121182; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1,2,3,4,5],ymm9[6,7]1183; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm8[0,1,2,3],ymm9[4,5,6,7]1184; AVX-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]1185; AVX-NEXT: vunpcklps {{.*#+}} ymm12 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]1186; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm12[0,1],ymm8[2,0],ymm12[4,5],ymm8[6,4]1187; AVX-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]1188; AVX-NEXT: vextractf128 $1, %ymm12, %xmm121189; AVX-NEXT: vunpcklps {{.*#+}} ymm13 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]1190; AVX-NEXT: vextractf128 $1, %ymm13, %xmm131191; AVX-NEXT: vshufps {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,0]1192; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm8[4,5,6,7]1193; AVX-NEXT: vunpcklps {{.*#+}} ymm8 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[4],ymm10[4],ymm11[5],ymm10[5]1194; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm6[1,0],ymm7[1,0],ymm6[5,4],ymm7[5,4]1195; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm13[2,0],ymm8[2,3],ymm13[6,4],ymm8[6,7]1196; AVX-NEXT: vunpcklps {{.*#+}} ymm13 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]1197; AVX-NEXT: vextractf128 $1, %ymm13, %xmm131198; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]1199; AVX-NEXT: vextractf128 $1, %ymm15, %xmm151200; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm15[2,0],xmm13[2,3]1201; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]1202; AVX-NEXT: vunpckhpd {{.*#+}} ymm13 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]1203; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[6],ymm6[6],ymm7[7],ymm6[7]1204; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm15[0,1],ymm13[2,0],ymm15[4,5],ymm13[6,4]1205; AVX-NEXT: vunpckhpd {{.*#+}} ymm15 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]1206; AVX-NEXT: vextractf128 $1, %ymm15, %xmm151207; AVX-NEXT: vunpckhps {{.*#+}} ymm4 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]1208; AVX-NEXT: vextractf128 $1, %ymm4, %xmm41209; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1],xmm15[2,0]1210; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm13[4,5,6,7]1211; AVX-NEXT: vunpckhps {{.*#+}} ymm10 = ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[6],ymm10[6],ymm11[7],ymm10[7]1212; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm6[3,0],ymm7[3,0],ymm6[7,4],ymm7[7,4]1213; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm6[2,0],ymm10[2,3],ymm6[6,4],ymm10[6,7]1214; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]1215; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]1216; AVX-NEXT: vextractf128 $1, %ymm2, %xmm11217; AVX-NEXT: vextractf128 $1, %ymm0, %xmm01218; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]1219; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]1220; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1221; AVX-NEXT: vmovaps %ymm1, (%rsi)1222; AVX-NEXT: vmovaps %ymm5, (%rdx)1223; AVX-NEXT: vmovaps %ymm14, (%rcx)1224; AVX-NEXT: vmovaps %ymm9, (%r8)1225; AVX-NEXT: vmovaps %ymm12, (%r9)1226; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax1227; AVX-NEXT: vmovaps %ymm8, (%rax)1228; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax1229; AVX-NEXT: vmovaps %ymm4, (%rax)1230; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax1231; AVX-NEXT: vmovaps %ymm0, (%rax)1232; AVX-NEXT: vzeroupper1233; AVX-NEXT: retq1234;1235; AVX2-LABEL: load_i32_stride8_vf8:1236; AVX2: # %bb.0:1237; AVX2-NEXT: vmovaps 96(%rdi), %ymm01238; AVX2-NEXT: vmovaps 64(%rdi), %ymm11239; AVX2-NEXT: vmovaps 32(%rdi), %ymm21240; AVX2-NEXT: vmovaps (%rdi), %ymm31241; AVX2-NEXT: vmovaps 160(%rdi), %xmm71242; AVX2-NEXT: vmovaps 128(%rdi), %xmm111243; AVX2-NEXT: vunpcklps {{.*#+}} xmm4 = xmm11[0],xmm7[0],xmm11[1],xmm7[1]1244; AVX2-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm41245; AVX2-NEXT: vmovaps 224(%rdi), %xmm81246; AVX2-NEXT: vbroadcastss %xmm8, %xmm51247; AVX2-NEXT: vmovaps 192(%rdi), %xmm101248; AVX2-NEXT: vbroadcastss %xmm10, %xmm61249; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]1250; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm51251; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]1252; AVX2-NEXT: vmovaps 96(%rdi), %xmm91253; AVX2-NEXT: vbroadcastss %xmm9, %xmm51254; AVX2-NEXT: vmovaps (%rdi), %xmm131255; AVX2-NEXT: vmovaps 32(%rdi), %xmm141256; AVX2-NEXT: vmovaps 64(%rdi), %xmm121257; AVX2-NEXT: vbroadcastss %xmm12, %xmm61258; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]1259; AVX2-NEXT: vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm14[0],xmm13[1],xmm14[1]1260; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]1261; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1262; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm12[0],xmm9[0],xmm12[1],xmm9[1]1263; AVX2-NEXT: vshufps {{.*#+}} xmm6 = xmm13[1,1,1,1]1264; AVX2-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0],xmm14[1],xmm6[2,3]1265; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]1266; AVX2-NEXT: vunpcklps {{.*#+}} xmm6 = xmm10[0],xmm8[0],xmm10[1],xmm8[1]1267; AVX2-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm61268; AVX2-NEXT: vshufps {{.*#+}} xmm15 = xmm11[1,1,1,1]1269; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0],xmm7[1],xmm15[2,3]1270; AVX2-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm151271; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],ymm6[6,7]1272; AVX2-NEXT: vmovaps 224(%rdi), %ymm61273; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm15[4,5,6,7]1274; AVX2-NEXT: vunpckhps {{.*#+}} xmm15 = xmm11[2],xmm7[2],xmm11[3],xmm7[3]1275; AVX2-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm71276; AVX2-NEXT: vshufps {{.*#+}} xmm11 = xmm8[2,2,2,2]1277; AVX2-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]1278; AVX2-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm111279; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm11[6,7]1280; AVX2-NEXT: vmovaps 192(%rdi), %ymm111281; AVX2-NEXT: vunpckhps {{.*#+}} xmm14 = xmm13[2],xmm14[2],xmm13[3],xmm14[3]1282; AVX2-NEXT: vshufps {{.*#+}} xmm13 = xmm9[2,2,2,2]1283; AVX2-NEXT: vblendps {{.*#+}} xmm13 = xmm12[0,1,2],xmm13[3]1284; AVX2-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]1285; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm13[0,1,2,3],ymm7[4,5,6,7]1286; AVX2-NEXT: vmovaps 160(%rdi), %ymm131287; AVX2-NEXT: vunpckhps {{.*#+}} xmm8 = xmm10[2],xmm8[2],xmm10[3],xmm8[3]1288; AVX2-NEXT: vmovaps 128(%rdi), %ymm101289; AVX2-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm81290; AVX2-NEXT: vshufps {{.*#+}} xmm15 = xmm15[2,3,2,3]1291; AVX2-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm151292; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0,1,2,3,4,5],ymm8[6,7]1293; AVX2-NEXT: vunpckhps {{.*#+}} xmm9 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]1294; AVX2-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm14[1],xmm9[1]1295; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1296; AVX2-NEXT: vunpcklps {{.*#+}} ymm9 = ymm10[0],ymm13[0],ymm10[1],ymm13[1],ymm10[4],ymm13[4],ymm10[5],ymm13[5]1297; AVX2-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]1298; AVX2-NEXT: vextractf128 $1, %ymm12, %xmm121299; AVX2-NEXT: vunpcklps {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]1300; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]1301; AVX2-NEXT: vblendps {{.*#+}} xmm12 = xmm12[0,1],xmm15[2,3]1302; AVX2-NEXT: vunpcklps {{.*#+}} ymm15 = ymm11[0],ymm6[0],ymm11[1],ymm6[1],ymm11[4],ymm6[4],ymm11[5],ymm6[5]1303; AVX2-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm15[0],ymm9[2],ymm15[2]1304; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1,2,3],ymm9[4,5,6,7]1305; AVX2-NEXT: vbroadcastss 148(%rdi), %ymm121306; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm13[5],ymm12[6,7]1307; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm15[6,7]1308; AVX2-NEXT: vextractf128 $1, %ymm14, %xmm141309; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm3[1,1,1,1,5,5,5,5]1310; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm2[1],ymm15[2,3,4],ymm2[5],ymm15[6,7]1311; AVX2-NEXT: vextractf128 $1, %ymm15, %xmm151312; AVX2-NEXT: vblendps {{.*#+}} xmm14 = xmm15[0,1],xmm14[2,3]1313; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm14[0,1,2,3],ymm12[4,5,6,7]1314; AVX2-NEXT: vbroadcastss 248(%rdi), %ymm141315; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm14[7]1316; AVX2-NEXT: vunpckhps {{.*#+}} ymm10 = ymm10[2],ymm13[2],ymm10[3],ymm13[3],ymm10[6],ymm13[6],ymm10[7],ymm13[7]1317; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]1318; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm31319; AVX2-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]1320; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]1321; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]1322; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4,5],ymm11[6,7]1323; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]1324; AVX2-NEXT: vbroadcastss 220(%rdi), %ymm31325; AVX2-NEXT: vunpckhps {{.*#+}} ymm3 = ymm3[2],ymm6[2],ymm3[3],ymm6[3],ymm3[6],ymm6[6],ymm3[7],ymm6[7]1326; AVX2-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm3[1],ymm10[3],ymm3[3]1327; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm01328; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]1329; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm21330; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]1331; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]1332; AVX2-NEXT: vmovaps %ymm4, (%rsi)1333; AVX2-NEXT: vmovaps %ymm5, (%rdx)1334; AVX2-NEXT: vmovaps %ymm7, (%rcx)1335; AVX2-NEXT: vmovaps %ymm8, (%r8)1336; AVX2-NEXT: vmovaps %ymm9, (%r9)1337; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax1338; AVX2-NEXT: vmovaps %ymm12, (%rax)1339; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax1340; AVX2-NEXT: vmovaps %ymm1, (%rax)1341; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax1342; AVX2-NEXT: vmovaps %ymm0, (%rax)1343; AVX2-NEXT: vzeroupper1344; AVX2-NEXT: retq1345;1346; AVX2-FP-LABEL: load_i32_stride8_vf8:1347; AVX2-FP: # %bb.0:1348; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm01349; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm11350; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm21351; AVX2-FP-NEXT: vmovaps (%rdi), %ymm31352; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm71353; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm111354; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm4 = xmm11[0],xmm7[0],xmm11[1],xmm7[1]1355; AVX2-FP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm41356; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm81357; AVX2-FP-NEXT: vbroadcastss %xmm8, %xmm51358; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm101359; AVX2-FP-NEXT: vbroadcastss %xmm10, %xmm61360; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]1361; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm51362; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]1363; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm91364; AVX2-FP-NEXT: vbroadcastss %xmm9, %xmm51365; AVX2-FP-NEXT: vmovaps (%rdi), %xmm131366; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm141367; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm121368; AVX2-FP-NEXT: vbroadcastss %xmm12, %xmm61369; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]1370; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm14[0],xmm13[1],xmm14[1]1371; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]1372; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1373; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm12[0],xmm9[0],xmm12[1],xmm9[1]1374; AVX2-FP-NEXT: vshufps {{.*#+}} xmm6 = xmm13[1,1,1,1]1375; AVX2-FP-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0],xmm14[1],xmm6[2,3]1376; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]1377; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm10[0],xmm8[0],xmm10[1],xmm8[1]1378; AVX2-FP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm61379; AVX2-FP-NEXT: vshufps {{.*#+}} xmm15 = xmm11[1,1,1,1]1380; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0],xmm7[1],xmm15[2,3]1381; AVX2-FP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm151382; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],ymm6[6,7]1383; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm61384; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm15[4,5,6,7]1385; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm15 = xmm11[2],xmm7[2],xmm11[3],xmm7[3]1386; AVX2-FP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm71387; AVX2-FP-NEXT: vshufps {{.*#+}} xmm11 = xmm8[2,2,2,2]1388; AVX2-FP-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]1389; AVX2-FP-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm111390; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm11[6,7]1391; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm111392; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm14 = xmm13[2],xmm14[2],xmm13[3],xmm14[3]1393; AVX2-FP-NEXT: vshufps {{.*#+}} xmm13 = xmm9[2,2,2,2]1394; AVX2-FP-NEXT: vblendps {{.*#+}} xmm13 = xmm12[0,1,2],xmm13[3]1395; AVX2-FP-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]1396; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm13[0,1,2,3],ymm7[4,5,6,7]1397; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm131398; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm8 = xmm10[2],xmm8[2],xmm10[3],xmm8[3]1399; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm101400; AVX2-FP-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm81401; AVX2-FP-NEXT: vshufps {{.*#+}} xmm15 = xmm15[2,3,2,3]1402; AVX2-FP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm151403; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0,1,2,3,4,5],ymm8[6,7]1404; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm9 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]1405; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm14[1],xmm9[1]1406; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1407; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm9 = ymm10[0],ymm13[0],ymm10[1],ymm13[1],ymm10[4],ymm13[4],ymm10[5],ymm13[5]1408; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]1409; AVX2-FP-NEXT: vextractf128 $1, %ymm12, %xmm121410; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]1411; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]1412; AVX2-FP-NEXT: vblendps {{.*#+}} xmm12 = xmm12[0,1],xmm15[2,3]1413; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm15 = ymm11[0],ymm6[0],ymm11[1],ymm6[1],ymm11[4],ymm6[4],ymm11[5],ymm6[5]1414; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm15[0],ymm9[2],ymm15[2]1415; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1,2,3],ymm9[4,5,6,7]1416; AVX2-FP-NEXT: vbroadcastss 148(%rdi), %ymm121417; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm13[5],ymm12[6,7]1418; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm15[6,7]1419; AVX2-FP-NEXT: vextractf128 $1, %ymm14, %xmm141420; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm3[1,1,1,1,5,5,5,5]1421; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm2[1],ymm15[2,3,4],ymm2[5],ymm15[6,7]1422; AVX2-FP-NEXT: vextractf128 $1, %ymm15, %xmm151423; AVX2-FP-NEXT: vblendps {{.*#+}} xmm14 = xmm15[0,1],xmm14[2,3]1424; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm14[0,1,2,3],ymm12[4,5,6,7]1425; AVX2-FP-NEXT: vbroadcastss 248(%rdi), %ymm141426; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm14[7]1427; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm10 = ymm10[2],ymm13[2],ymm10[3],ymm13[3],ymm10[6],ymm13[6],ymm10[7],ymm13[7]1428; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]1429; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm31430; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]1431; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]1432; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]1433; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4,5],ymm11[6,7]1434; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]1435; AVX2-FP-NEXT: vbroadcastss 220(%rdi), %ymm31436; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm3[2],ymm6[2],ymm3[3],ymm6[3],ymm3[6],ymm6[6],ymm3[7],ymm6[7]1437; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm3[1],ymm10[3],ymm3[3]1438; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm01439; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]1440; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm21441; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]1442; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]1443; AVX2-FP-NEXT: vmovaps %ymm4, (%rsi)1444; AVX2-FP-NEXT: vmovaps %ymm5, (%rdx)1445; AVX2-FP-NEXT: vmovaps %ymm7, (%rcx)1446; AVX2-FP-NEXT: vmovaps %ymm8, (%r8)1447; AVX2-FP-NEXT: vmovaps %ymm9, (%r9)1448; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax1449; AVX2-FP-NEXT: vmovaps %ymm12, (%rax)1450; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax1451; AVX2-FP-NEXT: vmovaps %ymm1, (%rax)1452; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax1453; AVX2-FP-NEXT: vmovaps %ymm0, (%rax)1454; AVX2-FP-NEXT: vzeroupper1455; AVX2-FP-NEXT: retq1456;1457; AVX2-FCP-LABEL: load_i32_stride8_vf8:1458; AVX2-FCP: # %bb.0:1459; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm01460; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm11461; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm21462; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm31463; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm71464; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm111465; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm4 = xmm11[0],xmm7[0],xmm11[1],xmm7[1]1466; AVX2-FCP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm41467; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm81468; AVX2-FCP-NEXT: vbroadcastss %xmm8, %xmm51469; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm101470; AVX2-FCP-NEXT: vbroadcastss %xmm10, %xmm61471; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]1472; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm51473; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]1474; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm91475; AVX2-FCP-NEXT: vbroadcastss %xmm9, %xmm51476; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm131477; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm141478; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm121479; AVX2-FCP-NEXT: vbroadcastss %xmm12, %xmm61480; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]1481; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm14[0],xmm13[1],xmm14[1]1482; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]1483; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1484; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm12[0],xmm9[0],xmm12[1],xmm9[1]1485; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm6 = xmm13[1,1,1,1]1486; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0],xmm14[1],xmm6[2,3]1487; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]1488; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm10[0],xmm8[0],xmm10[1],xmm8[1]1489; AVX2-FCP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm61490; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm15 = xmm11[1,1,1,1]1491; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0],xmm7[1],xmm15[2,3]1492; AVX2-FCP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm151493; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],ymm6[6,7]1494; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm61495; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm15[4,5,6,7]1496; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm15 = xmm11[2],xmm7[2],xmm11[3],xmm7[3]1497; AVX2-FCP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm71498; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm11 = xmm8[2,2,2,2]1499; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]1500; AVX2-FCP-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm111501; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm11[6,7]1502; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm111503; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm14 = xmm13[2],xmm14[2],xmm13[3],xmm14[3]1504; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm13 = xmm9[2,2,2,2]1505; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm13 = xmm12[0,1,2],xmm13[3]1506; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]1507; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm13[0,1,2,3],ymm7[4,5,6,7]1508; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm131509; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm8 = xmm10[2],xmm8[2],xmm10[3],xmm8[3]1510; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm101511; AVX2-FCP-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm81512; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm15 = xmm15[2,3,2,3]1513; AVX2-FCP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm151514; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0,1,2,3,4,5],ymm8[6,7]1515; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm9 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]1516; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm14[1],xmm9[1]1517; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1518; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm9 = ymm10[0],ymm13[0],ymm10[1],ymm13[1],ymm10[4],ymm13[4],ymm10[5],ymm13[5]1519; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]1520; AVX2-FCP-NEXT: vextractf128 $1, %ymm12, %xmm121521; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]1522; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]1523; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm12 = xmm12[0,1],xmm15[2,3]1524; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm15 = ymm11[0],ymm6[0],ymm11[1],ymm6[1],ymm11[4],ymm6[4],ymm11[5],ymm6[5]1525; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm15[0],ymm9[2],ymm15[2]1526; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1,2,3],ymm9[4,5,6,7]1527; AVX2-FCP-NEXT: vbroadcastss 148(%rdi), %ymm121528; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm13[5],ymm12[6,7]1529; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm15[6,7]1530; AVX2-FCP-NEXT: vextractf128 $1, %ymm14, %xmm141531; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm15 = ymm3[1,1,1,1,5,5,5,5]1532; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm2[1],ymm15[2,3,4],ymm2[5],ymm15[6,7]1533; AVX2-FCP-NEXT: vextractf128 $1, %ymm15, %xmm151534; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm14 = xmm15[0,1],xmm14[2,3]1535; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm14[0,1,2,3],ymm12[4,5,6,7]1536; AVX2-FCP-NEXT: vbroadcastss 248(%rdi), %ymm141537; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm14[7]1538; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm10 = ymm10[2],ymm13[2],ymm10[3],ymm13[3],ymm10[6],ymm13[6],ymm10[7],ymm13[7]1539; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]1540; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm31541; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]1542; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]1543; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]1544; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4,5],ymm11[6,7]1545; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]1546; AVX2-FCP-NEXT: vbroadcastss 220(%rdi), %ymm31547; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm3[2],ymm6[2],ymm3[3],ymm6[3],ymm3[6],ymm6[6],ymm3[7],ymm6[7]1548; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm10[1],ymm3[1],ymm10[3],ymm3[3]1549; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm01550; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]1551; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm21552; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]1553; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]1554; AVX2-FCP-NEXT: vmovaps %ymm4, (%rsi)1555; AVX2-FCP-NEXT: vmovaps %ymm5, (%rdx)1556; AVX2-FCP-NEXT: vmovaps %ymm7, (%rcx)1557; AVX2-FCP-NEXT: vmovaps %ymm8, (%r8)1558; AVX2-FCP-NEXT: vmovaps %ymm9, (%r9)1559; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1560; AVX2-FCP-NEXT: vmovaps %ymm12, (%rax)1561; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1562; AVX2-FCP-NEXT: vmovaps %ymm1, (%rax)1563; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1564; AVX2-FCP-NEXT: vmovaps %ymm0, (%rax)1565; AVX2-FCP-NEXT: vzeroupper1566; AVX2-FCP-NEXT: retq1567;1568; AVX512-LABEL: load_i32_stride8_vf8:1569; AVX512: # %bb.0:1570; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax1571; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r101572; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r111573; AVX512-NEXT: vmovdqa64 (%rdi), %zmm01574; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm11575; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm21576; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm31577; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1578; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm41579; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1580; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm51581; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1582; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1583; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm51584; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1585; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm61586; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1587; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1588; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm61589; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1590; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm71591; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1592; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1593; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm71594; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1595; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm81596; AVX512-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1597; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1598; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm81599; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1600; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm91601; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1602; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1603; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm91604; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1605; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm101606; AVX512-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1607; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1608; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm101609; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1610; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm111611; AVX512-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1612; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1613; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm111614; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1615; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm21616; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1617; AVX512-NEXT: vmovdqa %ymm4, (%rsi)1618; AVX512-NEXT: vmovdqa %ymm5, (%rdx)1619; AVX512-NEXT: vmovdqa %ymm6, (%rcx)1620; AVX512-NEXT: vmovdqa %ymm7, (%r8)1621; AVX512-NEXT: vmovdqa %ymm8, (%r9)1622; AVX512-NEXT: vmovdqa %ymm9, (%r11)1623; AVX512-NEXT: vmovdqa %ymm10, (%r10)1624; AVX512-NEXT: vmovdqa %ymm0, (%rax)1625; AVX512-NEXT: vzeroupper1626; AVX512-NEXT: retq1627;1628; AVX512-FCP-LABEL: load_i32_stride8_vf8:1629; AVX512-FCP: # %bb.0:1630; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1631; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r101632; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r111633; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm01634; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11635; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21636; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31637; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1638; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm41639; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1640; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm51641; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1642; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1643; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm51644; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1645; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm61646; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1647; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1648; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm61649; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1650; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm71651; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1652; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1653; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm71654; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1655; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm81656; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1657; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1658; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm81659; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1660; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm91661; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1662; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1663; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm91664; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1665; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm101666; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1667; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1668; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm101669; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1670; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm111671; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1672; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1673; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm111674; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1675; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm21676; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1677; AVX512-FCP-NEXT: vmovdqa %ymm4, (%rsi)1678; AVX512-FCP-NEXT: vmovdqa %ymm5, (%rdx)1679; AVX512-FCP-NEXT: vmovdqa %ymm6, (%rcx)1680; AVX512-FCP-NEXT: vmovdqa %ymm7, (%r8)1681; AVX512-FCP-NEXT: vmovdqa %ymm8, (%r9)1682; AVX512-FCP-NEXT: vmovdqa %ymm9, (%r11)1683; AVX512-FCP-NEXT: vmovdqa %ymm10, (%r10)1684; AVX512-FCP-NEXT: vmovdqa %ymm0, (%rax)1685; AVX512-FCP-NEXT: vzeroupper1686; AVX512-FCP-NEXT: retq1687;1688; AVX512DQ-LABEL: load_i32_stride8_vf8:1689; AVX512DQ: # %bb.0:1690; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax1691; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r101692; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r111693; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm01694; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm11695; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm21696; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm31697; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1698; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm41699; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1700; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm51701; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1702; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1703; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm51704; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1705; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm61706; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1707; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1708; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm61709; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1710; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm71711; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1712; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1713; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm71714; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1715; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm81716; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1717; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1718; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm81719; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1720; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm91721; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1722; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1723; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm91724; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1725; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm101726; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1727; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1728; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm101729; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1730; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm111731; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1732; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1733; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm111734; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1735; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm21736; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1737; AVX512DQ-NEXT: vmovdqa %ymm4, (%rsi)1738; AVX512DQ-NEXT: vmovdqa %ymm5, (%rdx)1739; AVX512DQ-NEXT: vmovdqa %ymm6, (%rcx)1740; AVX512DQ-NEXT: vmovdqa %ymm7, (%r8)1741; AVX512DQ-NEXT: vmovdqa %ymm8, (%r9)1742; AVX512DQ-NEXT: vmovdqa %ymm9, (%r11)1743; AVX512DQ-NEXT: vmovdqa %ymm10, (%r10)1744; AVX512DQ-NEXT: vmovdqa %ymm0, (%rax)1745; AVX512DQ-NEXT: vzeroupper1746; AVX512DQ-NEXT: retq1747;1748; AVX512DQ-FCP-LABEL: load_i32_stride8_vf8:1749; AVX512DQ-FCP: # %bb.0:1750; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1751; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r101752; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r111753; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm01754; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11755; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21756; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31757; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1758; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm41759; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1760; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm51761; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1762; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1763; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm51764; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1765; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm61766; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1767; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1768; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm61769; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1770; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm71771; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1772; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1773; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm71774; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1775; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm81776; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1777; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1778; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm81779; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1780; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm91781; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1782; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1783; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm91784; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1785; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm101786; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1787; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1788; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm101789; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1790; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm111791; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1792; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1793; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm111794; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1795; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm21796; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1797; AVX512DQ-FCP-NEXT: vmovdqa %ymm4, (%rsi)1798; AVX512DQ-FCP-NEXT: vmovdqa %ymm5, (%rdx)1799; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, (%rcx)1800; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, (%r8)1801; AVX512DQ-FCP-NEXT: vmovdqa %ymm8, (%r9)1802; AVX512DQ-FCP-NEXT: vmovdqa %ymm9, (%r11)1803; AVX512DQ-FCP-NEXT: vmovdqa %ymm10, (%r10)1804; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, (%rax)1805; AVX512DQ-FCP-NEXT: vzeroupper1806; AVX512DQ-FCP-NEXT: retq1807;1808; AVX512BW-LABEL: load_i32_stride8_vf8:1809; AVX512BW: # %bb.0:1810; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax1811; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r101812; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r111813; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm01814; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm11815; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm21816; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm31817; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1818; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm41819; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1820; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm51821; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1822; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1823; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm51824; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1825; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm61826; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1827; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1828; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm61829; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1830; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm71831; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1832; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1833; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm71834; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1835; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm81836; AVX512BW-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1837; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1838; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm81839; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1840; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm91841; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1842; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1843; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm91844; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1845; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm101846; AVX512BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1847; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1848; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm101849; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1850; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm111851; AVX512BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1852; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1853; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm111854; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1855; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm21856; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1857; AVX512BW-NEXT: vmovdqa %ymm4, (%rsi)1858; AVX512BW-NEXT: vmovdqa %ymm5, (%rdx)1859; AVX512BW-NEXT: vmovdqa %ymm6, (%rcx)1860; AVX512BW-NEXT: vmovdqa %ymm7, (%r8)1861; AVX512BW-NEXT: vmovdqa %ymm8, (%r9)1862; AVX512BW-NEXT: vmovdqa %ymm9, (%r11)1863; AVX512BW-NEXT: vmovdqa %ymm10, (%r10)1864; AVX512BW-NEXT: vmovdqa %ymm0, (%rax)1865; AVX512BW-NEXT: vzeroupper1866; AVX512BW-NEXT: retq1867;1868; AVX512BW-FCP-LABEL: load_i32_stride8_vf8:1869; AVX512BW-FCP: # %bb.0:1870; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1871; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r101872; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r111873; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm01874; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11875; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21876; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31877; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1878; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm41879; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1880; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm51881; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1882; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1883; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm51884; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1885; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm61886; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1887; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1888; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm61889; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1890; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm71891; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1892; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1893; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm71894; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1895; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm81896; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1897; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1898; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm81899; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1900; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm91901; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1902; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1903; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm91904; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1905; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm101906; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1907; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1908; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm101909; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1910; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm111911; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1912; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1913; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm111914; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1915; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm21916; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1917; AVX512BW-FCP-NEXT: vmovdqa %ymm4, (%rsi)1918; AVX512BW-FCP-NEXT: vmovdqa %ymm5, (%rdx)1919; AVX512BW-FCP-NEXT: vmovdqa %ymm6, (%rcx)1920; AVX512BW-FCP-NEXT: vmovdqa %ymm7, (%r8)1921; AVX512BW-FCP-NEXT: vmovdqa %ymm8, (%r9)1922; AVX512BW-FCP-NEXT: vmovdqa %ymm9, (%r11)1923; AVX512BW-FCP-NEXT: vmovdqa %ymm10, (%r10)1924; AVX512BW-FCP-NEXT: vmovdqa %ymm0, (%rax)1925; AVX512BW-FCP-NEXT: vzeroupper1926; AVX512BW-FCP-NEXT: retq1927;1928; AVX512DQ-BW-LABEL: load_i32_stride8_vf8:1929; AVX512DQ-BW: # %bb.0:1930; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax1931; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r101932; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r111933; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm01934; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm11935; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm21936; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm31937; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1938; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm41939; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]1940; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm51941; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]1942; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]1943; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm51944; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]1945; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm61946; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1947; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]1948; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm61949; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]1950; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm71951; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]1952; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]1953; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm71954; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]1955; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm81956; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1957; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]1958; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm81959; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]1960; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm91961; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]1962; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]1963; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm91964; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]1965; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm101966; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]1967; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]1968; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm101969; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]1970; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm111971; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]1972; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]1973; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm111974; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]1975; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm21976; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]1977; AVX512DQ-BW-NEXT: vmovdqa %ymm4, (%rsi)1978; AVX512DQ-BW-NEXT: vmovdqa %ymm5, (%rdx)1979; AVX512DQ-BW-NEXT: vmovdqa %ymm6, (%rcx)1980; AVX512DQ-BW-NEXT: vmovdqa %ymm7, (%r8)1981; AVX512DQ-BW-NEXT: vmovdqa %ymm8, (%r9)1982; AVX512DQ-BW-NEXT: vmovdqa %ymm9, (%r11)1983; AVX512DQ-BW-NEXT: vmovdqa %ymm10, (%r10)1984; AVX512DQ-BW-NEXT: vmovdqa %ymm0, (%rax)1985; AVX512DQ-BW-NEXT: vzeroupper1986; AVX512DQ-BW-NEXT: retq1987;1988; AVX512DQ-BW-FCP-LABEL: load_i32_stride8_vf8:1989; AVX512DQ-BW-FCP: # %bb.0:1990; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1991; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r101992; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r111993; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm01994; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11995; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21996; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31997; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,8,16,24]1998; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm41999; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [0,8,16,24]2000; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm52001; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]2002; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,0,0,0,1,9,17,25]2003; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm52004; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [1,9,17,25]2005; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm62006; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]2007; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,2,10,18,26]2008; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm62009; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [2,10,18,26]2010; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm72011; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]2012; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,3,11,19,27]2013; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm72014; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,11,19,27]2015; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82016; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]2017; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,4,12,20,28]2018; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm82019; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,12,20,28]2020; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm92021; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]2022; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,0,0,0,5,13,21,29]2023; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm92024; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm10 = [5,13,21,29]2025; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm102026; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]2027; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,6,14,22,30]2028; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm102029; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm11 = [6,14,22,30]2030; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm112031; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]2032; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,0,0,0,7,15,23,31]2033; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm112034; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [7,15,23,31]2035; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm22036; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm11[4,5,6,7]2037; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm4, (%rsi)2038; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm5, (%rdx)2039; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm6, (%rcx)2040; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm7, (%r8)2041; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm8, (%r9)2042; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm9, (%r11)2043; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm10, (%r10)2044; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, (%rax)2045; AVX512DQ-BW-FCP-NEXT: vzeroupper2046; AVX512DQ-BW-FCP-NEXT: retq2047 %wide.vec = load <64 x i32>, ptr %in.vec, align 642048 %strided.vec0 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56>2049 %strided.vec1 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57>2050 %strided.vec2 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58>2051 %strided.vec3 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59>2052 %strided.vec4 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60>2053 %strided.vec5 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61>2054 %strided.vec6 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62>2055 %strided.vec7 = shufflevector <64 x i32> %wide.vec, <64 x i32> poison, <8 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>2056 store <8 x i32> %strided.vec0, ptr %out.vec0, align 642057 store <8 x i32> %strided.vec1, ptr %out.vec1, align 642058 store <8 x i32> %strided.vec2, ptr %out.vec2, align 642059 store <8 x i32> %strided.vec3, ptr %out.vec3, align 642060 store <8 x i32> %strided.vec4, ptr %out.vec4, align 642061 store <8 x i32> %strided.vec5, ptr %out.vec5, align 642062 store <8 x i32> %strided.vec6, ptr %out.vec6, align 642063 store <8 x i32> %strided.vec7, ptr %out.vec7, align 642064 ret void2065}2066 2067define void @load_i32_stride8_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {2068; SSE-LABEL: load_i32_stride8_vf16:2069; SSE: # %bb.0:2070; SSE-NEXT: subq $296, %rsp # imm = 0x1282071; SSE-NEXT: movaps 288(%rdi), %xmm62072; SSE-NEXT: movaps 352(%rdi), %xmm02073; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2074; SSE-NEXT: movaps 320(%rdi), %xmm52075; SSE-NEXT: movaps 416(%rdi), %xmm22076; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2077; SSE-NEXT: movaps 384(%rdi), %xmm122078; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2079; SSE-NEXT: movaps 480(%rdi), %xmm132080; SSE-NEXT: movaps 448(%rdi), %xmm42081; SSE-NEXT: movaps 160(%rdi), %xmm72082; SSE-NEXT: movaps 128(%rdi), %xmm102083; SSE-NEXT: movaps 224(%rdi), %xmm82084; SSE-NEXT: movaps 192(%rdi), %xmm32085; SSE-NEXT: movaps %xmm3, %xmm92086; SSE-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]2087; SSE-NEXT: movaps %xmm10, %xmm112088; SSE-NEXT: movaps %xmm10, %xmm142089; SSE-NEXT: unpcklps {{.*#+}} xmm11 = xmm11[0],xmm7[0],xmm11[1],xmm7[1]2090; SSE-NEXT: movaps %xmm11, %xmm102091; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm9[0]2092; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2093; SSE-NEXT: movaps %xmm4, %xmm102094; SSE-NEXT: unpcklps {{.*#+}} xmm10 = xmm10[0],xmm13[0],xmm10[1],xmm13[1]2095; SSE-NEXT: unpcklps {{.*#+}} xmm12 = xmm12[0],xmm2[0],xmm12[1],xmm2[1]2096; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm9[1]2097; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2098; SSE-NEXT: movaps %xmm12, %xmm92099; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm10[0]2100; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2101; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm10[1]2102; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2103; SSE-NEXT: movaps %xmm5, %xmm92104; SSE-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]2105; SSE-NEXT: movaps 256(%rdi), %xmm152106; SSE-NEXT: movaps %xmm15, %xmm02107; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]2108; SSE-NEXT: movaps %xmm0, %xmm102109; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm9[0]2110; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2111; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]2112; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2113; SSE-NEXT: movaps 96(%rdi), %xmm102114; SSE-NEXT: movaps 64(%rdi), %xmm92115; SSE-NEXT: movaps %xmm9, %xmm112116; SSE-NEXT: unpcklps {{.*#+}} xmm11 = xmm11[0],xmm10[0],xmm11[1],xmm10[1]2117; SSE-NEXT: movaps (%rdi), %xmm22118; SSE-NEXT: movaps 32(%rdi), %xmm122119; SSE-NEXT: movaps %xmm2, %xmm12120; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm12[0],xmm1[1],xmm12[1]2121; SSE-NEXT: movaps %xmm1, %xmm02122; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm11[0]2123; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2124; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm11[1]2125; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2126; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm8[2],xmm3[3],xmm8[3]2127; SSE-NEXT: unpckhps {{.*#+}} xmm14 = xmm14[2],xmm7[2],xmm14[3],xmm7[3]2128; SSE-NEXT: unpckhps {{.*#+}} xmm4 = xmm4[2],xmm13[2],xmm4[3],xmm13[3]2129; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload2130; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload2131; SSE-NEXT: # xmm13 = xmm13[2],mem[2],xmm13[3],mem[3]2132; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload2133; SSE-NEXT: # xmm5 = xmm5[2],mem[2],xmm5[3],mem[3]2134; SSE-NEXT: unpckhps {{.*#+}} xmm15 = xmm15[2],xmm6[2],xmm15[3],xmm6[3]2135; SSE-NEXT: unpckhps {{.*#+}} xmm9 = xmm9[2],xmm10[2],xmm9[3],xmm10[3]2136; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm12[2],xmm2[3],xmm12[3]2137; SSE-NEXT: movaps %xmm14, %xmm02138; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]2139; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2140; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm3[1]2141; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2142; SSE-NEXT: movaps %xmm13, %xmm02143; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm4[0]2144; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2145; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm4[1]2146; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2147; SSE-NEXT: movaps %xmm15, %xmm02148; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm5[0]2149; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2150; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm5[1]2151; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2152; SSE-NEXT: movaps %xmm2, %xmm02153; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm9[0]2154; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2155; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm9[1]2156; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2157; SSE-NEXT: movaps 240(%rdi), %xmm12158; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill2159; SSE-NEXT: movaps 208(%rdi), %xmm152160; SSE-NEXT: movaps %xmm15, %xmm02161; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]2162; SSE-NEXT: movaps 176(%rdi), %xmm22163; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2164; SSE-NEXT: movaps 144(%rdi), %xmm12165; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2166; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]2167; SSE-NEXT: movaps %xmm1, %xmm22168; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]2169; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2170; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]2171; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2172; SSE-NEXT: movaps 496(%rdi), %xmm12173; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2174; SSE-NEXT: movaps 464(%rdi), %xmm52175; SSE-NEXT: movaps %xmm5, %xmm02176; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]2177; SSE-NEXT: movaps 432(%rdi), %xmm12178; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2179; SSE-NEXT: movaps 400(%rdi), %xmm62180; SSE-NEXT: movaps %xmm6, %xmm102181; SSE-NEXT: unpcklps {{.*#+}} xmm10 = xmm10[0],xmm1[0],xmm10[1],xmm1[1]2182; SSE-NEXT: movaps %xmm10, %xmm12183; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2184; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2185; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]2186; SSE-NEXT: movaps 368(%rdi), %xmm142187; SSE-NEXT: movaps 336(%rdi), %xmm22188; SSE-NEXT: movaps %xmm2, %xmm02189; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1]2190; SSE-NEXT: movaps 304(%rdi), %xmm122191; SSE-NEXT: movaps 272(%rdi), %xmm72192; SSE-NEXT: movaps %xmm7, %xmm42193; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1]2194; SSE-NEXT: movaps %xmm4, %xmm12195; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2196; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2197; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]2198; SSE-NEXT: movaps 112(%rdi), %xmm132199; SSE-NEXT: movaps 80(%rdi), %xmm12200; SSE-NEXT: movaps %xmm1, %xmm02201; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm13[0],xmm0[1],xmm13[1]2202; SSE-NEXT: movaps 16(%rdi), %xmm82203; SSE-NEXT: movaps 48(%rdi), %xmm112204; SSE-NEXT: movaps %xmm8, %xmm32205; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm11[0],xmm3[1],xmm11[1]2206; SSE-NEXT: movaps %xmm3, %xmm92207; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]2208; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2209; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]2210; SSE-NEXT: unpckhps (%rsp), %xmm15 # 16-byte Folded Reload2211; SSE-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]2212; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2213; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2214; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]2215; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm14[2],xmm2[3],xmm14[3]2216; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm12[2],xmm7[3],xmm12[3]2217; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload2218; SSE-NEXT: # xmm5 = xmm5[2],mem[2],xmm5[3],mem[3]2219; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload2220; SSE-NEXT: # xmm6 = xmm6[2],mem[2],xmm6[3],mem[3]2221; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm13[2],xmm1[3],xmm13[3]2222; SSE-NEXT: unpckhps {{.*#+}} xmm8 = xmm8[2],xmm11[2],xmm8[3],xmm11[3]2223; SSE-NEXT: movaps %xmm0, %xmm112224; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm15[0]2225; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]2226; SSE-NEXT: movaps %xmm0, %xmm122227; SSE-NEXT: movaps %xmm7, %xmm92228; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm2[0]2229; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm2[1]2230; SSE-NEXT: movaps %xmm6, %xmm02231; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm5[0]2232; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm5[1]2233; SSE-NEXT: movaps %xmm8, %xmm22234; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]2235; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]2236; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2237; SSE-NEXT: movaps %xmm1, 32(%rsi)2238; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2239; SSE-NEXT: movaps %xmm1, 48(%rsi)2240; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2241; SSE-NEXT: movaps %xmm1, (%rsi)2242; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2243; SSE-NEXT: movaps %xmm1, 16(%rsi)2244; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2245; SSE-NEXT: movaps %xmm1, 32(%rdx)2246; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2247; SSE-NEXT: movaps %xmm1, 48(%rdx)2248; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2249; SSE-NEXT: movaps %xmm1, (%rdx)2250; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2251; SSE-NEXT: movaps %xmm1, 16(%rdx)2252; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2253; SSE-NEXT: movaps %xmm1, 32(%rcx)2254; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2255; SSE-NEXT: movaps %xmm1, 48(%rcx)2256; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2257; SSE-NEXT: movaps %xmm1, (%rcx)2258; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2259; SSE-NEXT: movaps %xmm1, 16(%rcx)2260; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2261; SSE-NEXT: movaps %xmm1, 32(%r8)2262; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2263; SSE-NEXT: movaps %xmm1, 48(%r8)2264; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2265; SSE-NEXT: movaps %xmm1, (%r8)2266; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2267; SSE-NEXT: movaps %xmm1, 16(%r8)2268; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2269; SSE-NEXT: movaps %xmm1, 32(%r9)2270; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2271; SSE-NEXT: movaps %xmm1, 48(%r9)2272; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2273; SSE-NEXT: movaps %xmm1, (%r9)2274; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2275; SSE-NEXT: movaps %xmm1, 16(%r9)2276; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax2277; SSE-NEXT: movaps %xmm4, 32(%rax)2278; SSE-NEXT: movaps %xmm10, 48(%rax)2279; SSE-NEXT: movaps %xmm3, (%rax)2280; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2281; SSE-NEXT: movaps %xmm1, 16(%rax)2282; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax2283; SSE-NEXT: movaps %xmm0, 48(%rax)2284; SSE-NEXT: movaps %xmm9, 32(%rax)2285; SSE-NEXT: movaps %xmm11, 16(%rax)2286; SSE-NEXT: movaps %xmm2, (%rax)2287; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax2288; SSE-NEXT: movaps %xmm6, 48(%rax)2289; SSE-NEXT: movaps %xmm7, 32(%rax)2290; SSE-NEXT: movaps %xmm12, 16(%rax)2291; SSE-NEXT: movaps %xmm8, (%rax)2292; SSE-NEXT: addq $296, %rsp # imm = 0x1282293; SSE-NEXT: retq2294;2295; AVX-LABEL: load_i32_stride8_vf16:2296; AVX: # %bb.0:2297; AVX-NEXT: subq $584, %rsp # imm = 0x2482298; AVX-NEXT: vmovaps 32(%rdi), %xmm02299; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2300; AVX-NEXT: vmovaps (%rdi), %xmm122301; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm12[0],xmm0[0],xmm12[1],xmm0[1]2302; AVX-NEXT: vmovaps 96(%rdi), %xmm12303; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2304; AVX-NEXT: vmovaps 64(%rdi), %xmm22305; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2306; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2307; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm0[0],xmm9[0]2308; AVX-NEXT: vmovaps 160(%rdi), %xmm82309; AVX-NEXT: vmovaps 128(%rdi), %xmm102310; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm10[0],xmm8[0],xmm10[1],xmm8[1]2311; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm62312; AVX-NEXT: vmovaps 224(%rdi), %xmm02313; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2314; AVX-NEXT: vmovaps 192(%rdi), %xmm12315; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2316; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2317; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm5[0,1,0,1]2318; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm72319; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]2320; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm6[4,5,6,7]2321; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2322; AVX-NEXT: vmovaps 416(%rdi), %xmm112323; AVX-NEXT: vmovaps 384(%rdi), %xmm132324; AVX-NEXT: vunpcklps {{.*#+}} xmm4 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]2325; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm42326; AVX-NEXT: vmovaps 480(%rdi), %xmm02327; AVX-NEXT: vmovaps %xmm0, (%rsp) # 16-byte Spill2328; AVX-NEXT: vmovaps 448(%rdi), %xmm12329; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2330; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2331; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm3[0,1,0,1]2332; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm62333; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm6[6,7]2334; AVX-NEXT: vmovaps 288(%rdi), %xmm142335; AVX-NEXT: vmovaps 256(%rdi), %xmm152336; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]2337; AVX-NEXT: vmovaps 352(%rdi), %xmm72338; AVX-NEXT: vmovaps 320(%rdi), %xmm62339; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]2340; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]2341; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2342; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2343; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm12[1,1,1,1]2344; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload2345; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3]2346; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm9[2,3]2347; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm22348; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm10[1,1,1,1]2349; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0],xmm8[1],xmm5[2,3]2350; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52351; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]2352; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2353; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2354; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[1,1,1,1]2355; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm14[1],xmm1[2,3]2356; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2357; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm12358; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm13[1,1,1,1]2359; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm11[1],xmm2[2,3]2360; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22361; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2362; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2363; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2364; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm12[2],xmm4[2],xmm12[3],xmm4[3]2365; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm10[2],xmm8[2],xmm10[3],xmm8[3]2366; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2367; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm0[2,2,2,2]2368; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload2369; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm12[0,1,2],xmm2[3]2370; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22371; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm32372; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]2373; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload2374; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm10[2,2,2,2]2375; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload2376; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm9[0,1,2],xmm3[3]2377; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]2378; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]2379; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2380; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm15[2],xmm14[2],xmm15[3],xmm14[3]2381; AVX-NEXT: vunpckhps {{.*#+}} xmm3 = xmm13[2],xmm11[2],xmm13[3],xmm11[3]2382; AVX-NEXT: vmovaps (%rsp), %xmm14 # 16-byte Reload2383; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm14[2,2,2,2]2384; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload2385; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm11[0,1,2],xmm5[3]2386; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52387; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm82388; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2389; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm7[2,2,2,2]2390; AVX-NEXT: vblendps {{.*#+}} xmm8 = xmm6[0,1,2],xmm8[3]2391; AVX-NEXT: vblendps {{.*#+}} xmm8 = xmm2[0,1],xmm8[2,3]2392; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3],ymm5[4,5,6,7]2393; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2394; AVX-NEXT: vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm10[2],xmm9[3],xmm10[3]2395; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm5[1]2396; AVX-NEXT: vunpckhps {{.*#+}} xmm5 = xmm12[2],xmm0[2],xmm12[3],xmm0[3]2397; AVX-NEXT: vmovaps 320(%rdi), %ymm82398; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52399; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm4[2,3,2,3]2400; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm02401; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm5[6,7]2402; AVX-NEXT: vmovaps 352(%rdi), %ymm52403; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2404; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2405; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2406; AVX-NEXT: vmovaps 416(%rdi), %ymm42407; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2408; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]2409; AVX-NEXT: vmovaps 384(%rdi), %ymm62410; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2411; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm2[1],xmm0[1]2412; AVX-NEXT: vmovaps 448(%rdi), %ymm72413; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2414; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm11[2],xmm14[2],xmm11[3],xmm14[3]2415; AVX-NEXT: vmovaps 480(%rdi), %ymm92416; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12417; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm3[2,3,2,3]2418; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22419; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2420; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2421; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2422; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm7[0],ymm9[2],ymm7[2]2423; AVX-NEXT: vmovaps %ymm9, %ymm32424; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm6[0],ymm4[0],ymm6[1],ymm4[1],ymm6[4],ymm4[4],ymm6[5],ymm4[5]2425; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]2426; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm5[0],ymm8[0],ymm5[2],ymm8[2]2427; AVX-NEXT: vmovaps %ymm8, %ymm62428; AVX-NEXT: vextractf128 $1, %ymm1, %xmm12429; AVX-NEXT: vmovaps 288(%rdi), %ymm72430; AVX-NEXT: vmovaps 256(%rdi), %ymm92431; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]2432; AVX-NEXT: vextractf128 $1, %ymm2, %xmm22433; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]2434; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2435; AVX-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill2436; AVX-NEXT: vmovaps 160(%rdi), %ymm112437; AVX-NEXT: vmovaps 128(%rdi), %ymm122438; AVX-NEXT: vmovaps 192(%rdi), %ymm102439; AVX-NEXT: vmovaps 224(%rdi), %ymm132440; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm10[0],ymm13[2],ymm10[2]2441; AVX-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2442; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[4],ymm11[4],ymm12[5],ymm11[5]2443; AVX-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2444; AVX-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2445; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]2446; AVX-NEXT: vmovaps 64(%rdi), %ymm12447; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2448; AVX-NEXT: vmovaps 96(%rdi), %ymm02449; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2450; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]2451; AVX-NEXT: vextractf128 $1, %ymm0, %xmm02452; AVX-NEXT: vmovaps (%rdi), %ymm12453; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2454; AVX-NEXT: vmovaps 32(%rdi), %ymm152455; AVX-NEXT: vunpcklps {{.*#+}} ymm14 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]2456; AVX-NEXT: vextractf128 $1, %ymm14, %xmm142457; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm14[0,1],xmm0[2,0]2458; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]2459; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2460; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2461; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]2462; AVX-NEXT: vmovaps %ymm3, %ymm82463; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2464; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2465; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm0[1,0],ymm3[1,0],ymm0[5,4],ymm3[5,4]2466; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm4[2,0],ymm2[2,3],ymm4[6,4],ymm2[6,7]2467; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2468; AVX-NEXT: vunpcklps {{.*#+}} ymm4 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[4],ymm2[4],ymm6[5],ymm2[5]2469; AVX-NEXT: vextractf128 $1, %ymm4, %xmm42470; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm7[1,0],ymm9[1,0],ymm7[5,4],ymm9[5,4]2471; AVX-NEXT: vextractf128 $1, %ymm14, %xmm142472; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm14[2,0],xmm4[2,3]2473; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]2474; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2475; AVX-NEXT: vunpcklps {{.*#+}} ymm5 = ymm10[0],ymm13[0],ymm10[1],ymm13[1],ymm10[4],ymm13[4],ymm10[5],ymm13[5]2476; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm11[1,0],ymm12[1,0],ymm11[5,4],ymm12[5,4]2477; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm4[2,0],ymm5[2,3],ymm4[6,4],ymm5[6,7]2478; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload2479; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload2480; AVX-NEXT: vunpcklps {{.*#+}} ymm4 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]2481; AVX-NEXT: vextractf128 $1, %ymm4, %xmm42482; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload2483; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm15[1,0],ymm11[1,0],ymm15[5,4],ymm11[5,4]2484; AVX-NEXT: vextractf128 $1, %ymm14, %xmm142485; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm14[2,0],xmm4[2,3]2486; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]2487; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2488; AVX-NEXT: vunpckhpd {{.*#+}} ymm5 = ymm8[1],ymm1[1],ymm8[3],ymm1[3]2489; AVX-NEXT: vunpckhps {{.*#+}} ymm4 = ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[6],ymm0[6],ymm3[7],ymm0[7]2490; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,1],ymm5[2,0],ymm4[4,5],ymm5[6,4]2491; AVX-NEXT: vmovaps %ymm6, %ymm32492; AVX-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm2[1],ymm6[1],ymm2[3],ymm6[3]2493; AVX-NEXT: vextractf128 $1, %ymm4, %xmm42494; AVX-NEXT: vmovaps %ymm7, %ymm22495; AVX-NEXT: vunpckhps {{.*#+}} ymm14 = ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[6],ymm7[6],ymm9[7],ymm7[7]2496; AVX-NEXT: vextractf128 $1, %ymm14, %xmm142497; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm14[0,1],xmm4[2,0]2498; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm4[0,1,2,3],ymm1[4,5,6,7]2499; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2500; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]2501; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2502; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2503; AVX-NEXT: vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[6],ymm1[6],ymm6[7],ymm1[7]2504; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,0],ymm4[4,5],ymm0[6,4]2505; AVX-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]2506; AVX-NEXT: vextractf128 $1, %ymm4, %xmm42507; AVX-NEXT: vunpckhps {{.*#+}} ymm14 = ymm11[2],ymm15[2],ymm11[3],ymm15[3],ymm11[6],ymm15[6],ymm11[7],ymm15[7]2508; AVX-NEXT: vextractf128 $1, %ymm14, %xmm142509; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm14[0,1],xmm4[2,0]2510; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm4[0,1,2,3],ymm0[4,5,6,7]2511; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2512; AVX-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm8[2],ymm0[3],ymm8[3],ymm0[6],ymm8[6],ymm0[7],ymm8[7]2513; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2514; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload2515; AVX-NEXT: # ymm4 = ymm4[3,0],mem[3,0],ymm4[7,4],mem[7,4]2516; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm4[2,0],ymm0[2,3],ymm4[6,4],ymm0[6,7]2517; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm4 # 32-byte Folded Reload2518; AVX-NEXT: # ymm4 = ymm3[2],mem[2],ymm3[3],mem[3],ymm3[6],mem[6],ymm3[7],mem[7]2519; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm2[3,0],ymm9[3,0],ymm2[7,4],ymm9[7,4]2520; AVX-NEXT: vextractf128 $1, %ymm4, %xmm42521; AVX-NEXT: vextractf128 $1, %ymm9, %xmm92522; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm9[2,0],xmm4[2,3]2523; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]2524; AVX-NEXT: vunpckhps {{.*#+}} ymm4 = ymm7[2],ymm13[2],ymm7[3],ymm13[3],ymm7[6],ymm13[6],ymm7[7],ymm13[7]2525; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm1[3,0],ymm6[3,0],ymm1[7,4],ymm6[7,4]2526; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm6[2,0],ymm4[2,3],ymm6[6,4],ymm4[6,7]2527; AVX-NEXT: vunpckhps {{.*#+}} ymm3 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]2528; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm15[3,0],ymm11[3,0],ymm15[7,4],ymm11[7,4]2529; AVX-NEXT: vextractf128 $1, %ymm3, %xmm22530; AVX-NEXT: vextractf128 $1, %ymm1, %xmm12531; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,0],xmm2[2,3]2532; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]2533; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2534; AVX-NEXT: vmovaps %ymm2, 32(%rsi)2535; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2536; AVX-NEXT: vmovaps %ymm2, (%rsi)2537; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2538; AVX-NEXT: vmovaps %ymm2, 32(%rdx)2539; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2540; AVX-NEXT: vmovaps %ymm2, (%rdx)2541; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2542; AVX-NEXT: vmovaps %ymm2, 32(%rcx)2543; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2544; AVX-NEXT: vmovaps %ymm2, (%rcx)2545; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2546; AVX-NEXT: vmovaps %ymm2, 32(%r8)2547; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2548; AVX-NEXT: vmovaps %ymm2, (%r8)2549; AVX-NEXT: vmovups (%rsp), %ymm2 # 32-byte Reload2550; AVX-NEXT: vmovaps %ymm2, 32(%r9)2551; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2552; AVX-NEXT: vmovaps %ymm2, (%r9)2553; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax2554; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2555; AVX-NEXT: vmovaps %ymm2, 32(%rax)2556; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload2557; AVX-NEXT: vmovaps %ymm2, (%rax)2558; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax2559; AVX-NEXT: vmovaps %ymm5, 32(%rax)2560; AVX-NEXT: vmovaps %ymm14, (%rax)2561; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax2562; AVX-NEXT: vmovaps %ymm0, 32(%rax)2563; AVX-NEXT: vmovaps %ymm1, (%rax)2564; AVX-NEXT: addq $584, %rsp # imm = 0x2482565; AVX-NEXT: vzeroupper2566; AVX-NEXT: retq2567;2568; AVX2-LABEL: load_i32_stride8_vf16:2569; AVX2: # %bb.0:2570; AVX2-NEXT: subq $456, %rsp # imm = 0x1C82571; AVX2-NEXT: vmovaps 288(%rdi), %xmm82572; AVX2-NEXT: vmovaps 256(%rdi), %xmm92573; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]2574; AVX2-NEXT: vmovaps 352(%rdi), %xmm12575; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2576; AVX2-NEXT: vbroadcastss %xmm1, %xmm22577; AVX2-NEXT: vmovaps 320(%rdi), %xmm12578; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2579; AVX2-NEXT: vbroadcastss %xmm1, %xmm32580; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]2581; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]2582; AVX2-NEXT: vmovaps 416(%rdi), %xmm12583; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2584; AVX2-NEXT: vmovaps 384(%rdi), %xmm22585; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2586; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2587; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22588; AVX2-NEXT: vmovaps 480(%rdi), %xmm12589; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2590; AVX2-NEXT: vbroadcastss %xmm1, %xmm32591; AVX2-NEXT: vmovaps 448(%rdi), %xmm12592; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2593; AVX2-NEXT: vbroadcastss %xmm1, %xmm102594; AVX2-NEXT: vunpcklps {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1]2595; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm32596; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]2597; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]2598; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2599; AVX2-NEXT: vmovaps 160(%rdi), %xmm62600; AVX2-NEXT: vmovaps 128(%rdi), %xmm152601; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm15[0],xmm6[0],xmm15[1],xmm6[1]2602; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2603; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22604; AVX2-NEXT: vmovaps 224(%rdi), %xmm42605; AVX2-NEXT: vbroadcastss %xmm4, %xmm32606; AVX2-NEXT: vmovaps %xmm4, (%rsp) # 16-byte Spill2607; AVX2-NEXT: vmovaps 192(%rdi), %xmm122608; AVX2-NEXT: vbroadcastss %xmm12, %xmm112609; AVX2-NEXT: vunpcklps {{.*#+}} xmm3 = xmm11[0],xmm3[0],xmm11[1],xmm3[1]2610; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm32611; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm3[6,7]2612; AVX2-NEXT: vmovaps 96(%rdi), %xmm52613; AVX2-NEXT: vbroadcastss %xmm5, %xmm22614; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2615; AVX2-NEXT: vmovaps 64(%rdi), %xmm132616; AVX2-NEXT: vbroadcastss %xmm13, %xmm32617; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]2618; AVX2-NEXT: vmovaps (%rdi), %xmm112619; AVX2-NEXT: vmovaps 32(%rdi), %xmm102620; AVX2-NEXT: vunpcklps {{.*#+}} xmm7 = xmm11[0],xmm10[0],xmm11[1],xmm10[1]2621; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm7[0,1],xmm0[2,3]2622; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2623; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2624; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1,1,1]2625; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]2626; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm5[0],xmm13[1],xmm5[1]2627; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]2628; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm4[0],xmm12[1],xmm4[1]2629; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12630; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm15[1,1,1,1]2631; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm6[1],xmm7[2,3]2632; AVX2-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm72633; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]2634; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2635; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2636; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm9[1,1,1,1]2637; AVX2-NEXT: vmovaps %xmm8, %xmm62638; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm8[1],xmm0[2,3]2639; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload2640; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload2641; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm14[0],xmm8[0],xmm14[1],xmm8[1]2642; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]2643; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2644; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2645; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]2646; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12647; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload2648; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]2649; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload2650; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]2651; AVX2-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm72652; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]2653; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2654; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2655; AVX2-NEXT: vunpckhps {{.*#+}} xmm0 = xmm9[2],xmm6[2],xmm9[3],xmm6[3]2656; AVX2-NEXT: vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]2657; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm2[2,2,2,2]2658; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm3[0,1,2],xmm5[3]2659; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52660; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm62661; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]2662; AVX2-NEXT: vmovaps %xmm8, %xmm72663; AVX2-NEXT: vshufps {{.*#+}} xmm6 = xmm8[2,2,2,2]2664; AVX2-NEXT: vblendps {{.*#+}} xmm6 = xmm14[0,1,2],xmm6[3]2665; AVX2-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1],xmm6[2,3]2666; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]2667; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2668; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm4 # 16-byte Folded Reload2669; AVX2-NEXT: # xmm4 = xmm15[2],mem[2],xmm15[3],mem[3]2670; AVX2-NEXT: vmovaps (%rsp), %xmm9 # 16-byte Reload2671; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm9[2,2,2,2]2672; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm12[0,1,2],xmm5[3]2673; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52674; AVX2-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm62675; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]2676; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]2677; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload2678; AVX2-NEXT: vshufps {{.*#+}} xmm3 = xmm8[2,2,2,2]2679; AVX2-NEXT: vblendps {{.*#+}} xmm3 = xmm13[0,1,2],xmm3[3]2680; AVX2-NEXT: vblendps {{.*#+}} xmm3 = xmm2[0,1],xmm3[2,3]2681; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]2682; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2683; AVX2-NEXT: vunpckhps {{.*#+}} xmm3 = xmm14[2],xmm7[2],xmm14[3],xmm7[3]2684; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]2685; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2686; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload2687; AVX2-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]2688; AVX2-NEXT: vmovaps 96(%rdi), %ymm52689; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2690; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm32691; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]2692; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12693; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]2694; AVX2-NEXT: vmovaps 64(%rdi), %ymm62695; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2696; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2697; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2698; AVX2-NEXT: vmovaps 32(%rdi), %ymm152699; AVX2-NEXT: vunpckhps {{.*#+}} xmm0 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]2700; AVX2-NEXT: vmovaps (%rdi), %ymm142701; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm02702; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]2703; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12704; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]2705; AVX2-NEXT: vmovaps 224(%rdi), %ymm32706; AVX2-NEXT: vmovups %ymm3, (%rsp) # 32-byte Spill2707; AVX2-NEXT: vunpckhps {{.*#+}} xmm1 = xmm13[2],xmm8[2],xmm13[3],xmm8[3]2708; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]2709; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2710; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2711; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[4],ymm15[4],ymm14[5],ymm15[5]2712; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm02713; AVX2-NEXT: vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[4],ymm5[4],ymm6[5],ymm5[5]2714; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm11[2,2,2,2]2715; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]2716; AVX2-NEXT: vmovaps 192(%rdi), %ymm12717; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2718; AVX2-NEXT: vmovaps 160(%rdi), %ymm72719; AVX2-NEXT: vmovaps 128(%rdi), %ymm132720; AVX2-NEXT: vunpcklps {{.*#+}} ymm2 = ymm13[0],ymm7[0],ymm13[1],ymm7[1],ymm13[4],ymm7[4],ymm13[5],ymm7[5]2721; AVX2-NEXT: vunpcklps {{.*#+}} ymm5 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]2722; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm5[0],ymm2[2],ymm5[2]2723; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]2724; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2725; AVX2-NEXT: vmovaps 288(%rdi), %ymm92726; AVX2-NEXT: vmovaps 256(%rdi), %ymm82727; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]2728; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm02729; AVX2-NEXT: vmovaps 352(%rdi), %ymm12730; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2731; AVX2-NEXT: vmovaps 320(%rdi), %ymm22732; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2733; AVX2-NEXT: vunpcklps {{.*#+}} ymm4 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]2734; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[2,2,2,2]2735; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1],xmm6[2,3]2736; AVX2-NEXT: vmovaps 480(%rdi), %ymm02737; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2738; AVX2-NEXT: vmovaps 448(%rdi), %ymm122739; AVX2-NEXT: vmovaps 416(%rdi), %ymm62740; AVX2-NEXT: vmovaps 384(%rdi), %ymm102741; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[4],ymm6[4],ymm10[5],ymm6[5]2742; AVX2-NEXT: vunpcklps {{.*#+}} ymm3 = ymm12[0],ymm0[0],ymm12[1],ymm0[1],ymm12[4],ymm0[4],ymm12[5],ymm0[5]2743; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm3[0],ymm1[2],ymm3[2]2744; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]2745; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2746; AVX2-NEXT: vbroadcastss 148(%rdi), %ymm02747; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]2748; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm5[6,7]2749; AVX2-NEXT: vextractf128 $1, %ymm11, %xmm22750; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm14[1,1,1,1,5,5,5,5]2751; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm15[1],ymm5[2,3,4],ymm15[5],ymm5[6,7]2752; AVX2-NEXT: vextractf128 $1, %ymm5, %xmm52753; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]2754; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]2755; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2756; AVX2-NEXT: vbroadcastss 404(%rdi), %ymm02757; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5],ymm0[6,7]2758; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]2759; AVX2-NEXT: vextractf128 $1, %ymm4, %xmm12760; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm8[1,1,1,1,5,5,5,5]2761; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm9[1],ymm2[2,3,4],ymm9[5],ymm2[6,7]2762; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm22763; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]2764; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm1[0,1,2,3],ymm0[4,5,6,7]2765; AVX2-NEXT: vbroadcastss 248(%rdi), %ymm02766; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload2767; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]2768; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm13[2],ymm7[2],ymm13[3],ymm7[3],ymm13[6],ymm7[6],ymm13[7],ymm7[7]2769; AVX2-NEXT: vunpckhps {{.*#+}} ymm4 = ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[6],ymm15[6],ymm14[7],ymm15[7]2770; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2771; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload2772; AVX2-NEXT: # ymm5 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]2773; AVX2-NEXT: vextractf128 $1, %ymm4, %xmm12774; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm5[2,2,2,2]2775; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2,3]2776; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]2777; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]2778; AVX2-NEXT: vbroadcastss 504(%rdi), %ymm02779; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5,6],ymm0[7]2780; AVX2-NEXT: vunpckhps {{.*#+}} ymm3 = ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[6],ymm6[6],ymm10[7],ymm6[7]2781; AVX2-NEXT: vunpckhps {{.*#+}} ymm6 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]2782; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2783; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload2784; AVX2-NEXT: # ymm7 = ymm7[2],mem[2],ymm7[3],mem[3],ymm7[6],mem[6],ymm7[7],mem[7]2785; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm82786; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm7[2,2,2,2]2787; AVX2-NEXT: vblendps {{.*#+}} xmm8 = xmm8[0,1],xmm9[2,3]2788; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm0[6,7]2789; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]2790; AVX2-NEXT: vbroadcastss 220(%rdi), %ymm82791; AVX2-NEXT: vunpckhps (%rsp), %ymm8, %ymm8 # 32-byte Folded Reload2792; AVX2-NEXT: # ymm8 = ymm8[2],mem[2],ymm8[3],mem[3],ymm8[6],mem[6],ymm8[7],mem[7]2793; AVX2-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm8[1],ymm2[3],ymm8[3]2794; AVX2-NEXT: vextractf128 $1, %ymm5, %xmm52795; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]2796; AVX2-NEXT: vextractf128 $1, %ymm4, %xmm42797; AVX2-NEXT: vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm5[2,3]2798; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]2799; AVX2-NEXT: vbroadcastss 476(%rdi), %ymm42800; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload2801; AVX2-NEXT: # ymm4 = ymm4[2],mem[2],ymm4[3],mem[3],ymm4[6],mem[6],ymm4[7],mem[7]2802; AVX2-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]2803; AVX2-NEXT: vextractf128 $1, %ymm7, %xmm42804; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm6[2,3,2,3,6,7,6,7]2805; AVX2-NEXT: vextractf128 $1, %ymm5, %xmm52806; AVX2-NEXT: vblendps {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]2807; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2808; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2809; AVX2-NEXT: vmovaps %ymm4, 32(%rsi)2810; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2811; AVX2-NEXT: vmovaps %ymm4, (%rsi)2812; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2813; AVX2-NEXT: vmovaps %ymm4, 32(%rdx)2814; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2815; AVX2-NEXT: vmovaps %ymm4, (%rdx)2816; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2817; AVX2-NEXT: vmovaps %ymm4, 32(%rcx)2818; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2819; AVX2-NEXT: vmovaps %ymm4, (%rcx)2820; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2821; AVX2-NEXT: vmovaps %ymm4, 32(%r8)2822; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2823; AVX2-NEXT: vmovaps %ymm4, (%r8)2824; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2825; AVX2-NEXT: vmovaps %ymm4, 32(%r9)2826; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2827; AVX2-NEXT: vmovaps %ymm4, (%r9)2828; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax2829; AVX2-NEXT: vmovaps %ymm11, 32(%rax)2830; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2831; AVX2-NEXT: vmovaps %ymm4, (%rax)2832; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax2833; AVX2-NEXT: vmovaps %ymm0, 32(%rax)2834; AVX2-NEXT: vmovaps %ymm1, (%rax)2835; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax2836; AVX2-NEXT: vmovaps %ymm3, 32(%rax)2837; AVX2-NEXT: vmovaps %ymm2, (%rax)2838; AVX2-NEXT: addq $456, %rsp # imm = 0x1C82839; AVX2-NEXT: vzeroupper2840; AVX2-NEXT: retq2841;2842; AVX2-FP-LABEL: load_i32_stride8_vf16:2843; AVX2-FP: # %bb.0:2844; AVX2-FP-NEXT: subq $456, %rsp # imm = 0x1C82845; AVX2-FP-NEXT: vmovaps 288(%rdi), %xmm82846; AVX2-FP-NEXT: vmovaps 256(%rdi), %xmm92847; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]2848; AVX2-FP-NEXT: vmovaps 352(%rdi), %xmm12849; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2850; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm22851; AVX2-FP-NEXT: vmovaps 320(%rdi), %xmm12852; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2853; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm32854; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]2855; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]2856; AVX2-FP-NEXT: vmovaps 416(%rdi), %xmm12857; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2858; AVX2-FP-NEXT: vmovaps 384(%rdi), %xmm22859; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2860; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2861; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22862; AVX2-FP-NEXT: vmovaps 480(%rdi), %xmm12863; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2864; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm32865; AVX2-FP-NEXT: vmovaps 448(%rdi), %xmm12866; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2867; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm102868; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1]2869; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm32870; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]2871; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]2872; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2873; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm62874; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm152875; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm15[0],xmm6[0],xmm15[1],xmm6[1]2876; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2877; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm22878; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm42879; AVX2-FP-NEXT: vbroadcastss %xmm4, %xmm32880; AVX2-FP-NEXT: vmovaps %xmm4, (%rsp) # 16-byte Spill2881; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm122882; AVX2-FP-NEXT: vbroadcastss %xmm12, %xmm112883; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm11[0],xmm3[0],xmm11[1],xmm3[1]2884; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm32885; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm3[6,7]2886; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm52887; AVX2-FP-NEXT: vbroadcastss %xmm5, %xmm22888; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2889; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm132890; AVX2-FP-NEXT: vbroadcastss %xmm13, %xmm32891; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]2892; AVX2-FP-NEXT: vmovaps (%rdi), %xmm112893; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm102894; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm7 = xmm11[0],xmm10[0],xmm11[1],xmm10[1]2895; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm7[0,1],xmm0[2,3]2896; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2897; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2898; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1,1,1]2899; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]2900; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm5[0],xmm13[1],xmm5[1]2901; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]2902; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm4[0],xmm12[1],xmm4[1]2903; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12904; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm15[1,1,1,1]2905; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm6[1],xmm7[2,3]2906; AVX2-FP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm72907; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]2908; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2909; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2910; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm9[1,1,1,1]2911; AVX2-FP-NEXT: vmovaps %xmm8, %xmm62912; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm8[1],xmm0[2,3]2913; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload2914; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload2915; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm14[0],xmm8[0],xmm14[1],xmm8[1]2916; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]2917; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2918; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2919; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]2920; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12921; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload2922; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]2923; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload2924; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]2925; AVX2-FP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm72926; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]2927; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2928; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2929; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm9[2],xmm6[2],xmm9[3],xmm6[3]2930; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]2931; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm2[2,2,2,2]2932; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm3[0,1,2],xmm5[3]2933; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52934; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm62935; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]2936; AVX2-FP-NEXT: vmovaps %xmm8, %xmm72937; AVX2-FP-NEXT: vshufps {{.*#+}} xmm6 = xmm8[2,2,2,2]2938; AVX2-FP-NEXT: vblendps {{.*#+}} xmm6 = xmm14[0,1,2],xmm6[3]2939; AVX2-FP-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1],xmm6[2,3]2940; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]2941; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2942; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm4 # 16-byte Folded Reload2943; AVX2-FP-NEXT: # xmm4 = xmm15[2],mem[2],xmm15[3],mem[3]2944; AVX2-FP-NEXT: vmovaps (%rsp), %xmm9 # 16-byte Reload2945; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm9[2,2,2,2]2946; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm12[0,1,2],xmm5[3]2947; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm52948; AVX2-FP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm62949; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]2950; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]2951; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload2952; AVX2-FP-NEXT: vshufps {{.*#+}} xmm3 = xmm8[2,2,2,2]2953; AVX2-FP-NEXT: vblendps {{.*#+}} xmm3 = xmm13[0,1,2],xmm3[3]2954; AVX2-FP-NEXT: vblendps {{.*#+}} xmm3 = xmm2[0,1],xmm3[2,3]2955; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]2956; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2957; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm3 = xmm14[2],xmm7[2],xmm14[3],xmm7[3]2958; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]2959; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2960; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload2961; AVX2-FP-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]2962; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm52963; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2964; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm32965; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]2966; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12967; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]2968; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm62969; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2970; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2971; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2972; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm152973; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]2974; AVX2-FP-NEXT: vmovaps (%rdi), %ymm142975; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm02976; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]2977; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm12978; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]2979; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm32980; AVX2-FP-NEXT: vmovups %ymm3, (%rsp) # 32-byte Spill2981; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm1 = xmm13[2],xmm8[2],xmm13[3],xmm8[3]2982; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]2983; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2984; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2985; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[4],ymm15[4],ymm14[5],ymm15[5]2986; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm02987; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[4],ymm5[4],ymm6[5],ymm5[5]2988; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm11[2,2,2,2]2989; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]2990; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm12991; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2992; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm72993; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm132994; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm13[0],ymm7[0],ymm13[1],ymm7[1],ymm13[4],ymm7[4],ymm13[5],ymm7[5]2995; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm5 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]2996; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm5[0],ymm2[2],ymm5[2]2997; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]2998; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2999; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm93000; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm83001; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]3002; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm03003; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm13004; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3005; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm23006; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3007; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm4 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]3008; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[2,2,2,2]3009; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1],xmm6[2,3]3010; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm03011; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3012; AVX2-FP-NEXT: vmovaps 448(%rdi), %ymm123013; AVX2-FP-NEXT: vmovaps 416(%rdi), %ymm63014; AVX2-FP-NEXT: vmovaps 384(%rdi), %ymm103015; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[4],ymm6[4],ymm10[5],ymm6[5]3016; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm3 = ymm12[0],ymm0[0],ymm12[1],ymm0[1],ymm12[4],ymm0[4],ymm12[5],ymm0[5]3017; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm3[0],ymm1[2],ymm3[2]3018; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]3019; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3020; AVX2-FP-NEXT: vbroadcastss 148(%rdi), %ymm03021; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]3022; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm5[6,7]3023; AVX2-FP-NEXT: vextractf128 $1, %ymm11, %xmm23024; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm14[1,1,1,1,5,5,5,5]3025; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm15[1],ymm5[2,3,4],ymm15[5],ymm5[6,7]3026; AVX2-FP-NEXT: vextractf128 $1, %ymm5, %xmm53027; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]3028; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]3029; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3030; AVX2-FP-NEXT: vbroadcastss 404(%rdi), %ymm03031; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5],ymm0[6,7]3032; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]3033; AVX2-FP-NEXT: vextractf128 $1, %ymm4, %xmm13034; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm8[1,1,1,1,5,5,5,5]3035; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm9[1],ymm2[2,3,4],ymm9[5],ymm2[6,7]3036; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm23037; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]3038; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm1[0,1,2,3],ymm0[4,5,6,7]3039; AVX2-FP-NEXT: vbroadcastss 248(%rdi), %ymm03040; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3041; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]3042; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm13[2],ymm7[2],ymm13[3],ymm7[3],ymm13[6],ymm7[6],ymm13[7],ymm7[7]3043; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm4 = ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[6],ymm15[6],ymm14[7],ymm15[7]3044; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload3045; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload3046; AVX2-FP-NEXT: # ymm5 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]3047; AVX2-FP-NEXT: vextractf128 $1, %ymm4, %xmm13048; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm5[2,2,2,2]3049; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2,3]3050; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]3051; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]3052; AVX2-FP-NEXT: vbroadcastss 504(%rdi), %ymm03053; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5,6],ymm0[7]3054; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[6],ymm6[6],ymm10[7],ymm6[7]3055; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm6 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]3056; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload3057; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload3058; AVX2-FP-NEXT: # ymm7 = ymm7[2],mem[2],ymm7[3],mem[3],ymm7[6],mem[6],ymm7[7],mem[7]3059; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm83060; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm7[2,2,2,2]3061; AVX2-FP-NEXT: vblendps {{.*#+}} xmm8 = xmm8[0,1],xmm9[2,3]3062; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm0[6,7]3063; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]3064; AVX2-FP-NEXT: vbroadcastss 220(%rdi), %ymm83065; AVX2-FP-NEXT: vunpckhps (%rsp), %ymm8, %ymm8 # 32-byte Folded Reload3066; AVX2-FP-NEXT: # ymm8 = ymm8[2],mem[2],ymm8[3],mem[3],ymm8[6],mem[6],ymm8[7],mem[7]3067; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm8[1],ymm2[3],ymm8[3]3068; AVX2-FP-NEXT: vextractf128 $1, %ymm5, %xmm53069; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]3070; AVX2-FP-NEXT: vextractf128 $1, %ymm4, %xmm43071; AVX2-FP-NEXT: vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm5[2,3]3072; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]3073; AVX2-FP-NEXT: vbroadcastss 476(%rdi), %ymm43074; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload3075; AVX2-FP-NEXT: # ymm4 = ymm4[2],mem[2],ymm4[3],mem[3],ymm4[6],mem[6],ymm4[7],mem[7]3076; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]3077; AVX2-FP-NEXT: vextractf128 $1, %ymm7, %xmm43078; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm6[2,3,2,3,6,7,6,7]3079; AVX2-FP-NEXT: vextractf128 $1, %ymm5, %xmm53080; AVX2-FP-NEXT: vblendps {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]3081; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]3082; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3083; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rsi)3084; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3085; AVX2-FP-NEXT: vmovaps %ymm4, (%rsi)3086; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3087; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rdx)3088; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3089; AVX2-FP-NEXT: vmovaps %ymm4, (%rdx)3090; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3091; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rcx)3092; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3093; AVX2-FP-NEXT: vmovaps %ymm4, (%rcx)3094; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3095; AVX2-FP-NEXT: vmovaps %ymm4, 32(%r8)3096; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3097; AVX2-FP-NEXT: vmovaps %ymm4, (%r8)3098; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3099; AVX2-FP-NEXT: vmovaps %ymm4, 32(%r9)3100; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3101; AVX2-FP-NEXT: vmovaps %ymm4, (%r9)3102; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax3103; AVX2-FP-NEXT: vmovaps %ymm11, 32(%rax)3104; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3105; AVX2-FP-NEXT: vmovaps %ymm4, (%rax)3106; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax3107; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rax)3108; AVX2-FP-NEXT: vmovaps %ymm1, (%rax)3109; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax3110; AVX2-FP-NEXT: vmovaps %ymm3, 32(%rax)3111; AVX2-FP-NEXT: vmovaps %ymm2, (%rax)3112; AVX2-FP-NEXT: addq $456, %rsp # imm = 0x1C83113; AVX2-FP-NEXT: vzeroupper3114; AVX2-FP-NEXT: retq3115;3116; AVX2-FCP-LABEL: load_i32_stride8_vf16:3117; AVX2-FCP: # %bb.0:3118; AVX2-FCP-NEXT: subq $456, %rsp # imm = 0x1C83119; AVX2-FCP-NEXT: vmovaps 288(%rdi), %xmm83120; AVX2-FCP-NEXT: vmovaps 256(%rdi), %xmm93121; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]3122; AVX2-FCP-NEXT: vmovaps 352(%rdi), %xmm13123; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3124; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm23125; AVX2-FCP-NEXT: vmovaps 320(%rdi), %xmm13126; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3127; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm33128; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]3129; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]3130; AVX2-FCP-NEXT: vmovaps 416(%rdi), %xmm13131; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3132; AVX2-FCP-NEXT: vmovaps 384(%rdi), %xmm23133; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3134; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3135; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm23136; AVX2-FCP-NEXT: vmovaps 480(%rdi), %xmm13137; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3138; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm33139; AVX2-FCP-NEXT: vmovaps 448(%rdi), %xmm13140; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3141; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm103142; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1]3143; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm33144; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]3145; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]3146; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3147; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm63148; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm153149; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm15[0],xmm6[0],xmm15[1],xmm6[1]3150; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3151; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm23152; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm43153; AVX2-FCP-NEXT: vbroadcastss %xmm4, %xmm33154; AVX2-FCP-NEXT: vmovaps %xmm4, (%rsp) # 16-byte Spill3155; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm123156; AVX2-FCP-NEXT: vbroadcastss %xmm12, %xmm113157; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm11[0],xmm3[0],xmm11[1],xmm3[1]3158; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm33159; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm3[6,7]3160; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm53161; AVX2-FCP-NEXT: vbroadcastss %xmm5, %xmm23162; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3163; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm133164; AVX2-FCP-NEXT: vbroadcastss %xmm13, %xmm33165; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]3166; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm113167; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm103168; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm7 = xmm11[0],xmm10[0],xmm11[1],xmm10[1]3169; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm7[0,1],xmm0[2,3]3170; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3171; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3172; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1,1,1]3173; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]3174; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm5[0],xmm13[1],xmm5[1]3175; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]3176; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm4[0],xmm12[1],xmm4[1]3177; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13178; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm7 = xmm15[1,1,1,1]3179; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm6[1],xmm7[2,3]3180; AVX2-FCP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm73181; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]3182; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3183; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3184; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm9[1,1,1,1]3185; AVX2-FCP-NEXT: vmovaps %xmm8, %xmm63186; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm8[1],xmm0[2,3]3187; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload3188; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload3189; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm14[0],xmm8[0],xmm14[1],xmm8[1]3190; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]3191; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload3192; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload3193; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]3194; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13195; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload3196; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]3197; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload3198; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]3199; AVX2-FCP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm73200; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]3201; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3202; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3203; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm9[2],xmm6[2],xmm9[3],xmm6[3]3204; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]3205; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm2[2,2,2,2]3206; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm3[0,1,2],xmm5[3]3207; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm53208; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm63209; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]3210; AVX2-FCP-NEXT: vmovaps %xmm8, %xmm73211; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm6 = xmm8[2,2,2,2]3212; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm6 = xmm14[0,1,2],xmm6[3]3213; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1],xmm6[2,3]3214; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]3215; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3216; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm4 # 16-byte Folded Reload3217; AVX2-FCP-NEXT: # xmm4 = xmm15[2],mem[2],xmm15[3],mem[3]3218; AVX2-FCP-NEXT: vmovaps (%rsp), %xmm9 # 16-byte Reload3219; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm9[2,2,2,2]3220; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm12[0,1,2],xmm5[3]3221; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm53222; AVX2-FCP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm63223; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]3224; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]3225; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload3226; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm3 = xmm8[2,2,2,2]3227; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm13[0,1,2],xmm3[3]3228; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm2[0,1],xmm3[2,3]3229; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]3230; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3231; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm3 = xmm14[2],xmm7[2],xmm14[3],xmm7[3]3232; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]3233; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload3234; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload3235; AVX2-FCP-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]3236; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm53237; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3238; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm33239; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]3240; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13241; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]3242; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm63243; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3244; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3245; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3246; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm153247; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]3248; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm143249; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03250; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]3251; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13252; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]3253; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm33254; AVX2-FCP-NEXT: vmovups %ymm3, (%rsp) # 32-byte Spill3255; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm1 = xmm13[2],xmm8[2],xmm13[3],xmm8[3]3256; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]3257; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3258; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3259; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[4],ymm15[4],ymm14[5],ymm15[5]3260; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm03261; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[4],ymm5[4],ymm6[5],ymm5[5]3262; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm11[2,2,2,2]3263; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]3264; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm13265; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3266; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm73267; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm133268; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm13[0],ymm7[0],ymm13[1],ymm7[1],ymm13[4],ymm7[4],ymm13[5],ymm7[5]3269; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm5 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]3270; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm5[0],ymm2[2],ymm5[2]3271; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]3272; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3273; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm93274; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm83275; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]3276; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm03277; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm13278; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3279; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm23280; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3281; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm4 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]3282; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[2,2,2,2]3283; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1],xmm6[2,3]3284; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm03285; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3286; AVX2-FCP-NEXT: vmovaps 448(%rdi), %ymm123287; AVX2-FCP-NEXT: vmovaps 416(%rdi), %ymm63288; AVX2-FCP-NEXT: vmovaps 384(%rdi), %ymm103289; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[4],ymm6[4],ymm10[5],ymm6[5]3290; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm3 = ymm12[0],ymm0[0],ymm12[1],ymm0[1],ymm12[4],ymm0[4],ymm12[5],ymm0[5]3291; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm3[0],ymm1[2],ymm3[2]3292; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]3293; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3294; AVX2-FCP-NEXT: vbroadcastss 148(%rdi), %ymm03295; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]3296; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm5[6,7]3297; AVX2-FCP-NEXT: vextractf128 $1, %ymm11, %xmm23298; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm14[1,1,1,1,5,5,5,5]3299; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm15[1],ymm5[2,3,4],ymm15[5],ymm5[6,7]3300; AVX2-FCP-NEXT: vextractf128 $1, %ymm5, %xmm53301; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]3302; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]3303; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3304; AVX2-FCP-NEXT: vbroadcastss 404(%rdi), %ymm03305; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5],ymm0[6,7]3306; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]3307; AVX2-FCP-NEXT: vextractf128 $1, %ymm4, %xmm13308; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm8[1,1,1,1,5,5,5,5]3309; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm9[1],ymm2[2,3,4],ymm9[5],ymm2[6,7]3310; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm23311; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]3312; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm1[0,1,2,3],ymm0[4,5,6,7]3313; AVX2-FCP-NEXT: vbroadcastss 248(%rdi), %ymm03314; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3315; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]3316; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm13[2],ymm7[2],ymm13[3],ymm7[3],ymm13[6],ymm7[6],ymm13[7],ymm7[7]3317; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm4 = ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[6],ymm15[6],ymm14[7],ymm15[7]3318; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload3319; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload3320; AVX2-FCP-NEXT: # ymm5 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]3321; AVX2-FCP-NEXT: vextractf128 $1, %ymm4, %xmm13322; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm5[2,2,2,2]3323; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2,3]3324; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]3325; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]3326; AVX2-FCP-NEXT: vbroadcastss 504(%rdi), %ymm03327; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5,6],ymm0[7]3328; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[6],ymm6[6],ymm10[7],ymm6[7]3329; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm6 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]3330; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload3331; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload3332; AVX2-FCP-NEXT: # ymm7 = ymm7[2],mem[2],ymm7[3],mem[3],ymm7[6],mem[6],ymm7[7],mem[7]3333; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm83334; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm9 = ymm7[2,2,2,2]3335; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm8 = xmm8[0,1],xmm9[2,3]3336; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm0[6,7]3337; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]3338; AVX2-FCP-NEXT: vbroadcastss 220(%rdi), %ymm83339; AVX2-FCP-NEXT: vunpckhps (%rsp), %ymm8, %ymm8 # 32-byte Folded Reload3340; AVX2-FCP-NEXT: # ymm8 = ymm8[2],mem[2],ymm8[3],mem[3],ymm8[6],mem[6],ymm8[7],mem[7]3341; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm8[1],ymm2[3],ymm8[3]3342; AVX2-FCP-NEXT: vextractf128 $1, %ymm5, %xmm53343; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]3344; AVX2-FCP-NEXT: vextractf128 $1, %ymm4, %xmm43345; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm5[2,3]3346; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]3347; AVX2-FCP-NEXT: vbroadcastss 476(%rdi), %ymm43348; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload3349; AVX2-FCP-NEXT: # ymm4 = ymm4[2],mem[2],ymm4[3],mem[3],ymm4[6],mem[6],ymm4[7],mem[7]3350; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]3351; AVX2-FCP-NEXT: vextractf128 $1, %ymm7, %xmm43352; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm6[2,3,2,3,6,7,6,7]3353; AVX2-FCP-NEXT: vextractf128 $1, %ymm5, %xmm53354; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]3355; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]3356; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3357; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%rsi)3358; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3359; AVX2-FCP-NEXT: vmovaps %ymm4, (%rsi)3360; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3361; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%rdx)3362; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3363; AVX2-FCP-NEXT: vmovaps %ymm4, (%rdx)3364; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3365; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%rcx)3366; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3367; AVX2-FCP-NEXT: vmovaps %ymm4, (%rcx)3368; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3369; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%r8)3370; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3371; AVX2-FCP-NEXT: vmovaps %ymm4, (%r8)3372; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3373; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%r9)3374; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3375; AVX2-FCP-NEXT: vmovaps %ymm4, (%r9)3376; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3377; AVX2-FCP-NEXT: vmovaps %ymm11, 32(%rax)3378; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3379; AVX2-FCP-NEXT: vmovaps %ymm4, (%rax)3380; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3381; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rax)3382; AVX2-FCP-NEXT: vmovaps %ymm1, (%rax)3383; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3384; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rax)3385; AVX2-FCP-NEXT: vmovaps %ymm2, (%rax)3386; AVX2-FCP-NEXT: addq $456, %rsp # imm = 0x1C83387; AVX2-FCP-NEXT: vzeroupper3388; AVX2-FCP-NEXT: retq3389;3390; AVX512-LABEL: load_i32_stride8_vf16:3391; AVX512: # %bb.0:3392; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax3393; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r103394; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r113395; AVX512-NEXT: vmovdqa64 (%rdi), %zmm03396; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm23397; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm13398; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm43399; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm53400; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm33401; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm73402; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm63403; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]3404; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3405; AVX512-NEXT: vmovdqa64 %zmm6, %zmm93406; AVX512-NEXT: vpermt2d %zmm7, %zmm8, %zmm93407; AVX512-NEXT: vmovdqa64 %zmm3, %zmm103408; AVX512-NEXT: vpermt2d %zmm5, %zmm8, %zmm103409; AVX512-NEXT: movb $-64, %dil3410; AVX512-NEXT: kmovw %edi, %k13411; AVX512-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}3412; AVX512-NEXT: vmovdqa64 %zmm1, %zmm93413; AVX512-NEXT: vpermt2d %zmm4, %zmm8, %zmm93414; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm83415; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]3416; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm83417; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]3418; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3419; AVX512-NEXT: vmovdqa64 %zmm6, %zmm103420; AVX512-NEXT: vpermt2d %zmm7, %zmm9, %zmm103421; AVX512-NEXT: vmovdqa64 %zmm3, %zmm113422; AVX512-NEXT: vpermt2d %zmm5, %zmm9, %zmm113423; AVX512-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}3424; AVX512-NEXT: vmovdqa64 %zmm1, %zmm103425; AVX512-NEXT: vpermt2d %zmm4, %zmm9, %zmm103426; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm93427; AVX512-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]3428; AVX512-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm93429; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]3430; AVX512-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3431; AVX512-NEXT: vmovdqa64 %zmm6, %zmm113432; AVX512-NEXT: vpermt2d %zmm7, %zmm10, %zmm113433; AVX512-NEXT: vmovdqa64 %zmm3, %zmm123434; AVX512-NEXT: vpermt2d %zmm5, %zmm10, %zmm123435; AVX512-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}3436; AVX512-NEXT: vmovdqa64 %zmm1, %zmm113437; AVX512-NEXT: vpermt2d %zmm4, %zmm10, %zmm113438; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm103439; AVX512-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]3440; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm103441; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]3442; AVX512-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3443; AVX512-NEXT: vmovdqa64 %zmm6, %zmm123444; AVX512-NEXT: vpermt2d %zmm7, %zmm11, %zmm123445; AVX512-NEXT: vmovdqa64 %zmm3, %zmm133446; AVX512-NEXT: vpermt2d %zmm5, %zmm11, %zmm133447; AVX512-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}3448; AVX512-NEXT: vmovdqa64 %zmm1, %zmm123449; AVX512-NEXT: vpermt2d %zmm4, %zmm11, %zmm123450; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm113451; AVX512-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]3452; AVX512-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm113453; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]3454; AVX512-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3455; AVX512-NEXT: vmovdqa64 %zmm6, %zmm133456; AVX512-NEXT: vpermt2d %zmm7, %zmm12, %zmm133457; AVX512-NEXT: vmovdqa64 %zmm3, %zmm143458; AVX512-NEXT: vpermt2d %zmm5, %zmm12, %zmm143459; AVX512-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}3460; AVX512-NEXT: vmovdqa64 %zmm1, %zmm133461; AVX512-NEXT: vpermt2d %zmm4, %zmm12, %zmm133462; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm123463; AVX512-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]3464; AVX512-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm123465; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]3466; AVX512-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3467; AVX512-NEXT: vmovdqa64 %zmm6, %zmm143468; AVX512-NEXT: vpermt2d %zmm7, %zmm13, %zmm143469; AVX512-NEXT: vmovdqa64 %zmm3, %zmm153470; AVX512-NEXT: vpermt2d %zmm5, %zmm13, %zmm153471; AVX512-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}3472; AVX512-NEXT: vmovdqa64 %zmm1, %zmm143473; AVX512-NEXT: vpermt2d %zmm4, %zmm13, %zmm143474; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm133475; AVX512-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]3476; AVX512-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm133477; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]3478; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3479; AVX512-NEXT: vmovdqa64 %zmm6, %zmm153480; AVX512-NEXT: vpermt2d %zmm7, %zmm14, %zmm153481; AVX512-NEXT: vmovdqa64 %zmm3, %zmm163482; AVX512-NEXT: vpermt2d %zmm5, %zmm14, %zmm163483; AVX512-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}3484; AVX512-NEXT: vmovdqa64 %zmm1, %zmm153485; AVX512-NEXT: vpermt2d %zmm4, %zmm14, %zmm153486; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm143487; AVX512-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]3488; AVX512-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm143489; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]3490; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3491; AVX512-NEXT: vpermt2d %zmm7, %zmm15, %zmm63492; AVX512-NEXT: vpermt2d %zmm5, %zmm15, %zmm33493; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}3494; AVX512-NEXT: vpermt2d %zmm4, %zmm15, %zmm13495; AVX512-NEXT: vpermt2d %zmm2, %zmm15, %zmm03496; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3497; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm03498; AVX512-NEXT: vmovdqa64 %zmm8, (%rsi)3499; AVX512-NEXT: vmovdqa64 %zmm9, (%rdx)3500; AVX512-NEXT: vmovdqa64 %zmm10, (%rcx)3501; AVX512-NEXT: vmovdqa64 %zmm11, (%r8)3502; AVX512-NEXT: vmovdqa64 %zmm12, (%r9)3503; AVX512-NEXT: vmovdqa64 %zmm13, (%r11)3504; AVX512-NEXT: vmovdqa64 %zmm14, (%r10)3505; AVX512-NEXT: vmovdqa64 %zmm0, (%rax)3506; AVX512-NEXT: vzeroupper3507; AVX512-NEXT: retq3508;3509; AVX512-FCP-LABEL: load_i32_stride8_vf16:3510; AVX512-FCP: # %bb.0:3511; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3512; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r103513; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r113514; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm03515; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm23516; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm13517; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm43518; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm53519; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm33520; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm73521; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm63522; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]3523; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3524; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm93525; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm93526; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm103527; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm8, %zmm103528; AVX512-FCP-NEXT: movb $-64, %dil3529; AVX512-FCP-NEXT: kmovw %edi, %k13530; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}3531; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm93532; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm8, %zmm93533; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm83534; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]3535; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm83536; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]3537; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3538; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm103539; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm9, %zmm103540; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm113541; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm9, %zmm113542; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}3543; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm103544; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm9, %zmm103545; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm93546; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]3547; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm93548; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]3549; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3550; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm113551; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm113552; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm123553; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm10, %zmm123554; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}3555; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm113556; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm10, %zmm113557; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm103558; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]3559; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm103560; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]3561; AVX512-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3562; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm123563; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm11, %zmm123564; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm133565; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm11, %zmm133566; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}3567; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm123568; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm11, %zmm123569; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm113570; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]3571; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm113572; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]3573; AVX512-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3574; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm133575; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm12, %zmm133576; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm143577; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm12, %zmm143578; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}3579; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm133580; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm12, %zmm133581; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm123582; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]3583; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm123584; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]3585; AVX512-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3586; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm143587; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm143588; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm153589; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm13, %zmm153590; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}3591; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm143592; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm13, %zmm143593; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm133594; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]3595; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm133596; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]3597; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3598; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm153599; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm14, %zmm153600; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm163601; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm14, %zmm163602; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}3603; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm153604; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm14, %zmm153605; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm143606; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]3607; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm143608; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]3609; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3610; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm63611; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm15, %zmm33612; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}3613; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm15, %zmm13614; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm03615; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3616; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm03617; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)3618; AVX512-FCP-NEXT: vmovdqa64 %zmm9, (%rdx)3619; AVX512-FCP-NEXT: vmovdqa64 %zmm10, (%rcx)3620; AVX512-FCP-NEXT: vmovdqa64 %zmm11, (%r8)3621; AVX512-FCP-NEXT: vmovdqa64 %zmm12, (%r9)3622; AVX512-FCP-NEXT: vmovdqa64 %zmm13, (%r11)3623; AVX512-FCP-NEXT: vmovdqa64 %zmm14, (%r10)3624; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%rax)3625; AVX512-FCP-NEXT: vzeroupper3626; AVX512-FCP-NEXT: retq3627;3628; AVX512DQ-LABEL: load_i32_stride8_vf16:3629; AVX512DQ: # %bb.0:3630; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax3631; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r103632; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r113633; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm03634; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm23635; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm13636; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm43637; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm53638; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm33639; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm73640; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm63641; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]3642; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3643; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm93644; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm8, %zmm93645; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm103646; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm8, %zmm103647; AVX512DQ-NEXT: movb $-64, %dil3648; AVX512DQ-NEXT: kmovw %edi, %k13649; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}3650; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm93651; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm8, %zmm93652; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm83653; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]3654; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm83655; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]3656; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3657; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm103658; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm9, %zmm103659; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm113660; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm9, %zmm113661; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}3662; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm103663; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm9, %zmm103664; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm93665; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]3666; AVX512DQ-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm93667; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]3668; AVX512DQ-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3669; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm113670; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm10, %zmm113671; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm123672; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm10, %zmm123673; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}3674; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm113675; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm10, %zmm113676; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm103677; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]3678; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm103679; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]3680; AVX512DQ-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3681; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm123682; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm11, %zmm123683; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm133684; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm11, %zmm133685; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}3686; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm123687; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm11, %zmm123688; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm113689; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]3690; AVX512DQ-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm113691; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]3692; AVX512DQ-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3693; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm133694; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm12, %zmm133695; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm143696; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm12, %zmm143697; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}3698; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm133699; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm12, %zmm133700; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm123701; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]3702; AVX512DQ-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm123703; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]3704; AVX512DQ-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3705; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm143706; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm13, %zmm143707; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm153708; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm13, %zmm153709; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}3710; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm143711; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm13, %zmm143712; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm133713; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]3714; AVX512DQ-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm133715; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]3716; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3717; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm153718; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm14, %zmm153719; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm163720; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm14, %zmm163721; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}3722; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm153723; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm14, %zmm153724; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm143725; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]3726; AVX512DQ-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm143727; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]3728; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3729; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm15, %zmm63730; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm15, %zmm33731; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}3732; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm15, %zmm13733; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm15, %zmm03734; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3735; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm03736; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rsi)3737; AVX512DQ-NEXT: vmovdqa64 %zmm9, (%rdx)3738; AVX512DQ-NEXT: vmovdqa64 %zmm10, (%rcx)3739; AVX512DQ-NEXT: vmovdqa64 %zmm11, (%r8)3740; AVX512DQ-NEXT: vmovdqa64 %zmm12, (%r9)3741; AVX512DQ-NEXT: vmovdqa64 %zmm13, (%r11)3742; AVX512DQ-NEXT: vmovdqa64 %zmm14, (%r10)3743; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%rax)3744; AVX512DQ-NEXT: vzeroupper3745; AVX512DQ-NEXT: retq3746;3747; AVX512DQ-FCP-LABEL: load_i32_stride8_vf16:3748; AVX512DQ-FCP: # %bb.0:3749; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3750; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r103751; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r113752; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm03753; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm23754; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm13755; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm43756; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm53757; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm33758; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm73759; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm63760; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]3761; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3762; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm93763; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm93764; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm103765; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm8, %zmm103766; AVX512DQ-FCP-NEXT: movb $-64, %dil3767; AVX512DQ-FCP-NEXT: kmovw %edi, %k13768; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}3769; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm93770; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm8, %zmm93771; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm83772; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]3773; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm83774; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]3775; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3776; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm103777; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm9, %zmm103778; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm113779; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm9, %zmm113780; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}3781; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm103782; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm9, %zmm103783; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm93784; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]3785; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm93786; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]3787; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3788; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm113789; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm113790; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm123791; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm10, %zmm123792; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}3793; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm113794; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm10, %zmm113795; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm103796; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]3797; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm103798; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]3799; AVX512DQ-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3800; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm123801; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm11, %zmm123802; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm133803; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm11, %zmm133804; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}3805; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm123806; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm11, %zmm123807; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm113808; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]3809; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm113810; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]3811; AVX512DQ-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3812; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm133813; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm12, %zmm133814; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm143815; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm12, %zmm143816; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}3817; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm133818; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm12, %zmm133819; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm123820; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]3821; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm123822; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]3823; AVX512DQ-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3824; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm143825; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm143826; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm153827; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm13, %zmm153828; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}3829; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm143830; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm13, %zmm143831; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm133832; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]3833; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm133834; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]3835; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3836; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm153837; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm14, %zmm153838; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm163839; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm14, %zmm163840; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}3841; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm153842; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm14, %zmm153843; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm143844; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]3845; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm143846; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]3847; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3848; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm63849; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm15, %zmm33850; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}3851; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm15, %zmm13852; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm03853; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3854; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm03855; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)3856; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, (%rdx)3857; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, (%rcx)3858; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, (%r8)3859; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, (%r9)3860; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, (%r11)3861; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, (%r10)3862; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%rax)3863; AVX512DQ-FCP-NEXT: vzeroupper3864; AVX512DQ-FCP-NEXT: retq3865;3866; AVX512BW-LABEL: load_i32_stride8_vf16:3867; AVX512BW: # %bb.0:3868; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax3869; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r103870; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r113871; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm03872; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm23873; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm13874; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm43875; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm53876; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm33877; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm73878; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm63879; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]3880; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3881; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm93882; AVX512BW-NEXT: vpermt2d %zmm7, %zmm8, %zmm93883; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm103884; AVX512BW-NEXT: vpermt2d %zmm5, %zmm8, %zmm103885; AVX512BW-NEXT: movb $-64, %dil3886; AVX512BW-NEXT: kmovd %edi, %k13887; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}3888; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm93889; AVX512BW-NEXT: vpermt2d %zmm4, %zmm8, %zmm93890; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm83891; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]3892; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm83893; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]3894; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3895; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm103896; AVX512BW-NEXT: vpermt2d %zmm7, %zmm9, %zmm103897; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm113898; AVX512BW-NEXT: vpermt2d %zmm5, %zmm9, %zmm113899; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}3900; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm103901; AVX512BW-NEXT: vpermt2d %zmm4, %zmm9, %zmm103902; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm93903; AVX512BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]3904; AVX512BW-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm93905; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]3906; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3907; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm113908; AVX512BW-NEXT: vpermt2d %zmm7, %zmm10, %zmm113909; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm123910; AVX512BW-NEXT: vpermt2d %zmm5, %zmm10, %zmm123911; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}3912; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm113913; AVX512BW-NEXT: vpermt2d %zmm4, %zmm10, %zmm113914; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm103915; AVX512BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]3916; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm103917; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]3918; AVX512BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3919; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm123920; AVX512BW-NEXT: vpermt2d %zmm7, %zmm11, %zmm123921; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm133922; AVX512BW-NEXT: vpermt2d %zmm5, %zmm11, %zmm133923; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}3924; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm123925; AVX512BW-NEXT: vpermt2d %zmm4, %zmm11, %zmm123926; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm113927; AVX512BW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]3928; AVX512BW-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm113929; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]3930; AVX512BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3931; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm133932; AVX512BW-NEXT: vpermt2d %zmm7, %zmm12, %zmm133933; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm143934; AVX512BW-NEXT: vpermt2d %zmm5, %zmm12, %zmm143935; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}3936; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm133937; AVX512BW-NEXT: vpermt2d %zmm4, %zmm12, %zmm133938; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm123939; AVX512BW-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]3940; AVX512BW-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm123941; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]3942; AVX512BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3943; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm143944; AVX512BW-NEXT: vpermt2d %zmm7, %zmm13, %zmm143945; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm153946; AVX512BW-NEXT: vpermt2d %zmm5, %zmm13, %zmm153947; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}3948; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm143949; AVX512BW-NEXT: vpermt2d %zmm4, %zmm13, %zmm143950; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm133951; AVX512BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]3952; AVX512BW-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm133953; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]3954; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3955; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm153956; AVX512BW-NEXT: vpermt2d %zmm7, %zmm14, %zmm153957; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm163958; AVX512BW-NEXT: vpermt2d %zmm5, %zmm14, %zmm163959; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}3960; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm153961; AVX512BW-NEXT: vpermt2d %zmm4, %zmm14, %zmm153962; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm143963; AVX512BW-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]3964; AVX512BW-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm143965; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]3966; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]3967; AVX512BW-NEXT: vpermt2d %zmm7, %zmm15, %zmm63968; AVX512BW-NEXT: vpermt2d %zmm5, %zmm15, %zmm33969; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}3970; AVX512BW-NEXT: vpermt2d %zmm4, %zmm15, %zmm13971; AVX512BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm03972; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]3973; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm03974; AVX512BW-NEXT: vmovdqa64 %zmm8, (%rsi)3975; AVX512BW-NEXT: vmovdqa64 %zmm9, (%rdx)3976; AVX512BW-NEXT: vmovdqa64 %zmm10, (%rcx)3977; AVX512BW-NEXT: vmovdqa64 %zmm11, (%r8)3978; AVX512BW-NEXT: vmovdqa64 %zmm12, (%r9)3979; AVX512BW-NEXT: vmovdqa64 %zmm13, (%r11)3980; AVX512BW-NEXT: vmovdqa64 %zmm14, (%r10)3981; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rax)3982; AVX512BW-NEXT: vzeroupper3983; AVX512BW-NEXT: retq3984;3985; AVX512BW-FCP-LABEL: load_i32_stride8_vf16:3986; AVX512BW-FCP: # %bb.0:3987; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3988; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r103989; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r113990; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm03991; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm23992; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm13993; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm43994; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm53995; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm33996; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm73997; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm63998; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]3999; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4000; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm94001; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm94002; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm104003; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm8, %zmm104004; AVX512BW-FCP-NEXT: movb $-64, %dil4005; AVX512BW-FCP-NEXT: kmovd %edi, %k14006; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}4007; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm94008; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm8, %zmm94009; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm84010; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]4011; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm84012; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]4013; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4014; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm104015; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm9, %zmm104016; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm114017; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm9, %zmm114018; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}4019; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm104020; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm9, %zmm104021; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm94022; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]4023; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm94024; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]4025; AVX512BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4026; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm114027; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm114028; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm124029; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm10, %zmm124030; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}4031; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm114032; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm10, %zmm114033; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm104034; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]4035; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm104036; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]4037; AVX512BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4038; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm124039; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm11, %zmm124040; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm134041; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm11, %zmm134042; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}4043; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm124044; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm11, %zmm124045; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm114046; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]4047; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm114048; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]4049; AVX512BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4050; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm134051; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm12, %zmm134052; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm144053; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm12, %zmm144054; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}4055; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm134056; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm12, %zmm134057; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm124058; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]4059; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm124060; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]4061; AVX512BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4062; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm144063; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm144064; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm154065; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm13, %zmm154066; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}4067; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm144068; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm13, %zmm144069; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm134070; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]4071; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm134072; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]4073; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4074; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm154075; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm14, %zmm154076; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm164077; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm14, %zmm164078; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}4079; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm154080; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm14, %zmm154081; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm144082; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]4083; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm144084; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]4085; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4086; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm64087; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm15, %zmm34088; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}4089; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm15, %zmm14090; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm04091; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4092; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm04093; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)4094; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, (%rdx)4095; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, (%rcx)4096; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, (%r8)4097; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, (%r9)4098; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, (%r11)4099; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, (%r10)4100; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)4101; AVX512BW-FCP-NEXT: vzeroupper4102; AVX512BW-FCP-NEXT: retq4103;4104; AVX512DQ-BW-LABEL: load_i32_stride8_vf16:4105; AVX512DQ-BW: # %bb.0:4106; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax4107; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r104108; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r114109; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm04110; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm24111; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm14112; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm44113; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm54114; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm34115; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm74116; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm64117; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]4118; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4119; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm94120; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm8, %zmm94121; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm104122; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm8, %zmm104123; AVX512DQ-BW-NEXT: movb $-64, %dil4124; AVX512DQ-BW-NEXT: kmovd %edi, %k14125; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}4126; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm94127; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm8, %zmm94128; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm84129; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]4130; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm84131; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]4132; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4133; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm104134; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm9, %zmm104135; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm114136; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm9, %zmm114137; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}4138; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm104139; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm9, %zmm104140; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm94141; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]4142; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm94143; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]4144; AVX512DQ-BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4145; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm114146; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm10, %zmm114147; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm124148; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm10, %zmm124149; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}4150; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm114151; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm10, %zmm114152; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm104153; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]4154; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm104155; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]4156; AVX512DQ-BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4157; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm124158; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm11, %zmm124159; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm134160; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm11, %zmm134161; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}4162; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm124163; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm11, %zmm124164; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm114165; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]4166; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm114167; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]4168; AVX512DQ-BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4169; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm134170; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm12, %zmm134171; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm144172; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm12, %zmm144173; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}4174; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm134175; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm12, %zmm134176; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm124177; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]4178; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm124179; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]4180; AVX512DQ-BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4181; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm144182; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm13, %zmm144183; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm154184; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm13, %zmm154185; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}4186; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm144187; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm13, %zmm144188; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm134189; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]4190; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm134191; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]4192; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4193; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm154194; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm14, %zmm154195; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm164196; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm14, %zmm164197; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}4198; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm154199; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm14, %zmm154200; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm144201; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]4202; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm144203; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]4204; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4205; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm15, %zmm64206; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm15, %zmm34207; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}4208; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm15, %zmm14209; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm04210; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4211; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm04212; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%rsi)4213; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, (%rdx)4214; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, (%rcx)4215; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, (%r8)4216; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, (%r9)4217; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, (%r11)4218; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, (%r10)4219; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%rax)4220; AVX512DQ-BW-NEXT: vzeroupper4221; AVX512DQ-BW-NEXT: retq4222;4223; AVX512DQ-BW-FCP-LABEL: load_i32_stride8_vf16:4224; AVX512DQ-BW-FCP: # %bb.0:4225; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax4226; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r104227; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r114228; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm04229; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm24230; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm14231; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm44232; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm54233; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm34234; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm74235; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm64236; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]4237; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4238; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm94239; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm94240; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm104241; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm8, %zmm104242; AVX512DQ-BW-FCP-NEXT: movb $-64, %dil4243; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k14244; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm10 {%k1}4245; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm94246; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm8, %zmm94247; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm84248; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]4249; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm84250; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]4251; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4252; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm104253; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm9, %zmm104254; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm114255; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm9, %zmm114256; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm11 {%k1}4257; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm104258; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm9, %zmm104259; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm94260; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]4261; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm11, %zmm94262; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]4263; AVX512DQ-BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4264; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm114265; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm114266; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm124267; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm10, %zmm124268; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm12 {%k1}4269; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm114270; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm10, %zmm114271; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm104272; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]4273; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm12, %zmm104274; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]4275; AVX512DQ-BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4276; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm124277; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm11, %zmm124278; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm134279; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm11, %zmm134280; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm13 {%k1}4281; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm124282; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm11, %zmm124283; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm114284; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]4285; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm13, %zmm114286; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]4287; AVX512DQ-BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4288; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm134289; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm12, %zmm134290; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm144291; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm12, %zmm144292; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm14 {%k1}4293; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm134294; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm12, %zmm134295; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm124296; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6,7]4297; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm12, %zmm14, %zmm124298; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]4299; AVX512DQ-BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4300; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm144301; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm144302; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm154303; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm13, %zmm154304; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}4305; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm144306; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm13, %zmm144307; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm134308; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]4309; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm15, %zmm134310; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]4311; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4312; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm154313; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm14, %zmm154314; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm164315; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm14, %zmm164316; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm16 {%k1}4317; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm154318; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm14, %zmm154319; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm144320; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]4321; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm16, %zmm144322; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm15 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]4323; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]4324; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm64325; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm15, %zmm34326; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}4327; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm15, %zmm14328; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm04329; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4330; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm04331; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)4332; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, (%rdx)4333; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, (%rcx)4334; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, (%r8)4335; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, (%r9)4336; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, (%r11)4337; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, (%r10)4338; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)4339; AVX512DQ-BW-FCP-NEXT: vzeroupper4340; AVX512DQ-BW-FCP-NEXT: retq4341 %wide.vec = load <128 x i32>, ptr %in.vec, align 644342 %strided.vec0 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120>4343 %strided.vec1 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121>4344 %strided.vec2 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122>4345 %strided.vec3 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123>4346 %strided.vec4 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124>4347 %strided.vec5 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125>4348 %strided.vec6 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126>4349 %strided.vec7 = shufflevector <128 x i32> %wide.vec, <128 x i32> poison, <16 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127>4350 store <16 x i32> %strided.vec0, ptr %out.vec0, align 644351 store <16 x i32> %strided.vec1, ptr %out.vec1, align 644352 store <16 x i32> %strided.vec2, ptr %out.vec2, align 644353 store <16 x i32> %strided.vec3, ptr %out.vec3, align 644354 store <16 x i32> %strided.vec4, ptr %out.vec4, align 644355 store <16 x i32> %strided.vec5, ptr %out.vec5, align 644356 store <16 x i32> %strided.vec6, ptr %out.vec6, align 644357 store <16 x i32> %strided.vec7, ptr %out.vec7, align 644358 ret void4359}4360 4361define void @load_i32_stride8_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {4362; SSE-LABEL: load_i32_stride8_vf32:4363; SSE: # %bb.0:4364; SSE-NEXT: subq $952, %rsp # imm = 0x3B84365; SSE-NEXT: movaps 544(%rdi), %xmm54366; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4367; SSE-NEXT: movaps 608(%rdi), %xmm64368; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4369; SSE-NEXT: movaps 576(%rdi), %xmm74370; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4371; SSE-NEXT: movaps 672(%rdi), %xmm84372; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4373; SSE-NEXT: movaps 640(%rdi), %xmm44374; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4375; SSE-NEXT: movaps 736(%rdi), %xmm94376; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4377; SSE-NEXT: movaps 704(%rdi), %xmm34378; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4379; SSE-NEXT: movaps 160(%rdi), %xmm104380; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4381; SSE-NEXT: movaps 128(%rdi), %xmm14382; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4383; SSE-NEXT: movaps 224(%rdi), %xmm24384; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4385; SSE-NEXT: movaps 192(%rdi), %xmm04386; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4387; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]4388; SSE-NEXT: movaps %xmm1, %xmm24389; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1]4390; SSE-NEXT: movaps %xmm2, %xmm14391; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4392; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4393; SSE-NEXT: movaps %xmm3, %xmm14394; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]4395; SSE-NEXT: movaps %xmm4, %xmm34396; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1]4397; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]4398; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4399; SSE-NEXT: movaps %xmm3, %xmm04400; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]4401; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4402; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm1[1]4403; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4404; SSE-NEXT: movaps %xmm7, %xmm04405; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]4406; SSE-NEXT: movaps 512(%rdi), %xmm14407; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4408; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]4409; SSE-NEXT: movaps %xmm1, %xmm24410; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4411; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4412; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4413; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4414; SSE-NEXT: movaps 480(%rdi), %xmm14415; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4416; SSE-NEXT: movaps 448(%rdi), %xmm104417; SSE-NEXT: movaps %xmm10, %xmm04418; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4419; SSE-NEXT: movaps 416(%rdi), %xmm34420; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4421; SSE-NEXT: movaps 384(%rdi), %xmm24422; SSE-NEXT: movaps %xmm2, %xmm14423; SSE-NEXT: movaps %xmm2, %xmm144424; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]4425; 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 992(%rdi), %xmm14431; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4432; SSE-NEXT: movaps 960(%rdi), %xmm154433; SSE-NEXT: movaps %xmm15, %xmm04434; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4435; SSE-NEXT: movaps 928(%rdi), %xmm24436; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4437; SSE-NEXT: movaps 896(%rdi), %xmm14438; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4439; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]4440; SSE-NEXT: movaps %xmm1, %xmm24441; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4442; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4443; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4444; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4445; SSE-NEXT: movaps 352(%rdi), %xmm14446; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4447; SSE-NEXT: movaps 320(%rdi), %xmm124448; SSE-NEXT: movaps %xmm12, %xmm04449; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4450; SSE-NEXT: movaps 288(%rdi), %xmm34451; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4452; SSE-NEXT: movaps 256(%rdi), %xmm14453; SSE-NEXT: movaps %xmm1, %xmm24454; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]4455; SSE-NEXT: movaps %xmm2, %xmm34456; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]4457; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4458; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]4459; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4460; SSE-NEXT: movaps 864(%rdi), %xmm24461; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4462; SSE-NEXT: movaps 832(%rdi), %xmm114463; SSE-NEXT: movaps %xmm11, %xmm04464; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]4465; SSE-NEXT: movaps 800(%rdi), %xmm44466; SSE-NEXT: movaps %xmm4, (%rsp) # 16-byte Spill4467; SSE-NEXT: movaps 768(%rdi), %xmm24468; SSE-NEXT: movaps %xmm2, %xmm34469; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]4470; SSE-NEXT: movaps %xmm3, %xmm44471; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]4472; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4473; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]4474; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4475; SSE-NEXT: movaps 96(%rdi), %xmm64476; SSE-NEXT: movaps 64(%rdi), %xmm94477; SSE-NEXT: movaps %xmm9, %xmm134478; SSE-NEXT: unpcklps {{.*#+}} xmm13 = xmm13[0],xmm6[0],xmm13[1],xmm6[1]4479; SSE-NEXT: movaps (%rdi), %xmm84480; SSE-NEXT: movaps 32(%rdi), %xmm34481; SSE-NEXT: movaps %xmm8, %xmm74482; SSE-NEXT: unpcklps {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1]4483; SSE-NEXT: movaps %xmm7, %xmm54484; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm13[0]4485; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4486; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm13[1]4487; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4488; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload4489; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload4490; SSE-NEXT: # xmm4 = xmm4[2],mem[2],xmm4[3],mem[3]4491; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload4492; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload4493; SSE-NEXT: # xmm5 = xmm5[2],mem[2],xmm5[3],mem[3]4494; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload4495; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload4496; SSE-NEXT: # xmm13 = xmm13[2],mem[2],xmm13[3],mem[3]4497; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload4498; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload4499; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]4500; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4501; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload4502; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload4503; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]4504; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4505; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload4506; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]4507; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload4508; SSE-NEXT: # xmm10 = xmm10[2],mem[2],xmm10[3],mem[3]4509; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload4510; SSE-NEXT: # xmm14 = xmm14[2],mem[2],xmm14[3],mem[3]4511; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4512; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload4513; SSE-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]4514; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload4515; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload4516; SSE-NEXT: # xmm14 = xmm14[2],mem[2],xmm14[3],mem[3]4517; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload4518; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]4519; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload4520; SSE-NEXT: # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]4521; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload4522; SSE-NEXT: # xmm11 = xmm11[2],mem[2],xmm11[3],mem[3]4523; SSE-NEXT: unpckhps (%rsp), %xmm2 # 16-byte Folded Reload4524; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]4525; SSE-NEXT: unpckhps {{.*#+}} xmm9 = xmm9[2],xmm6[2],xmm9[3],xmm6[3]4526; SSE-NEXT: unpckhps {{.*#+}} xmm8 = xmm8[2],xmm3[2],xmm8[3],xmm3[3]4527; SSE-NEXT: movaps %xmm5, %xmm34528; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm4[0]4529; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4530; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm4[1]4531; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4532; SSE-NEXT: movaps %xmm1, %xmm34533; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm12[0]4534; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4535; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm12[1]4536; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4537; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload4538; SSE-NEXT: movaps %xmm6, %xmm14539; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm10[0]4540; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4541; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm10[1]4542; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4543; SSE-NEXT: movaps %xmm0, %xmm34544; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm7[0]4545; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4546; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm7[1]4547; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4548; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4549; SSE-NEXT: movaps %xmm1, %xmm34550; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm13[0]4551; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4552; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm13[1]4553; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4554; SSE-NEXT: movaps %xmm2, %xmm14555; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm11[0]4556; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4557; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm11[1]4558; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4559; SSE-NEXT: movaps %xmm14, %xmm14560; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm15[0]4561; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4562; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm15[1]4563; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4564; SSE-NEXT: movaps %xmm8, %xmm14565; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm9[0]4566; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4567; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm9[1]4568; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4569; SSE-NEXT: movaps 240(%rdi), %xmm24570; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4571; SSE-NEXT: movaps 208(%rdi), %xmm124572; SSE-NEXT: movaps %xmm12, %xmm04573; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]4574; SSE-NEXT: movaps 176(%rdi), %xmm134575; SSE-NEXT: movaps 144(%rdi), %xmm24576; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4577; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1]4578; SSE-NEXT: movaps %xmm2, %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 {{.*#+}} xmm2 = xmm2[1],xmm0[1]4582; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4583; SSE-NEXT: movaps 368(%rdi), %xmm44584; SSE-NEXT: movaps 336(%rdi), %xmm14585; SSE-NEXT: movaps %xmm1, %xmm04586; SSE-NEXT: movaps %xmm1, %xmm94587; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]4588; SSE-NEXT: movaps 304(%rdi), %xmm54589; SSE-NEXT: movaps 272(%rdi), %xmm84590; SSE-NEXT: movaps %xmm8, %xmm14591; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]4592; SSE-NEXT: movaps %xmm1, %xmm24593; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4594; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4595; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4596; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill4597; SSE-NEXT: movaps 496(%rdi), %xmm74598; SSE-NEXT: movaps 464(%rdi), %xmm04599; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4600; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]4601; SSE-NEXT: movaps 432(%rdi), %xmm24602; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4603; SSE-NEXT: movaps 400(%rdi), %xmm14604; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4605; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]4606; SSE-NEXT: movaps %xmm1, %xmm24607; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4608; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4609; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4610; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4611; SSE-NEXT: movaps 624(%rdi), %xmm24612; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4613; SSE-NEXT: movaps 592(%rdi), %xmm14614; SSE-NEXT: movaps %xmm1, %xmm04615; SSE-NEXT: movaps %xmm1, %xmm34616; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]4617; SSE-NEXT: movaps 560(%rdi), %xmm64618; SSE-NEXT: movaps 528(%rdi), %xmm14619; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4620; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1]4621; SSE-NEXT: movaps %xmm1, %xmm24622; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4623; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4624; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4625; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4626; SSE-NEXT: movaps 752(%rdi), %xmm14627; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4628; SSE-NEXT: movaps 720(%rdi), %xmm04629; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4630; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4631; SSE-NEXT: movaps 688(%rdi), %xmm24632; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4633; SSE-NEXT: movaps 656(%rdi), %xmm14634; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4635; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]4636; SSE-NEXT: movaps %xmm1, %xmm24637; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4638; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4639; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]4640; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4641; SSE-NEXT: movaps 880(%rdi), %xmm14642; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4643; SSE-NEXT: movaps 848(%rdi), %xmm144644; SSE-NEXT: movaps %xmm14, %xmm04645; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4646; SSE-NEXT: movaps 816(%rdi), %xmm14647; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4648; SSE-NEXT: movaps 784(%rdi), %xmm24649; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4650; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4651; SSE-NEXT: movaps %xmm2, %xmm14652; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4653; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4654; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]4655; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4656; SSE-NEXT: movaps 1008(%rdi), %xmm14657; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4658; SSE-NEXT: movaps 976(%rdi), %xmm114659; SSE-NEXT: movaps %xmm11, %xmm04660; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4661; SSE-NEXT: movaps 944(%rdi), %xmm14662; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4663; SSE-NEXT: movaps 912(%rdi), %xmm24664; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4665; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4666; SSE-NEXT: movaps %xmm2, %xmm14667; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4668; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4669; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]4670; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4671; SSE-NEXT: movaps 112(%rdi), %xmm14672; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4673; SSE-NEXT: movaps 80(%rdi), %xmm104674; SSE-NEXT: movaps %xmm10, %xmm04675; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4676; SSE-NEXT: movaps 16(%rdi), %xmm154677; SSE-NEXT: movaps 48(%rdi), %xmm14678; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4679; SSE-NEXT: movaps %xmm15, %xmm24680; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4681; SSE-NEXT: movaps %xmm2, %xmm14682; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]4683; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4684; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]4685; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4686; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload4687; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]4688; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4689; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm13[2],xmm1[3],xmm13[3]4690; SSE-NEXT: movaps %xmm9, %xmm134691; SSE-NEXT: unpckhps {{.*#+}} xmm13 = xmm13[2],xmm4[2],xmm13[3],xmm4[3]4692; SSE-NEXT: unpckhps {{.*#+}} xmm8 = xmm8[2],xmm5[2],xmm8[3],xmm5[3]4693; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload4694; SSE-NEXT: unpckhps {{.*#+}} xmm9 = xmm9[2],xmm7[2],xmm9[3],xmm7[3]4695; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload4696; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload4697; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]4698; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload4699; SSE-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]4700; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4701; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload4702; SSE-NEXT: unpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]4703; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload4704; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload4705; SSE-NEXT: # xmm6 = xmm6[2],mem[2],xmm6[3],mem[3]4706; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload4707; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload4708; SSE-NEXT: # xmm4 = xmm4[2],mem[2],xmm4[3],mem[3]4709; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload4710; SSE-NEXT: # xmm14 = xmm14[2],mem[2],xmm14[3],mem[3]4711; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload4712; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload4713; SSE-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]4714; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload4715; SSE-NEXT: # xmm11 = xmm11[2],mem[2],xmm11[3],mem[3]4716; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload4717; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload4718; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]4719; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload4720; SSE-NEXT: # xmm10 = xmm10[2],mem[2],xmm10[3],mem[3]4721; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload4722; SSE-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]4723; SSE-NEXT: movaps %xmm1, %xmm04724; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm12[0]4725; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4726; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm12[1]4727; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4728; SSE-NEXT: movaps %xmm8, %xmm14729; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm13[0]4730; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4731; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm13[1]4732; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4733; SSE-NEXT: movaps %xmm7, %xmm124734; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm9[0]4735; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm9[1]4736; SSE-NEXT: movaps %xmm5, %xmm134737; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4738; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm1[0]4739; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm1[1]4740; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4741; SSE-NEXT: movaps %xmm4, %xmm04742; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm6[0]4743; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm6[1]4744; SSE-NEXT: movaps %xmm0, %xmm54745; SSE-NEXT: movaps %xmm3, %xmm84746; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm14[0]4747; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm14[1]4748; SSE-NEXT: movaps %xmm3, %xmm64749; SSE-NEXT: movaps %xmm2, %xmm34750; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm11[0]4751; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm11[1]4752; SSE-NEXT: movaps %xmm15, %xmm04753; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm10[0]4754; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm10[1]4755; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4756; SSE-NEXT: movaps %xmm1, 96(%rsi)4757; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4758; SSE-NEXT: movaps %xmm1, 32(%rsi)4759; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4760; SSE-NEXT: movaps %xmm1, 112(%rsi)4761; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4762; SSE-NEXT: movaps %xmm1, 48(%rsi)4763; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4764; SSE-NEXT: movaps %xmm1, 64(%rsi)4765; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4766; SSE-NEXT: movaps %xmm1, (%rsi)4767; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4768; SSE-NEXT: movaps %xmm1, 80(%rsi)4769; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4770; SSE-NEXT: movaps %xmm1, 16(%rsi)4771; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4772; SSE-NEXT: movaps %xmm1, 96(%rdx)4773; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4774; SSE-NEXT: movaps %xmm1, 32(%rdx)4775; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4776; SSE-NEXT: movaps %xmm1, 112(%rdx)4777; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4778; SSE-NEXT: movaps %xmm1, 48(%rdx)4779; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4780; SSE-NEXT: movaps %xmm1, 64(%rdx)4781; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4782; SSE-NEXT: movaps %xmm1, (%rdx)4783; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4784; SSE-NEXT: movaps %xmm1, 80(%rdx)4785; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4786; SSE-NEXT: movaps %xmm1, 16(%rdx)4787; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4788; SSE-NEXT: movaps %xmm1, 96(%rcx)4789; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4790; SSE-NEXT: movaps %xmm1, 32(%rcx)4791; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4792; SSE-NEXT: movaps %xmm1, 112(%rcx)4793; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4794; SSE-NEXT: movaps %xmm1, 48(%rcx)4795; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4796; SSE-NEXT: movaps %xmm1, 64(%rcx)4797; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4798; SSE-NEXT: movaps %xmm1, (%rcx)4799; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4800; SSE-NEXT: movaps %xmm1, 80(%rcx)4801; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4802; SSE-NEXT: movaps %xmm1, 16(%rcx)4803; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4804; SSE-NEXT: movaps %xmm1, 112(%r8)4805; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4806; SSE-NEXT: movaps %xmm1, 96(%r8)4807; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4808; SSE-NEXT: movaps %xmm1, 80(%r8)4809; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4810; SSE-NEXT: movaps %xmm1, 64(%r8)4811; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4812; SSE-NEXT: movaps %xmm1, 48(%r8)4813; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4814; SSE-NEXT: movaps %xmm1, 32(%r8)4815; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4816; SSE-NEXT: movaps %xmm1, 16(%r8)4817; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4818; SSE-NEXT: movaps %xmm1, (%r8)4819; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4820; SSE-NEXT: movaps %xmm1, 112(%r9)4821; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4822; SSE-NEXT: movaps %xmm1, 96(%r9)4823; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4824; SSE-NEXT: movaps %xmm1, 80(%r9)4825; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4826; SSE-NEXT: movaps %xmm1, 64(%r9)4827; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4828; SSE-NEXT: movaps %xmm1, 48(%r9)4829; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4830; SSE-NEXT: movaps %xmm1, 32(%r9)4831; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4832; SSE-NEXT: movaps %xmm1, 16(%r9)4833; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4834; SSE-NEXT: movaps %xmm1, (%r9)4835; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax4836; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4837; SSE-NEXT: movaps %xmm1, 112(%rax)4838; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4839; SSE-NEXT: movaps %xmm1, 96(%rax)4840; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4841; SSE-NEXT: movaps %xmm1, 80(%rax)4842; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4843; SSE-NEXT: movaps %xmm1, 64(%rax)4844; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4845; SSE-NEXT: movaps %xmm1, 48(%rax)4846; SSE-NEXT: movaps (%rsp), %xmm1 # 16-byte Reload4847; SSE-NEXT: movaps %xmm1, 32(%rax)4848; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4849; SSE-NEXT: movaps %xmm1, 16(%rax)4850; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4851; SSE-NEXT: movaps %xmm1, (%rax)4852; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax4853; SSE-NEXT: movaps %xmm3, 112(%rax)4854; SSE-NEXT: movaps %xmm8, 96(%rax)4855; SSE-NEXT: movaps %xmm4, 80(%rax)4856; SSE-NEXT: movaps %xmm13, 64(%rax)4857; SSE-NEXT: movaps %xmm12, 48(%rax)4858; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4859; SSE-NEXT: movaps %xmm1, 32(%rax)4860; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4861; SSE-NEXT: movaps %xmm1, 16(%rax)4862; SSE-NEXT: movaps %xmm0, (%rax)4863; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax4864; SSE-NEXT: movaps %xmm2, 112(%rax)4865; SSE-NEXT: movaps %xmm6, 96(%rax)4866; SSE-NEXT: movaps %xmm5, 80(%rax)4867; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4868; SSE-NEXT: movaps %xmm0, 64(%rax)4869; SSE-NEXT: movaps %xmm7, 48(%rax)4870; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4871; SSE-NEXT: movaps %xmm0, 32(%rax)4872; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4873; SSE-NEXT: movaps %xmm0, 16(%rax)4874; SSE-NEXT: movaps %xmm15, (%rax)4875; SSE-NEXT: addq $952, %rsp # imm = 0x3B84876; SSE-NEXT: retq4877;4878; AVX-LABEL: load_i32_stride8_vf32:4879; AVX: # %bb.0:4880; AVX-NEXT: subq $1768, %rsp # imm = 0x6E84881; AVX-NEXT: vmovaps 288(%rdi), %xmm144882; AVX-NEXT: vmovaps 256(%rdi), %xmm104883; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm10[0],xmm14[0],xmm10[1],xmm14[1]4884; AVX-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4885; AVX-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4886; AVX-NEXT: vmovaps 352(%rdi), %xmm14887; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4888; AVX-NEXT: vmovaps 320(%rdi), %xmm24889; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4890; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4891; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4892; AVX-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]4893; AVX-NEXT: vmovaps 416(%rdi), %xmm14894; AVX-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill4895; AVX-NEXT: vmovaps 384(%rdi), %xmm24896; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4897; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4898; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14899; AVX-NEXT: vmovaps 480(%rdi), %xmm24900; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4901; AVX-NEXT: vmovaps 448(%rdi), %xmm34902; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4903; AVX-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]4904; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4905; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,0,1]4906; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm24907; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]4908; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4909; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4910; AVX-NEXT: vmovaps 928(%rdi), %xmm04911; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4912; AVX-NEXT: vmovaps 896(%rdi), %xmm14913; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4914; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4915; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm04916; AVX-NEXT: vmovaps 992(%rdi), %xmm14917; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4918; AVX-NEXT: vmovaps 960(%rdi), %xmm24919; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4920; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4921; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm9[0,1,0,1]4922; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14923; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4924; AVX-NEXT: vmovaps 800(%rdi), %xmm14925; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4926; AVX-NEXT: vmovaps 768(%rdi), %xmm114927; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm11[0],xmm1[0],xmm11[1],xmm1[1]4928; AVX-NEXT: vmovaps 864(%rdi), %xmm24929; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4930; AVX-NEXT: vmovaps 832(%rdi), %xmm34931; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4932; AVX-NEXT: vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]4933; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm7[0]4934; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4935; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4936; AVX-NEXT: vmovaps 160(%rdi), %xmm14937; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4938; AVX-NEXT: vmovaps 128(%rdi), %xmm04939; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4940; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]4941; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm04942; AVX-NEXT: vmovaps 224(%rdi), %xmm14943; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4944; AVX-NEXT: vmovaps 192(%rdi), %xmm24945; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4946; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4947; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4948; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1]4949; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14950; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4951; AVX-NEXT: vmovaps 32(%rdi), %xmm14952; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4953; AVX-NEXT: vmovaps (%rdi), %xmm134954; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1]4955; AVX-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4956; AVX-NEXT: vmovaps 96(%rdi), %xmm24957; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4958; AVX-NEXT: vmovaps 64(%rdi), %xmm34959; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4960; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]4961; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]4962; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4963; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4964; AVX-NEXT: vmovaps 672(%rdi), %xmm124965; AVX-NEXT: vmovaps 640(%rdi), %xmm04966; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4967; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1]4968; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm04969; AVX-NEXT: vmovaps 736(%rdi), %xmm14970; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4971; AVX-NEXT: vmovaps 704(%rdi), %xmm24972; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4973; AVX-NEXT: vunpcklps {{.*#+}} xmm8 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4974; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm8[0,1,0,1]4975; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm44976; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm4[6,7]4977; AVX-NEXT: vmovaps 544(%rdi), %xmm64978; AVX-NEXT: vmovaps 512(%rdi), %xmm34979; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm3[0],xmm6[0],xmm3[1],xmm6[1]4980; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4981; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4982; AVX-NEXT: vmovaps 608(%rdi), %xmm44983; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4984; AVX-NEXT: vmovaps 576(%rdi), %xmm24985; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4986; AVX-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]4987; AVX-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]4988; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4989; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4990; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm10[1,1,1,1]4991; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]4992; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload4993; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]4994; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload4995; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload4996; AVX-NEXT: vshufps {{.*#+}} xmm15 = xmm4[1,1,1,1]4997; AVX-NEXT: vmovaps (%rsp), %xmm14 # 16-byte Reload4998; AVX-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0],xmm14[1],xmm15[2,3]4999; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm155000; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm15[0,1,2,3,4,5],ymm1[6,7]5001; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5002; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5003; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1,1,1]5004; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload5005; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]5006; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm7[2,3]5007; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm15008; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload5009; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm9[1,1,1,1]5010; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5011; AVX-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0],xmm15[1],xmm7[2,3]5012; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm75013; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm1[6,7]5014; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5015; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5016; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm13[1,1,1,1]5017; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5018; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]5019; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]5020; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload5021; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload5022; AVX-NEXT: # xmm5 = mem[1,1,1,1]5023; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload5024; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0],xmm13[1],xmm5[2,3]5025; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm55026; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5],ymm1[6,7]5027; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5028; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5029; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1,1,1]5030; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm6[1],xmm0[2,3]5031; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]5032; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm15033; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload5034; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm5[1,1,1,1]5035; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm12[1],xmm2[2,3]5036; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25037; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5038; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5039; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5040; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5041; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload5042; AVX-NEXT: # xmm3 = xmm0[2],mem[2],xmm0[3],mem[3]5043; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5044; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm4[2],xmm14[2],xmm4[3],xmm14[3]5045; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5046; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5047; AVX-NEXT: # xmm0 = mem[2,2,2,2]5048; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5049; AVX-NEXT: # xmm0 = mem[0,1,2],xmm0[3]5050; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm05051; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm25052; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]5053; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload5054; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm7[2,2,2,2]5055; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5056; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm1[0,1,2],xmm2[3]5057; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]5058; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]5059; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5060; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]5061; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5062; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm9[2],xmm15[2],xmm9[3],xmm15[3]5063; AVX-NEXT: vmovaps %xmm0, (%rsp) # 16-byte Spill5064; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload5065; AVX-NEXT: # xmm2 = mem[2,2,2,2]5066; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload5067; AVX-NEXT: # xmm2 = mem[0,1,2],xmm2[3]5068; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25069; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm35070; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]5071; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload5072; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm14[2,2,2,2]5073; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5074; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm0[0,1,2],xmm3[3]5075; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]5076; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]5077; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5078; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5079; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm4 # 16-byte Folded Reload5080; AVX-NEXT: # xmm4 = xmm2[2],mem[2],xmm2[3],mem[3]5081; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5082; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5083; AVX-NEXT: vunpckhps {{.*#+}} xmm3 = xmm2[2],xmm13[2],xmm2[3],xmm13[3]5084; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5085; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5086; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,2,2,2]5087; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload5088; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm13[0,1,2],xmm2[3]5089; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25090; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm65091; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm2[6,7]5092; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload5093; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm9[2,2,2,2]5094; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5095; AVX-NEXT: vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]5096; AVX-NEXT: vblendps {{.*#+}} xmm6 = xmm4[0,1],xmm6[2,3]5097; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm2[4,5,6,7]5098; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5099; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5100; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm11 # 16-byte Folded Reload5101; AVX-NEXT: # xmm11 = xmm2[2],mem[2],xmm2[3],mem[3]5102; AVX-NEXT: vunpckhps {{.*#+}} xmm12 = xmm5[2],xmm12[2],xmm5[3],xmm12[3]5103; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload5104; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm10[2,2,2,2]5105; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload5106; AVX-NEXT: vblendps {{.*#+}} xmm6 = xmm5[0,1,2],xmm6[3]5107; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm65108; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm85109; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]5110; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload5111; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm4[2,2,2,2]5112; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5113; AVX-NEXT: vblendps {{.*#+}} xmm8 = xmm2[0,1,2],xmm8[3]5114; AVX-NEXT: vblendps {{.*#+}} xmm8 = xmm11[0,1],xmm8[2,3]5115; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]5116; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5117; AVX-NEXT: vunpckhps {{.*#+}} xmm6 = xmm1[2],xmm7[2],xmm1[3],xmm7[3]5118; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5119; AVX-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm1[1],xmm6[1]5120; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5121; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm7 # 16-byte Folded Reload5122; AVX-NEXT: # xmm7 = xmm1[2],mem[2],xmm1[3],mem[3]5123; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm75124; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload5125; AVX-NEXT: # xmm1 = mem[2,3,2,3]5126; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15127; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm7[6,7]5128; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]5129; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5130; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm0[2],xmm14[2],xmm0[3],xmm14[3]5131; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5132; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm0[1],xmm1[1]5133; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5134; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm6 # 16-byte Folded Reload5135; AVX-NEXT: # xmm6 = xmm0[2],mem[2],xmm0[3],mem[3]5136; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm65137; AVX-NEXT: vpermilps $238, (%rsp), %xmm0 # 16-byte Folded Reload5138; AVX-NEXT: # xmm0 = mem[2,3,2,3]5139; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm05140; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm6[6,7]5141; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5142; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5143; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm3[2],xmm9[2],xmm3[3],xmm9[3]5144; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5145; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]5146; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm13[2],xmm15[2],xmm13[3],xmm15[3]5147; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15148; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload5149; AVX-NEXT: # xmm3 = mem[2,3,2,3]5150; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm35151; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]5152; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5153; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5154; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]5155; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm11[1],xmm0[1]5156; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm10[2],xmm5[3],xmm10[3]5157; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15158; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm12[2,3,2,3]5159; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25160; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5161; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5162; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5163; AVX-NEXT: vmovaps 416(%rdi), %ymm125164; AVX-NEXT: vmovaps 384(%rdi), %ymm95165; AVX-NEXT: vmovaps 448(%rdi), %ymm75166; AVX-NEXT: vmovaps 480(%rdi), %ymm115167; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm7[0],ymm11[2],ymm7[2]5168; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5169; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm9[0],ymm12[0],ymm9[1],ymm12[1],ymm9[4],ymm12[4],ymm9[5],ymm12[5]5170; AVX-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5171; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5172; AVX-NEXT: vmovaps 320(%rdi), %ymm25173; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5174; AVX-NEXT: vmovaps 352(%rdi), %ymm15175; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5176; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]5177; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15178; AVX-NEXT: vmovaps 288(%rdi), %ymm25179; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5180; AVX-NEXT: vmovaps 256(%rdi), %ymm105181; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm10[0],ymm2[0],ymm10[1],ymm2[1],ymm10[4],ymm2[4],ymm10[5],ymm2[5]5182; AVX-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5183; AVX-NEXT: vextractf128 $1, %ymm2, %xmm25184; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]5185; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5186; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5187; AVX-NEXT: vmovaps 928(%rdi), %ymm55188; AVX-NEXT: vmovaps 896(%rdi), %ymm35189; AVX-NEXT: vmovups %ymm3, (%rsp) # 32-byte Spill5190; AVX-NEXT: vmovaps 960(%rdi), %ymm15191; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5192; AVX-NEXT: vmovaps 992(%rdi), %ymm05193; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5194; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5195; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm5[0],ymm3[1],ymm5[1],ymm3[4],ymm5[4],ymm3[5],ymm5[5]5196; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5197; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5198; AVX-NEXT: vmovaps 832(%rdi), %ymm25199; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5200; AVX-NEXT: vmovaps 864(%rdi), %ymm15201; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5202; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]5203; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15204; AVX-NEXT: vmovaps 800(%rdi), %ymm35205; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5206; AVX-NEXT: vmovaps 768(%rdi), %ymm25207; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5208; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]5209; AVX-NEXT: vextractf128 $1, %ymm2, %xmm25210; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]5211; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5212; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5213; AVX-NEXT: vmovaps 672(%rdi), %ymm25214; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5215; AVX-NEXT: vmovaps 640(%rdi), %ymm35216; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5217; AVX-NEXT: vmovaps 704(%rdi), %ymm135218; AVX-NEXT: vmovaps 736(%rdi), %ymm05219; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5220; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm13[0],ymm0[2],ymm13[2]5221; AVX-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5222; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]5223; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5224; AVX-NEXT: vmovaps 576(%rdi), %ymm15225; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5226; AVX-NEXT: vmovaps 608(%rdi), %ymm05227; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5228; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]5229; AVX-NEXT: vextractf128 $1, %ymm0, %xmm65230; AVX-NEXT: vmovaps 544(%rdi), %ymm05231; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5232; AVX-NEXT: vmovaps 512(%rdi), %ymm15233; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5234; AVX-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]5235; AVX-NEXT: vextractf128 $1, %ymm8, %xmm85236; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm8[0,1],xmm6[2,0]5237; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm4[4,5,6,7]5238; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5239; AVX-NEXT: vmovaps 160(%rdi), %ymm05240; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5241; AVX-NEXT: vmovaps 128(%rdi), %ymm15242; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5243; AVX-NEXT: vmovaps 192(%rdi), %ymm25244; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5245; AVX-NEXT: vmovaps 224(%rdi), %ymm35246; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5247; AVX-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]5248; AVX-NEXT: vunpcklps {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]5249; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm6[0,1],ymm4[2,0],ymm6[4,5],ymm4[6,4]5250; AVX-NEXT: vmovaps 64(%rdi), %ymm05251; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5252; AVX-NEXT: vmovaps 96(%rdi), %ymm25253; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5254; AVX-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]5255; AVX-NEXT: vextractf128 $1, %ymm4, %xmm05256; AVX-NEXT: vmovaps (%rdi), %ymm25257; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5258; AVX-NEXT: vmovaps 32(%rdi), %ymm145259; AVX-NEXT: vunpcklps {{.*#+}} ymm15 = ymm2[0],ymm14[0],ymm2[1],ymm14[1],ymm2[4],ymm14[4],ymm2[5],ymm14[5]5260; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155261; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm15[0,1],xmm0[2,0]5262; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5263; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5264; AVX-NEXT: vmovaps %ymm11, %ymm65265; AVX-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5266; AVX-NEXT: vunpcklps {{.*#+}} ymm0 = ymm7[0],ymm11[0],ymm7[1],ymm11[1],ymm7[4],ymm11[4],ymm7[5],ymm11[5]5267; AVX-NEXT: vmovaps %ymm12, %ymm75268; AVX-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5269; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,0],ymm9[1,0],ymm12[5,4],ymm9[5,4]5270; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]5271; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload5272; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5273; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm9[0],ymm3[1],ymm9[1],ymm3[4],ymm9[4],ymm3[5],ymm9[5]5274; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15275; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload5276; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm11[1,0],ymm10[1,0],ymm11[5,4],ymm10[5,4]5277; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155278; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[2,0],xmm1[2,3]5279; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5280; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5281; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload5282; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5283; AVX-NEXT: vunpcklps {{.*#+}} ymm0 = ymm2[0],ymm10[0],ymm2[1],ymm10[1],ymm2[4],ymm10[4],ymm2[5],ymm10[5]5284; AVX-NEXT: vmovups (%rsp), %ymm12 # 32-byte Reload5285; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm5[1,0],ymm12[1,0],ymm5[5,4],ymm12[5,4]5286; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]5287; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5288; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm1 # 32-byte Folded Reload5289; AVX-NEXT: # ymm1 = ymm5[0],mem[0],ymm5[1],mem[1],ymm5[4],mem[4],ymm5[5],mem[5]5290; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15291; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload5292; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm15 # 32-byte Folded Reload5293; AVX-NEXT: # ymm15 = ymm8[1,0],mem[1,0],ymm8[5,4],mem[5,4]5294; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155295; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[2,0],xmm1[2,3]5296; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5297; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5298; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5299; AVX-NEXT: vunpcklps {{.*#+}} ymm0 = ymm13[0],ymm4[0],ymm13[1],ymm4[1],ymm13[4],ymm4[4],ymm13[5],ymm4[5]5300; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload5301; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm1 # 32-byte Folded Reload5302; AVX-NEXT: # ymm1 = ymm13[1,0],mem[1,0],ymm13[5,4],mem[5,4]5303; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]5304; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5305; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5306; AVX-NEXT: # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[4],mem[4],ymm1[5],mem[5]5307; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15308; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload5309; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload5310; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]5311; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155312; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[2,0],xmm1[2,3]5313; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5314; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5315; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5316; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5317; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]5318; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5319; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5320; AVX-NEXT: # ymm1 = ymm1[1,0],mem[1,0],ymm1[5,4],mem[5,4]5321; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]5322; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5323; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5324; AVX-NEXT: # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[4],mem[4],ymm1[5],mem[5]5325; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15326; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm15 # 32-byte Folded Reload5327; AVX-NEXT: # ymm15 = ymm14[1,0],mem[1,0],ymm14[5,4],mem[5,4]5328; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155329; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[2,0],xmm1[2,3]5330; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5331; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5332; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm0 # 32-byte Folded Reload5333; AVX-NEXT: # ymm0 = ymm6[1],mem[1],ymm6[3],mem[3]5334; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5335; AVX-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]5336; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5337; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm3[1],ymm9[3],ymm3[3]5338; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15339; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload5340; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm9[2],ymm11[2],ymm9[3],ymm11[3],ymm9[6],ymm11[6],ymm9[7],ymm11[7]5341; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155342; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,0]5343; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5344; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5345; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm2[1],ymm10[3],ymm2[3]5346; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm1 # 32-byte Folded Reload5347; AVX-NEXT: # ymm1 = ymm12[2],mem[2],ymm12[3],mem[3],ymm12[6],mem[6],ymm12[7],mem[7]5348; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5349; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5350; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm5[1],ymm1[3],ymm5[3]5351; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15352; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5353; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm2[2],ymm8[2],ymm2[3],ymm8[3],ymm2[6],ymm8[6],ymm2[7],ymm8[7]5354; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155355; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,0]5356; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5357; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5358; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5359; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]5360; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5361; AVX-NEXT: vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm13[2],ymm4[3],ymm13[3],ymm4[6],ymm13[6],ymm4[7],ymm13[7]5362; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5363; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload5364; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5365; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm6[1],ymm2[1],ymm6[3],ymm2[3]5366; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15367; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload5368; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5369; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm7[2],ymm3[2],ymm7[3],ymm3[3],ymm7[6],ymm3[6],ymm7[7],ymm3[7]5370; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155371; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,0]5372; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5373; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5374; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload5375; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5376; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]5377; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload5378; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload5379; AVX-NEXT: vunpckhps {{.*#+}} ymm1 = ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[6],ymm10[6],ymm12[7],ymm10[7]5380; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]5381; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload5382; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5383; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]5384; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15385; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload5386; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[6],ymm14[6],ymm13[7],ymm14[7]5387; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155388; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,0]5389; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5390; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5391; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5392; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5393; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]5394; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5395; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5396; AVX-NEXT: # ymm1 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]5397; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]5398; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5399; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5400; AVX-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]5401; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload5402; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm15[3,0],ymm9[3,0],ymm15[7,4],ymm9[7,4]5403; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15404; AVX-NEXT: vextractf128 $1, %ymm15, %xmm155405; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm15[2,0],xmm1[2,3]5406; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]5407; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload5408; AVX-NEXT: # ymm0 = ymm5[2],mem[2],ymm5[3],mem[3],ymm5[6],mem[6],ymm5[7],mem[7]5409; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5410; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,0],ymm4[3,0],ymm1[7,4],ymm4[7,4]5411; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]5412; AVX-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm6[2],ymm2[3],ymm6[3],ymm2[6],ymm6[6],ymm2[7],ymm6[7]5413; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,0],ymm7[3,0],ymm3[7,4],ymm7[7,4]5414; AVX-NEXT: vextractf128 $1, %ymm1, %xmm15415; AVX-NEXT: vextractf128 $1, %ymm2, %xmm25416; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]5417; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5418; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5419; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5420; AVX-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]5421; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5422; AVX-NEXT: vshufps $51, (%rsp), %ymm2, %ymm2 # 32-byte Folded Reload5423; AVX-NEXT: # ymm2 = ymm2[3,0],mem[3,0],ymm2[7,4],mem[7,4]5424; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[2,3],ymm2[6,4],ymm1[6,7]5425; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5426; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload5427; AVX-NEXT: # ymm2 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]5428; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5429; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload5430; AVX-NEXT: # ymm3 = ymm3[3,0],mem[3,0],ymm3[7,4],mem[7,4]5431; AVX-NEXT: vextractf128 $1, %ymm2, %xmm25432; AVX-NEXT: vextractf128 $1, %ymm3, %xmm35433; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm3[2,0],xmm2[2,3]5434; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]5435; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm2 # 32-byte Folded Reload5436; AVX-NEXT: # ymm2 = ymm8[2],mem[2],ymm8[3],mem[3],ymm8[6],mem[6],ymm8[7],mem[7]5437; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm10[3,0],ymm12[3,0],ymm10[7,4],ymm12[7,4]5438; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,0],ymm2[2,3],ymm3[6,4],ymm2[6,7]5439; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm3 # 32-byte Folded Reload5440; AVX-NEXT: # ymm3 = ymm11[2],mem[2],ymm11[3],mem[3],ymm11[6],mem[6],ymm11[7],mem[7]5441; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm14[3,0],ymm13[3,0],ymm14[7,4],ymm13[7,4]5442; AVX-NEXT: vextractf128 $1, %ymm3, %xmm35443; AVX-NEXT: vextractf128 $1, %ymm4, %xmm45444; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm4[2,0],xmm3[2,3]5445; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]5446; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5447; AVX-NEXT: vmovaps %ymm3, 64(%rsi)5448; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5449; AVX-NEXT: vmovaps %ymm3, (%rsi)5450; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5451; AVX-NEXT: vmovaps %ymm3, 96(%rsi)5452; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5453; AVX-NEXT: vmovaps %ymm3, 32(%rsi)5454; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5455; AVX-NEXT: vmovaps %ymm3, 64(%rdx)5456; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5457; AVX-NEXT: vmovaps %ymm3, (%rdx)5458; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5459; AVX-NEXT: vmovaps %ymm3, 96(%rdx)5460; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5461; AVX-NEXT: vmovaps %ymm3, 32(%rdx)5462; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5463; AVX-NEXT: vmovaps %ymm3, 64(%rcx)5464; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5465; AVX-NEXT: vmovaps %ymm3, (%rcx)5466; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5467; AVX-NEXT: vmovaps %ymm3, 96(%rcx)5468; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5469; AVX-NEXT: vmovaps %ymm3, 32(%rcx)5470; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5471; AVX-NEXT: vmovaps %ymm3, 64(%r8)5472; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5473; AVX-NEXT: vmovaps %ymm3, (%r8)5474; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5475; AVX-NEXT: vmovaps %ymm3, 96(%r8)5476; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5477; AVX-NEXT: vmovaps %ymm3, 32(%r8)5478; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5479; AVX-NEXT: vmovaps %ymm3, 64(%r9)5480; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5481; AVX-NEXT: vmovaps %ymm3, (%r9)5482; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5483; AVX-NEXT: vmovaps %ymm3, 96(%r9)5484; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5485; AVX-NEXT: vmovaps %ymm3, 32(%r9)5486; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax5487; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5488; AVX-NEXT: vmovaps %ymm3, 64(%rax)5489; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5490; AVX-NEXT: vmovaps %ymm3, (%rax)5491; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5492; AVX-NEXT: vmovaps %ymm3, 96(%rax)5493; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5494; AVX-NEXT: vmovaps %ymm3, 32(%rax)5495; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax5496; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5497; AVX-NEXT: vmovaps %ymm3, 64(%rax)5498; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5499; AVX-NEXT: vmovaps %ymm3, (%rax)5500; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5501; AVX-NEXT: vmovaps %ymm3, 96(%rax)5502; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5503; AVX-NEXT: vmovaps %ymm3, 32(%rax)5504; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax5505; AVX-NEXT: vmovaps %ymm1, 96(%rax)5506; AVX-NEXT: vmovaps %ymm0, 64(%rax)5507; AVX-NEXT: vmovaps %ymm9, 32(%rax)5508; AVX-NEXT: vmovaps %ymm2, (%rax)5509; AVX-NEXT: addq $1768, %rsp # imm = 0x6E85510; AVX-NEXT: vzeroupper5511; AVX-NEXT: retq5512;5513; AVX2-LABEL: load_i32_stride8_vf32:5514; AVX2: # %bb.0:5515; AVX2-NEXT: subq $1544, %rsp # imm = 0x6085516; AVX2-NEXT: vmovaps 288(%rdi), %xmm05517; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5518; AVX2-NEXT: vmovaps 256(%rdi), %xmm105519; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm10[0],xmm0[0],xmm10[1],xmm0[1]5520; AVX2-NEXT: vmovaps 352(%rdi), %xmm145521; AVX2-NEXT: vbroadcastss %xmm14, %xmm15522; AVX2-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5523; AVX2-NEXT: vmovaps 320(%rdi), %xmm155524; AVX2-NEXT: vbroadcastss %xmm15, %xmm25525; AVX2-NEXT: vmovaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5526; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]5527; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5528; AVX2-NEXT: vmovaps 416(%rdi), %xmm15529; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5530; AVX2-NEXT: vmovaps 384(%rdi), %xmm125531; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]5532; AVX2-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5533; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15534; AVX2-NEXT: vmovaps 480(%rdi), %xmm25535; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5536; AVX2-NEXT: vbroadcastss %xmm2, %xmm25537; AVX2-NEXT: vmovaps 448(%rdi), %xmm35538; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5539; AVX2-NEXT: vbroadcastss %xmm3, %xmm35540; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5541; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25542; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]5543; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5544; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5545; AVX2-NEXT: vmovaps 800(%rdi), %xmm15546; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5547; AVX2-NEXT: vmovaps 768(%rdi), %xmm05548; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5549; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5550; AVX2-NEXT: vmovaps 864(%rdi), %xmm15551; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5552; AVX2-NEXT: vbroadcastss %xmm1, %xmm15553; AVX2-NEXT: vmovaps 832(%rdi), %xmm25554; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5555; AVX2-NEXT: vbroadcastss %xmm2, %xmm25556; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]5557; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5558; AVX2-NEXT: vmovaps 992(%rdi), %xmm15559; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5560; AVX2-NEXT: vbroadcastss %xmm1, %xmm15561; AVX2-NEXT: vmovaps 960(%rdi), %xmm115562; AVX2-NEXT: vbroadcastss %xmm11, %xmm25563; AVX2-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5564; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]5565; AVX2-NEXT: vmovaps 928(%rdi), %xmm25566; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5567; AVX2-NEXT: vmovaps 896(%rdi), %xmm35568; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5569; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5570; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25571; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15572; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5573; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5574; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5575; AVX2-NEXT: vmovaps 608(%rdi), %xmm05576; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5577; AVX2-NEXT: vbroadcastss %xmm0, %xmm05578; AVX2-NEXT: vmovaps 576(%rdi), %xmm15579; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5580; AVX2-NEXT: vbroadcastss %xmm1, %xmm15581; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5582; AVX2-NEXT: vmovaps 544(%rdi), %xmm15583; AVX2-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill5584; AVX2-NEXT: vmovaps 512(%rdi), %xmm25585; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5586; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]5587; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]5588; AVX2-NEXT: vmovaps 736(%rdi), %xmm15589; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5590; AVX2-NEXT: vbroadcastss %xmm1, %xmm15591; AVX2-NEXT: vmovaps 704(%rdi), %xmm25592; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5593; AVX2-NEXT: vbroadcastss %xmm2, %xmm25594; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]5595; AVX2-NEXT: vmovaps 672(%rdi), %xmm25596; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5597; AVX2-NEXT: vmovaps 640(%rdi), %xmm35598; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5599; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5600; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25601; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15602; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5603; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5604; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5605; AVX2-NEXT: vmovaps 224(%rdi), %xmm05606; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5607; AVX2-NEXT: vbroadcastss %xmm0, %xmm05608; AVX2-NEXT: vmovaps 192(%rdi), %xmm15609; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5610; AVX2-NEXT: vbroadcastss %xmm1, %xmm15611; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5612; AVX2-NEXT: vmovaps 160(%rdi), %xmm95613; AVX2-NEXT: vmovaps 128(%rdi), %xmm85614; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm9[0],xmm8[1],xmm9[1]5615; AVX2-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5616; AVX2-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5617; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15618; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm05619; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm1[0,1,2,3,4,5],ymm0[6,7]5620; AVX2-NEXT: vmovaps 96(%rdi), %xmm75621; AVX2-NEXT: vbroadcastss %xmm7, %xmm15622; AVX2-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5623; AVX2-NEXT: vmovaps 64(%rdi), %xmm65624; AVX2-NEXT: vbroadcastss %xmm6, %xmm25625; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5626; AVX2-NEXT: vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]5627; AVX2-NEXT: vmovaps (%rdi), %xmm55628; AVX2-NEXT: vmovaps 32(%rdi), %xmm45629; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5630; AVX2-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5631; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5632; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm3[2,3]5633; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm13[4,5,6,7]5634; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5635; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm10[1,1,1,1]5636; AVX2-NEXT: vmovaps %xmm10, %xmm35637; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload5638; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm13[1],xmm0[2,3]5639; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]5640; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5641; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload5642; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5643; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]5644; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15645; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm12[1,1,1,1]5646; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload5647; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]5648; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25649; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5650; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5651; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5652; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5653; AVX2-NEXT: # xmm0 = mem[1,1,1,1]5654; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5655; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]5656; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload5657; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm1 # 16-byte Folded Reload5658; AVX2-NEXT: # xmm1 = xmm12[0],mem[0],xmm12[1],mem[1]5659; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5660; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload5661; AVX2-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1]5662; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15663; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload5664; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm11[1,1,1,1]5665; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload5666; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm10[1],xmm2[2,3]5667; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25668; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5669; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5670; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5671; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]5672; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2,3]5673; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]5674; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5675; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5676; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload5677; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]5678; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15679; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm8[1,1,1,1]5680; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm9[1],xmm2[2,3]5681; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25682; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5683; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5684; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5685; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload5686; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm8[1,1,1,1]5687; AVX2-NEXT: vmovaps (%rsp), %xmm7 # 16-byte Reload5688; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm7[1],xmm0[2,3]5689; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload5690; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5691; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]5692; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5693; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload5694; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5695; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]5696; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15697; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload5698; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm6[1,1,1,1]5699; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload5700; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm5[1],xmm2[2,3]5701; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25702; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5703; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5704; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5705; AVX2-NEXT: vunpckhps {{.*#+}} xmm3 = xmm3[2],xmm13[2],xmm3[3],xmm13[3]5706; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5707; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5708; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload5709; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]5710; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5711; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm14[2,2,2,2]5712; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3]5713; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm05714; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15715; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]5716; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload5717; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm13[2,2,2,2]5718; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5719; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]5720; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]5721; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5722; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5723; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5724; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload5725; AVX2-NEXT: # xmm3 = xmm0[2],mem[2],xmm0[3],mem[3]5726; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5727; AVX2-NEXT: vunpckhps {{.*#+}} xmm0 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]5728; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5729; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload5730; AVX2-NEXT: # xmm15 = mem[2,2,2,2]5731; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5732; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]5733; AVX2-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm155734; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm145735; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]5736; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5737; AVX2-NEXT: vshufps {{.*#+}} xmm15 = xmm0[2,2,2,2]5738; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm12[0,1,2],xmm15[3]5739; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm3[0,1],xmm15[2,3]5740; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]5741; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5742; AVX2-NEXT: vunpckhps {{.*#+}} xmm14 = xmm8[2],xmm7[2],xmm8[3],xmm7[3]5743; AVX2-NEXT: vunpckhps {{.*#+}} xmm12 = xmm6[2],xmm5[2],xmm6[3],xmm5[3]5744; AVX2-NEXT: vshufps {{.*#+}} xmm11 = xmm4[2,2,2,2]5745; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload5746; AVX2-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]5747; AVX2-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm115748; AVX2-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm155749; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]5750; AVX2-NEXT: vshufps {{.*#+}} xmm15 = xmm9[2,2,2,2]5751; AVX2-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload5752; AVX2-NEXT: # xmm15 = mem[0,1,2],xmm15[3]5753; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm14[0,1],xmm15[2,3]5754; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]5755; AVX2-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5756; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5757; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload5758; AVX2-NEXT: # xmm9 = xmm3[2],mem[2],xmm3[3],mem[3]5759; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5760; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm15[2,2,2,2]5761; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload5762; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm4[0,1,2],xmm7[3]5763; AVX2-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm75764; AVX2-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm115765; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3,4,5],ymm7[6,7]5766; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5767; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm6 # 16-byte Folded Reload5768; AVX2-NEXT: # xmm6 = xmm3[2],mem[2],xmm3[3],mem[3]5769; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload5770; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm11[2,2,2,2]5771; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload5772; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm8[0,1,2],xmm5[3]5773; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]5774; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]5775; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5776; AVX2-NEXT: vunpckhps {{.*#+}} xmm5 = xmm2[2],xmm13[2],xmm2[3],xmm13[3]5777; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5778; AVX2-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]5779; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5780; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm5 # 16-byte Folded Reload5781; AVX2-NEXT: # xmm5 = xmm3[2],mem[2],xmm3[3],mem[3]5782; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm55783; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload5784; AVX2-NEXT: # xmm3 = mem[2,3,2,3]5785; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm35786; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]5787; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5788; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5789; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5790; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]5791; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5792; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]5793; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm2 # 16-byte Folded Reload5794; AVX2-NEXT: # xmm2 = xmm1[2],mem[2],xmm1[3],mem[3]5795; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25796; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload5797; AVX2-NEXT: # xmm1 = mem[2,3,2,3]5798; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15799; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]5800; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5801; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5802; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5803; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload5804; AVX2-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]5805; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm14[1],xmm0[1]5806; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload5807; AVX2-NEXT: # xmm1 = xmm10[2],mem[2],xmm10[3],mem[3]5808; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15809; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm12[2,3,2,3]5810; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm25811; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5812; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5813; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5814; AVX2-NEXT: vunpckhps {{.*#+}} xmm0 = xmm4[2],xmm15[2],xmm4[3],xmm15[3]5815; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm05816; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm9[2,3,2,3]5817; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm15818; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]5819; AVX2-NEXT: vunpckhps {{.*#+}} xmm1 = xmm8[2],xmm11[2],xmm8[3],xmm11[3]5820; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm1[1]5821; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5822; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5823; AVX2-NEXT: vmovaps 288(%rdi), %ymm05824; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5825; AVX2-NEXT: vmovaps 256(%rdi), %ymm15826; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5827; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]5828; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm05829; AVX2-NEXT: vmovaps 352(%rdi), %ymm15830; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5831; AVX2-NEXT: vmovaps 320(%rdi), %ymm25832; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5833; AVX2-NEXT: vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]5834; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm6[2,2,2,2]5835; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5836; AVX2-NEXT: vmovaps 480(%rdi), %ymm25837; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5838; AVX2-NEXT: vmovaps 448(%rdi), %ymm35839; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5840; AVX2-NEXT: vmovaps 416(%rdi), %ymm85841; AVX2-NEXT: vmovaps 384(%rdi), %ymm15842; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5843; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[1],ymm8[1],ymm1[4],ymm8[4],ymm1[5],ymm8[5]5844; AVX2-NEXT: vunpcklps {{.*#+}} ymm10 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]5845; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm10[0],ymm1[2],ymm10[2]5846; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5847; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5848; AVX2-NEXT: vmovaps 800(%rdi), %ymm05849; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5850; AVX2-NEXT: vmovaps 768(%rdi), %ymm15851; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5852; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]5853; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm05854; AVX2-NEXT: vmovaps 864(%rdi), %ymm15855; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5856; AVX2-NEXT: vmovaps 832(%rdi), %ymm25857; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5858; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]5859; AVX2-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill5860; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]5861; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5862; AVX2-NEXT: vmovaps 992(%rdi), %ymm25863; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5864; AVX2-NEXT: vmovaps 960(%rdi), %ymm35865; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5866; AVX2-NEXT: vmovaps 928(%rdi), %ymm75867; AVX2-NEXT: vmovaps 896(%rdi), %ymm15868; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5869; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]5870; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5871; AVX2-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]5872; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm12[0],ymm1[2],ymm12[2]5873; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5874; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5875; AVX2-NEXT: vmovaps 32(%rdi), %ymm05876; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5877; AVX2-NEXT: vmovaps (%rdi), %ymm15878; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5879; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]5880; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm05881; AVX2-NEXT: vmovaps 96(%rdi), %ymm15882; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5883; AVX2-NEXT: vmovaps 64(%rdi), %ymm25884; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5885; AVX2-NEXT: vunpcklps {{.*#+}} ymm13 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]5886; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[2,2,2,2]5887; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5888; AVX2-NEXT: vmovaps 224(%rdi), %ymm25889; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5890; AVX2-NEXT: vmovaps 192(%rdi), %ymm35891; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5892; AVX2-NEXT: vmovaps 160(%rdi), %ymm155893; AVX2-NEXT: vmovaps 128(%rdi), %ymm15894; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5895; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]5896; AVX2-NEXT: vunpcklps {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]5897; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[2],ymm4[2]5898; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5899; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5900; AVX2-NEXT: vmovaps 544(%rdi), %ymm145901; AVX2-NEXT: vmovaps 512(%rdi), %ymm115902; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm11[0],ymm14[0],ymm11[1],ymm14[1],ymm11[4],ymm14[4],ymm11[5],ymm14[5]5903; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5904; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm05905; AVX2-NEXT: vmovaps 608(%rdi), %ymm15906; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5907; AVX2-NEXT: vmovaps 576(%rdi), %ymm25908; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5909; AVX2-NEXT: vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]5910; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm9[2,2,2,2]5911; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm5[2,3]5912; AVX2-NEXT: vmovaps 736(%rdi), %ymm25913; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5914; AVX2-NEXT: vmovaps 704(%rdi), %ymm35915; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5916; AVX2-NEXT: vmovaps 672(%rdi), %ymm55917; AVX2-NEXT: vmovaps 640(%rdi), %ymm05918; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5919; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[4],ymm5[4],ymm0[5],ymm5[5]5920; AVX2-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]5921; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]5922; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5923; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5924; AVX2-NEXT: vbroadcastss 404(%rdi), %ymm05925; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm8[5],ymm0[6,7]5926; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm10[6,7]5927; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm15928; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload5929; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm6[1,1,1,1,5,5,5,5]5930; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5931; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm3[1],ymm10[2,3,4],ymm3[5],ymm10[6,7]5932; AVX2-NEXT: vextractf128 $1, %ymm10, %xmm105933; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]5934; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5935; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5936; AVX2-NEXT: vbroadcastss 916(%rdi), %ymm05937; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]5938; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm12[6,7]5939; AVX2-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload5940; AVX2-NEXT: vextractf128 $1, %ymm1, %xmm15941; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload5942; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm10[1,1,1,1,5,5,5,5]5943; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload5944; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm12[1],ymm7[2,3,4],ymm12[5],ymm7[6,7]5945; AVX2-NEXT: vextractf128 $1, %ymm7, %xmm75946; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]5947; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5948; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill5949; AVX2-NEXT: vbroadcastss 148(%rdi), %ymm05950; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm15[5],ymm0[6,7]5951; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]5952; AVX2-NEXT: vextractf128 $1, %ymm13, %xmm15953; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload5954; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm13[1,1,1,1,5,5,5,5]5955; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload5956; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm7[1],ymm4[2,3,4],ymm7[5],ymm4[6,7]5957; AVX2-NEXT: vextractf128 $1, %ymm4, %xmm45958; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]5959; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5960; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5961; AVX2-NEXT: vbroadcastss 660(%rdi), %ymm05962; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm5[5],ymm0[6,7]5963; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]5964; AVX2-NEXT: vextractf128 $1, %ymm9, %xmm15965; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm11[1,1,1,1,5,5,5,5]5966; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1],ymm2[2,3,4],ymm14[5],ymm2[6,7]5967; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm25968; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]5969; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5970; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5971; AVX2-NEXT: vbroadcastss 504(%rdi), %ymm05972; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload5973; AVX2-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm0[7]5974; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5975; AVX2-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm8[2],ymm0[3],ymm8[3],ymm0[6],ymm8[6],ymm0[7],ymm8[7]5976; AVX2-NEXT: vunpckhps {{.*#+}} ymm3 = ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[6],ymm3[6],ymm6[7],ymm3[7]5977; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5978; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5979; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm6 # 32-byte Folded Reload5980; AVX2-NEXT: # ymm6 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]5981; AVX2-NEXT: vextractf128 $1, %ymm3, %xmm25982; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm6[2,2,2,2]5983; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]5984; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm1[6,7]5985; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]5986; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5987; AVX2-NEXT: vbroadcastss 1016(%rdi), %ymm15988; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5989; AVX2-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm1[7]5990; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5991; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm3 # 32-byte Folded Reload5992; AVX2-NEXT: # ymm3 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]5993; AVX2-NEXT: vunpckhps {{.*#+}} ymm12 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]5994; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5995; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm10 # 32-byte Folded Reload5996; AVX2-NEXT: # ymm10 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]5997; AVX2-NEXT: vextractf128 $1, %ymm12, %xmm25998; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm10[2,2,2,2]5999; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]6000; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]6001; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]6002; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6003; AVX2-NEXT: vbroadcastss 248(%rdi), %ymm26004; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload6005; AVX2-NEXT: # ymm2 = mem[0,1,2,3,4,5,6],ymm2[7]6006; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6007; AVX2-NEXT: vunpckhps {{.*#+}} ymm4 = ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[6],ymm15[6],ymm1[7],ymm15[7]6008; AVX2-NEXT: vunpckhps {{.*#+}} ymm9 = ymm13[2],ymm7[2],ymm13[3],ymm7[3],ymm13[6],ymm7[6],ymm13[7],ymm7[7]6009; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6010; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload6011; AVX2-NEXT: # ymm14 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6012; AVX2-NEXT: vextractf128 $1, %ymm9, %xmm86013; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]6014; AVX2-NEXT: vblendps {{.*#+}} xmm8 = xmm8[0,1],xmm15[2,3]6015; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]6016; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm2[4,5,6,7]6017; AVX2-NEXT: vbroadcastss 760(%rdi), %ymm26018; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload6019; AVX2-NEXT: # ymm2 = mem[0,1,2,3,4,5,6],ymm2[7]6020; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6021; AVX2-NEXT: vunpckhps {{.*#+}} ymm5 = ymm1[2],ymm5[2],ymm1[3],ymm5[3],ymm1[6],ymm5[6],ymm1[7],ymm5[7]6022; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload6023; AVX2-NEXT: # ymm11 = ymm11[2],mem[2],ymm11[3],mem[3],ymm11[6],mem[6],ymm11[7],mem[7]6024; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6025; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm13 # 32-byte Folded Reload6026; AVX2-NEXT: # ymm13 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6027; AVX2-NEXT: vextractf128 $1, %ymm11, %xmm156028; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[2,2,2,2]6029; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]6030; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]6031; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]6032; AVX2-NEXT: vbroadcastss 220(%rdi), %ymm16033; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6034; AVX2-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6035; AVX2-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]6036; AVX2-NEXT: vextractf128 $1, %ymm14, %xmm46037; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]6038; AVX2-NEXT: vextractf128 $1, %ymm9, %xmm96039; AVX2-NEXT: vblendps {{.*#+}} xmm4 = xmm9[0,1],xmm4[2,3]6040; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm1[4,5,6,7]6041; AVX2-NEXT: vbroadcastss 476(%rdi), %ymm16042; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6043; AVX2-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6044; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]6045; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm16046; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload6047; AVX2-NEXT: # ymm6 = mem[2,3,2,3,6,7,6,7]6048; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm66049; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]6050; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6051; AVX2-NEXT: vbroadcastss 732(%rdi), %ymm16052; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6053; AVX2-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6054; AVX2-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm1[1],ymm5[3],ymm1[3]6055; AVX2-NEXT: vextractf128 $1, %ymm13, %xmm56056; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm11[2,3,2,3,6,7,6,7]6057; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm66058; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]6059; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm1[4,5,6,7]6060; AVX2-NEXT: vbroadcastss 988(%rdi), %ymm16061; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6062; AVX2-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6063; AVX2-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm1[1],ymm3[3],ymm1[3]6064; AVX2-NEXT: vextractf128 $1, %ymm10, %xmm36065; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm12[2,3,2,3,6,7,6,7]6066; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm66067; AVX2-NEXT: vblendps {{.*#+}} xmm3 = xmm6[0,1],xmm3[2,3]6068; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]6069; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6070; AVX2-NEXT: vmovaps %ymm3, 64(%rsi)6071; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6072; AVX2-NEXT: vmovaps %ymm3, (%rsi)6073; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6074; AVX2-NEXT: vmovaps %ymm3, 96(%rsi)6075; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6076; AVX2-NEXT: vmovaps %ymm3, 32(%rsi)6077; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6078; AVX2-NEXT: vmovaps %ymm3, 64(%rdx)6079; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6080; AVX2-NEXT: vmovaps %ymm3, (%rdx)6081; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6082; AVX2-NEXT: vmovaps %ymm3, 96(%rdx)6083; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6084; AVX2-NEXT: vmovaps %ymm3, 32(%rdx)6085; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6086; AVX2-NEXT: vmovaps %ymm3, 64(%rcx)6087; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6088; AVX2-NEXT: vmovaps %ymm3, (%rcx)6089; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6090; AVX2-NEXT: vmovaps %ymm3, 96(%rcx)6091; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6092; AVX2-NEXT: vmovaps %ymm3, 32(%rcx)6093; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6094; AVX2-NEXT: vmovaps %ymm3, 64(%r8)6095; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6096; AVX2-NEXT: vmovaps %ymm3, (%r8)6097; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6098; AVX2-NEXT: vmovaps %ymm3, 96(%r8)6099; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6100; AVX2-NEXT: vmovaps %ymm3, 32(%r8)6101; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6102; AVX2-NEXT: vmovaps %ymm3, 64(%r9)6103; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6104; AVX2-NEXT: vmovaps %ymm3, (%r9)6105; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6106; AVX2-NEXT: vmovaps %ymm3, 96(%r9)6107; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6108; AVX2-NEXT: vmovaps %ymm3, 32(%r9)6109; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax6110; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6111; AVX2-NEXT: vmovaps %ymm3, 64(%rax)6112; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6113; AVX2-NEXT: vmovaps %ymm3, (%rax)6114; AVX2-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload6115; AVX2-NEXT: vmovaps %ymm3, 96(%rax)6116; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6117; AVX2-NEXT: vmovaps %ymm3, 32(%rax)6118; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax6119; AVX2-NEXT: vmovaps %ymm2, 64(%rax)6120; AVX2-NEXT: vmovaps %ymm8, (%rax)6121; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6122; AVX2-NEXT: vmovaps %ymm2, 96(%rax)6123; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6124; AVX2-NEXT: vmovaps %ymm2, 32(%rax)6125; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax6126; AVX2-NEXT: vmovaps %ymm1, 96(%rax)6127; AVX2-NEXT: vmovaps %ymm5, 64(%rax)6128; AVX2-NEXT: vmovaps %ymm0, 32(%rax)6129; AVX2-NEXT: vmovaps %ymm4, (%rax)6130; AVX2-NEXT: addq $1544, %rsp # imm = 0x6086131; AVX2-NEXT: vzeroupper6132; AVX2-NEXT: retq6133;6134; AVX2-FP-LABEL: load_i32_stride8_vf32:6135; AVX2-FP: # %bb.0:6136; AVX2-FP-NEXT: subq $1544, %rsp # imm = 0x6086137; AVX2-FP-NEXT: vmovaps 288(%rdi), %xmm06138; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6139; AVX2-FP-NEXT: vmovaps 256(%rdi), %xmm106140; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm10[0],xmm0[0],xmm10[1],xmm0[1]6141; AVX2-FP-NEXT: vmovaps 352(%rdi), %xmm146142; AVX2-FP-NEXT: vbroadcastss %xmm14, %xmm16143; AVX2-FP-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6144; AVX2-FP-NEXT: vmovaps 320(%rdi), %xmm156145; AVX2-FP-NEXT: vbroadcastss %xmm15, %xmm26146; AVX2-FP-NEXT: vmovaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6147; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6148; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6149; AVX2-FP-NEXT: vmovaps 416(%rdi), %xmm16150; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6151; AVX2-FP-NEXT: vmovaps 384(%rdi), %xmm126152; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]6153; AVX2-FP-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6154; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16155; AVX2-FP-NEXT: vmovaps 480(%rdi), %xmm26156; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6157; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm26158; AVX2-FP-NEXT: vmovaps 448(%rdi), %xmm36159; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6160; AVX2-FP-NEXT: vbroadcastss %xmm3, %xmm36161; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6162; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26163; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]6164; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6165; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6166; AVX2-FP-NEXT: vmovaps 800(%rdi), %xmm16167; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6168; AVX2-FP-NEXT: vmovaps 768(%rdi), %xmm06169; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6170; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]6171; AVX2-FP-NEXT: vmovaps 864(%rdi), %xmm16172; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6173; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm16174; AVX2-FP-NEXT: vmovaps 832(%rdi), %xmm26175; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6176; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm26177; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6178; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6179; AVX2-FP-NEXT: vmovaps 992(%rdi), %xmm16180; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6181; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm16182; AVX2-FP-NEXT: vmovaps 960(%rdi), %xmm116183; AVX2-FP-NEXT: vbroadcastss %xmm11, %xmm26184; AVX2-FP-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6185; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6186; AVX2-FP-NEXT: vmovaps 928(%rdi), %xmm26187; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6188; AVX2-FP-NEXT: vmovaps 896(%rdi), %xmm36189; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6190; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6191; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26192; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16193; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6194; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6195; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6196; AVX2-FP-NEXT: vmovaps 608(%rdi), %xmm06197; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6198; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm06199; AVX2-FP-NEXT: vmovaps 576(%rdi), %xmm16200; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6201; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm16202; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6203; AVX2-FP-NEXT: vmovaps 544(%rdi), %xmm16204; AVX2-FP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill6205; AVX2-FP-NEXT: vmovaps 512(%rdi), %xmm26206; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6207; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6208; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]6209; AVX2-FP-NEXT: vmovaps 736(%rdi), %xmm16210; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6211; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm16212; AVX2-FP-NEXT: vmovaps 704(%rdi), %xmm26213; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6214; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm26215; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6216; AVX2-FP-NEXT: vmovaps 672(%rdi), %xmm26217; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6218; AVX2-FP-NEXT: vmovaps 640(%rdi), %xmm36219; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6220; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6221; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26222; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16223; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6224; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6225; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6226; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm06227; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6228; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm06229; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm16230; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6231; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm16232; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6233; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm96234; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm86235; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm9[0],xmm8[1],xmm9[1]6236; AVX2-FP-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6237; AVX2-FP-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6238; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16239; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm06240; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm1[0,1,2,3,4,5],ymm0[6,7]6241; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm76242; AVX2-FP-NEXT: vbroadcastss %xmm7, %xmm16243; AVX2-FP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6244; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm66245; AVX2-FP-NEXT: vbroadcastss %xmm6, %xmm26246; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6247; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6248; AVX2-FP-NEXT: vmovaps (%rdi), %xmm56249; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm46250; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]6251; AVX2-FP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6252; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6253; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm3[2,3]6254; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm13[4,5,6,7]6255; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6256; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm10[1,1,1,1]6257; AVX2-FP-NEXT: vmovaps %xmm10, %xmm36258; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload6259; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm13[1],xmm0[2,3]6260; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]6261; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6262; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload6263; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload6264; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]6265; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16266; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm12[1,1,1,1]6267; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload6268; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]6269; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26270; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6271; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6272; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6273; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload6274; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]6275; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6276; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]6277; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload6278; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm1 # 16-byte Folded Reload6279; AVX2-FP-NEXT: # xmm1 = xmm12[0],mem[0],xmm12[1],mem[1]6280; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6281; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload6282; AVX2-FP-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1]6283; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16284; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload6285; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm11[1,1,1,1]6286; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload6287; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm10[1],xmm2[2,3]6288; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26289; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6290; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6291; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6292; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]6293; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2,3]6294; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]6295; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6296; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6297; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload6298; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]6299; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16300; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm8[1,1,1,1]6301; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm9[1],xmm2[2,3]6302; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26303; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6304; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6305; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6306; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload6307; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm8[1,1,1,1]6308; AVX2-FP-NEXT: vmovaps (%rsp), %xmm7 # 16-byte Reload6309; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm7[1],xmm0[2,3]6310; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload6311; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6312; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]6313; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6314; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload6315; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6316; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]6317; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16318; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload6319; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm6[1,1,1,1]6320; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload6321; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm5[1],xmm2[2,3]6322; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26323; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6324; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6325; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6326; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm3 = xmm3[2],xmm13[2],xmm3[3],xmm13[3]6327; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6328; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6329; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload6330; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]6331; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6332; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm14[2,2,2,2]6333; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3]6334; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm06335; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16336; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]6337; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload6338; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm13[2,2,2,2]6339; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload6340; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]6341; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]6342; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6343; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6344; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6345; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload6346; AVX2-FP-NEXT: # xmm3 = xmm0[2],mem[2],xmm0[3],mem[3]6347; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6348; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]6349; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6350; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload6351; AVX2-FP-NEXT: # xmm15 = mem[2,2,2,2]6352; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6353; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]6354; AVX2-FP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm156355; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm146356; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]6357; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6358; AVX2-FP-NEXT: vshufps {{.*#+}} xmm15 = xmm0[2,2,2,2]6359; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm12[0,1,2],xmm15[3]6360; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm3[0,1],xmm15[2,3]6361; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]6362; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6363; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm14 = xmm8[2],xmm7[2],xmm8[3],xmm7[3]6364; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm12 = xmm6[2],xmm5[2],xmm6[3],xmm5[3]6365; AVX2-FP-NEXT: vshufps {{.*#+}} xmm11 = xmm4[2,2,2,2]6366; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload6367; AVX2-FP-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]6368; AVX2-FP-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm116369; AVX2-FP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm156370; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]6371; AVX2-FP-NEXT: vshufps {{.*#+}} xmm15 = xmm9[2,2,2,2]6372; AVX2-FP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload6373; AVX2-FP-NEXT: # xmm15 = mem[0,1,2],xmm15[3]6374; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm14[0,1],xmm15[2,3]6375; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]6376; AVX2-FP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6377; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload6378; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload6379; AVX2-FP-NEXT: # xmm9 = xmm3[2],mem[2],xmm3[3],mem[3]6380; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload6381; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm15[2,2,2,2]6382; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload6383; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm4[0,1,2],xmm7[3]6384; AVX2-FP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm76385; AVX2-FP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm116386; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3,4,5],ymm7[6,7]6387; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload6388; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm6 # 16-byte Folded Reload6389; AVX2-FP-NEXT: # xmm6 = xmm3[2],mem[2],xmm3[3],mem[3]6390; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload6391; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm11[2,2,2,2]6392; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload6393; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm8[0,1,2],xmm5[3]6394; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]6395; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]6396; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6397; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm2[2],xmm13[2],xmm2[3],xmm13[3]6398; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload6399; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]6400; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload6401; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm5 # 16-byte Folded Reload6402; AVX2-FP-NEXT: # xmm5 = xmm3[2],mem[2],xmm3[3],mem[3]6403; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm56404; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload6405; AVX2-FP-NEXT: # xmm3 = mem[2,3,2,3]6406; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm36407; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]6408; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6409; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6410; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload6411; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]6412; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6413; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]6414; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm2 # 16-byte Folded Reload6415; AVX2-FP-NEXT: # xmm2 = xmm1[2],mem[2],xmm1[3],mem[3]6416; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26417; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload6418; AVX2-FP-NEXT: # xmm1 = mem[2,3,2,3]6419; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16420; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]6421; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6422; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6423; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6424; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6425; AVX2-FP-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]6426; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm14[1],xmm0[1]6427; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload6428; AVX2-FP-NEXT: # xmm1 = xmm10[2],mem[2],xmm10[3],mem[3]6429; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16430; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm12[2,3,2,3]6431; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26432; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6433; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6434; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6435; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm4[2],xmm15[2],xmm4[3],xmm15[3]6436; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm06437; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm9[2,3,2,3]6438; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16439; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]6440; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm1 = xmm8[2],xmm11[2],xmm8[3],xmm11[3]6441; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm1[1]6442; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6443; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6444; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm06445; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6446; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm16447; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6448; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]6449; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm06450; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm16451; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6452; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm26453; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6454; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]6455; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm6[2,2,2,2]6456; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6457; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm26458; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6459; AVX2-FP-NEXT: vmovaps 448(%rdi), %ymm36460; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6461; AVX2-FP-NEXT: vmovaps 416(%rdi), %ymm86462; AVX2-FP-NEXT: vmovaps 384(%rdi), %ymm16463; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6464; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[1],ymm8[1],ymm1[4],ymm8[4],ymm1[5],ymm8[5]6465; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm10 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]6466; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm10[0],ymm1[2],ymm10[2]6467; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6468; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6469; AVX2-FP-NEXT: vmovaps 800(%rdi), %ymm06470; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6471; AVX2-FP-NEXT: vmovaps 768(%rdi), %ymm16472; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6473; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]6474; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm06475; AVX2-FP-NEXT: vmovaps 864(%rdi), %ymm16476; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6477; AVX2-FP-NEXT: vmovaps 832(%rdi), %ymm26478; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6479; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]6480; AVX2-FP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill6481; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]6482; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6483; AVX2-FP-NEXT: vmovaps 992(%rdi), %ymm26484; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6485; AVX2-FP-NEXT: vmovaps 960(%rdi), %ymm36486; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6487; AVX2-FP-NEXT: vmovaps 928(%rdi), %ymm76488; AVX2-FP-NEXT: vmovaps 896(%rdi), %ymm16489; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6490; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]6491; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6492; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]6493; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm12[0],ymm1[2],ymm12[2]6494; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6495; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6496; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm06497; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6498; AVX2-FP-NEXT: vmovaps (%rdi), %ymm16499; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6500; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]6501; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm06502; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm16503; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6504; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm26505; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6506; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm13 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]6507; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[2,2,2,2]6508; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6509; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm26510; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6511; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm36512; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6513; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm156514; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm16515; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6516; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]6517; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]6518; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[2],ymm4[2]6519; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6520; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6521; AVX2-FP-NEXT: vmovaps 544(%rdi), %ymm146522; AVX2-FP-NEXT: vmovaps 512(%rdi), %ymm116523; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm11[0],ymm14[0],ymm11[1],ymm14[1],ymm11[4],ymm14[4],ymm11[5],ymm14[5]6524; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6525; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm06526; AVX2-FP-NEXT: vmovaps 608(%rdi), %ymm16527; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6528; AVX2-FP-NEXT: vmovaps 576(%rdi), %ymm26529; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6530; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]6531; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm9[2,2,2,2]6532; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm5[2,3]6533; AVX2-FP-NEXT: vmovaps 736(%rdi), %ymm26534; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6535; AVX2-FP-NEXT: vmovaps 704(%rdi), %ymm36536; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6537; AVX2-FP-NEXT: vmovaps 672(%rdi), %ymm56538; AVX2-FP-NEXT: vmovaps 640(%rdi), %ymm06539; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6540; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[4],ymm5[4],ymm0[5],ymm5[5]6541; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]6542; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]6543; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6544; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6545; AVX2-FP-NEXT: vbroadcastss 404(%rdi), %ymm06546; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm8[5],ymm0[6,7]6547; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm10[6,7]6548; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm16549; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload6550; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm6[1,1,1,1,5,5,5,5]6551; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6552; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm3[1],ymm10[2,3,4],ymm3[5],ymm10[6,7]6553; AVX2-FP-NEXT: vextractf128 $1, %ymm10, %xmm106554; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]6555; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6556; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6557; AVX2-FP-NEXT: vbroadcastss 916(%rdi), %ymm06558; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]6559; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm12[6,7]6560; AVX2-FP-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload6561; AVX2-FP-NEXT: vextractf128 $1, %ymm1, %xmm16562; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload6563; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm10[1,1,1,1,5,5,5,5]6564; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload6565; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm12[1],ymm7[2,3,4],ymm12[5],ymm7[6,7]6566; AVX2-FP-NEXT: vextractf128 $1, %ymm7, %xmm76567; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]6568; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6569; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill6570; AVX2-FP-NEXT: vbroadcastss 148(%rdi), %ymm06571; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm15[5],ymm0[6,7]6572; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]6573; AVX2-FP-NEXT: vextractf128 $1, %ymm13, %xmm16574; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload6575; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm13[1,1,1,1,5,5,5,5]6576; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload6577; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm7[1],ymm4[2,3,4],ymm7[5],ymm4[6,7]6578; AVX2-FP-NEXT: vextractf128 $1, %ymm4, %xmm46579; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]6580; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6581; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6582; AVX2-FP-NEXT: vbroadcastss 660(%rdi), %ymm06583; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm5[5],ymm0[6,7]6584; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]6585; AVX2-FP-NEXT: vextractf128 $1, %ymm9, %xmm16586; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm11[1,1,1,1,5,5,5,5]6587; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1],ymm2[2,3,4],ymm14[5],ymm2[6,7]6588; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm26589; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]6590; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6591; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6592; AVX2-FP-NEXT: vbroadcastss 504(%rdi), %ymm06593; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload6594; AVX2-FP-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm0[7]6595; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6596; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm8[2],ymm0[3],ymm8[3],ymm0[6],ymm8[6],ymm0[7],ymm8[7]6597; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[6],ymm3[6],ymm6[7],ymm3[7]6598; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6599; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6600; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm6 # 32-byte Folded Reload6601; AVX2-FP-NEXT: # ymm6 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]6602; AVX2-FP-NEXT: vextractf128 $1, %ymm3, %xmm26603; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm6[2,2,2,2]6604; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]6605; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm1[6,7]6606; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]6607; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6608; AVX2-FP-NEXT: vbroadcastss 1016(%rdi), %ymm16609; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6610; AVX2-FP-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm1[7]6611; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6612; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm3 # 32-byte Folded Reload6613; AVX2-FP-NEXT: # ymm3 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]6614; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm12 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]6615; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6616; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm10 # 32-byte Folded Reload6617; AVX2-FP-NEXT: # ymm10 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]6618; AVX2-FP-NEXT: vextractf128 $1, %ymm12, %xmm26619; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm10[2,2,2,2]6620; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]6621; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]6622; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]6623; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6624; AVX2-FP-NEXT: vbroadcastss 248(%rdi), %ymm26625; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload6626; AVX2-FP-NEXT: # ymm2 = mem[0,1,2,3,4,5,6],ymm2[7]6627; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6628; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm4 = ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[6],ymm15[6],ymm1[7],ymm15[7]6629; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm9 = ymm13[2],ymm7[2],ymm13[3],ymm7[3],ymm13[6],ymm7[6],ymm13[7],ymm7[7]6630; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6631; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload6632; AVX2-FP-NEXT: # ymm14 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6633; AVX2-FP-NEXT: vextractf128 $1, %ymm9, %xmm86634; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]6635; AVX2-FP-NEXT: vblendps {{.*#+}} xmm8 = xmm8[0,1],xmm15[2,3]6636; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]6637; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm2[4,5,6,7]6638; AVX2-FP-NEXT: vbroadcastss 760(%rdi), %ymm26639; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload6640; AVX2-FP-NEXT: # ymm2 = mem[0,1,2,3,4,5,6],ymm2[7]6641; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6642; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm5 = ymm1[2],ymm5[2],ymm1[3],ymm5[3],ymm1[6],ymm5[6],ymm1[7],ymm5[7]6643; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload6644; AVX2-FP-NEXT: # ymm11 = ymm11[2],mem[2],ymm11[3],mem[3],ymm11[6],mem[6],ymm11[7],mem[7]6645; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload6646; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm13 # 32-byte Folded Reload6647; AVX2-FP-NEXT: # ymm13 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6648; AVX2-FP-NEXT: vextractf128 $1, %ymm11, %xmm156649; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[2,2,2,2]6650; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]6651; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]6652; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]6653; AVX2-FP-NEXT: vbroadcastss 220(%rdi), %ymm16654; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6655; AVX2-FP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6656; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]6657; AVX2-FP-NEXT: vextractf128 $1, %ymm14, %xmm46658; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]6659; AVX2-FP-NEXT: vextractf128 $1, %ymm9, %xmm96660; AVX2-FP-NEXT: vblendps {{.*#+}} xmm4 = xmm9[0,1],xmm4[2,3]6661; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm1[4,5,6,7]6662; AVX2-FP-NEXT: vbroadcastss 476(%rdi), %ymm16663; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6664; AVX2-FP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6665; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]6666; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm16667; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload6668; AVX2-FP-NEXT: # ymm6 = mem[2,3,2,3,6,7,6,7]6669; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm66670; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]6671; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6672; AVX2-FP-NEXT: vbroadcastss 732(%rdi), %ymm16673; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6674; AVX2-FP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6675; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm1[1],ymm5[3],ymm1[3]6676; AVX2-FP-NEXT: vextractf128 $1, %ymm13, %xmm56677; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm11[2,3,2,3,6,7,6,7]6678; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm66679; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]6680; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm1[4,5,6,7]6681; AVX2-FP-NEXT: vbroadcastss 988(%rdi), %ymm16682; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload6683; AVX2-FP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]6684; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm1[1],ymm3[3],ymm1[3]6685; AVX2-FP-NEXT: vextractf128 $1, %ymm10, %xmm36686; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm12[2,3,2,3,6,7,6,7]6687; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm66688; AVX2-FP-NEXT: vblendps {{.*#+}} xmm3 = xmm6[0,1],xmm3[2,3]6689; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]6690; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6691; AVX2-FP-NEXT: vmovaps %ymm3, 64(%rsi)6692; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6693; AVX2-FP-NEXT: vmovaps %ymm3, (%rsi)6694; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6695; AVX2-FP-NEXT: vmovaps %ymm3, 96(%rsi)6696; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6697; AVX2-FP-NEXT: vmovaps %ymm3, 32(%rsi)6698; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6699; AVX2-FP-NEXT: vmovaps %ymm3, 64(%rdx)6700; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6701; AVX2-FP-NEXT: vmovaps %ymm3, (%rdx)6702; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6703; AVX2-FP-NEXT: vmovaps %ymm3, 96(%rdx)6704; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6705; AVX2-FP-NEXT: vmovaps %ymm3, 32(%rdx)6706; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6707; AVX2-FP-NEXT: vmovaps %ymm3, 64(%rcx)6708; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6709; AVX2-FP-NEXT: vmovaps %ymm3, (%rcx)6710; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6711; AVX2-FP-NEXT: vmovaps %ymm3, 96(%rcx)6712; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6713; AVX2-FP-NEXT: vmovaps %ymm3, 32(%rcx)6714; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6715; AVX2-FP-NEXT: vmovaps %ymm3, 64(%r8)6716; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6717; AVX2-FP-NEXT: vmovaps %ymm3, (%r8)6718; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6719; AVX2-FP-NEXT: vmovaps %ymm3, 96(%r8)6720; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6721; AVX2-FP-NEXT: vmovaps %ymm3, 32(%r8)6722; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6723; AVX2-FP-NEXT: vmovaps %ymm3, 64(%r9)6724; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6725; AVX2-FP-NEXT: vmovaps %ymm3, (%r9)6726; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6727; AVX2-FP-NEXT: vmovaps %ymm3, 96(%r9)6728; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6729; AVX2-FP-NEXT: vmovaps %ymm3, 32(%r9)6730; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax6731; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6732; AVX2-FP-NEXT: vmovaps %ymm3, 64(%rax)6733; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6734; AVX2-FP-NEXT: vmovaps %ymm3, (%rax)6735; AVX2-FP-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload6736; AVX2-FP-NEXT: vmovaps %ymm3, 96(%rax)6737; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload6738; AVX2-FP-NEXT: vmovaps %ymm3, 32(%rax)6739; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax6740; AVX2-FP-NEXT: vmovaps %ymm2, 64(%rax)6741; AVX2-FP-NEXT: vmovaps %ymm8, (%rax)6742; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6743; AVX2-FP-NEXT: vmovaps %ymm2, 96(%rax)6744; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload6745; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rax)6746; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax6747; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rax)6748; AVX2-FP-NEXT: vmovaps %ymm5, 64(%rax)6749; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rax)6750; AVX2-FP-NEXT: vmovaps %ymm4, (%rax)6751; AVX2-FP-NEXT: addq $1544, %rsp # imm = 0x6086752; AVX2-FP-NEXT: vzeroupper6753; AVX2-FP-NEXT: retq6754;6755; AVX2-FCP-LABEL: load_i32_stride8_vf32:6756; AVX2-FCP: # %bb.0:6757; AVX2-FCP-NEXT: subq $1544, %rsp # imm = 0x6086758; AVX2-FCP-NEXT: vmovaps 288(%rdi), %xmm06759; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6760; AVX2-FCP-NEXT: vmovaps 256(%rdi), %xmm106761; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm10[0],xmm0[0],xmm10[1],xmm0[1]6762; AVX2-FCP-NEXT: vmovaps 352(%rdi), %xmm146763; AVX2-FCP-NEXT: vbroadcastss %xmm14, %xmm16764; AVX2-FCP-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6765; AVX2-FCP-NEXT: vmovaps 320(%rdi), %xmm156766; AVX2-FCP-NEXT: vbroadcastss %xmm15, %xmm26767; AVX2-FCP-NEXT: vmovaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6768; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6769; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6770; AVX2-FCP-NEXT: vmovaps 416(%rdi), %xmm16771; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6772; AVX2-FCP-NEXT: vmovaps 384(%rdi), %xmm126773; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]6774; AVX2-FCP-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6775; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16776; AVX2-FCP-NEXT: vmovaps 480(%rdi), %xmm26777; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6778; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm26779; AVX2-FCP-NEXT: vmovaps 448(%rdi), %xmm36780; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6781; AVX2-FCP-NEXT: vbroadcastss %xmm3, %xmm36782; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6783; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26784; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]6785; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6786; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6787; AVX2-FCP-NEXT: vmovaps 800(%rdi), %xmm16788; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6789; AVX2-FCP-NEXT: vmovaps 768(%rdi), %xmm06790; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6791; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]6792; AVX2-FCP-NEXT: vmovaps 864(%rdi), %xmm16793; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6794; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm16795; AVX2-FCP-NEXT: vmovaps 832(%rdi), %xmm26796; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6797; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm26798; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6799; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6800; AVX2-FCP-NEXT: vmovaps 992(%rdi), %xmm16801; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6802; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm16803; AVX2-FCP-NEXT: vmovaps 960(%rdi), %xmm116804; AVX2-FCP-NEXT: vbroadcastss %xmm11, %xmm26805; AVX2-FCP-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6806; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6807; AVX2-FCP-NEXT: vmovaps 928(%rdi), %xmm26808; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6809; AVX2-FCP-NEXT: vmovaps 896(%rdi), %xmm36810; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6811; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6812; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26813; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16814; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6815; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6816; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6817; AVX2-FCP-NEXT: vmovaps 608(%rdi), %xmm06818; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6819; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm06820; AVX2-FCP-NEXT: vmovaps 576(%rdi), %xmm16821; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6822; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm16823; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6824; AVX2-FCP-NEXT: vmovaps 544(%rdi), %xmm16825; AVX2-FCP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill6826; AVX2-FCP-NEXT: vmovaps 512(%rdi), %xmm26827; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6828; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6829; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]6830; AVX2-FCP-NEXT: vmovaps 736(%rdi), %xmm16831; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6832; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm16833; AVX2-FCP-NEXT: vmovaps 704(%rdi), %xmm26834; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6835; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm26836; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6837; AVX2-FCP-NEXT: vmovaps 672(%rdi), %xmm26838; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6839; AVX2-FCP-NEXT: vmovaps 640(%rdi), %xmm36840; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6841; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6842; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26843; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16844; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6845; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6846; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6847; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm06848; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6849; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm06850; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm16851; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6852; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm16853; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6854; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm96855; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm86856; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm9[0],xmm8[1],xmm9[1]6857; AVX2-FCP-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6858; AVX2-FCP-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6859; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16860; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm06861; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm1[0,1,2,3,4,5],ymm0[6,7]6862; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm76863; AVX2-FCP-NEXT: vbroadcastss %xmm7, %xmm16864; AVX2-FCP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6865; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm66866; AVX2-FCP-NEXT: vbroadcastss %xmm6, %xmm26867; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6868; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]6869; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm56870; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm46871; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]6872; AVX2-FCP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6873; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6874; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm3[2,3]6875; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm13[4,5,6,7]6876; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6877; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm10[1,1,1,1]6878; AVX2-FCP-NEXT: vmovaps %xmm10, %xmm36879; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload6880; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm13[1],xmm0[2,3]6881; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]6882; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6883; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload6884; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload6885; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]6886; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16887; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm12[1,1,1,1]6888; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload6889; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]6890; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26891; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6892; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6893; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6894; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload6895; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]6896; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload6897; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]6898; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload6899; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm1 # 16-byte Folded Reload6900; AVX2-FCP-NEXT: # xmm1 = xmm12[0],mem[0],xmm12[1],mem[1]6901; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6902; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload6903; AVX2-FCP-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1]6904; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16905; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload6906; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm11[1,1,1,1]6907; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload6908; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm10[1],xmm2[2,3]6909; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26910; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6911; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6912; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6913; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]6914; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2,3]6915; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]6916; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6917; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6918; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload6919; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]6920; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16921; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm8[1,1,1,1]6922; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm9[1],xmm2[2,3]6923; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26924; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6925; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6926; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6927; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload6928; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm8[1,1,1,1]6929; AVX2-FCP-NEXT: vmovaps (%rsp), %xmm7 # 16-byte Reload6930; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm7[1],xmm0[2,3]6931; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload6932; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6933; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]6934; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6935; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload6936; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6937; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]6938; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16939; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload6940; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm6[1,1,1,1]6941; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload6942; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0],xmm5[1],xmm2[2,3]6943; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm26944; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6945; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6946; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6947; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm3 = xmm3[2],xmm13[2],xmm3[3],xmm13[3]6948; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6949; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6950; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload6951; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]6952; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6953; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm14[2,2,2,2]6954; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3]6955; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm06956; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm16957; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]6958; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload6959; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm13[2,2,2,2]6960; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload6961; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]6962; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]6963; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]6964; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6965; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6966; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload6967; AVX2-FCP-NEXT: # xmm3 = xmm0[2],mem[2],xmm0[3],mem[3]6968; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6969; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm11[2],xmm10[2],xmm11[3],xmm10[3]6970; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6971; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload6972; AVX2-FCP-NEXT: # xmm15 = mem[2,2,2,2]6973; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload6974; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]6975; AVX2-FCP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm156976; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm146977; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]6978; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload6979; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm15 = xmm0[2,2,2,2]6980; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm12[0,1,2],xmm15[3]6981; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm3[0,1],xmm15[2,3]6982; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]6983; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6984; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm14 = xmm8[2],xmm7[2],xmm8[3],xmm7[3]6985; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm12 = xmm6[2],xmm5[2],xmm6[3],xmm5[3]6986; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm11 = xmm4[2,2,2,2]6987; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload6988; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]6989; AVX2-FCP-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm116990; AVX2-FCP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm156991; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]6992; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm15 = xmm9[2,2,2,2]6993; AVX2-FCP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload6994; AVX2-FCP-NEXT: # xmm15 = mem[0,1,2],xmm15[3]6995; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm14[0,1],xmm15[2,3]6996; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]6997; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6998; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload6999; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload7000; AVX2-FCP-NEXT: # xmm9 = xmm3[2],mem[2],xmm3[3],mem[3]7001; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7002; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm7 = xmm15[2,2,2,2]7003; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload7004; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm4[0,1,2],xmm7[3]7005; AVX2-FCP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm77006; AVX2-FCP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm117007; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3,4,5],ymm7[6,7]7008; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload7009; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm6 # 16-byte Folded Reload7010; AVX2-FCP-NEXT: # xmm6 = xmm3[2],mem[2],xmm3[3],mem[3]7011; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload7012; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm11[2,2,2,2]7013; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload7014; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm8[0,1,2],xmm5[3]7015; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]7016; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]7017; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7018; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm2[2],xmm13[2],xmm2[3],xmm13[3]7019; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload7020; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]7021; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload7022; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm5 # 16-byte Folded Reload7023; AVX2-FCP-NEXT: # xmm5 = xmm3[2],mem[2],xmm3[3],mem[3]7024; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm57025; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7026; AVX2-FCP-NEXT: # xmm3 = mem[2,3,2,3]7027; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm37028; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]7029; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]7030; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7031; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload7032; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]7033; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7034; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]7035; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm2 # 16-byte Folded Reload7036; AVX2-FCP-NEXT: # xmm2 = xmm1[2],mem[2],xmm1[3],mem[3]7037; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm27038; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7039; AVX2-FCP-NEXT: # xmm1 = mem[2,3,2,3]7040; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm17041; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]7042; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7043; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7044; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7045; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7046; AVX2-FCP-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]7047; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm14[1],xmm0[1]7048; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload7049; AVX2-FCP-NEXT: # xmm1 = xmm10[2],mem[2],xmm10[3],mem[3]7050; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm17051; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm12[2,3,2,3]7052; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm27053; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]7054; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7055; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7056; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm4[2],xmm15[2],xmm4[3],xmm15[3]7057; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07058; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm9[2,3,2,3]7059; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm17060; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]7061; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm1 = xmm8[2],xmm11[2],xmm8[3],xmm11[3]7062; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm1[1]7063; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7064; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7065; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm07066; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7067; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm17068; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7069; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]7070; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm07071; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm17072; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7073; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm27074; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7075; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]7076; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm6[2,2,2,2]7077; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]7078; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm27079; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7080; AVX2-FCP-NEXT: vmovaps 448(%rdi), %ymm37081; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7082; AVX2-FCP-NEXT: vmovaps 416(%rdi), %ymm87083; AVX2-FCP-NEXT: vmovaps 384(%rdi), %ymm17084; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7085; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[1],ymm8[1],ymm1[4],ymm8[4],ymm1[5],ymm8[5]7086; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm10 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]7087; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm10[0],ymm1[2],ymm10[2]7088; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7089; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7090; AVX2-FCP-NEXT: vmovaps 800(%rdi), %ymm07091; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7092; AVX2-FCP-NEXT: vmovaps 768(%rdi), %ymm17093; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7094; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]7095; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm07096; AVX2-FCP-NEXT: vmovaps 864(%rdi), %ymm17097; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7098; AVX2-FCP-NEXT: vmovaps 832(%rdi), %ymm27099; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7100; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]7101; AVX2-FCP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill7102; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]7103; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]7104; AVX2-FCP-NEXT: vmovaps 992(%rdi), %ymm27105; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7106; AVX2-FCP-NEXT: vmovaps 960(%rdi), %ymm37107; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7108; AVX2-FCP-NEXT: vmovaps 928(%rdi), %ymm77109; AVX2-FCP-NEXT: vmovaps 896(%rdi), %ymm17110; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7111; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]7112; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7113; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]7114; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm12[0],ymm1[2],ymm12[2]7115; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7116; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7117; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm07118; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7119; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm17120; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7121; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]7122; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm07123; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm17124; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7125; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm27126; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7127; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm13 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]7128; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[2,2,2,2]7129; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]7130; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm27131; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7132; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm37133; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7134; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm157135; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm17136; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7137; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]7138; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]7139; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[2],ymm4[2]7140; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7141; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7142; AVX2-FCP-NEXT: vmovaps 544(%rdi), %ymm147143; AVX2-FCP-NEXT: vmovaps 512(%rdi), %ymm117144; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm11[0],ymm14[0],ymm11[1],ymm14[1],ymm11[4],ymm14[4],ymm11[5],ymm14[5]7145; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7146; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm07147; AVX2-FCP-NEXT: vmovaps 608(%rdi), %ymm17148; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7149; AVX2-FCP-NEXT: vmovaps 576(%rdi), %ymm27150; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7151; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]7152; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm9[2,2,2,2]7153; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm5[2,3]7154; AVX2-FCP-NEXT: vmovaps 736(%rdi), %ymm27155; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7156; AVX2-FCP-NEXT: vmovaps 704(%rdi), %ymm37157; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7158; AVX2-FCP-NEXT: vmovaps 672(%rdi), %ymm57159; AVX2-FCP-NEXT: vmovaps 640(%rdi), %ymm07160; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7161; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[4],ymm5[4],ymm0[5],ymm5[5]7162; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]7163; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]7164; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7165; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7166; AVX2-FCP-NEXT: vbroadcastss 404(%rdi), %ymm07167; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm8[5],ymm0[6,7]7168; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm10[6,7]7169; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm17170; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload7171; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm10 = ymm6[1,1,1,1,5,5,5,5]7172; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7173; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm3[1],ymm10[2,3,4],ymm3[5],ymm10[6,7]7174; AVX2-FCP-NEXT: vextractf128 $1, %ymm10, %xmm107175; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]7176; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7177; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7178; AVX2-FCP-NEXT: vbroadcastss 916(%rdi), %ymm07179; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]7180; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm12[6,7]7181; AVX2-FCP-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload7182; AVX2-FCP-NEXT: vextractf128 $1, %ymm1, %xmm17183; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload7184; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm10[1,1,1,1,5,5,5,5]7185; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload7186; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm12[1],ymm7[2,3,4],ymm12[5],ymm7[6,7]7187; AVX2-FCP-NEXT: vextractf128 $1, %ymm7, %xmm77188; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]7189; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7190; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill7191; AVX2-FCP-NEXT: vbroadcastss 148(%rdi), %ymm07192; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm15[5],ymm0[6,7]7193; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]7194; AVX2-FCP-NEXT: vextractf128 $1, %ymm13, %xmm17195; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload7196; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm13[1,1,1,1,5,5,5,5]7197; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload7198; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm7[1],ymm4[2,3,4],ymm7[5],ymm4[6,7]7199; AVX2-FCP-NEXT: vextractf128 $1, %ymm4, %xmm47200; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]7201; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7202; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7203; AVX2-FCP-NEXT: vbroadcastss 660(%rdi), %ymm07204; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm5[5],ymm0[6,7]7205; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]7206; AVX2-FCP-NEXT: vextractf128 $1, %ymm9, %xmm17207; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm11[1,1,1,1,5,5,5,5]7208; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1],ymm2[2,3,4],ymm14[5],ymm2[6,7]7209; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm27210; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]7211; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7212; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7213; AVX2-FCP-NEXT: vbroadcastss 504(%rdi), %ymm07214; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload7215; AVX2-FCP-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm0[7]7216; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7217; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm8[2],ymm0[3],ymm8[3],ymm0[6],ymm8[6],ymm0[7],ymm8[7]7218; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[6],ymm3[6],ymm6[7],ymm3[7]7219; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7220; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload7221; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm6 # 32-byte Folded Reload7222; AVX2-FCP-NEXT: # ymm6 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]7223; AVX2-FCP-NEXT: vextractf128 $1, %ymm3, %xmm27224; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm6[2,2,2,2]7225; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]7226; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm1[6,7]7227; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]7228; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7229; AVX2-FCP-NEXT: vbroadcastss 1016(%rdi), %ymm17230; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload7231; AVX2-FCP-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm1[7]7232; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload7233; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm3 # 32-byte Folded Reload7234; AVX2-FCP-NEXT: # ymm3 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]7235; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm12 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]7236; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload7237; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm10 # 32-byte Folded Reload7238; AVX2-FCP-NEXT: # ymm10 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]7239; AVX2-FCP-NEXT: vextractf128 $1, %ymm12, %xmm27240; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm4 = ymm10[2,2,2,2]7241; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]7242; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]7243; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]7244; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7245; AVX2-FCP-NEXT: vbroadcastss 248(%rdi), %ymm27246; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload7247; AVX2-FCP-NEXT: # ymm2 = mem[0,1,2,3,4,5,6],ymm2[7]7248; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7249; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm4 = ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[6],ymm15[6],ymm1[7],ymm15[7]7250; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm9 = ymm13[2],ymm7[2],ymm13[3],ymm7[3],ymm13[6],ymm7[6],ymm13[7],ymm7[7]7251; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7252; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload7253; AVX2-FCP-NEXT: # ymm14 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]7254; AVX2-FCP-NEXT: vextractf128 $1, %ymm9, %xmm87255; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]7256; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm8 = xmm8[0,1],xmm15[2,3]7257; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]7258; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm2[4,5,6,7]7259; AVX2-FCP-NEXT: vbroadcastss 760(%rdi), %ymm27260; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload7261; AVX2-FCP-NEXT: # ymm2 = mem[0,1,2,3,4,5,6],ymm2[7]7262; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7263; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm5 = ymm1[2],ymm5[2],ymm1[3],ymm5[3],ymm1[6],ymm5[6],ymm1[7],ymm5[7]7264; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload7265; AVX2-FCP-NEXT: # ymm11 = ymm11[2],mem[2],ymm11[3],mem[3],ymm11[6],mem[6],ymm11[7],mem[7]7266; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7267; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm13 # 32-byte Folded Reload7268; AVX2-FCP-NEXT: # ymm13 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]7269; AVX2-FCP-NEXT: vextractf128 $1, %ymm11, %xmm157270; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[2,2,2,2]7271; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]7272; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]7273; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]7274; AVX2-FCP-NEXT: vbroadcastss 220(%rdi), %ymm17275; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload7276; AVX2-FCP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]7277; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]7278; AVX2-FCP-NEXT: vextractf128 $1, %ymm14, %xmm47279; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]7280; AVX2-FCP-NEXT: vextractf128 $1, %ymm9, %xmm97281; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm4 = xmm9[0,1],xmm4[2,3]7282; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm1[4,5,6,7]7283; AVX2-FCP-NEXT: vbroadcastss 476(%rdi), %ymm17284; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload7285; AVX2-FCP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]7286; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]7287; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm17288; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload7289; AVX2-FCP-NEXT: # ymm6 = mem[2,3,2,3,6,7,6,7]7290; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm67291; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]7292; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7293; AVX2-FCP-NEXT: vbroadcastss 732(%rdi), %ymm17294; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload7295; AVX2-FCP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]7296; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm1[1],ymm5[3],ymm1[3]7297; AVX2-FCP-NEXT: vextractf128 $1, %ymm13, %xmm57298; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm11[2,3,2,3,6,7,6,7]7299; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm67300; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]7301; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm1[4,5,6,7]7302; AVX2-FCP-NEXT: vbroadcastss 988(%rdi), %ymm17303; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload7304; AVX2-FCP-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]7305; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm1[1],ymm3[3],ymm1[3]7306; AVX2-FCP-NEXT: vextractf128 $1, %ymm10, %xmm37307; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm12[2,3,2,3,6,7,6,7]7308; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm67309; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm6[0,1],xmm3[2,3]7310; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]7311; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7312; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%rsi)7313; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7314; AVX2-FCP-NEXT: vmovaps %ymm3, (%rsi)7315; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7316; AVX2-FCP-NEXT: vmovaps %ymm3, 96(%rsi)7317; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7318; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rsi)7319; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7320; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%rdx)7321; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7322; AVX2-FCP-NEXT: vmovaps %ymm3, (%rdx)7323; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7324; AVX2-FCP-NEXT: vmovaps %ymm3, 96(%rdx)7325; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7326; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rdx)7327; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7328; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%rcx)7329; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7330; AVX2-FCP-NEXT: vmovaps %ymm3, (%rcx)7331; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7332; AVX2-FCP-NEXT: vmovaps %ymm3, 96(%rcx)7333; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7334; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rcx)7335; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7336; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%r8)7337; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7338; AVX2-FCP-NEXT: vmovaps %ymm3, (%r8)7339; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7340; AVX2-FCP-NEXT: vmovaps %ymm3, 96(%r8)7341; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7342; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%r8)7343; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7344; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%r9)7345; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7346; AVX2-FCP-NEXT: vmovaps %ymm3, (%r9)7347; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7348; AVX2-FCP-NEXT: vmovaps %ymm3, 96(%r9)7349; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7350; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%r9)7351; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax7352; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7353; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%rax)7354; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7355; AVX2-FCP-NEXT: vmovaps %ymm3, (%rax)7356; AVX2-FCP-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload7357; AVX2-FCP-NEXT: vmovaps %ymm3, 96(%rax)7358; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload7359; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rax)7360; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax7361; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%rax)7362; AVX2-FCP-NEXT: vmovaps %ymm8, (%rax)7363; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload7364; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%rax)7365; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload7366; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rax)7367; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax7368; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rax)7369; AVX2-FCP-NEXT: vmovaps %ymm5, 64(%rax)7370; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rax)7371; AVX2-FCP-NEXT: vmovaps %ymm4, (%rax)7372; AVX2-FCP-NEXT: addq $1544, %rsp # imm = 0x6087373; AVX2-FCP-NEXT: vzeroupper7374; AVX2-FCP-NEXT: retq7375;7376; AVX512-LABEL: load_i32_stride8_vf32:7377; AVX512: # %bb.0:7378; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax7379; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r107380; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r117381; AVX512-NEXT: vmovdqa64 (%rdi), %zmm07382; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm297383; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm17384; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm307385; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm317386; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm37387; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm77388; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm67389; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm97390; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm57391; AVX512-NEXT: vmovdqa64 704(%rdi), %zmm127392; AVX512-NEXT: vmovdqa64 640(%rdi), %zmm27393; AVX512-NEXT: vmovdqa64 832(%rdi), %zmm147394; AVX512-NEXT: vmovdqa64 768(%rdi), %zmm117395; AVX512-NEXT: vmovdqa64 960(%rdi), %zmm167396; AVX512-NEXT: vmovdqa64 896(%rdi), %zmm157397; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]7398; AVX512-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7399; AVX512-NEXT: vmovdqa64 %zmm15, %zmm177400; AVX512-NEXT: vpermt2d %zmm16, %zmm13, %zmm177401; AVX512-NEXT: vmovdqa64 %zmm11, %zmm187402; AVX512-NEXT: vpermt2d %zmm14, %zmm13, %zmm187403; AVX512-NEXT: movb $-64, %dil7404; AVX512-NEXT: kmovw %edi, %k17405; AVX512-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}7406; AVX512-NEXT: vmovdqa64 %zmm2, %zmm107407; AVX512-NEXT: vpermt2d %zmm12, %zmm13, %zmm107408; AVX512-NEXT: vmovdqa64 %zmm5, %zmm87409; AVX512-NEXT: vpermt2d %zmm9, %zmm13, %zmm87410; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]7411; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm287412; AVX512-NEXT: vmovdqa64 %zmm6, %zmm87413; AVX512-NEXT: vpermt2d %zmm7, %zmm13, %zmm87414; AVX512-NEXT: vmovdqa64 %zmm3, %zmm107415; AVX512-NEXT: vpermt2d %zmm31, %zmm13, %zmm107416; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7417; AVX512-NEXT: vmovdqa64 %zmm1, %zmm87418; AVX512-NEXT: vpermt2d %zmm30, %zmm13, %zmm87419; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm137420; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7421; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm177422; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]7423; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7424; AVX512-NEXT: vmovdqa64 %zmm15, %zmm107425; AVX512-NEXT: vpermt2d %zmm16, %zmm8, %zmm107426; AVX512-NEXT: vmovdqa64 %zmm11, %zmm137427; AVX512-NEXT: vpermt2d %zmm14, %zmm8, %zmm137428; AVX512-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}7429; AVX512-NEXT: vmovdqa64 %zmm2, %zmm107430; AVX512-NEXT: vpermt2d %zmm12, %zmm8, %zmm107431; AVX512-NEXT: vmovdqa64 %zmm5, %zmm47432; AVX512-NEXT: vpermt2d %zmm9, %zmm8, %zmm47433; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]7434; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm187435; AVX512-NEXT: vmovdqa64 %zmm6, %zmm47436; AVX512-NEXT: vpermt2d %zmm7, %zmm8, %zmm47437; AVX512-NEXT: vmovdqa64 %zmm3, %zmm107438; AVX512-NEXT: vpermt2d %zmm31, %zmm8, %zmm107439; AVX512-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}7440; AVX512-NEXT: vmovdqa64 %zmm1, %zmm47441; AVX512-NEXT: vpermt2d %zmm30, %zmm8, %zmm47442; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm87443; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]7444; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm197445; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]7446; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7447; AVX512-NEXT: vmovdqa64 %zmm15, %zmm87448; AVX512-NEXT: vpermt2d %zmm16, %zmm4, %zmm87449; AVX512-NEXT: vmovdqa64 %zmm11, %zmm107450; AVX512-NEXT: vpermt2d %zmm14, %zmm4, %zmm107451; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7452; AVX512-NEXT: vmovdqa64 %zmm2, %zmm87453; AVX512-NEXT: vpermt2d %zmm12, %zmm4, %zmm87454; AVX512-NEXT: vmovdqa64 %zmm5, %zmm137455; AVX512-NEXT: vpermt2d %zmm9, %zmm4, %zmm137456; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7457; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm207458; AVX512-NEXT: vmovdqa64 %zmm6, %zmm87459; AVX512-NEXT: vpermt2d %zmm7, %zmm4, %zmm87460; AVX512-NEXT: vmovdqa64 %zmm3, %zmm107461; AVX512-NEXT: vpermt2d %zmm31, %zmm4, %zmm107462; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7463; AVX512-NEXT: vmovdqa64 %zmm1, %zmm87464; AVX512-NEXT: vpermt2d %zmm30, %zmm4, %zmm87465; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm47466; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7467; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm217468; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]7469; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7470; AVX512-NEXT: vmovdqa64 %zmm15, %zmm87471; AVX512-NEXT: vpermt2d %zmm16, %zmm4, %zmm87472; AVX512-NEXT: vmovdqa64 %zmm11, %zmm107473; AVX512-NEXT: vpermt2d %zmm14, %zmm4, %zmm107474; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7475; AVX512-NEXT: vmovdqa64 %zmm2, %zmm87476; AVX512-NEXT: vpermt2d %zmm12, %zmm4, %zmm87477; AVX512-NEXT: vmovdqa64 %zmm5, %zmm137478; AVX512-NEXT: vpermt2d %zmm9, %zmm4, %zmm137479; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7480; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm227481; AVX512-NEXT: vmovdqa64 %zmm6, %zmm87482; AVX512-NEXT: vpermt2d %zmm7, %zmm4, %zmm87483; AVX512-NEXT: vmovdqa64 %zmm3, %zmm107484; AVX512-NEXT: vpermt2d %zmm31, %zmm4, %zmm107485; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7486; AVX512-NEXT: vmovdqa64 %zmm1, %zmm87487; AVX512-NEXT: vpermt2d %zmm30, %zmm4, %zmm87488; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm47489; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7490; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm237491; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]7492; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7493; AVX512-NEXT: vmovdqa64 %zmm15, %zmm87494; AVX512-NEXT: vpermt2d %zmm16, %zmm4, %zmm87495; AVX512-NEXT: vmovdqa64 %zmm11, %zmm107496; AVX512-NEXT: vpermt2d %zmm14, %zmm4, %zmm107497; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7498; AVX512-NEXT: vmovdqa64 %zmm2, %zmm87499; AVX512-NEXT: vpermt2d %zmm12, %zmm4, %zmm87500; AVX512-NEXT: vmovdqa64 %zmm5, %zmm137501; AVX512-NEXT: vpermt2d %zmm9, %zmm4, %zmm137502; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7503; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm247504; AVX512-NEXT: vmovdqa64 %zmm6, %zmm87505; AVX512-NEXT: vpermt2d %zmm7, %zmm4, %zmm87506; AVX512-NEXT: vmovdqa64 %zmm3, %zmm107507; AVX512-NEXT: vpermt2d %zmm31, %zmm4, %zmm107508; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7509; AVX512-NEXT: vmovdqa64 %zmm1, %zmm87510; AVX512-NEXT: vpermt2d %zmm30, %zmm4, %zmm87511; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm47512; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7513; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm257514; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]7515; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7516; AVX512-NEXT: vmovdqa64 %zmm15, %zmm87517; AVX512-NEXT: vpermt2d %zmm16, %zmm4, %zmm87518; AVX512-NEXT: vmovdqa64 %zmm11, %zmm107519; AVX512-NEXT: vpermt2d %zmm14, %zmm4, %zmm107520; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7521; AVX512-NEXT: vmovdqa64 %zmm2, %zmm87522; AVX512-NEXT: vpermt2d %zmm12, %zmm4, %zmm87523; AVX512-NEXT: vmovdqa64 %zmm5, %zmm137524; AVX512-NEXT: vpermt2d %zmm9, %zmm4, %zmm137525; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7526; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm267527; AVX512-NEXT: vmovdqa64 %zmm6, %zmm87528; AVX512-NEXT: vpermt2d %zmm7, %zmm4, %zmm87529; AVX512-NEXT: vmovdqa64 %zmm3, %zmm107530; AVX512-NEXT: vpermt2d %zmm31, %zmm4, %zmm107531; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7532; AVX512-NEXT: vmovdqa64 %zmm1, %zmm87533; AVX512-NEXT: vpermt2d %zmm30, %zmm4, %zmm87534; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm47535; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7536; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm277537; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]7538; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7539; AVX512-NEXT: vmovdqa64 %zmm15, %zmm87540; AVX512-NEXT: vpermt2d %zmm16, %zmm4, %zmm87541; AVX512-NEXT: vmovdqa64 %zmm11, %zmm107542; AVX512-NEXT: vpermt2d %zmm14, %zmm4, %zmm107543; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7544; AVX512-NEXT: vmovdqa64 %zmm2, %zmm87545; AVX512-NEXT: vpermt2d %zmm12, %zmm4, %zmm87546; AVX512-NEXT: vmovdqa64 %zmm5, %zmm137547; AVX512-NEXT: vpermt2d %zmm9, %zmm4, %zmm137548; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7549; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm87550; AVX512-NEXT: vmovdqa64 %zmm6, %zmm107551; AVX512-NEXT: vpermt2d %zmm7, %zmm4, %zmm107552; AVX512-NEXT: vmovdqa64 %zmm3, %zmm137553; AVX512-NEXT: vpermt2d %zmm31, %zmm4, %zmm137554; AVX512-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}7555; AVX512-NEXT: vmovdqa64 %zmm1, %zmm107556; AVX512-NEXT: vpermt2d %zmm30, %zmm4, %zmm107557; AVX512-NEXT: vpermi2d %zmm29, %zmm0, %zmm47558; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]7559; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm47560; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]7561; AVX512-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7562; AVX512-NEXT: vpermt2d %zmm16, %zmm10, %zmm157563; AVX512-NEXT: vpermt2d %zmm14, %zmm10, %zmm117564; AVX512-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}7565; AVX512-NEXT: vpermt2d %zmm12, %zmm10, %zmm27566; AVX512-NEXT: vpermt2d %zmm9, %zmm10, %zmm57567; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]7568; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm27569; AVX512-NEXT: vpermt2d %zmm7, %zmm10, %zmm67570; AVX512-NEXT: vpermt2d %zmm31, %zmm10, %zmm37571; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}7572; AVX512-NEXT: vpermt2d %zmm30, %zmm10, %zmm17573; AVX512-NEXT: vpermt2d %zmm29, %zmm10, %zmm07574; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7575; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm07576; AVX512-NEXT: vmovdqa64 %zmm28, 64(%rsi)7577; AVX512-NEXT: vmovdqa64 %zmm17, (%rsi)7578; AVX512-NEXT: vmovdqa64 %zmm18, 64(%rdx)7579; AVX512-NEXT: vmovdqa64 %zmm19, (%rdx)7580; AVX512-NEXT: vmovdqa64 %zmm20, 64(%rcx)7581; AVX512-NEXT: vmovdqa64 %zmm21, (%rcx)7582; AVX512-NEXT: vmovdqa64 %zmm22, 64(%r8)7583; AVX512-NEXT: vmovdqa64 %zmm23, (%r8)7584; AVX512-NEXT: vmovdqa64 %zmm24, 64(%r9)7585; AVX512-NEXT: vmovdqa64 %zmm25, (%r9)7586; AVX512-NEXT: vmovdqa64 %zmm26, 64(%r11)7587; AVX512-NEXT: vmovdqa64 %zmm27, (%r11)7588; AVX512-NEXT: vmovdqa64 %zmm8, 64(%r10)7589; AVX512-NEXT: vmovdqa64 %zmm4, (%r10)7590; AVX512-NEXT: vmovdqa64 %zmm2, 64(%rax)7591; AVX512-NEXT: vmovdqa64 %zmm0, (%rax)7592; AVX512-NEXT: vzeroupper7593; AVX512-NEXT: retq7594;7595; AVX512-FCP-LABEL: load_i32_stride8_vf32:7596; AVX512-FCP: # %bb.0:7597; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax7598; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r107599; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r117600; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm07601; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm297602; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm17603; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm307604; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm317605; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm37606; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm77607; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm67608; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm97609; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm57610; AVX512-FCP-NEXT: vmovdqa64 704(%rdi), %zmm127611; AVX512-FCP-NEXT: vmovdqa64 640(%rdi), %zmm27612; AVX512-FCP-NEXT: vmovdqa64 832(%rdi), %zmm147613; AVX512-FCP-NEXT: vmovdqa64 768(%rdi), %zmm117614; AVX512-FCP-NEXT: vmovdqa64 960(%rdi), %zmm167615; AVX512-FCP-NEXT: vmovdqa64 896(%rdi), %zmm157616; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]7617; AVX512-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7618; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm177619; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm13, %zmm177620; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm187621; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm13, %zmm187622; AVX512-FCP-NEXT: movb $-64, %dil7623; AVX512-FCP-NEXT: kmovw %edi, %k17624; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}7625; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm107626; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm13, %zmm107627; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm87628; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm13, %zmm87629; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]7630; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm287631; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm87632; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm87633; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm107634; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm13, %zmm107635; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7636; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm87637; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm13, %zmm87638; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm137639; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7640; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm177641; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]7642; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7643; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm107644; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm8, %zmm107645; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm137646; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm8, %zmm137647; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}7648; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm107649; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm8, %zmm107650; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm47651; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm8, %zmm47652; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]7653; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm187654; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm47655; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm47656; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm107657; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm8, %zmm107658; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}7659; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm47660; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm8, %zmm47661; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm87662; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]7663; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm197664; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]7665; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7666; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm87667; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm87668; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm107669; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm107670; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7671; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm87672; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm87673; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm137674; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm137675; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7676; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm207677; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm87678; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm87679; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm107680; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm107681; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7682; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm87683; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm87684; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm47685; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7686; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm217687; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]7688; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7689; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm87690; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm87691; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm107692; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm107693; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7694; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm87695; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm87696; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm137697; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm137698; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7699; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm227700; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm87701; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm87702; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm107703; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm107704; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7705; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm87706; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm87707; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm47708; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7709; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm237710; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]7711; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7712; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm87713; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm87714; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm107715; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm107716; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7717; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm87718; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm87719; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm137720; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm137721; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7722; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm247723; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm87724; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm87725; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm107726; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm107727; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7728; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm87729; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm87730; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm47731; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7732; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm257733; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]7734; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7735; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm87736; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm87737; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm107738; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm107739; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7740; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm87741; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm87742; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm137743; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm137744; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7745; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm267746; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm87747; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm87748; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm107749; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm107750; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7751; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm87752; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm87753; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm47754; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7755; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm277756; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]7757; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7758; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm87759; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm87760; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm107761; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm107762; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7763; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm87764; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm87765; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm137766; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm137767; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7768; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm87769; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm107770; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm107771; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm137772; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm137773; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}7774; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm107775; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm107776; AVX512-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm47777; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]7778; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm47779; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]7780; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7781; AVX512-FCP-NEXT: vpermt2d %zmm16, %zmm10, %zmm157782; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm10, %zmm117783; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}7784; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm27785; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm10, %zmm57786; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]7787; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm27788; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm67789; AVX512-FCP-NEXT: vpermt2d %zmm31, %zmm10, %zmm37790; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}7791; AVX512-FCP-NEXT: vpermt2d %zmm30, %zmm10, %zmm17792; AVX512-FCP-NEXT: vpermt2d %zmm29, %zmm10, %zmm07793; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]7794; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm07795; AVX512-FCP-NEXT: vmovdqa64 %zmm28, 64(%rsi)7796; AVX512-FCP-NEXT: vmovdqa64 %zmm17, (%rsi)7797; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 64(%rdx)7798; AVX512-FCP-NEXT: vmovdqa64 %zmm19, (%rdx)7799; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 64(%rcx)7800; AVX512-FCP-NEXT: vmovdqa64 %zmm21, (%rcx)7801; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 64(%r8)7802; AVX512-FCP-NEXT: vmovdqa64 %zmm23, (%r8)7803; AVX512-FCP-NEXT: vmovdqa64 %zmm24, 64(%r9)7804; AVX512-FCP-NEXT: vmovdqa64 %zmm25, (%r9)7805; AVX512-FCP-NEXT: vmovdqa64 %zmm26, 64(%r11)7806; AVX512-FCP-NEXT: vmovdqa64 %zmm27, (%r11)7807; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 64(%r10)7808; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%r10)7809; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)7810; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%rax)7811; AVX512-FCP-NEXT: vzeroupper7812; AVX512-FCP-NEXT: retq7813;7814; AVX512DQ-LABEL: load_i32_stride8_vf32:7815; AVX512DQ: # %bb.0:7816; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax7817; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r107818; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r117819; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm07820; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm297821; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm17822; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm307823; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm317824; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm37825; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm77826; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm67827; AVX512DQ-NEXT: vmovdqa64 576(%rdi), %zmm97828; AVX512DQ-NEXT: vmovdqa64 512(%rdi), %zmm57829; AVX512DQ-NEXT: vmovdqa64 704(%rdi), %zmm127830; AVX512DQ-NEXT: vmovdqa64 640(%rdi), %zmm27831; AVX512DQ-NEXT: vmovdqa64 832(%rdi), %zmm147832; AVX512DQ-NEXT: vmovdqa64 768(%rdi), %zmm117833; AVX512DQ-NEXT: vmovdqa64 960(%rdi), %zmm167834; AVX512DQ-NEXT: vmovdqa64 896(%rdi), %zmm157835; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]7836; AVX512DQ-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7837; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm177838; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm13, %zmm177839; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm187840; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm13, %zmm187841; AVX512DQ-NEXT: movb $-64, %dil7842; AVX512DQ-NEXT: kmovw %edi, %k17843; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}7844; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm107845; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm13, %zmm107846; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm87847; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm13, %zmm87848; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]7849; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm287850; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm87851; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm13, %zmm87852; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm107853; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm13, %zmm107854; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7855; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm87856; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm13, %zmm87857; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm137858; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7859; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm177860; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]7861; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7862; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm107863; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm8, %zmm107864; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm137865; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm8, %zmm137866; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}7867; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm107868; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm8, %zmm107869; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm47870; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm8, %zmm47871; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]7872; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm187873; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm47874; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm8, %zmm47875; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm107876; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm8, %zmm107877; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}7878; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm47879; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm8, %zmm47880; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm87881; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]7882; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm197883; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]7884; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7885; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm87886; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm4, %zmm87887; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm107888; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm4, %zmm107889; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7890; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm87891; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm4, %zmm87892; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm137893; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm4, %zmm137894; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7895; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm207896; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm87897; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm4, %zmm87898; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm107899; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm4, %zmm107900; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7901; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm87902; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm4, %zmm87903; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm47904; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7905; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm217906; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]7907; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7908; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm87909; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm4, %zmm87910; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm107911; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm4, %zmm107912; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7913; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm87914; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm4, %zmm87915; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm137916; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm4, %zmm137917; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7918; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm227919; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm87920; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm4, %zmm87921; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm107922; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm4, %zmm107923; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7924; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm87925; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm4, %zmm87926; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm47927; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7928; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm237929; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]7930; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7931; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm87932; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm4, %zmm87933; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm107934; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm4, %zmm107935; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7936; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm87937; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm4, %zmm87938; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm137939; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm4, %zmm137940; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7941; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm247942; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm87943; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm4, %zmm87944; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm107945; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm4, %zmm107946; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7947; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm87948; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm4, %zmm87949; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm47950; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7951; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm257952; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]7953; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7954; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm87955; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm4, %zmm87956; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm107957; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm4, %zmm107958; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7959; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm87960; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm4, %zmm87961; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm137962; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm4, %zmm137963; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7964; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm267965; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm87966; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm4, %zmm87967; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm107968; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm4, %zmm107969; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7970; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm87971; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm4, %zmm87972; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm47973; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]7974; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm277975; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]7976; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]7977; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm87978; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm4, %zmm87979; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm107980; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm4, %zmm107981; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}7982; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm87983; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm4, %zmm87984; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm137985; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm4, %zmm137986; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]7987; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm87988; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm107989; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm4, %zmm107990; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm137991; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm4, %zmm137992; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}7993; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm107994; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm4, %zmm107995; AVX512DQ-NEXT: vpermi2d %zmm29, %zmm0, %zmm47996; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]7997; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm47998; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]7999; AVX512DQ-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8000; AVX512DQ-NEXT: vpermt2d %zmm16, %zmm10, %zmm158001; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm10, %zmm118002; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}8003; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm10, %zmm28004; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm10, %zmm58005; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]8006; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm28007; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm10, %zmm68008; AVX512DQ-NEXT: vpermt2d %zmm31, %zmm10, %zmm38009; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}8010; AVX512DQ-NEXT: vpermt2d %zmm30, %zmm10, %zmm18011; AVX512DQ-NEXT: vpermt2d %zmm29, %zmm10, %zmm08012; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]8013; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm08014; AVX512DQ-NEXT: vmovdqa64 %zmm28, 64(%rsi)8015; AVX512DQ-NEXT: vmovdqa64 %zmm17, (%rsi)8016; AVX512DQ-NEXT: vmovdqa64 %zmm18, 64(%rdx)8017; AVX512DQ-NEXT: vmovdqa64 %zmm19, (%rdx)8018; AVX512DQ-NEXT: vmovdqa64 %zmm20, 64(%rcx)8019; AVX512DQ-NEXT: vmovdqa64 %zmm21, (%rcx)8020; AVX512DQ-NEXT: vmovdqa64 %zmm22, 64(%r8)8021; AVX512DQ-NEXT: vmovdqa64 %zmm23, (%r8)8022; AVX512DQ-NEXT: vmovdqa64 %zmm24, 64(%r9)8023; AVX512DQ-NEXT: vmovdqa64 %zmm25, (%r9)8024; AVX512DQ-NEXT: vmovdqa64 %zmm26, 64(%r11)8025; AVX512DQ-NEXT: vmovdqa64 %zmm27, (%r11)8026; AVX512DQ-NEXT: vmovdqa64 %zmm8, 64(%r10)8027; AVX512DQ-NEXT: vmovdqa64 %zmm4, (%r10)8028; AVX512DQ-NEXT: vmovdqa64 %zmm2, 64(%rax)8029; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%rax)8030; AVX512DQ-NEXT: vzeroupper8031; AVX512DQ-NEXT: retq8032;8033; AVX512DQ-FCP-LABEL: load_i32_stride8_vf32:8034; AVX512DQ-FCP: # %bb.0:8035; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax8036; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r108037; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r118038; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm08039; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm298040; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm18041; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm308042; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm318043; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm38044; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm78045; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm68046; AVX512DQ-FCP-NEXT: vmovdqa64 576(%rdi), %zmm98047; AVX512DQ-FCP-NEXT: vmovdqa64 512(%rdi), %zmm58048; AVX512DQ-FCP-NEXT: vmovdqa64 704(%rdi), %zmm128049; AVX512DQ-FCP-NEXT: vmovdqa64 640(%rdi), %zmm28050; AVX512DQ-FCP-NEXT: vmovdqa64 832(%rdi), %zmm148051; AVX512DQ-FCP-NEXT: vmovdqa64 768(%rdi), %zmm118052; AVX512DQ-FCP-NEXT: vmovdqa64 960(%rdi), %zmm168053; AVX512DQ-FCP-NEXT: vmovdqa64 896(%rdi), %zmm158054; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]8055; AVX512DQ-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8056; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm178057; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm13, %zmm178058; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm188059; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm13, %zmm188060; AVX512DQ-FCP-NEXT: movb $-64, %dil8061; AVX512DQ-FCP-NEXT: kmovw %edi, %k18062; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}8063; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm108064; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm13, %zmm108065; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm88066; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm13, %zmm88067; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]8068; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm288069; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm88070; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm88071; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm108072; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm13, %zmm108073; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8074; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm88075; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm13, %zmm88076; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm138077; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8078; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm178079; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]8080; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8081; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm108082; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm8, %zmm108083; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm138084; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm8, %zmm138085; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8086; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm108087; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm8, %zmm108088; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm48089; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm8, %zmm48090; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8091; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm188092; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm48093; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm48094; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm108095; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm8, %zmm108096; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}8097; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm48098; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm8, %zmm48099; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm88100; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]8101; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm198102; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]8103; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8104; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm88105; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88106; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm108107; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108108; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8109; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm88110; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88111; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm138112; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138113; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8114; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm208115; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm88116; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88117; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm108118; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108119; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8120; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm88121; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88122; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48123; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8124; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm218125; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]8126; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8127; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm88128; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88129; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm108130; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108131; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8132; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm88133; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88134; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm138135; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138136; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8137; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm228138; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm88139; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88140; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm108141; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108142; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8143; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm88144; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88145; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48146; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8147; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm238148; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]8149; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8150; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm88151; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88152; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm108153; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108154; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8155; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm88156; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88157; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm138158; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138159; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8160; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm248161; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm88162; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88163; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm108164; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108165; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8166; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm88167; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88168; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48169; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8170; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm258171; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]8172; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8173; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm88174; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88175; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm108176; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108177; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8178; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm88179; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88180; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm138181; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138182; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8183; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm268184; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm88185; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88186; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm108187; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108188; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8189; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm88190; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88191; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48192; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8193; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm278194; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]8195; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8196; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm88197; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88198; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm108199; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108200; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8201; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm88202; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88203; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm138204; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138205; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8206; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm88207; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm108208; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm108209; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm138210; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm138211; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8212; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm108213; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm108214; AVX512DQ-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48215; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8216; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm48217; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]8218; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8219; AVX512DQ-FCP-NEXT: vpermt2d %zmm16, %zmm10, %zmm158220; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm10, %zmm118221; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}8222; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm28223; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm10, %zmm58224; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]8225; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm28226; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm68227; AVX512DQ-FCP-NEXT: vpermt2d %zmm31, %zmm10, %zmm38228; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}8229; AVX512DQ-FCP-NEXT: vpermt2d %zmm30, %zmm10, %zmm18230; AVX512DQ-FCP-NEXT: vpermt2d %zmm29, %zmm10, %zmm08231; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]8232; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm08233; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, 64(%rsi)8234; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, (%rsi)8235; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 64(%rdx)8236; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, (%rdx)8237; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 64(%rcx)8238; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, (%rcx)8239; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 64(%r8)8240; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, (%r8)8241; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, 64(%r9)8242; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, (%r9)8243; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, 64(%r11)8244; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, (%r11)8245; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 64(%r10)8246; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%r10)8247; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)8248; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%rax)8249; AVX512DQ-FCP-NEXT: vzeroupper8250; AVX512DQ-FCP-NEXT: retq8251;8252; AVX512BW-LABEL: load_i32_stride8_vf32:8253; AVX512BW: # %bb.0:8254; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax8255; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r108256; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r118257; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm08258; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm298259; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm18260; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm308261; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm318262; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm38263; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm78264; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm68265; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm98266; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm58267; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm128268; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm28269; AVX512BW-NEXT: vmovdqa64 832(%rdi), %zmm148270; AVX512BW-NEXT: vmovdqa64 768(%rdi), %zmm118271; AVX512BW-NEXT: vmovdqa64 960(%rdi), %zmm168272; AVX512BW-NEXT: vmovdqa64 896(%rdi), %zmm158273; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]8274; AVX512BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8275; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm178276; AVX512BW-NEXT: vpermt2d %zmm16, %zmm13, %zmm178277; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm188278; AVX512BW-NEXT: vpermt2d %zmm14, %zmm13, %zmm188279; AVX512BW-NEXT: movb $-64, %dil8280; AVX512BW-NEXT: kmovd %edi, %k18281; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}8282; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm108283; AVX512BW-NEXT: vpermt2d %zmm12, %zmm13, %zmm108284; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm88285; AVX512BW-NEXT: vpermt2d %zmm9, %zmm13, %zmm88286; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]8287; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm288288; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm88289; AVX512BW-NEXT: vpermt2d %zmm7, %zmm13, %zmm88290; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm108291; AVX512BW-NEXT: vpermt2d %zmm31, %zmm13, %zmm108292; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8293; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm88294; AVX512BW-NEXT: vpermt2d %zmm30, %zmm13, %zmm88295; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm138296; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8297; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm178298; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]8299; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8300; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm108301; AVX512BW-NEXT: vpermt2d %zmm16, %zmm8, %zmm108302; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm138303; AVX512BW-NEXT: vpermt2d %zmm14, %zmm8, %zmm138304; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8305; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm108306; AVX512BW-NEXT: vpermt2d %zmm12, %zmm8, %zmm108307; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm48308; AVX512BW-NEXT: vpermt2d %zmm9, %zmm8, %zmm48309; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8310; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm188311; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm48312; AVX512BW-NEXT: vpermt2d %zmm7, %zmm8, %zmm48313; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm108314; AVX512BW-NEXT: vpermt2d %zmm31, %zmm8, %zmm108315; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}8316; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm48317; AVX512BW-NEXT: vpermt2d %zmm30, %zmm8, %zmm48318; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm88319; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]8320; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm198321; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]8322; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8323; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm88324; AVX512BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88325; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm108326; AVX512BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108327; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8328; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm88329; AVX512BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88330; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm138331; AVX512BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138332; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8333; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm208334; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm88335; AVX512BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88336; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm108337; AVX512BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108338; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8339; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm88340; AVX512BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88341; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48342; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8343; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm218344; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]8345; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8346; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm88347; AVX512BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88348; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm108349; AVX512BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108350; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8351; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm88352; AVX512BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88353; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm138354; AVX512BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138355; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8356; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm228357; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm88358; AVX512BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88359; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm108360; AVX512BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108361; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8362; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm88363; AVX512BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88364; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48365; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8366; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm238367; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]8368; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8369; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm88370; AVX512BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88371; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm108372; AVX512BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108373; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8374; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm88375; AVX512BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88376; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm138377; AVX512BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138378; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8379; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm248380; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm88381; AVX512BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88382; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm108383; AVX512BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108384; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8385; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm88386; AVX512BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88387; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48388; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8389; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm258390; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]8391; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8392; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm88393; AVX512BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88394; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm108395; AVX512BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108396; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8397; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm88398; AVX512BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88399; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm138400; AVX512BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138401; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8402; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm268403; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm88404; AVX512BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88405; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm108406; AVX512BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108407; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8408; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm88409; AVX512BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88410; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48411; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8412; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm278413; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]8414; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8415; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm88416; AVX512BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88417; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm108418; AVX512BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108419; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8420; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm88421; AVX512BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88422; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm138423; AVX512BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138424; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8425; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm88426; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm108427; AVX512BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm108428; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm138429; AVX512BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm138430; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8431; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm108432; AVX512BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm108433; AVX512BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48434; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8435; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm48436; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]8437; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8438; AVX512BW-NEXT: vpermt2d %zmm16, %zmm10, %zmm158439; AVX512BW-NEXT: vpermt2d %zmm14, %zmm10, %zmm118440; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}8441; AVX512BW-NEXT: vpermt2d %zmm12, %zmm10, %zmm28442; AVX512BW-NEXT: vpermt2d %zmm9, %zmm10, %zmm58443; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]8444; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm28445; AVX512BW-NEXT: vpermt2d %zmm7, %zmm10, %zmm68446; AVX512BW-NEXT: vpermt2d %zmm31, %zmm10, %zmm38447; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}8448; AVX512BW-NEXT: vpermt2d %zmm30, %zmm10, %zmm18449; AVX512BW-NEXT: vpermt2d %zmm29, %zmm10, %zmm08450; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]8451; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm08452; AVX512BW-NEXT: vmovdqa64 %zmm28, 64(%rsi)8453; AVX512BW-NEXT: vmovdqa64 %zmm17, (%rsi)8454; AVX512BW-NEXT: vmovdqa64 %zmm18, 64(%rdx)8455; AVX512BW-NEXT: vmovdqa64 %zmm19, (%rdx)8456; AVX512BW-NEXT: vmovdqa64 %zmm20, 64(%rcx)8457; AVX512BW-NEXT: vmovdqa64 %zmm21, (%rcx)8458; AVX512BW-NEXT: vmovdqa64 %zmm22, 64(%r8)8459; AVX512BW-NEXT: vmovdqa64 %zmm23, (%r8)8460; AVX512BW-NEXT: vmovdqa64 %zmm24, 64(%r9)8461; AVX512BW-NEXT: vmovdqa64 %zmm25, (%r9)8462; AVX512BW-NEXT: vmovdqa64 %zmm26, 64(%r11)8463; AVX512BW-NEXT: vmovdqa64 %zmm27, (%r11)8464; AVX512BW-NEXT: vmovdqa64 %zmm8, 64(%r10)8465; AVX512BW-NEXT: vmovdqa64 %zmm4, (%r10)8466; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rax)8467; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rax)8468; AVX512BW-NEXT: vzeroupper8469; AVX512BW-NEXT: retq8470;8471; AVX512BW-FCP-LABEL: load_i32_stride8_vf32:8472; AVX512BW-FCP: # %bb.0:8473; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax8474; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r108475; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r118476; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm08477; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm298478; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm18479; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm308480; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm318481; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm38482; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm78483; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm68484; AVX512BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm98485; AVX512BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm58486; AVX512BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm128487; AVX512BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm28488; AVX512BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm148489; AVX512BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm118490; AVX512BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm168491; AVX512BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm158492; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]8493; AVX512BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8494; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm178495; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm13, %zmm178496; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm188497; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm13, %zmm188498; AVX512BW-FCP-NEXT: movb $-64, %dil8499; AVX512BW-FCP-NEXT: kmovd %edi, %k18500; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}8501; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm108502; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm13, %zmm108503; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm88504; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm13, %zmm88505; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]8506; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm288507; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88508; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm88509; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108510; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm13, %zmm108511; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8512; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88513; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm13, %zmm88514; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm138515; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8516; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm178517; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]8518; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8519; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm108520; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm8, %zmm108521; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm138522; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm8, %zmm138523; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8524; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm108525; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm8, %zmm108526; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm48527; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm8, %zmm48528; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8529; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm188530; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm48531; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm48532; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108533; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm8, %zmm108534; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}8535; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm48536; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm8, %zmm48537; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm88538; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]8539; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm198540; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]8541; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8542; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm88543; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88544; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm108545; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108546; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8547; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm88548; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88549; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm138550; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138551; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8552; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm208553; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88554; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88555; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108556; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108557; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8558; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88559; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88560; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48561; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8562; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm218563; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]8564; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8565; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm88566; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88567; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm108568; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108569; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8570; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm88571; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88572; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm138573; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138574; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8575; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm228576; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88577; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88578; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108579; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108580; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8581; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88582; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88583; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48584; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8585; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm238586; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]8587; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8588; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm88589; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88590; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm108591; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108592; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8593; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm88594; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88595; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm138596; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138597; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8598; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm248599; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88600; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88601; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108602; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108603; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8604; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88605; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88606; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48607; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8608; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm258609; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]8610; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8611; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm88612; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88613; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm108614; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108615; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8616; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm88617; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88618; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm138619; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138620; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8621; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm268622; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88623; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88624; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108625; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108626; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8627; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88628; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88629; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48630; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8631; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm278632; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]8633; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8634; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm88635; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88636; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm108637; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108638; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8639; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm88640; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88641; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm138642; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138643; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8644; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm88645; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm108646; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm108647; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm138648; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm138649; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8650; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm108651; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm108652; AVX512BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48653; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8654; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm48655; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]8656; AVX512BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8657; AVX512BW-FCP-NEXT: vpermt2d %zmm16, %zmm10, %zmm158658; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm10, %zmm118659; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}8660; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm28661; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm10, %zmm58662; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]8663; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm28664; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm68665; AVX512BW-FCP-NEXT: vpermt2d %zmm31, %zmm10, %zmm38666; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}8667; AVX512BW-FCP-NEXT: vpermt2d %zmm30, %zmm10, %zmm18668; AVX512BW-FCP-NEXT: vpermt2d %zmm29, %zmm10, %zmm08669; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]8670; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm08671; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, 64(%rsi)8672; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, (%rsi)8673; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, 64(%rdx)8674; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, (%rdx)8675; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 64(%rcx)8676; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, (%rcx)8677; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 64(%r8)8678; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, (%r8)8679; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, 64(%r9)8680; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, (%r9)8681; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, 64(%r11)8682; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, (%r11)8683; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 64(%r10)8684; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, (%r10)8685; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)8686; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)8687; AVX512BW-FCP-NEXT: vzeroupper8688; AVX512BW-FCP-NEXT: retq8689;8690; AVX512DQ-BW-LABEL: load_i32_stride8_vf32:8691; AVX512DQ-BW: # %bb.0:8692; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax8693; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r108694; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r118695; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm08696; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm298697; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm18698; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm308699; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm318700; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm38701; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm78702; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm68703; AVX512DQ-BW-NEXT: vmovdqa64 576(%rdi), %zmm98704; AVX512DQ-BW-NEXT: vmovdqa64 512(%rdi), %zmm58705; AVX512DQ-BW-NEXT: vmovdqa64 704(%rdi), %zmm128706; AVX512DQ-BW-NEXT: vmovdqa64 640(%rdi), %zmm28707; AVX512DQ-BW-NEXT: vmovdqa64 832(%rdi), %zmm148708; AVX512DQ-BW-NEXT: vmovdqa64 768(%rdi), %zmm118709; AVX512DQ-BW-NEXT: vmovdqa64 960(%rdi), %zmm168710; AVX512DQ-BW-NEXT: vmovdqa64 896(%rdi), %zmm158711; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]8712; AVX512DQ-BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8713; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm178714; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm13, %zmm178715; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm188716; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm13, %zmm188717; AVX512DQ-BW-NEXT: movb $-64, %dil8718; AVX512DQ-BW-NEXT: kmovd %edi, %k18719; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}8720; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm108721; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm13, %zmm108722; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm88723; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm13, %zmm88724; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]8725; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm288726; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm88727; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm13, %zmm88728; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm108729; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm13, %zmm108730; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8731; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm88732; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm13, %zmm88733; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm138734; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8735; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm178736; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]8737; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8738; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm108739; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm8, %zmm108740; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm138741; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm8, %zmm138742; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8743; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm108744; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm8, %zmm108745; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm48746; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm8, %zmm48747; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8748; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm188749; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm48750; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm8, %zmm48751; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm108752; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm8, %zmm108753; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}8754; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm48755; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm8, %zmm48756; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm88757; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]8758; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm198759; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]8760; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8761; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm88762; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88763; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm108764; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108765; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8766; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm88767; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88768; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm138769; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138770; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8771; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm208772; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm88773; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88774; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm108775; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108776; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8777; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm88778; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88779; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48780; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8781; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm218782; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]8783; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8784; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm88785; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88786; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm108787; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108788; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8789; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm88790; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88791; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm138792; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138793; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8794; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm228795; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm88796; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88797; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm108798; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108799; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8800; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm88801; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88802; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48803; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8804; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm238805; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]8806; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8807; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm88808; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88809; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm108810; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108811; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8812; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm88813; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88814; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm138815; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138816; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8817; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm248818; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm88819; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88820; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm108821; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108822; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8823; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm88824; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88825; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48826; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8827; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm258828; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]8829; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8830; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm88831; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88832; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm108833; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108834; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8835; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm88836; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88837; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm138838; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138839; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8840; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm268841; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm88842; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm88843; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm108844; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm108845; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8846; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm88847; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm88848; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48849; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]8850; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm278851; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]8852; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8853; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm88854; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm4, %zmm88855; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm108856; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm4, %zmm108857; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8858; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm88859; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm4, %zmm88860; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm138861; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm4, %zmm138862; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8863; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm88864; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm108865; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm4, %zmm108866; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm138867; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm4, %zmm138868; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8869; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm108870; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm4, %zmm108871; AVX512DQ-BW-NEXT: vpermi2d %zmm29, %zmm0, %zmm48872; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8873; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm48874; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]8875; AVX512DQ-BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8876; AVX512DQ-BW-NEXT: vpermt2d %zmm16, %zmm10, %zmm158877; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm10, %zmm118878; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}8879; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm10, %zmm28880; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm10, %zmm58881; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]8882; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm28883; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm10, %zmm68884; AVX512DQ-BW-NEXT: vpermt2d %zmm31, %zmm10, %zmm38885; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}8886; AVX512DQ-BW-NEXT: vpermt2d %zmm30, %zmm10, %zmm18887; AVX512DQ-BW-NEXT: vpermt2d %zmm29, %zmm10, %zmm08888; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]8889; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm08890; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, 64(%rsi)8891; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, (%rsi)8892; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 64(%rdx)8893; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, (%rdx)8894; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 64(%rcx)8895; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, (%rcx)8896; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 64(%r8)8897; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, (%r8)8898; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, 64(%r9)8899; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, (%r9)8900; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, 64(%r11)8901; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, (%r11)8902; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 64(%r10)8903; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, (%r10)8904; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rax)8905; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%rax)8906; AVX512DQ-BW-NEXT: vzeroupper8907; AVX512DQ-BW-NEXT: retq8908;8909; AVX512DQ-BW-FCP-LABEL: load_i32_stride8_vf32:8910; AVX512DQ-BW-FCP: # %bb.0:8911; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax8912; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r108913; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r118914; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm08915; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm298916; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm18917; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm308918; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm318919; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm38920; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm78921; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm68922; AVX512DQ-BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm98923; AVX512DQ-BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm58924; AVX512DQ-BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm128925; AVX512DQ-BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm28926; AVX512DQ-BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm148927; AVX512DQ-BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm118928; AVX512DQ-BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm168929; AVX512DQ-BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm158930; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]8931; AVX512DQ-BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8932; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm178933; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm13, %zmm178934; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm188935; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm13, %zmm188936; AVX512DQ-BW-FCP-NEXT: movb $-64, %dil8937; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k18938; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}8939; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm108940; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm13, %zmm108941; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm88942; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm13, %zmm88943; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm10[4,5,6,7]8944; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm18, %zmm288945; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88946; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm13, %zmm88947; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108948; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm13, %zmm108949; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8950; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88951; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm13, %zmm88952; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm138953; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8954; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm178955; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]8956; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8957; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm108958; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm8, %zmm108959; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm138960; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm8, %zmm138961; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}8962; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm108963; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm8, %zmm108964; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm48965; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm8, %zmm48966; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]8967; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm188968; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm48969; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm8, %zmm48970; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108971; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm8, %zmm108972; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}8973; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm48974; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm8, %zmm48975; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm88976; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]8977; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm198978; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]8979; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]8980; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm88981; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm88982; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm108983; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm108984; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8985; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm88986; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm88987; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm138988; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm138989; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]8990; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm208991; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm88992; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm88993; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm108994; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm108995; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}8996; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm88997; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm88998; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm48999; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]9000; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm219001; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]9002; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9003; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm89004; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm89005; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm109006; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm109007; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9008; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm89009; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm89010; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm139011; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm139012; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]9013; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm229014; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm89015; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm89016; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm109017; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm109018; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9019; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm89020; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm89021; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm49022; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]9023; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm239024; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]9025; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9026; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm89027; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm89028; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm109029; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm109030; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9031; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm89032; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm89033; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm139034; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm139035; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]9036; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm249037; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm89038; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm89039; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm109040; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm109041; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9042; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm89043; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm89044; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm49045; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]9046; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm259047; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]9048; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9049; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm89050; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm89051; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm109052; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm109053; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9054; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm89055; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm89056; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm139057; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm139058; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]9059; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm269060; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm89061; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm89062; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm109063; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm109064; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9065; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm89066; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm89067; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm49068; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm8[4,5,6,7]9069; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm279070; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]9071; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9072; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm89073; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm4, %zmm89074; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm109075; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm4, %zmm109076; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}9077; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm89078; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm4, %zmm89079; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm139080; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm4, %zmm139081; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]9082; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm10, %zmm89083; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm109084; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm4, %zmm109085; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm139086; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm4, %zmm139087; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm13 {%k1}9088; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm109089; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm4, %zmm109090; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm29, %zmm0, %zmm49091; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]9092; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm13, %zmm49093; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]9094; AVX512DQ-BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]9095; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm16, %zmm10, %zmm159096; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm10, %zmm119097; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm11 {%k1}9098; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm29099; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm10, %zmm59100; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]9101; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm29102; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm10, %zmm69103; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm31, %zmm10, %zmm39104; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3 {%k1}9105; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm30, %zmm10, %zmm19106; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm29, %zmm10, %zmm09107; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]9108; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm09109; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, 64(%rsi)9110; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, (%rsi)9111; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, 64(%rdx)9112; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, (%rdx)9113; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 64(%rcx)9114; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, (%rcx)9115; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 64(%r8)9116; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, (%r8)9117; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, 64(%r9)9118; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, (%r9)9119; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, 64(%r11)9120; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, (%r11)9121; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 64(%r10)9122; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, (%r10)9123; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)9124; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)9125; AVX512DQ-BW-FCP-NEXT: vzeroupper9126; AVX512DQ-BW-FCP-NEXT: retq9127 %wide.vec = load <256 x i32>, ptr %in.vec, align 649128 %strided.vec0 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248>9129 %strided.vec1 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249>9130 %strided.vec2 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250>9131 %strided.vec3 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251>9132 %strided.vec4 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252>9133 %strided.vec5 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253>9134 %strided.vec6 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254>9135 %strided.vec7 = shufflevector <256 x i32> %wide.vec, <256 x i32> poison, <32 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255>9136 store <32 x i32> %strided.vec0, ptr %out.vec0, align 649137 store <32 x i32> %strided.vec1, ptr %out.vec1, align 649138 store <32 x i32> %strided.vec2, ptr %out.vec2, align 649139 store <32 x i32> %strided.vec3, ptr %out.vec3, align 649140 store <32 x i32> %strided.vec4, ptr %out.vec4, align 649141 store <32 x i32> %strided.vec5, ptr %out.vec5, align 649142 store <32 x i32> %strided.vec6, ptr %out.vec6, align 649143 store <32 x i32> %strided.vec7, ptr %out.vec7, align 649144 ret void9145}9146 9147define void @load_i32_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {9148; SSE-LABEL: load_i32_stride8_vf64:9149; SSE: # %bb.0:9150; SSE-NEXT: subq $2232, %rsp # imm = 0x8B89151; SSE-NEXT: movaps 288(%rdi), %xmm49152; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9153; SSE-NEXT: movaps 352(%rdi), %xmm59154; SSE-NEXT: movaps %xmm5, (%rsp) # 16-byte Spill9155; SSE-NEXT: movaps 320(%rdi), %xmm69156; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9157; SSE-NEXT: movaps 416(%rdi), %xmm79158; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9159; SSE-NEXT: movaps 384(%rdi), %xmm89160; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9161; SSE-NEXT: movaps 480(%rdi), %xmm99162; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9163; SSE-NEXT: movaps 448(%rdi), %xmm39164; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9165; SSE-NEXT: movaps 160(%rdi), %xmm109166; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9167; SSE-NEXT: movaps 128(%rdi), %xmm19168; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9169; SSE-NEXT: movaps 224(%rdi), %xmm29170; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9171; SSE-NEXT: movaps 192(%rdi), %xmm09172; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9173; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]9174; SSE-NEXT: movaps %xmm1, %xmm29175; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1]9176; SSE-NEXT: movaps %xmm2, %xmm19177; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]9178; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9179; SSE-NEXT: movaps %xmm3, %xmm19180; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]9181; SSE-NEXT: movaps %xmm8, %xmm39182; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1]9183; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]9184; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9185; SSE-NEXT: movaps %xmm3, %xmm09186; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]9187; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9188; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm1[1]9189; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9190; SSE-NEXT: movaps %xmm6, %xmm09191; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]9192; SSE-NEXT: movaps 256(%rdi), %xmm29193; SSE-NEXT: movaps %xmm2, %xmm19194; SSE-NEXT: movaps %xmm2, %xmm39195; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]9196; SSE-NEXT: movaps %xmm1, %xmm29197; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9198; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9199; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9200; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9201; SSE-NEXT: movaps 736(%rdi), %xmm99202; SSE-NEXT: movaps 704(%rdi), %xmm09203; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9204; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]9205; SSE-NEXT: movaps 672(%rdi), %xmm29206; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9207; SSE-NEXT: movaps 640(%rdi), %xmm19208; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9209; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9210; SSE-NEXT: movaps %xmm1, %xmm29211; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9212; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9213; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9214; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9215; SSE-NEXT: movaps 608(%rdi), %xmm29216; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9217; SSE-NEXT: movaps 576(%rdi), %xmm19218; SSE-NEXT: movaps %xmm1, %xmm09219; SSE-NEXT: movaps %xmm1, %xmm49220; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]9221; SSE-NEXT: movaps 544(%rdi), %xmm159222; SSE-NEXT: movaps 512(%rdi), %xmm29223; SSE-NEXT: movaps %xmm2, %xmm19224; SSE-NEXT: movaps %xmm2, %xmm69225; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1]9226; SSE-NEXT: movaps %xmm1, %xmm29227; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9228; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9229; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9230; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9231; SSE-NEXT: movaps 992(%rdi), %xmm19232; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9233; SSE-NEXT: movaps 960(%rdi), %xmm09234; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9235; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9236; SSE-NEXT: movaps 928(%rdi), %xmm29237; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9238; SSE-NEXT: movaps 896(%rdi), %xmm19239; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9240; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9241; SSE-NEXT: movaps %xmm1, %xmm29242; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9243; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9244; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9245; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9246; SSE-NEXT: movaps 864(%rdi), %xmm19247; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9248; SSE-NEXT: movaps 832(%rdi), %xmm09249; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9250; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9251; SSE-NEXT: movaps 800(%rdi), %xmm149252; SSE-NEXT: movaps 768(%rdi), %xmm29253; SSE-NEXT: movaps %xmm2, %xmm19254; SSE-NEXT: movaps %xmm2, %xmm89255; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1]9256; SSE-NEXT: movaps %xmm1, %xmm29257; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9258; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9259; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9260; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9261; SSE-NEXT: movaps 1248(%rdi), %xmm19262; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9263; SSE-NEXT: movaps 1216(%rdi), %xmm09264; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9265; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9266; SSE-NEXT: movaps 1184(%rdi), %xmm29267; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9268; SSE-NEXT: movaps 1152(%rdi), %xmm19269; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9270; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9271; SSE-NEXT: movaps %xmm1, %xmm29272; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9273; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9274; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9275; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9276; SSE-NEXT: movaps 1120(%rdi), %xmm29277; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9278; SSE-NEXT: movaps 1088(%rdi), %xmm19279; SSE-NEXT: movaps %xmm1, %xmm09280; SSE-NEXT: movaps %xmm1, %xmm79281; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]9282; SSE-NEXT: movaps 1056(%rdi), %xmm29283; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9284; SSE-NEXT: movaps 1024(%rdi), %xmm19285; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9286; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9287; SSE-NEXT: movaps %xmm1, %xmm29288; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9289; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9290; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9291; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9292; SSE-NEXT: movaps 1504(%rdi), %xmm19293; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9294; SSE-NEXT: movaps 1472(%rdi), %xmm09295; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9296; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9297; SSE-NEXT: movaps 1440(%rdi), %xmm29298; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9299; SSE-NEXT: movaps 1408(%rdi), %xmm19300; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9301; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9302; SSE-NEXT: movaps %xmm1, %xmm29303; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9304; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9305; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9306; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9307; SSE-NEXT: movaps 1376(%rdi), %xmm19308; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9309; SSE-NEXT: movaps 1344(%rdi), %xmm09310; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9311; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9312; SSE-NEXT: movaps 1312(%rdi), %xmm29313; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9314; SSE-NEXT: movaps 1280(%rdi), %xmm19315; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9316; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9317; SSE-NEXT: movaps %xmm1, %xmm29318; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9319; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9320; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9321; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9322; SSE-NEXT: movaps 1760(%rdi), %xmm19323; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9324; SSE-NEXT: movaps 1728(%rdi), %xmm09325; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9326; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9327; SSE-NEXT: movaps 1696(%rdi), %xmm29328; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9329; SSE-NEXT: movaps 1664(%rdi), %xmm19330; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9331; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9332; SSE-NEXT: movaps %xmm1, %xmm29333; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9334; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9335; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9336; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9337; SSE-NEXT: movaps 1632(%rdi), %xmm19338; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9339; SSE-NEXT: movaps 1600(%rdi), %xmm09340; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9341; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9342; SSE-NEXT: movaps 1568(%rdi), %xmm59343; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9344; SSE-NEXT: movaps 1536(%rdi), %xmm29345; SSE-NEXT: movaps %xmm2, %xmm19346; SSE-NEXT: movaps %xmm2, %xmm139347; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]9348; SSE-NEXT: movaps %xmm1, %xmm29349; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9350; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9351; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9352; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9353; SSE-NEXT: movaps 2016(%rdi), %xmm19354; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9355; SSE-NEXT: movaps 1984(%rdi), %xmm09356; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9357; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9358; SSE-NEXT: movaps 1952(%rdi), %xmm29359; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9360; SSE-NEXT: movaps 1920(%rdi), %xmm19361; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9362; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9363; SSE-NEXT: movaps %xmm1, %xmm29364; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9365; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9366; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9367; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9368; SSE-NEXT: movaps 1888(%rdi), %xmm19369; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9370; SSE-NEXT: movaps 1856(%rdi), %xmm09371; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9372; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9373; SSE-NEXT: movaps 1824(%rdi), %xmm29374; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9375; SSE-NEXT: movaps 1792(%rdi), %xmm19376; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9377; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9378; SSE-NEXT: movaps %xmm1, %xmm59379; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0]9380; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9381; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9382; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9383; SSE-NEXT: movaps 96(%rdi), %xmm29384; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9385; SSE-NEXT: movaps 64(%rdi), %xmm129386; SSE-NEXT: movaps %xmm12, %xmm09387; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]9388; SSE-NEXT: movaps (%rdi), %xmm109389; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9390; SSE-NEXT: movaps 32(%rdi), %xmm19391; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9392; SSE-NEXT: unpcklps {{.*#+}} xmm10 = xmm10[0],xmm1[0],xmm10[1],xmm1[1]9393; SSE-NEXT: movaps %xmm10, %xmm59394; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0]9395; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9396; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]9397; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9398; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload9399; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload9400; SSE-NEXT: # xmm11 = xmm11[2],mem[2],xmm11[3],mem[3]9401; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload9402; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload9403; SSE-NEXT: # xmm5 = xmm5[2],mem[2],xmm5[3],mem[3]9404; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload9405; SSE-NEXT: unpckhps (%rsp), %xmm10 # 16-byte Folded Reload9406; SSE-NEXT: # xmm10 = xmm10[2],mem[2],xmm10[3],mem[3]9407; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload9408; SSE-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]9409; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9410; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload9411; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload9412; SSE-NEXT: # xmm3 = xmm3[2],mem[2],xmm3[3],mem[3]9413; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9414; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload9415; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]9416; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload9417; SSE-NEXT: # xmm4 = xmm4[2],mem[2],xmm4[3],mem[3]9418; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9419; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm15[2],xmm6[3],xmm15[3]9420; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9421; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9422; SSE-NEXT: unpckhps {{.*#+}} xmm15 = xmm15[2],xmm9[2],xmm15[3],xmm9[3]9423; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9424; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload9425; SSE-NEXT: # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]9426; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9427; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9428; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9429; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9430; SSE-NEXT: unpckhps {{.*#+}} xmm8 = xmm8[2],xmm14[2],xmm8[3],xmm14[3]9431; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9432; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload9433; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload9434; SSE-NEXT: # xmm14 = xmm14[2],mem[2],xmm14[3],mem[3]9435; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload9436; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload9437; SSE-NEXT: # xmm9 = xmm9[2],mem[2],xmm9[3],mem[3]9438; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9439; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9440; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9441; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9442; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9443; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9444; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9445; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload9446; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload9447; SSE-NEXT: # xmm8 = xmm8[2],mem[2],xmm8[3],mem[3]9448; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload9449; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload9450; SSE-NEXT: # xmm6 = xmm6[2],mem[2],xmm6[3],mem[3]9451; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9452; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9453; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9454; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9455; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9456; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9457; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9458; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9459; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload9460; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9461; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9462; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9463; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload9464; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload9465; SSE-NEXT: # xmm4 = xmm4[2],mem[2],xmm4[3],mem[3]9466; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9467; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9468; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9469; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9470; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload9471; SSE-NEXT: # xmm13 = xmm13[2],mem[2],xmm13[3],mem[3]9472; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9473; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9474; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9475; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9476; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload9477; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9478; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9479; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9480; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload9481; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9482; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9483; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9484; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload9485; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9486; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9487; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9488; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload9489; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9490; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9491; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9492; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload9493; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload9494; SSE-NEXT: # xmm13 = xmm13[2],mem[2],xmm13[3],mem[3]9495; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload9496; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]9497; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9498; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload9499; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload9500; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]9501; SSE-NEXT: movaps %xmm5, %xmm79502; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm11[0]9503; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9504; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm11[1]9505; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9506; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload9507; SSE-NEXT: movaps %xmm5, %xmm79508; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm10[0]9509; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9510; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm10[1]9511; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9512; SSE-NEXT: movaps %xmm0, %xmm59513; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]9514; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9515; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]9516; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9517; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9518; SSE-NEXT: movaps %xmm0, %xmm59519; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload9520; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]9521; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9522; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]9523; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9524; SSE-NEXT: movaps %xmm1, %xmm59525; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm15[0]9526; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9527; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm15[1]9528; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9529; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9530; SSE-NEXT: movaps %xmm0, %xmm59531; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9532; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]9533; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9534; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9535; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9536; SSE-NEXT: movaps %xmm9, %xmm59537; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm14[0]9538; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9539; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm14[1]9540; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9541; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9542; SSE-NEXT: movaps %xmm0, %xmm59543; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9544; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]9545; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9546; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9547; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9548; SSE-NEXT: movaps %xmm6, %xmm59549; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm8[0]9550; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9551; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm8[1]9552; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9553; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9554; SSE-NEXT: movaps %xmm0, %xmm59555; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9556; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]9557; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9558; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9559; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9560; SSE-NEXT: movaps %xmm4, %xmm59561; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9562; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]9563; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9564; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9565; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9566; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9567; SSE-NEXT: movaps %xmm0, %xmm59568; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9569; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]9570; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9571; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9572; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9573; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9574; SSE-NEXT: movaps %xmm0, %xmm49575; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]9576; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9577; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]9578; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9579; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload9580; SSE-NEXT: movaps %xmm6, %xmm09581; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9582; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]9583; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9584; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm1[1]9585; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9586; SSE-NEXT: movaps %xmm13, %xmm09587; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9588; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]9589; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9590; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm1[1]9591; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9592; SSE-NEXT: movaps %xmm12, %xmm09593; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9594; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]9595; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9596; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm1[1]9597; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9598; SSE-NEXT: movaps 240(%rdi), %xmm29599; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9600; SSE-NEXT: movaps 208(%rdi), %xmm79601; SSE-NEXT: movaps %xmm7, %xmm09602; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]9603; SSE-NEXT: movaps 176(%rdi), %xmm39604; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9605; SSE-NEXT: movaps 144(%rdi), %xmm29606; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9607; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]9608; SSE-NEXT: movaps %xmm2, %xmm19609; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]9610; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9611; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]9612; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9613; SSE-NEXT: movaps 368(%rdi), %xmm39614; SSE-NEXT: movaps 336(%rdi), %xmm09615; SSE-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill9616; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]9617; SSE-NEXT: movaps 304(%rdi), %xmm49618; SSE-NEXT: movaps 272(%rdi), %xmm19619; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9620; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]9621; SSE-NEXT: movaps %xmm1, %xmm29622; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9623; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9624; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9625; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9626; SSE-NEXT: movaps 496(%rdi), %xmm59627; SSE-NEXT: movaps 464(%rdi), %xmm09628; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9629; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]9630; SSE-NEXT: movaps 432(%rdi), %xmm69631; SSE-NEXT: movaps 400(%rdi), %xmm19632; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9633; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1]9634; SSE-NEXT: movaps %xmm1, %xmm29635; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9636; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9637; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9638; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9639; SSE-NEXT: movaps 624(%rdi), %xmm99640; SSE-NEXT: movaps 592(%rdi), %xmm09641; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9642; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]9643; SSE-NEXT: movaps 560(%rdi), %xmm109644; SSE-NEXT: movaps 528(%rdi), %xmm19645; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9646; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1]9647; SSE-NEXT: movaps %xmm1, %xmm29648; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9649; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9650; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9651; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9652; SSE-NEXT: movaps 752(%rdi), %xmm129653; SSE-NEXT: movaps 720(%rdi), %xmm09654; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9655; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1]9656; SSE-NEXT: movaps 688(%rdi), %xmm139657; SSE-NEXT: movaps 656(%rdi), %xmm29658; SSE-NEXT: movaps %xmm2, %xmm19659; SSE-NEXT: movaps %xmm2, %xmm159660; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1]9661; SSE-NEXT: movaps %xmm1, %xmm29662; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9663; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9664; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9665; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9666; SSE-NEXT: movaps 880(%rdi), %xmm19667; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9668; SSE-NEXT: movaps 848(%rdi), %xmm09669; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9670; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9671; SSE-NEXT: movaps 816(%rdi), %xmm29672; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9673; SSE-NEXT: movaps 784(%rdi), %xmm19674; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9675; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9676; SSE-NEXT: movaps %xmm1, %xmm29677; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9678; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9679; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9680; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9681; SSE-NEXT: movaps 1008(%rdi), %xmm149682; SSE-NEXT: movaps 976(%rdi), %xmm09683; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9684; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1]9685; SSE-NEXT: movaps 944(%rdi), %xmm29686; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9687; SSE-NEXT: movaps 912(%rdi), %xmm19688; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9689; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9690; SSE-NEXT: movaps %xmm1, %xmm29691; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9692; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9693; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9694; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9695; SSE-NEXT: movaps 1136(%rdi), %xmm19696; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9697; SSE-NEXT: movaps 1104(%rdi), %xmm09698; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9699; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9700; SSE-NEXT: movaps 1072(%rdi), %xmm29701; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9702; SSE-NEXT: movaps 1040(%rdi), %xmm19703; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9704; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9705; SSE-NEXT: movaps %xmm1, %xmm29706; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9707; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9708; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9709; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9710; SSE-NEXT: movaps 1264(%rdi), %xmm119711; SSE-NEXT: movaps 1232(%rdi), %xmm09712; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9713; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1]9714; SSE-NEXT: movaps 1200(%rdi), %xmm29715; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9716; SSE-NEXT: movaps 1168(%rdi), %xmm19717; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9718; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9719; SSE-NEXT: movaps %xmm1, %xmm29720; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9721; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9722; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9723; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9724; SSE-NEXT: movaps 1392(%rdi), %xmm19725; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9726; SSE-NEXT: movaps 1360(%rdi), %xmm09727; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9728; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9729; SSE-NEXT: movaps 1328(%rdi), %xmm29730; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9731; SSE-NEXT: movaps 1296(%rdi), %xmm19732; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9733; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9734; 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 1520(%rdi), %xmm19740; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9741; SSE-NEXT: movaps 1488(%rdi), %xmm09742; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9743; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9744; SSE-NEXT: movaps 1456(%rdi), %xmm29745; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9746; SSE-NEXT: movaps 1424(%rdi), %xmm19747; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9748; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9749; SSE-NEXT: movaps %xmm1, %xmm29750; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9751; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9752; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9753; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9754; SSE-NEXT: movaps 1648(%rdi), %xmm19755; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9756; SSE-NEXT: movaps 1616(%rdi), %xmm09757; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9758; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9759; SSE-NEXT: movaps 1584(%rdi), %xmm29760; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9761; SSE-NEXT: movaps 1552(%rdi), %xmm19762; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9763; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9764; SSE-NEXT: movaps %xmm1, %xmm29765; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9766; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9767; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9768; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9769; SSE-NEXT: movaps 1776(%rdi), %xmm89770; SSE-NEXT: movaps 1744(%rdi), %xmm09771; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9772; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1]9773; SSE-NEXT: movaps 1712(%rdi), %xmm29774; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9775; SSE-NEXT: movaps 1680(%rdi), %xmm19776; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9777; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9778; SSE-NEXT: movaps %xmm1, %xmm29779; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9780; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9781; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9782; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9783; SSE-NEXT: movaps 1904(%rdi), %xmm19784; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9785; SSE-NEXT: movaps 1872(%rdi), %xmm09786; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9787; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9788; SSE-NEXT: movaps 1840(%rdi), %xmm29789; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9790; SSE-NEXT: movaps 1808(%rdi), %xmm19791; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9792; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9793; SSE-NEXT: movaps %xmm1, %xmm29794; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9795; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9796; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9797; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9798; SSE-NEXT: movaps 2032(%rdi), %xmm19799; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9800; SSE-NEXT: movaps 2000(%rdi), %xmm09801; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9802; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9803; SSE-NEXT: movaps 1968(%rdi), %xmm29804; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9805; SSE-NEXT: movaps 1936(%rdi), %xmm19806; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9807; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9808; SSE-NEXT: movaps %xmm1, %xmm29809; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9810; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9811; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9812; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9813; SSE-NEXT: movaps 112(%rdi), %xmm19814; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9815; SSE-NEXT: movaps 80(%rdi), %xmm09816; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9817; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]9818; SSE-NEXT: movaps 16(%rdi), %xmm19819; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9820; SSE-NEXT: movaps 48(%rdi), %xmm29821; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9822; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]9823; SSE-NEXT: movaps %xmm1, %xmm29824; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9825; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9826; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]9827; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9828; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload9829; SSE-NEXT: # xmm7 = xmm7[2],mem[2],xmm7[3],mem[3]9830; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload9831; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload9832; SSE-NEXT: # xmm2 = xmm2[2],mem[2],xmm2[3],mem[3]9833; SSE-NEXT: movaps (%rsp), %xmm1 # 16-byte Reload9834; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]9835; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill9836; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9837; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm4[2],xmm1[3],xmm4[3]9838; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9839; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9840; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]9841; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9842; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9843; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm6[2],xmm1[3],xmm6[3]9844; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9845; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm9[2],xmm0[3],xmm9[3]9846; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9847; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload9848; SSE-NEXT: unpckhps {{.*#+}} xmm5 = xmm5[2],xmm10[2],xmm5[3],xmm10[3]9849; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload9850; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm12[2],xmm6[3],xmm12[3]9851; SSE-NEXT: movaps %xmm15, %xmm49852; SSE-NEXT: unpckhps {{.*#+}} xmm4 = xmm4[2],xmm13[2],xmm4[3],xmm13[3]9853; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9854; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload9855; SSE-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]9856; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9857; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9858; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload9859; SSE-NEXT: # xmm15 = xmm15[2],mem[2],xmm15[3],mem[3]9860; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload9861; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm14[2],xmm3[3],xmm14[3]9862; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload9863; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload9864; SSE-NEXT: # xmm14 = xmm14[2],mem[2],xmm14[3],mem[3]9865; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9866; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload9867; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload9868; SSE-NEXT: # xmm14 = xmm14[2],mem[2],xmm14[3],mem[3]9869; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload9870; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload9871; SSE-NEXT: # xmm13 = xmm13[2],mem[2],xmm13[3],mem[3]9872; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9873; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload9874; SSE-NEXT: unpckhps {{.*#+}} xmm13 = xmm13[2],xmm11[2],xmm13[3],xmm11[3]9875; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload9876; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload9877; SSE-NEXT: # xmm11 = xmm11[2],mem[2],xmm11[3],mem[3]9878; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9879; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload9880; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload9881; SSE-NEXT: # xmm11 = xmm11[2],mem[2],xmm11[3],mem[3]9882; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload9883; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload9884; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]9885; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9886; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload9887; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload9888; SSE-NEXT: # xmm12 = xmm12[2],mem[2],xmm12[3],mem[3]9889; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload9890; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload9891; SSE-NEXT: # xmm10 = xmm10[2],mem[2],xmm10[3],mem[3]9892; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9893; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload9894; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload9895; SSE-NEXT: # xmm10 = xmm10[2],mem[2],xmm10[3],mem[3]9896; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload9897; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload9898; SSE-NEXT: # xmm9 = xmm9[2],mem[2],xmm9[3],mem[3]9899; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9900; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload9901; SSE-NEXT: unpckhps {{.*#+}} xmm9 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]9902; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload9903; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload9904; SSE-NEXT: # xmm8 = xmm8[2],mem[2],xmm8[3],mem[3]9905; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9906; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload9907; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload9908; SSE-NEXT: # xmm8 = xmm8[2],mem[2],xmm8[3],mem[3]9909; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9910; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload9911; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]9912; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9913; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9914; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload9915; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]9916; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9917; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9918; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload9919; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]9920; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9921; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9922; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload9923; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]9924; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9925; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9926; SSE-NEXT: unpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload9927; SSE-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]9928; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9929; SSE-NEXT: movaps %xmm2, %xmm09930; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm7[0]9931; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9932; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm7[1]9933; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9934; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9935; SSE-NEXT: movaps %xmm0, %xmm29936; SSE-NEXT: movaps (%rsp), %xmm7 # 16-byte Reload9937; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm7[0]9938; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9939; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm7[1]9940; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9941; SSE-NEXT: movaps %xmm1, %xmm09942; SSE-NEXT: movaps %xmm1, %xmm29943; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9944; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]9945; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9946; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9947; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9948; SSE-NEXT: movaps %xmm5, %xmm29949; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9950; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]9951; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9952; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm1[1]9953; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9954; SSE-NEXT: movaps %xmm4, %xmm29955; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm6[0]9956; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9957; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]9958; SSE-NEXT: movaps %xmm4, (%rsp) # 16-byte Spill9959; SSE-NEXT: movaps %xmm15, %xmm09960; SSE-NEXT: movaps %xmm15, %xmm29961; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload9962; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]9963; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9964; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9965; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9966; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9967; SSE-NEXT: movaps %xmm0, %xmm159968; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm3[0]9969; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]9970; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9971; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9972; SSE-NEXT: movaps %xmm0, %xmm29973; SSE-NEXT: movaps %xmm14, %xmm19974; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm14[0]9975; SSE-NEXT: movaps %xmm2, %xmm149976; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9977; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9978; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9979; SSE-NEXT: movaps %xmm0, %xmm29980; SSE-NEXT: movaps %xmm13, %xmm19981; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm13[0]9982; SSE-NEXT: movaps %xmm2, %xmm139983; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]9984; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9985; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9986; SSE-NEXT: movaps %xmm0, %xmm79987; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm11[0]9988; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm11[1]9989; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9990; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9991; SSE-NEXT: movaps %xmm0, %xmm69992; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm12[0]9993; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm12[1]9994; SSE-NEXT: movaps %xmm0, %xmm129995; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9996; SSE-NEXT: movaps %xmm0, %xmm29997; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm10[0]9998; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm10[1]9999; SSE-NEXT: movaps %xmm0, %xmm1010000; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10001; SSE-NEXT: movaps %xmm0, %xmm510002; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm9[0]10003; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm9[1]10004; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10005; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10006; SSE-NEXT: movaps %xmm0, %xmm410007; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm8[0]10008; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm8[1]10009; SSE-NEXT: movaps %xmm0, %xmm810010; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10011; SSE-NEXT: movaps %xmm0, %xmm310012; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10013; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]10014; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]10015; SSE-NEXT: movaps %xmm0, %xmm910016; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload10017; SSE-NEXT: movaps %xmm11, %xmm010018; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10019; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]10020; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm1[1]10021; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10022; SSE-NEXT: movaps %xmm1, 224(%rsi)10023; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10024; SSE-NEXT: movaps %xmm1, 160(%rsi)10025; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10026; SSE-NEXT: movaps %xmm1, 96(%rsi)10027; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10028; SSE-NEXT: movaps %xmm1, 32(%rsi)10029; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10030; SSE-NEXT: movaps %xmm1, 240(%rsi)10031; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10032; SSE-NEXT: movaps %xmm1, 176(%rsi)10033; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10034; SSE-NEXT: movaps %xmm1, 112(%rsi)10035; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10036; SSE-NEXT: movaps %xmm1, 48(%rsi)10037; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10038; SSE-NEXT: movaps %xmm1, 192(%rsi)10039; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10040; SSE-NEXT: movaps %xmm1, 128(%rsi)10041; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10042; SSE-NEXT: movaps %xmm1, 64(%rsi)10043; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10044; SSE-NEXT: movaps %xmm1, (%rsi)10045; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10046; SSE-NEXT: movaps %xmm1, 208(%rsi)10047; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10048; SSE-NEXT: movaps %xmm1, 144(%rsi)10049; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10050; SSE-NEXT: movaps %xmm1, 80(%rsi)10051; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10052; SSE-NEXT: movaps %xmm1, 16(%rsi)10053; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10054; SSE-NEXT: movaps %xmm1, 224(%rdx)10055; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10056; SSE-NEXT: movaps %xmm1, 240(%rdx)10057; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10058; SSE-NEXT: movaps %xmm1, 192(%rdx)10059; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10060; SSE-NEXT: movaps %xmm1, 208(%rdx)10061; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10062; SSE-NEXT: movaps %xmm1, 160(%rdx)10063; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10064; SSE-NEXT: movaps %xmm1, 176(%rdx)10065; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10066; SSE-NEXT: movaps %xmm1, 128(%rdx)10067; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10068; SSE-NEXT: movaps %xmm1, 144(%rdx)10069; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10070; SSE-NEXT: movaps %xmm1, 96(%rdx)10071; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10072; SSE-NEXT: movaps %xmm1, 112(%rdx)10073; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10074; SSE-NEXT: movaps %xmm1, 64(%rdx)10075; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10076; SSE-NEXT: movaps %xmm1, 80(%rdx)10077; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10078; SSE-NEXT: movaps %xmm1, 32(%rdx)10079; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10080; SSE-NEXT: movaps %xmm1, 48(%rdx)10081; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10082; SSE-NEXT: movaps %xmm1, (%rdx)10083; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10084; SSE-NEXT: movaps %xmm1, 16(%rdx)10085; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10086; SSE-NEXT: movaps %xmm1, 240(%rcx)10087; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10088; SSE-NEXT: movaps %xmm1, 224(%rcx)10089; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10090; SSE-NEXT: movaps %xmm1, 208(%rcx)10091; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10092; SSE-NEXT: movaps %xmm1, 192(%rcx)10093; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10094; SSE-NEXT: movaps %xmm1, 176(%rcx)10095; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10096; SSE-NEXT: movaps %xmm1, 160(%rcx)10097; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10098; SSE-NEXT: movaps %xmm1, 144(%rcx)10099; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10100; SSE-NEXT: movaps %xmm1, 128(%rcx)10101; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10102; SSE-NEXT: movaps %xmm1, 112(%rcx)10103; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10104; SSE-NEXT: movaps %xmm1, 96(%rcx)10105; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10106; SSE-NEXT: movaps %xmm1, 80(%rcx)10107; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10108; SSE-NEXT: movaps %xmm1, 64(%rcx)10109; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10110; SSE-NEXT: movaps %xmm1, 48(%rcx)10111; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10112; SSE-NEXT: movaps %xmm1, 32(%rcx)10113; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10114; SSE-NEXT: movaps %xmm1, 16(%rcx)10115; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10116; SSE-NEXT: movaps %xmm1, (%rcx)10117; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10118; SSE-NEXT: movaps %xmm1, 240(%r8)10119; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10120; SSE-NEXT: movaps %xmm1, 224(%r8)10121; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10122; SSE-NEXT: movaps %xmm1, 208(%r8)10123; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10124; SSE-NEXT: movaps %xmm1, 192(%r8)10125; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10126; SSE-NEXT: movaps %xmm1, 176(%r8)10127; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10128; SSE-NEXT: movaps %xmm1, 160(%r8)10129; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10130; SSE-NEXT: movaps %xmm1, 144(%r8)10131; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10132; SSE-NEXT: movaps %xmm1, 128(%r8)10133; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10134; SSE-NEXT: movaps %xmm1, 112(%r8)10135; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10136; SSE-NEXT: movaps %xmm1, 96(%r8)10137; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10138; SSE-NEXT: movaps %xmm1, 80(%r8)10139; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10140; SSE-NEXT: movaps %xmm1, 64(%r8)10141; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10142; SSE-NEXT: movaps %xmm1, 48(%r8)10143; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10144; SSE-NEXT: movaps %xmm1, 32(%r8)10145; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10146; SSE-NEXT: movaps %xmm1, 16(%r8)10147; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10148; SSE-NEXT: movaps %xmm1, (%r8)10149; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10150; SSE-NEXT: movaps %xmm1, 240(%r9)10151; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10152; SSE-NEXT: movaps %xmm1, 224(%r9)10153; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10154; SSE-NEXT: movaps %xmm1, 208(%r9)10155; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10156; SSE-NEXT: movaps %xmm1, 192(%r9)10157; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10158; SSE-NEXT: movaps %xmm1, 176(%r9)10159; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10160; SSE-NEXT: movaps %xmm1, 160(%r9)10161; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10162; SSE-NEXT: movaps %xmm1, 144(%r9)10163; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10164; SSE-NEXT: movaps %xmm1, 128(%r9)10165; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10166; SSE-NEXT: movaps %xmm1, 112(%r9)10167; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10168; SSE-NEXT: movaps %xmm1, 96(%r9)10169; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10170; SSE-NEXT: movaps %xmm1, 80(%r9)10171; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10172; SSE-NEXT: movaps %xmm1, 64(%r9)10173; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10174; SSE-NEXT: movaps %xmm1, 48(%r9)10175; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10176; SSE-NEXT: movaps %xmm1, 32(%r9)10177; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10178; SSE-NEXT: movaps %xmm1, 16(%r9)10179; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10180; SSE-NEXT: movaps %xmm1, (%r9)10181; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax10182; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10183; SSE-NEXT: movaps %xmm1, 240(%rax)10184; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10185; SSE-NEXT: movaps %xmm1, 224(%rax)10186; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10187; SSE-NEXT: movaps %xmm1, 208(%rax)10188; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10189; SSE-NEXT: movaps %xmm1, 192(%rax)10190; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10191; SSE-NEXT: movaps %xmm1, 176(%rax)10192; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10193; SSE-NEXT: movaps %xmm1, 160(%rax)10194; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10195; SSE-NEXT: movaps %xmm1, 144(%rax)10196; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10197; SSE-NEXT: movaps %xmm1, 128(%rax)10198; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10199; SSE-NEXT: movaps %xmm1, 112(%rax)10200; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10201; SSE-NEXT: movaps %xmm1, 96(%rax)10202; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10203; SSE-NEXT: movaps %xmm1, 80(%rax)10204; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10205; SSE-NEXT: movaps %xmm1, 64(%rax)10206; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10207; SSE-NEXT: movaps %xmm1, 48(%rax)10208; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10209; SSE-NEXT: movaps %xmm1, 32(%rax)10210; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10211; SSE-NEXT: movaps %xmm1, 16(%rax)10212; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10213; SSE-NEXT: movaps %xmm1, (%rax)10214; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax10215; SSE-NEXT: movaps %xmm3, 240(%rax)10216; SSE-NEXT: movaps %xmm4, 224(%rax)10217; SSE-NEXT: movaps %xmm5, 208(%rax)10218; SSE-NEXT: movaps %xmm2, 192(%rax)10219; SSE-NEXT: movaps %xmm6, 176(%rax)10220; SSE-NEXT: movaps %xmm7, 160(%rax)10221; SSE-NEXT: movaps %xmm13, 144(%rax)10222; SSE-NEXT: movaps %xmm14, 128(%rax)10223; SSE-NEXT: movaps %xmm15, 112(%rax)10224; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10225; SSE-NEXT: movaps %xmm1, 96(%rax)10226; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10227; SSE-NEXT: movaps %xmm1, 80(%rax)10228; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10229; SSE-NEXT: movaps %xmm1, 64(%rax)10230; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10231; SSE-NEXT: movaps %xmm1, 48(%rax)10232; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10233; SSE-NEXT: movaps %xmm1, 32(%rax)10234; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10235; SSE-NEXT: movaps %xmm1, 16(%rax)10236; SSE-NEXT: movaps %xmm0, (%rax)10237; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax10238; SSE-NEXT: movaps %xmm9, 240(%rax)10239; SSE-NEXT: movaps %xmm8, 224(%rax)10240; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10241; SSE-NEXT: movaps %xmm0, 208(%rax)10242; SSE-NEXT: movaps %xmm10, 192(%rax)10243; SSE-NEXT: movaps %xmm12, 176(%rax)10244; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10245; SSE-NEXT: movaps %xmm0, 160(%rax)10246; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10247; SSE-NEXT: movaps %xmm0, 144(%rax)10248; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10249; SSE-NEXT: movaps %xmm0, 128(%rax)10250; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10251; SSE-NEXT: movaps %xmm0, 112(%rax)10252; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10253; SSE-NEXT: movaps %xmm0, 96(%rax)10254; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload10255; SSE-NEXT: movaps %xmm0, 80(%rax)10256; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10257; SSE-NEXT: movaps %xmm0, 64(%rax)10258; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10259; SSE-NEXT: movaps %xmm0, 48(%rax)10260; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10261; SSE-NEXT: movaps %xmm0, 32(%rax)10262; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10263; SSE-NEXT: movaps %xmm0, 16(%rax)10264; SSE-NEXT: movaps %xmm11, (%rax)10265; SSE-NEXT: addq $2232, %rsp # imm = 0x8B810266; SSE-NEXT: retq10267;10268; AVX-LABEL: load_i32_stride8_vf64:10269; AVX: # %bb.0:10270; AVX-NEXT: subq $3720, %rsp # imm = 0xE8810271; AVX-NEXT: vmovaps 288(%rdi), %xmm1310272; AVX-NEXT: vmovaps 256(%rdi), %xmm1510273; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm15[0],xmm13[0],xmm15[1],xmm13[1]10274; AVX-NEXT: vmovaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10275; AVX-NEXT: vmovaps 352(%rdi), %xmm110276; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10277; AVX-NEXT: vmovaps 320(%rdi), %xmm210278; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10279; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10280; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10281; AVX-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]10282; AVX-NEXT: vmovaps 416(%rdi), %xmm110283; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10284; AVX-NEXT: vmovaps 384(%rdi), %xmm210285; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10286; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10287; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110288; AVX-NEXT: vmovaps 480(%rdi), %xmm210289; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10290; AVX-NEXT: vmovaps 448(%rdi), %xmm310291; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10292; AVX-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10293; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10294; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,0,1]10295; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210296; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]10297; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10298; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10299; AVX-NEXT: vmovaps 928(%rdi), %xmm110300; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10301; AVX-NEXT: vmovaps 896(%rdi), %xmm010302; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10303; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]10304; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010305; AVX-NEXT: vmovaps 992(%rdi), %xmm110306; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10307; AVX-NEXT: vmovaps 960(%rdi), %xmm210308; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10309; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10310; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10311; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1]10312; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110313; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]10314; AVX-NEXT: vmovaps 800(%rdi), %xmm110315; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10316; AVX-NEXT: vmovaps 768(%rdi), %xmm1410317; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]10318; AVX-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10319; AVX-NEXT: vmovaps 864(%rdi), %xmm210320; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10321; AVX-NEXT: vmovaps 832(%rdi), %xmm310322; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10323; AVX-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10324; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10325; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10326; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10327; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10328; AVX-NEXT: vmovaps 1440(%rdi), %xmm110329; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10330; AVX-NEXT: vmovaps 1408(%rdi), %xmm010331; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10332; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]10333; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010334; AVX-NEXT: vmovaps 1504(%rdi), %xmm110335; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10336; AVX-NEXT: vmovaps 1472(%rdi), %xmm210337; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10338; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10339; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10340; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1]10341; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110342; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]10343; AVX-NEXT: vmovaps 1312(%rdi), %xmm210344; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10345; AVX-NEXT: vmovaps 1280(%rdi), %xmm110346; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10347; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]10348; AVX-NEXT: vmovaps 1376(%rdi), %xmm210349; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10350; AVX-NEXT: vmovaps 1344(%rdi), %xmm310351; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10352; AVX-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10353; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10354; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]10355; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10356; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10357; AVX-NEXT: vmovaps 1952(%rdi), %xmm010358; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10359; AVX-NEXT: vmovaps 1920(%rdi), %xmm110360; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10361; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10362; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010363; AVX-NEXT: vmovaps 2016(%rdi), %xmm110364; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10365; AVX-NEXT: vmovaps 1984(%rdi), %xmm210366; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10367; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10368; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10369; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1]10370; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110371; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]10372; AVX-NEXT: vmovaps 1824(%rdi), %xmm110373; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10374; AVX-NEXT: vmovaps 1792(%rdi), %xmm510375; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]10376; AVX-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10377; AVX-NEXT: vmovaps 1888(%rdi), %xmm210378; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10379; AVX-NEXT: vmovaps 1856(%rdi), %xmm310380; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10381; AVX-NEXT: vunpcklps {{.*#+}} xmm10 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10382; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm10[0]10383; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10384; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10385; AVX-NEXT: vmovaps 160(%rdi), %xmm010386; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10387; AVX-NEXT: vmovaps 128(%rdi), %xmm110388; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10389; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10390; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010391; AVX-NEXT: vmovaps 224(%rdi), %xmm110392; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10393; AVX-NEXT: vmovaps 192(%rdi), %xmm210394; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10395; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10396; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10397; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1]10398; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110399; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]10400; AVX-NEXT: vmovaps 32(%rdi), %xmm110401; AVX-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill10402; AVX-NEXT: vmovaps (%rdi), %xmm410403; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]10404; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10405; AVX-NEXT: vmovaps 96(%rdi), %xmm210406; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10407; AVX-NEXT: vmovaps 64(%rdi), %xmm310408; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10409; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10410; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]10411; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10412; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10413; AVX-NEXT: vmovaps 672(%rdi), %xmm110414; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10415; AVX-NEXT: vmovaps 640(%rdi), %xmm010416; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10417; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]10418; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010419; AVX-NEXT: vmovaps 736(%rdi), %xmm110420; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10421; AVX-NEXT: vmovaps 704(%rdi), %xmm210422; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10423; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]10424; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10425; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm1[0,1,0,1]10426; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm610427; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]10428; AVX-NEXT: vmovaps 544(%rdi), %xmm010429; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10430; AVX-NEXT: vmovaps 512(%rdi), %xmm210431; AVX-NEXT: vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]10432; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10433; AVX-NEXT: vmovaps 608(%rdi), %xmm010434; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10435; AVX-NEXT: vmovaps 576(%rdi), %xmm110436; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10437; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10438; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]10439; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm6[4,5,6,7]10440; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10441; AVX-NEXT: vmovaps 1184(%rdi), %xmm010442; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10443; AVX-NEXT: vmovaps 1152(%rdi), %xmm610444; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10445; AVX-NEXT: vunpcklps {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]10446; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm710447; AVX-NEXT: vmovaps 1248(%rdi), %xmm010448; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10449; AVX-NEXT: vmovaps 1216(%rdi), %xmm610450; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10451; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]10452; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10453; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm0[0,1,0,1]10454; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm810455; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm7[0,1,2,3,4,5],ymm8[6,7]10456; AVX-NEXT: vmovaps 1056(%rdi), %xmm010457; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10458; AVX-NEXT: vmovaps 1024(%rdi), %xmm610459; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10460; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]10461; AVX-NEXT: vmovaps 1120(%rdi), %xmm010462; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10463; AVX-NEXT: vmovaps 1088(%rdi), %xmm610464; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10465; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]10466; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10467; AVX-NEXT: vmovlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]10468; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]10469; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10470; AVX-NEXT: vmovaps 1696(%rdi), %xmm010471; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10472; AVX-NEXT: vmovaps 1664(%rdi), %xmm610473; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10474; AVX-NEXT: vunpcklps {{.*#+}} xmm8 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]10475; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm910476; AVX-NEXT: vmovaps 1760(%rdi), %xmm010477; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10478; AVX-NEXT: vmovaps 1728(%rdi), %xmm610479; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10480; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]10481; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10482; AVX-NEXT: vshufps {{.*#+}} xmm11 = xmm0[0,1,0,1]10483; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm1110484; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3,4,5],ymm11[6,7]10485; AVX-NEXT: vmovaps 1568(%rdi), %xmm1210486; AVX-NEXT: vmovaps 1536(%rdi), %xmm810487; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm8[0],xmm12[0],xmm8[1],xmm12[1]10488; AVX-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10489; AVX-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10490; AVX-NEXT: vmovaps 1632(%rdi), %xmm610491; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10492; AVX-NEXT: vmovaps 1600(%rdi), %xmm710493; AVX-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10494; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]10495; AVX-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm9[0]10496; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]10497; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10498; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm15[1,1,1,1]10499; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm13[1],xmm0[2,3]10500; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10501; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]10502; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 16-byte Folded Reload10503; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload10504; AVX-NEXT: vshufps {{.*#+}} xmm15 = xmm7[1,1,1,1]10505; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload10506; AVX-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0],xmm6[1],xmm15[2,3]10507; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1510508; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]10509; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]10510; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10511; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm14[1,1,1,1]10512; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10513; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]10514; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10515; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]10516; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 16-byte Folded Reload10517; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload10518; AVX-NEXT: # xmm15 = mem[1,1,1,1]10519; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload10520; AVX-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0],xmm14[1],xmm15[2,3]10521; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1510522; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]10523; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]10524; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10525; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10526; AVX-NEXT: # xmm0 = mem[1,1,1,1]10527; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10528; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]10529; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10530; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]10531; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 16-byte Folded Reload10532; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload10533; AVX-NEXT: # xmm15 = mem[1,1,1,1]10534; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10535; AVX-NEXT: # xmm15 = xmm15[0],mem[1],xmm15[2,3]10536; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1510537; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]10538; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm11[4,5,6,7]10539; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10540; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]10541; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10542; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]10543; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm10[2,3]10544; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 16-byte Folded Reload10545; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload10546; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm11[1,1,1,1]10547; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10548; AVX-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm15[1],xmm10[2,3]10549; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm1010550; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0,1,2,3,4,5],ymm5[6,7]10551; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]10552; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10553; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm4[1,1,1,1]10554; AVX-NEXT: vblendps $2, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload10555; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]10556; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm3[2,3]10557; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 16-byte Folded Reload10558; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload10559; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm10[1,1,1,1]10560; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload10561; AVX-NEXT: # xmm4 = xmm4[0],mem[1],xmm4[2,3]10562; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm410563; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]10564; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]10565; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10566; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm2[1,1,1,1]10567; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10568; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]10569; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]10570; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload10571; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload10572; AVX-NEXT: # xmm2 = mem[1,1,1,1]10573; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload10574; AVX-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]10575; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210576; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]10577; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10578; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10579; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10580; AVX-NEXT: # xmm0 = mem[1,1,1,1]10581; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10582; AVX-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]10583; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10584; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]10585; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload10586; AVX-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload10587; AVX-NEXT: # xmm2 = mem[1,1,1,1]10588; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload10589; AVX-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]10590; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210591; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]10592; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10593; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10594; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm8[1,1,1,1]10595; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm12[1],xmm0[2,3]10596; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm9[2,3]10597; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload10598; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload10599; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm9[1,1,1,1]10600; AVX-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload10601; AVX-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]10602; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210603; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]10604; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10605; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10606; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10607; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm0[2],xmm13[2],xmm0[3],xmm13[3]10608; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10609; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]10610; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10611; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10612; AVX-NEXT: # xmm0 = mem[2,2,2,2]10613; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10614; AVX-NEXT: # xmm0 = mem[0,1,2],xmm0[3]10615; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010616; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110617; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10618; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload10619; AVX-NEXT: # xmm1 = mem[2,2,2,2]10620; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload10621; AVX-NEXT: # xmm1 = mem[0,1,2],xmm1[3]10622; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10623; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10624; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10625; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10626; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload10627; AVX-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]10628; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10629; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10630; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm0[2],xmm14[2],xmm0[3],xmm14[3]10631; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10632; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10633; AVX-NEXT: # xmm0 = mem[2,2,2,2]10634; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10635; AVX-NEXT: # xmm0 = mem[0,1,2],xmm0[3]10636; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010637; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110638; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10639; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload10640; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm7[2,2,2,2]10641; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload10642; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1,2],xmm1[3]10643; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10644; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10645; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10646; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10647; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload10648; AVX-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]10649; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10650; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10651; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload10652; AVX-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]10653; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10654; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10655; AVX-NEXT: # xmm0 = mem[2,2,2,2]10656; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10657; AVX-NEXT: # xmm0 = mem[0,1,2],xmm0[3]10658; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010659; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110660; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10661; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload10662; AVX-NEXT: # xmm1 = mem[2,2,2,2]10663; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload10664; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm4[0,1,2],xmm1[3]10665; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10666; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10667; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10668; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10669; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload10670; AVX-NEXT: # xmm3 = xmm0[2],mem[2],xmm0[3],mem[3]10671; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10672; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm11[2],xmm15[2],xmm11[3],xmm15[3]10673; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10674; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10675; AVX-NEXT: # xmm0 = mem[2,2,2,2]10676; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10677; AVX-NEXT: # xmm0 = mem[0,1,2],xmm0[3]10678; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010679; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110680; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10681; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload10682; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm5[2,2,2,2]10683; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload10684; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]10685; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]10686; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10687; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10688; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10689; AVX-NEXT: vunpckhps (%rsp), %xmm0, %xmm8 # 16-byte Folded Reload10690; AVX-NEXT: # xmm8 = xmm0[2],mem[2],xmm0[3],mem[3]10691; AVX-NEXT: vmovaps %xmm8, (%rsp) # 16-byte Spill10692; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload10693; AVX-NEXT: # xmm1 = xmm10[2],mem[2],xmm10[3],mem[3]10694; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10695; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload10696; AVX-NEXT: # xmm0 = mem[2,2,2,2]10697; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10698; AVX-NEXT: # xmm0 = mem[0,1,2],xmm0[3]10699; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm010700; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110701; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10702; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload10703; AVX-NEXT: # xmm1 = mem[2,2,2,2]10704; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload10705; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]10706; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]10707; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10708; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10709; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10710; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm8 # 16-byte Folded Reload10711; AVX-NEXT: # xmm8 = xmm0[2],mem[2],xmm0[3],mem[3]10712; AVX-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10713; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10714; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10715; AVX-NEXT: # xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]10716; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10717; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload10718; AVX-NEXT: # xmm15 = mem[2,2,2,2]10719; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10720; AVX-NEXT: # xmm15 = mem[0,1,2],xmm15[3]10721; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1510722; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm1410723; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]10724; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload10725; AVX-NEXT: # xmm15 = mem[2,2,2,2]10726; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10727; AVX-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]10728; AVX-NEXT: vblendps {{.*#+}} xmm15 = xmm8[0,1],xmm15[2,3]10729; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0,1,2,3],ymm14[4,5,6,7]10730; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10731; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload10732; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm0 # 16-byte Folded Reload10733; AVX-NEXT: # xmm0 = xmm8[2],mem[2],xmm8[3],mem[3]10734; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10735; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload10736; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload10737; AVX-NEXT: # xmm8 = xmm8[2],mem[2],xmm8[3],mem[3]10738; AVX-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10739; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload10740; AVX-NEXT: # xmm13 = mem[2,2,2,2]10741; AVX-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload10742; AVX-NEXT: # xmm13 = mem[0,1,2],xmm13[3]10743; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm1310744; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm1210745; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]10746; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10747; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm15[2,2,2,2]10748; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload10749; AVX-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]10750; AVX-NEXT: vblendps {{.*#+}} xmm13 = xmm0[0,1],xmm13[2,3]10751; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm12[4,5,6,7]10752; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10753; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload10754; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm0 # 16-byte Folded Reload10755; AVX-NEXT: # xmm0 = xmm8[2],mem[2],xmm8[3],mem[3]10756; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10757; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm8 # 16-byte Folded Reload10758; AVX-NEXT: # xmm8 = xmm9[2],mem[2],xmm9[3],mem[3]10759; AVX-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10760; AVX-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload10761; AVX-NEXT: # xmm11 = mem[2,2,2,2]10762; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload10763; AVX-NEXT: vblendps {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]10764; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm1110765; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm910766; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm11[6,7]10767; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload10768; AVX-NEXT: vshufps {{.*#+}} xmm11 = xmm12[2,2,2,2]10769; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload10770; AVX-NEXT: vblendps {{.*#+}} xmm11 = xmm10[0,1,2],xmm11[3]10771; AVX-NEXT: vblendps {{.*#+}} xmm11 = xmm0[0,1],xmm11[2,3]10772; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm11[0,1,2,3],ymm9[4,5,6,7]10773; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10774; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload10775; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm9 # 16-byte Folded Reload10776; AVX-NEXT: # xmm9 = xmm8[2],mem[2],xmm8[3],mem[3]10777; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload10778; AVX-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm8[1],xmm9[1]10779; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload10780; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm11 # 16-byte Folded Reload10781; AVX-NEXT: # xmm11 = xmm8[2],mem[2],xmm8[3],mem[3]10782; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm1110783; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload10784; AVX-NEXT: # xmm8 = mem[2,3,2,3]10785; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm810786; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm11[6,7]10787; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3],ymm8[4,5,6,7]10788; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10789; AVX-NEXT: vunpckhps {{.*#+}} xmm8 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]10790; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload10791; AVX-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm6[1],xmm8[1]10792; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload10793; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm8 # 16-byte Folded Reload10794; AVX-NEXT: # xmm8 = xmm6[2],mem[2],xmm6[3],mem[3]10795; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm810796; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload10797; AVX-NEXT: # xmm9 = mem[2,3,2,3]10798; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm910799; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3,4,5],ymm8[6,7]10800; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]10801; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10802; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm7 # 16-byte Folded Reload10803; AVX-NEXT: # xmm7 = xmm4[2],mem[2],xmm4[3],mem[3]10804; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload10805; AVX-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm4[1],xmm7[1]10806; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload10807; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm8 # 16-byte Folded Reload10808; AVX-NEXT: # xmm8 = xmm4[2],mem[2],xmm4[3],mem[3]10809; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm810810; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload10811; AVX-NEXT: # xmm6 = mem[2,3,2,3]10812; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm610813; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm8[6,7]10814; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10815; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10816; AVX-NEXT: vunpckhps {{.*#+}} xmm6 = xmm2[2],xmm5[2],xmm2[3],xmm5[3]10817; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload10818; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm2[1],xmm6[1]10819; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload10820; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm6 # 16-byte Folded Reload10821; AVX-NEXT: # xmm6 = xmm2[2],mem[2],xmm2[3],mem[3]10822; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm610823; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload10824; AVX-NEXT: # xmm5 = mem[2,3,2,3]10825; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm510826; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7]10827; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]10828; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10829; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm4 # 16-byte Folded Reload10830; AVX-NEXT: # xmm4 = xmm3[2],mem[2],xmm3[3],mem[3]10831; AVX-NEXT: vmovaps (%rsp), %xmm2 # 16-byte Reload10832; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm4[1]10833; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload10834; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm4 # 16-byte Folded Reload10835; AVX-NEXT: # xmm4 = xmm3[2],mem[2],xmm3[3],mem[3]10836; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm410837; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload10838; AVX-NEXT: # xmm3 = mem[2,3,2,3]10839; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm310840; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]10841; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]10842; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10843; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm2 # 16-byte Folded Reload10844; AVX-NEXT: # xmm2 = xmm1[2],mem[2],xmm1[3],mem[3]10845; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload10846; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]10847; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10848; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm2 # 16-byte Folded Reload10849; AVX-NEXT: # xmm2 = xmm1[2],mem[2],xmm1[3],mem[3]10850; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210851; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload10852; AVX-NEXT: # xmm1 = mem[2,3,2,3]10853; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110854; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]10855; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10856; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10857; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm14[2],xmm15[2],xmm14[3],xmm15[3]10858; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10859; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]10860; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10861; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload10862; AVX-NEXT: # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]10863; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110864; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload10865; AVX-NEXT: # xmm2 = mem[2,3,2,3]10866; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210867; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]10868; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10869; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10870; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm10[2],xmm12[2],xmm10[3],xmm12[3]10871; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload10872; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]10873; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm1 # 16-byte Folded Reload10874; AVX-NEXT: # xmm1 = xmm13[2],mem[2],xmm13[3],mem[3]10875; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110876; AVX-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload10877; AVX-NEXT: # xmm2 = mem[2,3,2,3]10878; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm210879; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]10880; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10881; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10882; AVX-NEXT: vmovaps 416(%rdi), %ymm210883; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10884; AVX-NEXT: vmovaps 384(%rdi), %ymm310885; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10886; AVX-NEXT: vmovaps 448(%rdi), %ymm410887; AVX-NEXT: vmovaps 480(%rdi), %ymm010888; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10889; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]10890; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]10891; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]10892; AVX-NEXT: vmovaps 320(%rdi), %ymm610893; AVX-NEXT: vmovaps 352(%rdi), %ymm1310894; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm6[0],ymm13[2],ymm6[2]10895; AVX-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10896; AVX-NEXT: vextractf128 $1, %ymm1, %xmm110897; AVX-NEXT: vmovaps 288(%rdi), %ymm810898; AVX-NEXT: vmovaps 256(%rdi), %ymm710899; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[4],ymm8[4],ymm7[5],ymm8[5]10900; AVX-NEXT: vextractf128 $1, %ymm2, %xmm210901; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]10902; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10903; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10904; AVX-NEXT: vmovaps 672(%rdi), %ymm210905; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10906; AVX-NEXT: vmovaps 640(%rdi), %ymm910907; AVX-NEXT: vmovaps 704(%rdi), %ymm1210908; AVX-NEXT: vmovaps 736(%rdi), %ymm1010909; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm12[0],ymm10[2],ymm12[2]10910; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm9[0],ymm2[0],ymm9[1],ymm2[1],ymm9[4],ymm2[4],ymm9[5],ymm2[5]10911; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]10912; AVX-NEXT: vmovaps 576(%rdi), %ymm1410913; AVX-NEXT: vmovaps 608(%rdi), %ymm1110914; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm11[0],ymm14[0],ymm11[2],ymm14[2]10915; AVX-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10916; AVX-NEXT: vextractf128 $1, %ymm1, %xmm110917; AVX-NEXT: vmovaps 544(%rdi), %ymm210918; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10919; AVX-NEXT: vmovaps 512(%rdi), %ymm310920; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10921; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]10922; AVX-NEXT: vextractf128 $1, %ymm2, %xmm210923; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]10924; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10925; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10926; AVX-NEXT: vmovaps 928(%rdi), %ymm210927; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10928; AVX-NEXT: vmovaps 896(%rdi), %ymm310929; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10930; AVX-NEXT: vmovaps 960(%rdi), %ymm110931; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10932; AVX-NEXT: vmovaps 992(%rdi), %ymm010933; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10934; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10935; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]10936; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]10937; AVX-NEXT: vmovaps 832(%rdi), %ymm210938; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10939; AVX-NEXT: vmovaps 864(%rdi), %ymm110940; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10941; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]10942; AVX-NEXT: vextractf128 $1, %ymm1, %xmm110943; AVX-NEXT: vmovaps 800(%rdi), %ymm310944; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10945; AVX-NEXT: vmovaps 768(%rdi), %ymm210946; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10947; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10948; AVX-NEXT: vextractf128 $1, %ymm2, %xmm210949; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]10950; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10951; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10952; AVX-NEXT: vmovaps 1184(%rdi), %ymm210953; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10954; AVX-NEXT: vmovaps 1152(%rdi), %ymm110955; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10956; AVX-NEXT: vmovaps 1216(%rdi), %ymm010957; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10958; AVX-NEXT: vmovaps 1248(%rdi), %ymm310959; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10960; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]10961; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5]10962; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]10963; AVX-NEXT: vmovaps 1088(%rdi), %ymm210964; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10965; AVX-NEXT: vmovaps 1120(%rdi), %ymm110966; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10967; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]10968; AVX-NEXT: vextractf128 $1, %ymm1, %xmm110969; AVX-NEXT: vmovaps 1056(%rdi), %ymm310970; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10971; AVX-NEXT: vmovaps 1024(%rdi), %ymm210972; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10973; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10974; AVX-NEXT: vextractf128 $1, %ymm2, %xmm210975; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]10976; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10977; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10978; AVX-NEXT: vmovaps 1440(%rdi), %ymm210979; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10980; AVX-NEXT: vmovaps 1408(%rdi), %ymm310981; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10982; AVX-NEXT: vmovaps 1472(%rdi), %ymm110983; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10984; AVX-NEXT: vmovaps 1504(%rdi), %ymm010985; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10986; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]10987; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]10988; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]10989; AVX-NEXT: vmovaps 1344(%rdi), %ymm210990; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10991; AVX-NEXT: vmovaps 1376(%rdi), %ymm110992; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10993; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]10994; AVX-NEXT: vextractf128 $1, %ymm1, %xmm110995; AVX-NEXT: vmovaps 1312(%rdi), %ymm310996; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10997; AVX-NEXT: vmovaps 1280(%rdi), %ymm210998; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10999; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]11000; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211001; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]11002; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11003; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11004; AVX-NEXT: vmovaps 1696(%rdi), %ymm211005; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11006; AVX-NEXT: vmovaps 1664(%rdi), %ymm311007; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11008; AVX-NEXT: vmovaps 1728(%rdi), %ymm111009; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11010; AVX-NEXT: vmovaps 1760(%rdi), %ymm011011; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11012; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11013; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]11014; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]11015; AVX-NEXT: vmovaps 1600(%rdi), %ymm211016; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11017; AVX-NEXT: vmovaps 1632(%rdi), %ymm111018; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11019; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]11020; AVX-NEXT: vextractf128 $1, %ymm1, %xmm111021; AVX-NEXT: vmovaps 1568(%rdi), %ymm311022; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11023; AVX-NEXT: vmovaps 1536(%rdi), %ymm211024; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11025; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]11026; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211027; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]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 1952(%rdi), %ymm211031; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11032; AVX-NEXT: vmovaps 1920(%rdi), %ymm311033; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11034; AVX-NEXT: vmovaps 1984(%rdi), %ymm111035; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11036; AVX-NEXT: vmovaps 2016(%rdi), %ymm011037; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11038; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11039; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]11040; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]11041; AVX-NEXT: vmovaps 1856(%rdi), %ymm211042; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11043; AVX-NEXT: vmovaps 1888(%rdi), %ymm111044; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11045; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]11046; AVX-NEXT: vextractf128 $1, %ymm1, %xmm111047; AVX-NEXT: vmovaps 1824(%rdi), %ymm311048; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11049; AVX-NEXT: vmovaps 1792(%rdi), %ymm211050; AVX-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill11051; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]11052; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211053; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,0]11054; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11055; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11056; AVX-NEXT: vmovaps 160(%rdi), %ymm211057; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11058; AVX-NEXT: vmovaps 128(%rdi), %ymm311059; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11060; AVX-NEXT: vmovaps 192(%rdi), %ymm111061; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11062; AVX-NEXT: vmovaps 224(%rdi), %ymm011063; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11064; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]11065; AVX-NEXT: vunpcklps {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]11066; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]11067; AVX-NEXT: vmovaps 64(%rdi), %ymm011068; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11069; AVX-NEXT: vmovaps 96(%rdi), %ymm111070; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11071; AVX-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]11072; AVX-NEXT: vextractf128 $1, %ymm1, %xmm011073; AVX-NEXT: vmovaps (%rdi), %ymm111074; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11075; AVX-NEXT: vmovaps 32(%rdi), %ymm311076; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11077; AVX-NEXT: vunpcklps {{.*#+}} ymm15 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]11078; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511079; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm15[0,1],xmm0[2,0]11080; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11081; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11082; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload11083; AVX-NEXT: vmovaps %ymm4, %ymm511084; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11085; AVX-NEXT: vunpcklps {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[4],ymm3[4],ymm4[5],ymm3[5]11086; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload11087; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11088; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm1[1,0],ymm4[1,0],ymm1[5,4],ymm4[5,4]11089; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11090; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11091; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm6[0],ymm13[0],ymm6[1],ymm13[1],ymm6[4],ymm13[4],ymm6[5],ymm13[5]11092; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211093; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11094; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11095; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm8[1,0],ymm7[1,0],ymm8[5,4],ymm7[5,4]11096; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511097; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11098; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11099; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11100; AVX-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11101; AVX-NEXT: vmovaps %ymm12, %ymm1311102; AVX-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11103; AVX-NEXT: vunpcklps {{.*#+}} ymm0 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[4],ymm10[4],ymm12[5],ymm10[5]11104; AVX-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11105; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload11106; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm12[1,0],ymm9[1,0],ymm12[5,4],ymm9[5,4]11107; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11108; AVX-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11109; AVX-NEXT: vunpcklps {{.*#+}} ymm2 = ymm14[0],ymm11[0],ymm14[1],ymm11[1],ymm14[4],ymm11[4],ymm14[5],ymm11[5]11110; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211111; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload11112; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11113; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm15[1,0],ymm14[1,0],ymm15[5,4],ymm14[5,4]11114; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511115; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11116; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11117; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11118; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11119; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11120; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]11121; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11122; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11123; AVX-NEXT: # ymm2 = ymm2[1,0],mem[1,0],ymm2[5,4],mem[5,4]11124; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11125; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11126; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11127; AVX-NEXT: # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[4],mem[4],ymm2[5],mem[5]11128; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211129; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11130; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11131; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]11132; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511133; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11134; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11135; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11136; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11137; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11138; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]11139; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11140; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11141; AVX-NEXT: # ymm2 = ymm2[1,0],mem[1,0],ymm2[5,4],mem[5,4]11142; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11143; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11144; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11145; AVX-NEXT: # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[4],mem[4],ymm2[5],mem[5]11146; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211147; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11148; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11149; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]11150; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511151; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11152; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[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: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11156; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]11157; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11158; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11159; AVX-NEXT: # ymm2 = ymm2[1,0],mem[1,0],ymm2[5,4],mem[5,4]11160; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11161; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11162; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11163; AVX-NEXT: # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[4],mem[4],ymm2[5],mem[5]11164; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211165; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11166; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11167; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]11168; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511169; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11170; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11171; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11172; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11173; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11174; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]11175; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11176; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11177; AVX-NEXT: # ymm2 = ymm2[1,0],mem[1,0],ymm2[5,4],mem[5,4]11178; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11179; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11180; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11181; AVX-NEXT: # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[4],mem[4],ymm2[5],mem[5]11182; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211183; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11184; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11185; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]11186; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511187; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11188; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11189; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11190; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11191; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11192; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]11193; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11194; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11195; AVX-NEXT: # ymm2 = ymm2[1,0],mem[1,0],ymm2[5,4],mem[5,4]11196; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11197; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11198; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11199; AVX-NEXT: # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[4],mem[4],ymm2[5],mem[5]11200; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211201; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11202; AVX-NEXT: vshufps $17, (%rsp), %ymm15, %ymm15 # 32-byte Folded Reload11203; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]11204; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511205; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11206; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11207; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11208; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11209; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11210; AVX-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5]11211; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11212; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11213; AVX-NEXT: # ymm2 = ymm2[1,0],mem[1,0],ymm2[5,4],mem[5,4]11214; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11215; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11216; AVX-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11217; AVX-NEXT: # ymm2 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[4],mem[4],ymm2[5],mem[5]11218; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211219; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11220; AVX-NEXT: vshufps $17, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11221; AVX-NEXT: # ymm15 = ymm15[1,0],mem[1,0],ymm15[5,4],mem[5,4]11222; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511223; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11224; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11225; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11226; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]11227; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm4[2],ymm1[2],ymm4[3],ymm1[3],ymm4[6],ymm1[6],ymm4[7],ymm1[7]11228; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11229; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload11230; AVX-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm4[1],ymm6[1],ymm4[3],ymm6[3]11231; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211232; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[6],ymm8[6],ymm7[7],ymm8[7]11233; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511234; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11235; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11236; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11237; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm13[1],ymm10[3],ymm13[3]11238; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm9[2],ymm12[2],ymm9[3],ymm12[3],ymm9[6],ymm12[6],ymm9[7],ymm12[7]11239; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11240; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm2 # 32-byte Folded Reload11241; AVX-NEXT: # ymm2 = ymm11[1],mem[1],ymm11[3],mem[3]11242; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211243; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11244; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm14[2],ymm1[2],ymm14[3],ymm1[3],ymm14[6],ymm1[6],ymm14[7],ymm1[7]11245; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511246; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11247; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11248; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11249; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload11250; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload11251; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]11252; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload11253; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload11254; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[6],ymm3[6],ymm9[7],ymm3[7]11255; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11256; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload11257; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload11258; AVX-NEXT: vunpckhpd {{.*#+}} ymm2 = ymm8[1],ymm6[1],ymm8[3],ymm6[3]11259; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211260; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload11261; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload11262; AVX-NEXT: vunpckhps {{.*#+}} ymm15 = ymm10[2],ymm7[2],ymm10[3],ymm7[3],ymm10[6],ymm7[6],ymm10[7],ymm7[7]11263; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511264; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11265; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11266; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11267; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload11268; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload11269; AVX-NEXT: # ymm0 = ymm13[1],mem[1],ymm13[3],mem[3]11270; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload11271; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload11272; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm14[2],ymm12[2],ymm14[3],ymm12[3],ymm14[6],ymm12[6],ymm14[7],ymm12[7]11273; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11274; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11275; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11276; AVX-NEXT: # ymm2 = ymm2[1],mem[1],ymm2[3],mem[3]11277; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211278; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11279; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11280; AVX-NEXT: # ymm15 = ymm15[2],mem[2],ymm15[3],mem[3],ymm15[6],mem[6],ymm15[7],mem[7]11281; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511282; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11283; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11284; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11285; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11286; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11287; AVX-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11288; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11289; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11290; AVX-NEXT: # ymm2 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]11291; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11292; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11293; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11294; AVX-NEXT: # ymm2 = ymm2[1],mem[1],ymm2[3],mem[3]11295; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211296; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11297; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11298; AVX-NEXT: # ymm15 = ymm15[2],mem[2],ymm15[3],mem[3],ymm15[6],mem[6],ymm15[7],mem[7]11299; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511300; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11301; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11302; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11303; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11304; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11305; AVX-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11306; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11307; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11308; AVX-NEXT: # ymm2 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]11309; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11310; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11311; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11312; AVX-NEXT: # ymm2 = ymm2[1],mem[1],ymm2[3],mem[3]11313; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211314; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11315; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11316; AVX-NEXT: # ymm15 = ymm15[2],mem[2],ymm15[3],mem[3],ymm15[6],mem[6],ymm15[7],mem[7]11317; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511318; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11319; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11320; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11321; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11322; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11323; AVX-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11324; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11325; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11326; AVX-NEXT: # ymm2 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]11327; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11328; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11329; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11330; AVX-NEXT: # ymm2 = ymm2[1],mem[1],ymm2[3],mem[3]11331; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211332; AVX-NEXT: vmovups (%rsp), %ymm15 # 32-byte Reload11333; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11334; AVX-NEXT: # ymm15 = ymm15[2],mem[2],ymm15[3],mem[3],ymm15[6],mem[6],ymm15[7],mem[7]11335; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511336; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11337; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11338; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11339; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11340; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11341; AVX-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]11342; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11343; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11344; AVX-NEXT: # ymm2 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]11345; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,0],ymm2[4,5],ymm0[6,4]11346; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11347; AVX-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11348; AVX-NEXT: # ymm2 = ymm2[1],mem[1],ymm2[3],mem[3]11349; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211350; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11351; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11352; AVX-NEXT: # ymm15 = ymm15[2],mem[2],ymm15[3],mem[3],ymm15[6],mem[6],ymm15[7],mem[7]11353; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511354; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[0,1],xmm2[2,0]11355; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11356; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11357; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11358; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11359; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]11360; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11361; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11362; AVX-NEXT: # ymm2 = ymm2[3,0],mem[3,0],ymm2[7,4],mem[7,4]11363; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11364; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11365; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]11366; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload11367; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm15 # 32-byte Folded Reload11368; AVX-NEXT: # ymm15 = ymm4[3,0],mem[3,0],ymm4[7,4],mem[7,4]11369; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211370; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511371; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11372; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm0[4,5,6,7]11373; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11374; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11375; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]11376; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11377; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11378; AVX-NEXT: # ymm2 = ymm2[3,0],mem[3,0],ymm2[7,4],mem[7,4]11379; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11380; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11381; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11382; AVX-NEXT: # ymm2 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7]11383; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload11384; AVX-NEXT: # ymm15 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11385; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211386; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511387; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11388; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11389; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11390; AVX-NEXT: vunpckhps {{.*#+}} ymm0 = ymm5[2],ymm11[2],ymm5[3],ymm11[3],ymm5[6],ymm11[6],ymm5[7],ymm11[7]11391; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,0],ymm9[3,0],ymm3[7,4],ymm9[7,4]11392; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11393; AVX-NEXT: vunpckhps {{.*#+}} ymm2 = ymm6[2],ymm8[2],ymm6[3],ymm8[3],ymm6[6],ymm8[6],ymm6[7],ymm8[7]11394; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm7[3,0],ymm10[3,0],ymm7[7,4],ymm10[7,4]11395; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211396; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511397; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11398; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm0[4,5,6,7]11399; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11400; AVX-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm13[2],ymm0[3],ymm13[3],ymm0[6],ymm13[6],ymm0[7],ymm13[7]11401; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm12[3,0],ymm14[3,0],ymm12[7,4],ymm14[7,4]11402; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11403; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11404; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11405; AVX-NEXT: # ymm2 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]11406; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11407; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload11408; AVX-NEXT: # ymm15 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11409; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211410; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511411; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11412; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm2[0,1,2,3],ymm0[4,5,6,7]11413; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11414; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11415; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]11416; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11417; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11418; AVX-NEXT: # ymm2 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11419; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11420; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11421; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11422; AVX-NEXT: # ymm2 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]11423; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11424; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload11425; AVX-NEXT: # ymm15 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11426; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211427; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511428; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11429; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm2[0,1,2,3],ymm0[4,5,6,7]11430; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11431; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11432; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]11433; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11434; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11435; AVX-NEXT: # ymm2 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11436; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11437; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11438; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11439; AVX-NEXT: # ymm2 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]11440; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11441; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload11442; AVX-NEXT: # ymm15 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11443; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211444; AVX-NEXT: vextractf128 $1, %ymm15, %xmm1511445; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,0],xmm2[2,3]11446; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm2[0,1,2,3],ymm0[4,5,6,7]11447; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11448; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11449; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]11450; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11451; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11452; AVX-NEXT: # ymm2 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11453; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11454; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11455; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11456; AVX-NEXT: # ymm2 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]11457; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11458; AVX-NEXT: vshufps $51, (%rsp), %ymm1, %ymm7 # 32-byte Folded Reload11459; AVX-NEXT: # ymm7 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11460; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211461; AVX-NEXT: vextractf128 $1, %ymm7, %xmm711462; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm7[2,0],xmm2[2,3]11463; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3],ymm0[4,5,6,7]11464; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11465; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11466; AVX-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]11467; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11468; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11469; AVX-NEXT: # ymm2 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11470; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[2,0],ymm0[2,3],ymm2[6,4],ymm0[6,7]11471; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11472; AVX-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload11473; AVX-NEXT: # ymm2 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]11474; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11475; AVX-NEXT: vshufps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload11476; AVX-NEXT: # ymm1 = ymm1[3,0],mem[3,0],ymm1[7,4],mem[7,4]11477; AVX-NEXT: vextractf128 $1, %ymm2, %xmm211478; AVX-NEXT: vextractf128 $1, %ymm1, %xmm111479; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,0],xmm2[2,3]11480; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]11481; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11482; AVX-NEXT: vmovaps %ymm1, 192(%rsi)11483; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11484; AVX-NEXT: vmovaps %ymm1, 128(%rsi)11485; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11486; AVX-NEXT: vmovaps %ymm1, 64(%rsi)11487; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11488; AVX-NEXT: vmovaps %ymm1, (%rsi)11489; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11490; AVX-NEXT: vmovaps %ymm1, 224(%rsi)11491; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11492; AVX-NEXT: vmovaps %ymm1, 160(%rsi)11493; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11494; AVX-NEXT: vmovaps %ymm1, 96(%rsi)11495; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11496; AVX-NEXT: vmovaps %ymm1, 32(%rsi)11497; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11498; AVX-NEXT: vmovaps %ymm1, 192(%rdx)11499; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11500; AVX-NEXT: vmovaps %ymm1, 128(%rdx)11501; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11502; AVX-NEXT: vmovaps %ymm1, 64(%rdx)11503; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11504; AVX-NEXT: vmovaps %ymm1, (%rdx)11505; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11506; AVX-NEXT: vmovaps %ymm1, 224(%rdx)11507; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11508; AVX-NEXT: vmovaps %ymm1, 160(%rdx)11509; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11510; AVX-NEXT: vmovaps %ymm1, 96(%rdx)11511; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11512; AVX-NEXT: vmovaps %ymm1, 32(%rdx)11513; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11514; AVX-NEXT: vmovaps %ymm1, 192(%rcx)11515; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11516; AVX-NEXT: vmovaps %ymm1, 128(%rcx)11517; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11518; AVX-NEXT: vmovaps %ymm1, 64(%rcx)11519; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11520; AVX-NEXT: vmovaps %ymm1, (%rcx)11521; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11522; AVX-NEXT: vmovaps %ymm1, 224(%rcx)11523; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11524; AVX-NEXT: vmovaps %ymm1, 160(%rcx)11525; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11526; AVX-NEXT: vmovaps %ymm1, 96(%rcx)11527; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11528; AVX-NEXT: vmovaps %ymm1, 32(%rcx)11529; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11530; AVX-NEXT: vmovaps %ymm1, 192(%r8)11531; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11532; AVX-NEXT: vmovaps %ymm1, 128(%r8)11533; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11534; AVX-NEXT: vmovaps %ymm1, 64(%r8)11535; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11536; AVX-NEXT: vmovaps %ymm1, (%r8)11537; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11538; AVX-NEXT: vmovaps %ymm1, 224(%r8)11539; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11540; AVX-NEXT: vmovaps %ymm1, 160(%r8)11541; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11542; AVX-NEXT: vmovaps %ymm1, 96(%r8)11543; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11544; AVX-NEXT: vmovaps %ymm1, 32(%r8)11545; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11546; AVX-NEXT: vmovaps %ymm1, 224(%r9)11547; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11548; AVX-NEXT: vmovaps %ymm1, 192(%r9)11549; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11550; AVX-NEXT: vmovaps %ymm1, 160(%r9)11551; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11552; AVX-NEXT: vmovaps %ymm1, 128(%r9)11553; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11554; AVX-NEXT: vmovaps %ymm1, 96(%r9)11555; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11556; AVX-NEXT: vmovaps %ymm1, 64(%r9)11557; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11558; AVX-NEXT: vmovaps %ymm1, 32(%r9)11559; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11560; AVX-NEXT: vmovaps %ymm1, (%r9)11561; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax11562; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11563; AVX-NEXT: vmovaps %ymm1, 224(%rax)11564; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11565; AVX-NEXT: vmovaps %ymm1, 192(%rax)11566; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11567; AVX-NEXT: vmovaps %ymm1, 160(%rax)11568; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11569; AVX-NEXT: vmovaps %ymm1, 128(%rax)11570; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11571; AVX-NEXT: vmovaps %ymm1, 96(%rax)11572; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11573; AVX-NEXT: vmovaps %ymm1, 64(%rax)11574; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11575; AVX-NEXT: vmovaps %ymm1, 32(%rax)11576; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11577; AVX-NEXT: vmovaps %ymm1, (%rax)11578; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax11579; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11580; AVX-NEXT: vmovaps %ymm1, 224(%rax)11581; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11582; AVX-NEXT: vmovaps %ymm1, 192(%rax)11583; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11584; AVX-NEXT: vmovaps %ymm1, 160(%rax)11585; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11586; AVX-NEXT: vmovaps %ymm1, 128(%rax)11587; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11588; AVX-NEXT: vmovaps %ymm1, 96(%rax)11589; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11590; AVX-NEXT: vmovaps %ymm1, 64(%rax)11591; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11592; AVX-NEXT: vmovaps %ymm1, 32(%rax)11593; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11594; AVX-NEXT: vmovaps %ymm1, (%rax)11595; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax11596; AVX-NEXT: vmovaps %ymm7, 224(%rax)11597; AVX-NEXT: vmovaps %ymm15, 192(%rax)11598; AVX-NEXT: vmovaps %ymm6, 160(%rax)11599; AVX-NEXT: vmovaps %ymm5, 128(%rax)11600; AVX-NEXT: vmovaps %ymm3, 96(%rax)11601; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11602; AVX-NEXT: vmovaps %ymm1, 64(%rax)11603; AVX-NEXT: vmovaps %ymm4, 32(%rax)11604; AVX-NEXT: vmovaps %ymm0, (%rax)11605; AVX-NEXT: addq $3720, %rsp # imm = 0xE8811606; AVX-NEXT: vzeroupper11607; AVX-NEXT: retq11608;11609; AVX2-LABEL: load_i32_stride8_vf64:11610; AVX2: # %bb.0:11611; AVX2-NEXT: subq $3528, %rsp # imm = 0xDC811612; AVX2-NEXT: vmovaps 288(%rdi), %xmm1011613; AVX2-NEXT: vmovaps 256(%rdi), %xmm011614; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11615; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]11616; AVX2-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11617; AVX2-NEXT: vmovaps 352(%rdi), %xmm911618; AVX2-NEXT: vbroadcastss %xmm9, %xmm111619; AVX2-NEXT: vmovaps 320(%rdi), %xmm211620; AVX2-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill11621; AVX2-NEXT: vbroadcastss %xmm2, %xmm211622; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11623; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11624; AVX2-NEXT: vmovaps 416(%rdi), %xmm111625; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11626; AVX2-NEXT: vmovaps 384(%rdi), %xmm211627; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11628; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11629; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111630; AVX2-NEXT: vmovaps 480(%rdi), %xmm1311631; AVX2-NEXT: vbroadcastss %xmm13, %xmm211632; AVX2-NEXT: vmovaps 448(%rdi), %xmm311633; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11634; AVX2-NEXT: vbroadcastss %xmm3, %xmm311635; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11636; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211637; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]11638; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11639; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11640; AVX2-NEXT: vmovaps 800(%rdi), %xmm011641; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11642; AVX2-NEXT: vmovaps 768(%rdi), %xmm111643; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11644; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11645; AVX2-NEXT: vmovaps 864(%rdi), %xmm1211646; AVX2-NEXT: vbroadcastss %xmm12, %xmm111647; AVX2-NEXT: vmovaps 832(%rdi), %xmm211648; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11649; AVX2-NEXT: vbroadcastss %xmm2, %xmm211650; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11651; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11652; AVX2-NEXT: vmovaps 992(%rdi), %xmm111653; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11654; AVX2-NEXT: vbroadcastss %xmm1, %xmm111655; AVX2-NEXT: vmovaps 960(%rdi), %xmm211656; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11657; AVX2-NEXT: vbroadcastss %xmm2, %xmm211658; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11659; AVX2-NEXT: vmovaps 928(%rdi), %xmm211660; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11661; AVX2-NEXT: vmovaps 896(%rdi), %xmm311662; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11663; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11664; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211665; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111666; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11667; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11668; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11669; AVX2-NEXT: vmovaps 1376(%rdi), %xmm011670; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11671; AVX2-NEXT: vbroadcastss %xmm0, %xmm011672; AVX2-NEXT: vmovaps 1344(%rdi), %xmm111673; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11674; AVX2-NEXT: vbroadcastss %xmm1, %xmm111675; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11676; AVX2-NEXT: vmovaps 1312(%rdi), %xmm111677; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11678; AVX2-NEXT: vmovaps 1280(%rdi), %xmm211679; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11680; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11681; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]11682; AVX2-NEXT: vmovaps 1504(%rdi), %xmm111683; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11684; AVX2-NEXT: vbroadcastss %xmm1, %xmm111685; AVX2-NEXT: vmovaps 1472(%rdi), %xmm211686; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11687; AVX2-NEXT: vbroadcastss %xmm2, %xmm211688; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11689; AVX2-NEXT: vmovaps 1440(%rdi), %xmm211690; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11691; AVX2-NEXT: vmovaps 1408(%rdi), %xmm311692; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11693; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11694; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211695; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111696; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11697; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11698; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11699; AVX2-NEXT: vmovaps 1888(%rdi), %xmm011700; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11701; AVX2-NEXT: vbroadcastss %xmm0, %xmm011702; AVX2-NEXT: vmovaps 1856(%rdi), %xmm111703; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11704; AVX2-NEXT: vbroadcastss %xmm1, %xmm111705; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11706; AVX2-NEXT: vmovaps 1824(%rdi), %xmm111707; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11708; AVX2-NEXT: vmovaps 1792(%rdi), %xmm211709; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11710; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11711; AVX2-NEXT: vblendps {{.*#+}} xmm3 = xmm1[0,1],xmm0[2,3]11712; AVX2-NEXT: vmovaps 2016(%rdi), %xmm011713; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11714; AVX2-NEXT: vbroadcastss %xmm0, %xmm111715; AVX2-NEXT: vmovaps 1984(%rdi), %xmm011716; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11717; AVX2-NEXT: vbroadcastss %xmm0, %xmm211718; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11719; AVX2-NEXT: vmovaps 1952(%rdi), %xmm011720; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11721; AVX2-NEXT: vmovaps 1920(%rdi), %xmm211722; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11723; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]11724; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211725; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111726; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11727; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm1[4,5,6,7]11728; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11729; AVX2-NEXT: vmovaps 608(%rdi), %xmm011730; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11731; AVX2-NEXT: vbroadcastss %xmm0, %xmm011732; AVX2-NEXT: vmovaps 576(%rdi), %xmm111733; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11734; AVX2-NEXT: vbroadcastss %xmm1, %xmm111735; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11736; AVX2-NEXT: vmovaps 544(%rdi), %xmm211737; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11738; AVX2-NEXT: vmovaps 512(%rdi), %xmm111739; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11740; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]11741; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]11742; AVX2-NEXT: vmovaps 736(%rdi), %xmm111743; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11744; AVX2-NEXT: vbroadcastss %xmm1, %xmm111745; AVX2-NEXT: vmovaps 704(%rdi), %xmm211746; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11747; AVX2-NEXT: vbroadcastss %xmm2, %xmm211748; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11749; AVX2-NEXT: vmovaps 672(%rdi), %xmm311750; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11751; AVX2-NEXT: vmovaps 640(%rdi), %xmm211752; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11753; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]11754; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211755; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111756; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11757; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11758; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11759; AVX2-NEXT: vmovaps 1120(%rdi), %xmm011760; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11761; AVX2-NEXT: vbroadcastss %xmm0, %xmm011762; AVX2-NEXT: vmovaps 1088(%rdi), %xmm111763; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11764; AVX2-NEXT: vbroadcastss %xmm1, %xmm111765; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11766; AVX2-NEXT: vmovaps 1056(%rdi), %xmm211767; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11768; AVX2-NEXT: vmovaps 1024(%rdi), %xmm111769; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11770; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]11771; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]11772; AVX2-NEXT: vmovaps 1248(%rdi), %xmm111773; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11774; AVX2-NEXT: vbroadcastss %xmm1, %xmm111775; AVX2-NEXT: vmovaps 1216(%rdi), %xmm211776; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11777; AVX2-NEXT: vbroadcastss %xmm2, %xmm211778; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11779; AVX2-NEXT: vmovaps 1184(%rdi), %xmm311780; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11781; AVX2-NEXT: vmovaps 1152(%rdi), %xmm211782; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11783; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]11784; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211785; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111786; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11787; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11788; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11789; AVX2-NEXT: vmovaps 1632(%rdi), %xmm011790; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11791; AVX2-NEXT: vbroadcastss %xmm0, %xmm011792; AVX2-NEXT: vmovaps 1600(%rdi), %xmm111793; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11794; AVX2-NEXT: vbroadcastss %xmm1, %xmm111795; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11796; AVX2-NEXT: vmovaps 1568(%rdi), %xmm211797; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11798; AVX2-NEXT: vmovaps 1536(%rdi), %xmm111799; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11800; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]11801; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]11802; AVX2-NEXT: vmovaps 1760(%rdi), %xmm111803; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11804; AVX2-NEXT: vbroadcastss %xmm1, %xmm111805; AVX2-NEXT: vmovaps 1728(%rdi), %xmm211806; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11807; AVX2-NEXT: vbroadcastss %xmm2, %xmm211808; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11809; AVX2-NEXT: vmovaps 1696(%rdi), %xmm311810; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11811; AVX2-NEXT: vmovaps 1664(%rdi), %xmm211812; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11813; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]11814; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211815; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111816; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11817; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11818; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11819; AVX2-NEXT: vmovaps 224(%rdi), %xmm011820; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11821; AVX2-NEXT: vbroadcastss %xmm0, %xmm011822; AVX2-NEXT: vmovaps 192(%rdi), %xmm1111823; AVX2-NEXT: vbroadcastss %xmm11, %xmm111824; AVX2-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11825; AVX2-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11826; AVX2-NEXT: vmovaps 160(%rdi), %xmm211827; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11828; AVX2-NEXT: vmovaps 128(%rdi), %xmm111829; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11830; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]11831; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111832; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm011833; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4,5],ymm0[6,7]11834; AVX2-NEXT: vmovaps 96(%rdi), %xmm811835; AVX2-NEXT: vbroadcastss %xmm8, %xmm111836; AVX2-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11837; AVX2-NEXT: vmovaps 64(%rdi), %xmm711838; AVX2-NEXT: vbroadcastss %xmm7, %xmm211839; AVX2-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11840; AVX2-NEXT: vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]11841; AVX2-NEXT: vmovaps (%rdi), %xmm511842; AVX2-NEXT: vmovaps 32(%rdi), %xmm611843; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]11844; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11845; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11846; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm3[2,3]11847; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm4[4,5,6,7]11848; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11849; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11850; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11851; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]11852; AVX2-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11853; AVX2-NEXT: vmovaps (%rsp), %xmm1 # 16-byte Reload11854; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]11855; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11856; AVX2-NEXT: vmovaps %xmm13, %xmm911857; AVX2-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11858; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload11859; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm10[0],xmm13[0],xmm10[1],xmm13[1]11860; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111861; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11862; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11863; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11864; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11865; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211866; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11867; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11868; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11869; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11870; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11871; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11872; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]11873; AVX2-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill11874; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11875; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm12[0],xmm1[1],xmm12[1]11876; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11877; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload11878; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload11879; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]11880; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111881; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11882; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11883; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11884; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11885; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211886; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11887; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11888; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11889; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11890; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11891; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11892; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]11893; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11894; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11895; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11896; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11897; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload11898; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload11899; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]11900; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111901; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11902; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11903; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11904; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11905; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211906; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11907; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11908; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11909; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11910; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11911; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11912; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]11913; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11914; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11915; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11916; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11917; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11918; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11919; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11920; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111921; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11922; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11923; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11924; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11925; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211926; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11927; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11928; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11929; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]11930; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm6[1],xmm0[2,3]11931; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm8[0],xmm7[1],xmm8[1]11932; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11933; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload11934; AVX2-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1]11935; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111936; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11937; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11938; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11939; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11940; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211941; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11942; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11943; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11944; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11945; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11946; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11947; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]11948; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11949; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11950; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11951; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11952; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload11953; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload11954; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]11955; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111956; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11957; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11958; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11959; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11960; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211961; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11962; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11963; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11964; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11965; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11966; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11967; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]11968; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11969; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11970; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11971; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11972; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload11973; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload11974; AVX2-NEXT: vunpcklps {{.*#+}} xmm1 = xmm11[0],xmm5[0],xmm11[1],xmm5[1]11975; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111976; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11977; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11978; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11979; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]11980; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm211981; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11982; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11983; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11984; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload11985; AVX2-NEXT: # xmm0 = mem[1,1,1,1]11986; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload11987; AVX2-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]11988; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11989; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11990; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11991; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11992; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload11993; AVX2-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload11994; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]11995; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm111996; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload11997; AVX2-NEXT: # xmm2 = mem[1,1,1,1]11998; AVX2-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload11999; AVX2-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]12000; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm212001; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12002; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12003; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12004; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12005; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm6 # 16-byte Folded Reload12006; AVX2-NEXT: # xmm6 = xmm0[2],mem[2],xmm0[3],mem[3]12007; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12008; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12009; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12010; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12011; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12012; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm9[2,2,2,2]12013; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]12014; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012015; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112016; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]12017; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload12018; AVX2-NEXT: # xmm1 = mem[2,2,2,2]12019; AVX2-NEXT: vblendps $7, (%rsp), %xmm1, %xmm1 # 16-byte Folded Reload12020; AVX2-NEXT: # xmm1 = mem[0,1,2],xmm1[3]12021; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]12022; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12023; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12024; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12025; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload12026; AVX2-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]12027; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12028; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12029; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12030; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12031; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12032; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm12[2,2,2,2]12033; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm13[0,1,2],xmm0[3]12034; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012035; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112036; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]12037; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload12038; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm9[2,2,2,2]12039; AVX2-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload12040; AVX2-NEXT: # xmm1 = mem[0,1,2],xmm1[3]12041; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]12042; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12043; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12044; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12045; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload12046; AVX2-NEXT: # xmm10 = xmm0[2],mem[2],xmm0[3],mem[3]12047; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12048; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12049; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12050; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12051; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm14[2,2,2,2]12052; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3]12053; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012054; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112055; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]12056; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload12057; AVX2-NEXT: # xmm1 = mem[2,2,2,2]12058; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload12059; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm7[0,1,2],xmm1[3]12060; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]12061; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12062; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12063; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12064; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm8 # 16-byte Folded Reload12065; AVX2-NEXT: # xmm8 = xmm0[2],mem[2],xmm0[3],mem[3]12066; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12067; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12068; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12069; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12070; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload12071; AVX2-NEXT: # xmm0 = mem[2,2,2,2]12072; AVX2-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12073; AVX2-NEXT: # xmm0 = mem[0,1,2],xmm0[3]12074; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012075; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112076; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]12077; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload12078; AVX2-NEXT: # xmm1 = mem[2,2,2,2]12079; AVX2-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload12080; AVX2-NEXT: # xmm1 = mem[0,1,2],xmm1[3]12081; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]12082; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12083; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12084; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12085; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload12086; AVX2-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]12087; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12088; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12089; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12090; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12091; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12092; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm3[2,2,2,2]12093; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm4[0,1,2],xmm0[3]12094; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012095; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112096; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]12097; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload12098; AVX2-NEXT: # xmm1 = mem[2,2,2,2]12099; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload12100; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1,2],xmm1[3]12101; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]12102; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12103; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12104; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12105; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload12106; AVX2-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]12107; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12108; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12109; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12110; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12111; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12112; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm5[2,2,2,2]12113; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm11[0,1,2],xmm0[3]12114; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012115; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112116; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]12117; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload12118; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm4[2,2,2,2]12119; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload12120; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]12121; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]12122; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12123; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12124; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12125; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload12126; AVX2-NEXT: # xmm5 = xmm0[2],mem[2],xmm0[3],mem[3]12127; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12128; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12129; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12130; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12131; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12132; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload12133; AVX2-NEXT: # xmm0 = mem[2,2,2,2]12134; AVX2-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload12135; AVX2-NEXT: # xmm0 = mem[0,1,2],xmm0[3]12136; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012137; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1512138; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5],ymm0[6,7]12139; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload12140; AVX2-NEXT: vshufps {{.*#+}} xmm15 = xmm2[2,2,2,2]12141; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload12142; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]12143; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm5[0,1],xmm15[2,3]12144; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]12145; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12146; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12147; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload12148; AVX2-NEXT: # xmm5 = xmm0[2],mem[2],xmm0[3],mem[3]12149; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12150; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload12151; AVX2-NEXT: # xmm15 = mem[2,2,2,2]12152; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload12153; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm0[0,1,2],xmm15[3]12154; AVX2-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1512155; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1412156; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]12157; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12158; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm15 # 16-byte Folded Reload12159; AVX2-NEXT: # xmm15 = xmm5[2],mem[2],xmm5[3],mem[3]12160; AVX2-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload12161; AVX2-NEXT: # xmm13 = mem[2,2,2,2]12162; AVX2-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload12163; AVX2-NEXT: # xmm13 = mem[0,1,2],xmm13[3]12164; AVX2-NEXT: vblendps {{.*#+}} xmm13 = xmm15[0,1],xmm13[2,3]12165; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm14[4,5,6,7]12166; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12167; AVX2-NEXT: vmovaps (%rsp), %xmm5 # 16-byte Reload12168; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm13 # 16-byte Folded Reload12169; AVX2-NEXT: # xmm13 = xmm5[2],mem[2],xmm5[3],mem[3]12170; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12171; AVX2-NEXT: vunpckhpd {{.*#+}} xmm13 = xmm5[1],xmm13[1]12172; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12173; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm14 # 16-byte Folded Reload12174; AVX2-NEXT: # xmm14 = xmm5[2],mem[2],xmm5[3],mem[3]12175; AVX2-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm1412176; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload12177; AVX2-NEXT: # xmm12 = mem[2,3,2,3]12178; AVX2-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1212179; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm14[6,7]12180; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]12181; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12182; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12183; AVX2-NEXT: vunpckhps {{.*#+}} xmm12 = xmm5[2],xmm9[2],xmm5[3],xmm9[3]12184; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12185; AVX2-NEXT: vunpckhpd {{.*#+}} xmm11 = xmm5[1],xmm12[1]12186; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12187; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm12 # 16-byte Folded Reload12188; AVX2-NEXT: # xmm12 = xmm5[2],mem[2],xmm5[3],mem[3]12189; AVX2-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1212190; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload12191; AVX2-NEXT: # xmm13 = mem[2,3,2,3]12192; AVX2-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm1312193; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]12194; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]12195; AVX2-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12196; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm11 # 16-byte Folded Reload12197; AVX2-NEXT: # xmm11 = xmm7[2],mem[2],xmm7[3],mem[3]12198; AVX2-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm10[1],xmm11[1]12199; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12200; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm11 # 16-byte Folded Reload12201; AVX2-NEXT: # xmm11 = xmm5[2],mem[2],xmm5[3],mem[3]12202; AVX2-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm1112203; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload12204; AVX2-NEXT: # xmm10 = mem[2,3,2,3]12205; AVX2-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm1012206; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm11[6,7]12207; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]12208; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12209; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12210; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm9 # 16-byte Folded Reload12211; AVX2-NEXT: # xmm9 = xmm5[2],mem[2],xmm5[3],mem[3]12212; AVX2-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm8[1],xmm9[1]12213; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12214; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm9 # 16-byte Folded Reload12215; AVX2-NEXT: # xmm9 = xmm5[2],mem[2],xmm5[3],mem[3]12216; AVX2-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm912217; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload12218; AVX2-NEXT: # xmm8 = mem[2,3,2,3]12219; AVX2-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm812220; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm9[6,7]12221; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]12222; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12223; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload12224; AVX2-NEXT: # xmm7 = xmm6[2],mem[2],xmm6[3],mem[3]12225; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload12226; AVX2-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm7[1]12227; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload12228; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload12229; AVX2-NEXT: # xmm7 = xmm6[2],mem[2],xmm6[3],mem[3]12230; AVX2-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm712231; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload12232; AVX2-NEXT: # xmm6 = mem[2,3,2,3]12233; AVX2-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm612234; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]12235; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7]12236; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12237; AVX2-NEXT: vunpckhps {{.*#+}} xmm5 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]12238; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload12239; AVX2-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm5[1]12240; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload12241; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm5 # 16-byte Folded Reload12242; AVX2-NEXT: # xmm5 = xmm4[2],mem[2],xmm4[3],mem[3]12243; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm512244; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload12245; AVX2-NEXT: # xmm4 = mem[2,3,2,3]12246; AVX2-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm412247; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]12248; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]12249; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12250; AVX2-NEXT: vunpckhps {{.*#+}} xmm3 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]12251; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload12252; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]12253; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload12254; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload12255; AVX2-NEXT: # xmm3 = xmm2[2],mem[2],xmm2[3],mem[3]12256; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm312257; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload12258; AVX2-NEXT: # xmm2 = mem[2,3,2,3]12259; AVX2-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm212260; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]12261; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]12262; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12263; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload12264; AVX2-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]12265; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112266; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload12267; AVX2-NEXT: # xmm0 = mem[2,3,2,3]12268; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm012269; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]12270; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload12271; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload12272; AVX2-NEXT: # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]12273; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm15[1],xmm1[1]12274; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12275; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12276; AVX2-NEXT: vmovaps 32(%rdi), %ymm012277; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12278; AVX2-NEXT: vmovaps (%rdi), %ymm112279; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12280; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]12281; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm012282; AVX2-NEXT: vmovaps 96(%rdi), %ymm112283; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12284; AVX2-NEXT: vmovaps 64(%rdi), %ymm212285; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12286; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]12287; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12288; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]12289; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12290; AVX2-NEXT: vmovaps 224(%rdi), %ymm212291; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12292; AVX2-NEXT: vmovaps 192(%rdi), %ymm312293; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12294; AVX2-NEXT: vmovaps 160(%rdi), %ymm1512295; AVX2-NEXT: vmovaps 128(%rdi), %ymm112296; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12297; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]12298; AVX2-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12299; AVX2-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]12300; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12301; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]12302; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12303; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12304; AVX2-NEXT: vmovaps 288(%rdi), %ymm012305; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12306; AVX2-NEXT: vmovaps 256(%rdi), %ymm112307; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12308; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]12309; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm012310; AVX2-NEXT: vmovaps 352(%rdi), %ymm112311; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12312; AVX2-NEXT: vmovaps 320(%rdi), %ymm212313; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12314; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]12315; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12316; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]12317; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12318; AVX2-NEXT: vmovaps 480(%rdi), %ymm212319; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12320; AVX2-NEXT: vmovaps 448(%rdi), %ymm312321; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12322; AVX2-NEXT: vmovaps 416(%rdi), %ymm412323; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12324; AVX2-NEXT: vmovaps 384(%rdi), %ymm112325; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12326; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[4],ymm4[4],ymm1[5],ymm4[5]12327; AVX2-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]12328; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm12[0],ymm1[2],ymm12[2]12329; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12330; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12331; AVX2-NEXT: vmovaps 544(%rdi), %ymm012332; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12333; AVX2-NEXT: vmovaps 512(%rdi), %ymm112334; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12335; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]12336; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm012337; AVX2-NEXT: vmovaps 608(%rdi), %ymm112338; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12339; AVX2-NEXT: vmovaps 576(%rdi), %ymm212340; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12341; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]12342; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12343; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]12344; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12345; AVX2-NEXT: vmovaps 736(%rdi), %ymm212346; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12347; AVX2-NEXT: vmovaps 704(%rdi), %ymm312348; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12349; AVX2-NEXT: vmovaps 672(%rdi), %ymm412350; AVX2-NEXT: vmovaps 640(%rdi), %ymm112351; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12352; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[4],ymm4[4],ymm1[5],ymm4[5]12353; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12354; AVX2-NEXT: vunpcklps {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]12355; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[2],ymm8[2]12356; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12357; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12358; AVX2-NEXT: vmovaps 800(%rdi), %ymm012359; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12360; AVX2-NEXT: vmovaps 768(%rdi), %ymm112361; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12362; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]12363; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm012364; AVX2-NEXT: vmovaps 864(%rdi), %ymm112365; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12366; AVX2-NEXT: vmovaps 832(%rdi), %ymm212367; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12368; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]12369; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12370; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]12371; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12372; AVX2-NEXT: vmovaps 992(%rdi), %ymm212373; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12374; AVX2-NEXT: vmovaps 960(%rdi), %ymm512375; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12376; AVX2-NEXT: vmovaps 928(%rdi), %ymm312377; AVX2-NEXT: vmovaps 896(%rdi), %ymm112378; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12379; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]12380; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12381; AVX2-NEXT: vunpcklps {{.*#+}} ymm5 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[4],ymm2[4],ymm5[5],ymm2[5]12382; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm5[0],ymm1[2],ymm5[2]12383; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12384; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12385; AVX2-NEXT: vmovaps 1056(%rdi), %ymm012386; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12387; AVX2-NEXT: vmovaps 1024(%rdi), %ymm112388; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12389; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]12390; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm012391; AVX2-NEXT: vmovaps 1120(%rdi), %ymm112392; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12393; AVX2-NEXT: vmovaps 1088(%rdi), %ymm212394; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12395; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]12396; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12397; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]12398; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12399; AVX2-NEXT: vmovaps 1248(%rdi), %ymm112400; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12401; AVX2-NEXT: vmovaps 1216(%rdi), %ymm712402; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12403; AVX2-NEXT: vmovaps 1184(%rdi), %ymm212404; AVX2-NEXT: vmovaps 1152(%rdi), %ymm612405; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12406; AVX2-NEXT: vunpcklps {{.*#+}} ymm6 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[4],ymm2[4],ymm6[5],ymm2[5]12407; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12408; AVX2-NEXT: vunpcklps {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[1],ymm1[1],ymm7[4],ymm1[4],ymm7[5],ymm1[5]12409; AVX2-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]12410; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]12411; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12412; AVX2-NEXT: vmovaps 1312(%rdi), %ymm012413; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12414; AVX2-NEXT: vmovaps 1280(%rdi), %ymm612415; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12416; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]12417; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm612418; AVX2-NEXT: vmovaps 1376(%rdi), %ymm012419; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12420; AVX2-NEXT: vmovaps 1344(%rdi), %ymm712421; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12422; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[4],ymm0[4],ymm7[5],ymm0[5]12423; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12424; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[2,2,2,2]12425; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]12426; AVX2-NEXT: vmovaps 1504(%rdi), %ymm612427; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12428; AVX2-NEXT: vmovaps 1472(%rdi), %ymm1012429; AVX2-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12430; AVX2-NEXT: vmovaps 1440(%rdi), %ymm012431; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill12432; AVX2-NEXT: vmovaps 1408(%rdi), %ymm912433; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12434; AVX2-NEXT: vunpcklps {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[1],ymm0[1],ymm9[4],ymm0[4],ymm9[5],ymm0[5]12435; AVX2-NEXT: vunpcklps {{.*#+}} ymm0 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[4],ymm6[4],ymm10[5],ymm6[5]12436; AVX2-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]12437; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]12438; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12439; AVX2-NEXT: vmovaps 1568(%rdi), %ymm612440; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12441; AVX2-NEXT: vmovaps 1536(%rdi), %ymm712442; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12443; AVX2-NEXT: vunpcklps {{.*#+}} ymm7 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]12444; AVX2-NEXT: vextractf128 $1, %ymm7, %xmm912445; AVX2-NEXT: vmovaps 1632(%rdi), %ymm612446; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12447; AVX2-NEXT: vmovaps 1600(%rdi), %ymm712448; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12449; AVX2-NEXT: vunpcklps {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]12450; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12451; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm6[2,2,2,2]12452; AVX2-NEXT: vblendps {{.*#+}} xmm10 = xmm9[0,1],xmm10[2,3]12453; AVX2-NEXT: vmovaps 1760(%rdi), %ymm912454; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12455; AVX2-NEXT: vmovaps 1728(%rdi), %ymm612456; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12457; AVX2-NEXT: vmovaps 1696(%rdi), %ymm712458; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12459; AVX2-NEXT: vmovaps 1664(%rdi), %ymm1112460; AVX2-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12461; AVX2-NEXT: vunpcklps {{.*#+}} ymm11 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[4],ymm7[4],ymm11[5],ymm7[5]12462; AVX2-NEXT: vunpcklps {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[4],ymm9[4],ymm6[5],ymm9[5]12463; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12464; AVX2-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm6[0],ymm11[2],ymm6[2]12465; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]12466; AVX2-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12467; AVX2-NEXT: vmovaps 1824(%rdi), %ymm612468; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12469; AVX2-NEXT: vmovaps 1792(%rdi), %ymm712470; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12471; AVX2-NEXT: vunpcklps {{.*#+}} ymm10 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]12472; AVX2-NEXT: vextractf128 $1, %ymm10, %xmm1112473; AVX2-NEXT: vmovaps 1888(%rdi), %ymm612474; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12475; AVX2-NEXT: vmovaps 1856(%rdi), %ymm712476; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12477; AVX2-NEXT: vunpcklps {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]12478; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12479; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm6[2,2,2,2]12480; AVX2-NEXT: vblendps {{.*#+}} xmm13 = xmm11[0,1],xmm13[2,3]12481; AVX2-NEXT: vmovaps 2016(%rdi), %ymm1112482; AVX2-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12483; AVX2-NEXT: vmovaps 1984(%rdi), %ymm612484; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12485; AVX2-NEXT: vmovaps 1952(%rdi), %ymm712486; AVX2-NEXT: vmovaps 1920(%rdi), %ymm912487; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12488; AVX2-NEXT: vunpcklps {{.*#+}} ymm14 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]12489; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12490; AVX2-NEXT: vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm11[0],ymm6[1],ymm11[1],ymm6[4],ymm11[4],ymm6[5],ymm11[5]12491; AVX2-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm14[0],ymm11[0],ymm14[2],ymm11[2]12492; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]12493; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12494; AVX2-NEXT: vbroadcastss 148(%rdi), %ymm1312495; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm15[5],ymm13[6,7]12496; AVX2-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload12497; AVX2-NEXT: # ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]12498; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload12499; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm1412500; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload12501; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm9[1,1,1,1,5,5,5,5]12502; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload12503; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm6[1],ymm15[2,3,4],ymm6[5],ymm15[6,7]12504; AVX2-NEXT: vextractf128 $1, %ymm15, %xmm1512505; AVX2-NEXT: vblendps {{.*#+}} xmm14 = xmm15[0,1],xmm14[2,3]12506; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]12507; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12508; AVX2-NEXT: vbroadcastss 404(%rdi), %ymm1312509; AVX2-NEXT: vblendps $32, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload12510; AVX2-NEXT: # ymm13 = ymm13[0,1,2,3,4],mem[5],ymm13[6,7]12511; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]12512; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload12513; AVX2-NEXT: vextractf128 $1, %ymm10, %xmm1312514; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload12515; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm15[1,1,1,1,5,5,5,5]12516; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload12517; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm10[1],ymm14[2,3,4],ymm10[5],ymm14[6,7]12518; AVX2-NEXT: vextractf128 $1, %ymm14, %xmm1412519; AVX2-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]12520; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]12521; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12522; AVX2-NEXT: vbroadcastss 660(%rdi), %ymm1212523; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm4[5],ymm12[6,7]12524; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm8[6,7]12525; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload12526; AVX2-NEXT: vextractf128 $1, %ymm8, %xmm812527; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload12528; AVX2-NEXT: vshufps {{.*#+}} ymm12 = ymm13[1,1,1,1,5,5,5,5]12529; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload12530; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3,4],ymm14[5],ymm12[6,7]12531; AVX2-NEXT: vextractf128 $1, %ymm12, %xmm1212532; AVX2-NEXT: vblendps {{.*#+}} xmm8 = xmm12[0,1],xmm8[2,3]12533; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]12534; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12535; AVX2-NEXT: vbroadcastss 916(%rdi), %ymm412536; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]12537; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm5[6,7]12538; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload12539; AVX2-NEXT: vextractf128 $1, %ymm4, %xmm412540; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload12541; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm8[1,1,1,1,5,5,5,5]12542; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload12543; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm12[1],ymm5[2,3,4],ymm12[5],ymm5[6,7]12544; AVX2-NEXT: vextractf128 $1, %ymm5, %xmm512545; AVX2-NEXT: vblendps {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]12546; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]12547; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12548; AVX2-NEXT: vbroadcastss 1172(%rdi), %ymm312549; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]12550; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]12551; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12552; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm212553; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload12554; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm4[1,1,1,1,5,5,5,5]12555; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload12556; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4],ymm5[5],ymm3[6,7]12557; AVX2-NEXT: vextractf128 $1, %ymm3, %xmm312558; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]12559; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]12560; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12561; AVX2-NEXT: vbroadcastss 1428(%rdi), %ymm112562; AVX2-NEXT: vblendps $32, (%rsp), %ymm1, %ymm1 # 32-byte Folded Reload12563; AVX2-NEXT: # ymm1 = ymm1[0,1,2,3,4],mem[5],ymm1[6,7]12564; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm0[6,7]12565; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12566; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm012567; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload12568; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm3[1,1,1,1,5,5,5,5]12569; AVX2-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload12570; AVX2-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]12571; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm212572; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]12573; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12574; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12575; AVX2-NEXT: vbroadcastss 1684(%rdi), %ymm012576; AVX2-NEXT: vblendps $32, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12577; AVX2-NEXT: # ymm0 = ymm0[0,1,2,3,4],mem[5],ymm0[6,7]12578; AVX2-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12579; AVX2-NEXT: # ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]12580; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12581; AVX2-NEXT: vextractf128 $1, %ymm1, %xmm112582; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload12583; AVX2-NEXT: # ymm2 = mem[1,1,1,1,5,5,5,5]12584; AVX2-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload12585; AVX2-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]12586; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm212587; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]12588; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12589; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12590; AVX2-NEXT: vbroadcastss 1940(%rdi), %ymm012591; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]12592; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm11[6,7]12593; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12594; AVX2-NEXT: vextractf128 $1, %ymm1, %xmm112595; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload12596; AVX2-NEXT: # ymm2 = mem[1,1,1,1,5,5,5,5]12597; AVX2-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload12598; AVX2-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]12599; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm212600; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]12601; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12602; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12603; AVX2-NEXT: vbroadcastss 248(%rdi), %ymm012604; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12605; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12606; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12607; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload12608; AVX2-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12609; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12610; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]12611; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12612; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12613; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload12614; AVX2-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12615; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12616; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm112617; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm6[2,2,2,2]12618; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12619; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]12620; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12621; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12622; AVX2-NEXT: vbroadcastss 504(%rdi), %ymm012623; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12624; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12625; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12626; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload12627; AVX2-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12628; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12629; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[6],ymm10[6],ymm15[7],ymm10[7]12630; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12631; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12632; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload12633; AVX2-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12634; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12635; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm112636; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm7[2,2,2,2]12637; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12638; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]12639; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12640; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12641; AVX2-NEXT: vbroadcastss 760(%rdi), %ymm012642; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12643; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12644; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12645; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload12646; AVX2-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12647; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12648; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[6],ymm14[6],ymm13[7],ymm14[7]12649; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12650; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12651; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload12652; AVX2-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12653; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12654; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm112655; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm7[2,2,2,2]12656; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12657; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]12658; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12659; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12660; AVX2-NEXT: vbroadcastss 1016(%rdi), %ymm012661; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12662; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12663; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12664; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload12665; AVX2-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12666; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12667; AVX2-NEXT: vunpckhps {{.*#+}} ymm2 = ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[6],ymm12[6],ymm8[7],ymm12[7]12668; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12669; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12670; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload12671; AVX2-NEXT: # ymm14 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12672; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm112673; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm14[2,2,2,2]12674; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12675; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]12676; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12677; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12678; AVX2-NEXT: vbroadcastss 1272(%rdi), %ymm012679; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12680; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12681; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12682; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm12 # 32-byte Folded Reload12683; AVX2-NEXT: # ymm12 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12684; AVX2-NEXT: vunpckhps {{.*#+}} ymm13 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]12685; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12686; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm11 # 32-byte Folded Reload12687; AVX2-NEXT: # ymm11 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12688; AVX2-NEXT: vextractf128 $1, %ymm13, %xmm112689; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm11[2,2,2,2]12690; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12691; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],ymm0[6,7]12692; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12693; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12694; AVX2-NEXT: vbroadcastss 1528(%rdi), %ymm012695; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12696; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12697; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12698; AVX2-NEXT: vunpckhps (%rsp), %ymm1, %ymm10 # 32-byte Folded Reload12699; AVX2-NEXT: # ymm10 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12700; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm9 # 32-byte Folded Reload12701; AVX2-NEXT: # ymm9 = ymm3[2],mem[2],ymm3[3],mem[3],ymm3[6],mem[6],ymm3[7],mem[7]12702; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12703; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload12704; AVX2-NEXT: # ymm8 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12705; AVX2-NEXT: vextractf128 $1, %ymm9, %xmm112706; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm8[2,2,2,2]12707; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12708; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],ymm0[6,7]12709; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12710; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12711; AVX2-NEXT: vbroadcastss 1784(%rdi), %ymm012712; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12713; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]12714; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12715; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload12716; AVX2-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12717; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12718; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload12719; AVX2-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12720; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12721; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload12722; AVX2-NEXT: # ymm5 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]12723; AVX2-NEXT: vextractf128 $1, %ymm7, %xmm112724; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm5[2,2,2,2]12725; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]12726; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]12727; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12728; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12729; AVX2-NEXT: vbroadcastss 2040(%rdi), %ymm012730; AVX2-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload12731; AVX2-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm0[7]12732; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12733; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload12734; AVX2-NEXT: # ymm4 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12735; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12736; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload12737; AVX2-NEXT: # ymm3 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12738; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12739; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload12740; AVX2-NEXT: # ymm2 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12741; AVX2-NEXT: vextractf128 $1, %ymm3, %xmm012742; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm2[2,2,2,2]12743; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm15[2,3]12744; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]12745; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12746; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill12747; AVX2-NEXT: vbroadcastss 220(%rdi), %ymm012748; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12749; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12750; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12751; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]12752; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12753; AVX2-NEXT: vextractf128 $1, %ymm1, %xmm112754; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload12755; AVX2-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]12756; AVX2-NEXT: vextractf128 $1, %ymm15, %xmm1512757; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]12758; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12759; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12760; AVX2-NEXT: vbroadcastss 476(%rdi), %ymm012761; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12762; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12763; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12764; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]12765; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12766; AVX2-NEXT: vextractf128 $1, %ymm1, %xmm112767; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload12768; AVX2-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]12769; AVX2-NEXT: vextractf128 $1, %ymm15, %xmm1512770; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]12771; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]12772; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12773; AVX2-NEXT: vbroadcastss 732(%rdi), %ymm012774; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12775; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12776; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12777; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]12778; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12779; AVX2-NEXT: vextractf128 $1, %ymm1, %xmm112780; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload12781; AVX2-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]12782; AVX2-NEXT: vextractf128 $1, %ymm15, %xmm1512783; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]12784; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm0[4,5,6,7]12785; AVX2-NEXT: vbroadcastss 988(%rdi), %ymm012786; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12787; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12788; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload12789; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]12790; AVX2-NEXT: vextractf128 $1, %ymm14, %xmm112791; AVX2-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload12792; AVX2-NEXT: # ymm14 = mem[2,3,2,3,6,7,6,7]12793; AVX2-NEXT: vextractf128 $1, %ymm14, %xmm1412794; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]12795; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm0[4,5,6,7]12796; AVX2-NEXT: vbroadcastss 1244(%rdi), %ymm012797; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12798; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12799; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm12[1],ymm0[1],ymm12[3],ymm0[3]12800; AVX2-NEXT: vextractf128 $1, %ymm11, %xmm112801; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm13[2,3,2,3,6,7,6,7]12802; AVX2-NEXT: vextractf128 $1, %ymm11, %xmm1112803; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm11[0,1],xmm1[2,3]12804; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]12805; AVX2-NEXT: vbroadcastss 1500(%rdi), %ymm012806; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12807; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12808; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm0[1],ymm10[3],ymm0[3]12809; AVX2-NEXT: vextractf128 $1, %ymm8, %xmm812810; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]12811; AVX2-NEXT: vextractf128 $1, %ymm9, %xmm912812; AVX2-NEXT: vblendps {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]12813; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm0[4,5,6,7]12814; AVX2-NEXT: vbroadcastss 1756(%rdi), %ymm012815; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12816; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12817; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm0[1],ymm6[3],ymm0[3]12818; AVX2-NEXT: vextractf128 $1, %ymm5, %xmm512819; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm7[2,3,2,3,6,7,6,7]12820; AVX2-NEXT: vextractf128 $1, %ymm6, %xmm612821; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]12822; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]12823; AVX2-NEXT: vbroadcastss 2012(%rdi), %ymm012824; AVX2-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload12825; AVX2-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]12826; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm0[1],ymm4[3],ymm0[3]12827; AVX2-NEXT: vextractf128 $1, %ymm2, %xmm212828; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]12829; AVX2-NEXT: vextractf128 $1, %ymm3, %xmm312830; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]12831; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]12832; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12833; AVX2-NEXT: vmovaps %ymm2, 192(%rsi)12834; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12835; AVX2-NEXT: vmovaps %ymm2, 128(%rsi)12836; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12837; AVX2-NEXT: vmovaps %ymm2, 64(%rsi)12838; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12839; AVX2-NEXT: vmovaps %ymm2, (%rsi)12840; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12841; AVX2-NEXT: vmovaps %ymm2, 224(%rsi)12842; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12843; AVX2-NEXT: vmovaps %ymm2, 160(%rsi)12844; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12845; AVX2-NEXT: vmovaps %ymm2, 96(%rsi)12846; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12847; AVX2-NEXT: vmovaps %ymm2, 32(%rsi)12848; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12849; AVX2-NEXT: vmovaps %ymm2, 192(%rdx)12850; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12851; AVX2-NEXT: vmovaps %ymm2, 128(%rdx)12852; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12853; AVX2-NEXT: vmovaps %ymm2, 64(%rdx)12854; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12855; AVX2-NEXT: vmovaps %ymm2, (%rdx)12856; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12857; AVX2-NEXT: vmovaps %ymm2, 224(%rdx)12858; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12859; AVX2-NEXT: vmovaps %ymm2, 160(%rdx)12860; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12861; AVX2-NEXT: vmovaps %ymm2, 96(%rdx)12862; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12863; AVX2-NEXT: vmovaps %ymm2, 32(%rdx)12864; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12865; AVX2-NEXT: vmovaps %ymm2, 192(%rcx)12866; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12867; AVX2-NEXT: vmovaps %ymm2, 128(%rcx)12868; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12869; AVX2-NEXT: vmovaps %ymm2, 64(%rcx)12870; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12871; AVX2-NEXT: vmovaps %ymm2, (%rcx)12872; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12873; AVX2-NEXT: vmovaps %ymm2, 224(%rcx)12874; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12875; AVX2-NEXT: vmovaps %ymm2, 160(%rcx)12876; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12877; AVX2-NEXT: vmovaps %ymm2, 96(%rcx)12878; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12879; AVX2-NEXT: vmovaps %ymm2, 32(%rcx)12880; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12881; AVX2-NEXT: vmovaps %ymm2, 192(%r8)12882; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12883; AVX2-NEXT: vmovaps %ymm2, 128(%r8)12884; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12885; AVX2-NEXT: vmovaps %ymm2, 64(%r8)12886; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12887; AVX2-NEXT: vmovaps %ymm2, (%r8)12888; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12889; AVX2-NEXT: vmovaps %ymm2, 224(%r8)12890; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12891; AVX2-NEXT: vmovaps %ymm2, 160(%r8)12892; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12893; AVX2-NEXT: vmovaps %ymm2, 96(%r8)12894; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12895; AVX2-NEXT: vmovaps %ymm2, 32(%r8)12896; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12897; AVX2-NEXT: vmovaps %ymm2, 224(%r9)12898; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12899; AVX2-NEXT: vmovaps %ymm2, 192(%r9)12900; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12901; AVX2-NEXT: vmovaps %ymm2, 160(%r9)12902; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12903; AVX2-NEXT: vmovaps %ymm2, 128(%r9)12904; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12905; AVX2-NEXT: vmovaps %ymm2, 96(%r9)12906; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12907; AVX2-NEXT: vmovaps %ymm2, 64(%r9)12908; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12909; AVX2-NEXT: vmovaps %ymm2, 32(%r9)12910; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12911; AVX2-NEXT: vmovaps %ymm2, (%r9)12912; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax12913; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12914; AVX2-NEXT: vmovaps %ymm2, 224(%rax)12915; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12916; AVX2-NEXT: vmovaps %ymm2, 192(%rax)12917; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12918; AVX2-NEXT: vmovaps %ymm2, 160(%rax)12919; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12920; AVX2-NEXT: vmovaps %ymm2, 128(%rax)12921; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12922; AVX2-NEXT: vmovaps %ymm2, 96(%rax)12923; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12924; AVX2-NEXT: vmovaps %ymm2, 64(%rax)12925; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12926; AVX2-NEXT: vmovaps %ymm2, 32(%rax)12927; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12928; AVX2-NEXT: vmovaps %ymm2, (%rax)12929; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax12930; AVX2-NEXT: vmovups (%rsp), %ymm2 # 32-byte Reload12931; AVX2-NEXT: vmovaps %ymm2, 224(%rax)12932; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12933; AVX2-NEXT: vmovaps %ymm2, 192(%rax)12934; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12935; AVX2-NEXT: vmovaps %ymm2, 160(%rax)12936; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12937; AVX2-NEXT: vmovaps %ymm2, 128(%rax)12938; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12939; AVX2-NEXT: vmovaps %ymm2, 96(%rax)12940; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12941; AVX2-NEXT: vmovaps %ymm2, 64(%rax)12942; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12943; AVX2-NEXT: vmovaps %ymm2, 32(%rax)12944; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload12945; AVX2-NEXT: vmovaps %ymm2, (%rax)12946; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax12947; AVX2-NEXT: vmovaps %ymm0, 224(%rax)12948; AVX2-NEXT: vmovaps %ymm5, 192(%rax)12949; AVX2-NEXT: vmovaps %ymm8, 160(%rax)12950; AVX2-NEXT: vmovaps %ymm1, 128(%rax)12951; AVX2-NEXT: vmovaps %ymm14, 96(%rax)12952; AVX2-NEXT: vmovaps %ymm15, 64(%rax)12953; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12954; AVX2-NEXT: vmovaps %ymm0, 32(%rax)12955; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12956; AVX2-NEXT: vmovaps %ymm0, (%rax)12957; AVX2-NEXT: addq $3528, %rsp # imm = 0xDC812958; AVX2-NEXT: vzeroupper12959; AVX2-NEXT: retq12960;12961; AVX2-FP-LABEL: load_i32_stride8_vf64:12962; AVX2-FP: # %bb.0:12963; AVX2-FP-NEXT: subq $3528, %rsp # imm = 0xDC812964; AVX2-FP-NEXT: vmovaps 288(%rdi), %xmm1012965; AVX2-FP-NEXT: vmovaps 256(%rdi), %xmm012966; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12967; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]12968; AVX2-FP-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12969; AVX2-FP-NEXT: vmovaps 352(%rdi), %xmm912970; AVX2-FP-NEXT: vbroadcastss %xmm9, %xmm112971; AVX2-FP-NEXT: vmovaps 320(%rdi), %xmm212972; AVX2-FP-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill12973; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm212974; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]12975; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12976; AVX2-FP-NEXT: vmovaps 416(%rdi), %xmm112977; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12978; AVX2-FP-NEXT: vmovaps 384(%rdi), %xmm212979; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12980; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]12981; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm112982; AVX2-FP-NEXT: vmovaps 480(%rdi), %xmm1312983; AVX2-FP-NEXT: vbroadcastss %xmm13, %xmm212984; AVX2-FP-NEXT: vmovaps 448(%rdi), %xmm312985; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12986; AVX2-FP-NEXT: vbroadcastss %xmm3, %xmm312987; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12988; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm212989; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]12990; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12991; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12992; AVX2-FP-NEXT: vmovaps 800(%rdi), %xmm012993; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12994; AVX2-FP-NEXT: vmovaps 768(%rdi), %xmm112995; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill12996; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12997; AVX2-FP-NEXT: vmovaps 864(%rdi), %xmm1212998; AVX2-FP-NEXT: vbroadcastss %xmm12, %xmm112999; AVX2-FP-NEXT: vmovaps 832(%rdi), %xmm213000; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13001; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm213002; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13003; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13004; AVX2-FP-NEXT: vmovaps 992(%rdi), %xmm113005; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13006; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113007; AVX2-FP-NEXT: vmovaps 960(%rdi), %xmm213008; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13009; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm213010; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13011; AVX2-FP-NEXT: vmovaps 928(%rdi), %xmm213012; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13013; AVX2-FP-NEXT: vmovaps 896(%rdi), %xmm313014; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13015; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13016; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213017; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113018; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13019; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13020; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13021; AVX2-FP-NEXT: vmovaps 1376(%rdi), %xmm013022; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13023; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm013024; AVX2-FP-NEXT: vmovaps 1344(%rdi), %xmm113025; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13026; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113027; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13028; AVX2-FP-NEXT: vmovaps 1312(%rdi), %xmm113029; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13030; AVX2-FP-NEXT: vmovaps 1280(%rdi), %xmm213031; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13032; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13033; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]13034; AVX2-FP-NEXT: vmovaps 1504(%rdi), %xmm113035; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13036; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113037; AVX2-FP-NEXT: vmovaps 1472(%rdi), %xmm213038; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13039; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm213040; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13041; AVX2-FP-NEXT: vmovaps 1440(%rdi), %xmm213042; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13043; AVX2-FP-NEXT: vmovaps 1408(%rdi), %xmm313044; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13045; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13046; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213047; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113048; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13049; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13050; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13051; AVX2-FP-NEXT: vmovaps 1888(%rdi), %xmm013052; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13053; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm013054; AVX2-FP-NEXT: vmovaps 1856(%rdi), %xmm113055; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13056; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113057; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13058; AVX2-FP-NEXT: vmovaps 1824(%rdi), %xmm113059; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13060; AVX2-FP-NEXT: vmovaps 1792(%rdi), %xmm213061; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13062; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13063; AVX2-FP-NEXT: vblendps {{.*#+}} xmm3 = xmm1[0,1],xmm0[2,3]13064; AVX2-FP-NEXT: vmovaps 2016(%rdi), %xmm013065; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13066; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm113067; AVX2-FP-NEXT: vmovaps 1984(%rdi), %xmm013068; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13069; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm213070; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13071; AVX2-FP-NEXT: vmovaps 1952(%rdi), %xmm013072; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13073; AVX2-FP-NEXT: vmovaps 1920(%rdi), %xmm213074; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13075; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]13076; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213077; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113078; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13079; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm1[4,5,6,7]13080; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13081; AVX2-FP-NEXT: vmovaps 608(%rdi), %xmm013082; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13083; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm013084; AVX2-FP-NEXT: vmovaps 576(%rdi), %xmm113085; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13086; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113087; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13088; AVX2-FP-NEXT: vmovaps 544(%rdi), %xmm213089; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13090; AVX2-FP-NEXT: vmovaps 512(%rdi), %xmm113091; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13092; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]13093; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]13094; AVX2-FP-NEXT: vmovaps 736(%rdi), %xmm113095; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13096; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113097; AVX2-FP-NEXT: vmovaps 704(%rdi), %xmm213098; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13099; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm213100; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13101; AVX2-FP-NEXT: vmovaps 672(%rdi), %xmm313102; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13103; AVX2-FP-NEXT: vmovaps 640(%rdi), %xmm213104; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13105; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]13106; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213107; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113108; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13109; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13110; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13111; AVX2-FP-NEXT: vmovaps 1120(%rdi), %xmm013112; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13113; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm013114; AVX2-FP-NEXT: vmovaps 1088(%rdi), %xmm113115; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13116; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113117; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13118; AVX2-FP-NEXT: vmovaps 1056(%rdi), %xmm213119; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13120; AVX2-FP-NEXT: vmovaps 1024(%rdi), %xmm113121; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13122; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]13123; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]13124; AVX2-FP-NEXT: vmovaps 1248(%rdi), %xmm113125; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13126; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113127; AVX2-FP-NEXT: vmovaps 1216(%rdi), %xmm213128; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13129; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm213130; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13131; AVX2-FP-NEXT: vmovaps 1184(%rdi), %xmm313132; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13133; AVX2-FP-NEXT: vmovaps 1152(%rdi), %xmm213134; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13135; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]13136; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213137; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113138; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13139; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13140; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13141; AVX2-FP-NEXT: vmovaps 1632(%rdi), %xmm013142; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13143; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm013144; AVX2-FP-NEXT: vmovaps 1600(%rdi), %xmm113145; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13146; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113147; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13148; AVX2-FP-NEXT: vmovaps 1568(%rdi), %xmm213149; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13150; AVX2-FP-NEXT: vmovaps 1536(%rdi), %xmm113151; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13152; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]13153; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]13154; AVX2-FP-NEXT: vmovaps 1760(%rdi), %xmm113155; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13156; AVX2-FP-NEXT: vbroadcastss %xmm1, %xmm113157; AVX2-FP-NEXT: vmovaps 1728(%rdi), %xmm213158; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13159; AVX2-FP-NEXT: vbroadcastss %xmm2, %xmm213160; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13161; AVX2-FP-NEXT: vmovaps 1696(%rdi), %xmm313162; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13163; AVX2-FP-NEXT: vmovaps 1664(%rdi), %xmm213164; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13165; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]13166; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213167; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113168; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13169; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13170; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13171; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm013172; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13173; AVX2-FP-NEXT: vbroadcastss %xmm0, %xmm013174; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm1113175; AVX2-FP-NEXT: vbroadcastss %xmm11, %xmm113176; AVX2-FP-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13177; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13178; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm213179; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13180; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm113181; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13182; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]13183; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113184; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013185; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4,5],ymm0[6,7]13186; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm813187; AVX2-FP-NEXT: vbroadcastss %xmm8, %xmm113188; AVX2-FP-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13189; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm713190; AVX2-FP-NEXT: vbroadcastss %xmm7, %xmm213191; AVX2-FP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13192; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]13193; AVX2-FP-NEXT: vmovaps (%rdi), %xmm513194; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm613195; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]13196; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13197; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13198; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm3[2,3]13199; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm4[4,5,6,7]13200; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13201; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13202; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13203; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]13204; AVX2-FP-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13205; AVX2-FP-NEXT: vmovaps (%rsp), %xmm1 # 16-byte Reload13206; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]13207; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13208; AVX2-FP-NEXT: vmovaps %xmm13, %xmm913209; AVX2-FP-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13210; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload13211; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm10[0],xmm13[0],xmm10[1],xmm13[1]13212; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113213; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13214; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13215; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13216; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13217; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213218; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13219; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13220; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13221; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13222; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13223; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13224; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]13225; AVX2-FP-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13226; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13227; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm12[0],xmm1[1],xmm12[1]13228; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13229; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload13230; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload13231; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]13232; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113233; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13234; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13235; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13236; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13237; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213238; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13239; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13240; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13241; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13242; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13243; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13244; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]13245; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13246; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13247; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13248; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13249; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload13250; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload13251; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]13252; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113253; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13254; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13255; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13256; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13257; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213258; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]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: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13262; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13263; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13264; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]13265; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13266; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13267; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13268; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13269; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13270; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13271; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13272; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113273; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13274; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13275; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13276; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13277; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213278; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13279; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13280; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13281; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]13282; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm6[1],xmm0[2,3]13283; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm8[0],xmm7[1],xmm8[1]13284; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13285; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload13286; AVX2-FP-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1]13287; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113288; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13289; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13290; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13291; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13292; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213293; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13294; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13295; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13296; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13297; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13298; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13299; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]13300; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13301; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13302; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13303; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13304; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload13305; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload13306; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]13307; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113308; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13309; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13310; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13311; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13312; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213313; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13314; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13315; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13316; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13317; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13318; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13319; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]13320; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13321; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13322; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13323; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13324; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13325; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload13326; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm11[0],xmm5[0],xmm11[1],xmm5[1]13327; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113328; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13329; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13330; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13331; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13332; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213333; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13334; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13335; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13336; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13337; AVX2-FP-NEXT: # xmm0 = mem[1,1,1,1]13338; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13339; AVX2-FP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]13340; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13341; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13342; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13343; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13344; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13345; AVX2-FP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13346; AVX2-FP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]13347; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113348; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13349; AVX2-FP-NEXT: # xmm2 = mem[1,1,1,1]13350; AVX2-FP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload13351; AVX2-FP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]13352; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213353; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13354; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13355; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13356; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13357; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm6 # 16-byte Folded Reload13358; AVX2-FP-NEXT: # xmm6 = xmm0[2],mem[2],xmm0[3],mem[3]13359; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13360; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13361; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13362; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13363; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13364; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm9[2,2,2,2]13365; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]13366; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013367; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113368; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]13369; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload13370; AVX2-FP-NEXT: # xmm1 = mem[2,2,2,2]13371; AVX2-FP-NEXT: vblendps $7, (%rsp), %xmm1, %xmm1 # 16-byte Folded Reload13372; AVX2-FP-NEXT: # xmm1 = mem[0,1,2],xmm1[3]13373; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]13374; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13375; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13376; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13377; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload13378; AVX2-FP-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]13379; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13380; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13381; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13382; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13383; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13384; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm12[2,2,2,2]13385; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm13[0,1,2],xmm0[3]13386; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013387; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113388; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]13389; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload13390; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm9[2,2,2,2]13391; AVX2-FP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13392; AVX2-FP-NEXT: # xmm1 = mem[0,1,2],xmm1[3]13393; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]13394; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13395; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13396; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13397; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload13398; AVX2-FP-NEXT: # xmm10 = xmm0[2],mem[2],xmm0[3],mem[3]13399; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13400; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13401; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13402; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13403; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm14[2,2,2,2]13404; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3]13405; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013406; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113407; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]13408; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload13409; AVX2-FP-NEXT: # xmm1 = mem[2,2,2,2]13410; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload13411; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm7[0,1,2],xmm1[3]13412; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]13413; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13414; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13415; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13416; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm8 # 16-byte Folded Reload13417; AVX2-FP-NEXT: # xmm8 = xmm0[2],mem[2],xmm0[3],mem[3]13418; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13419; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13420; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13421; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13422; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13423; AVX2-FP-NEXT: # xmm0 = mem[2,2,2,2]13424; AVX2-FP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13425; AVX2-FP-NEXT: # xmm0 = mem[0,1,2],xmm0[3]13426; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013427; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113428; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]13429; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload13430; AVX2-FP-NEXT: # xmm1 = mem[2,2,2,2]13431; AVX2-FP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13432; AVX2-FP-NEXT: # xmm1 = mem[0,1,2],xmm1[3]13433; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]13434; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13435; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13436; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13437; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload13438; AVX2-FP-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]13439; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13440; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13441; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13442; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13443; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13444; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm3[2,2,2,2]13445; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm4[0,1,2],xmm0[3]13446; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013447; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113448; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]13449; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload13450; AVX2-FP-NEXT: # xmm1 = mem[2,2,2,2]13451; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload13452; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1,2],xmm1[3]13453; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]13454; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13455; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13456; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13457; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload13458; AVX2-FP-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]13459; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13460; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13461; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13462; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13463; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13464; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm5[2,2,2,2]13465; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm11[0,1,2],xmm0[3]13466; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013467; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113468; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]13469; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload13470; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm4[2,2,2,2]13471; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload13472; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]13473; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]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: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13477; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload13478; AVX2-FP-NEXT: # xmm5 = xmm0[2],mem[2],xmm0[3],mem[3]13479; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13480; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13481; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13482; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13483; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13484; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13485; AVX2-FP-NEXT: # xmm0 = mem[2,2,2,2]13486; AVX2-FP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload13487; AVX2-FP-NEXT: # xmm0 = mem[0,1,2],xmm0[3]13488; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013489; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1513490; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5],ymm0[6,7]13491; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload13492; AVX2-FP-NEXT: vshufps {{.*#+}} xmm15 = xmm2[2,2,2,2]13493; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13494; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]13495; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm5[0,1],xmm15[2,3]13496; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]13497; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13498; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13499; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload13500; AVX2-FP-NEXT: # xmm5 = xmm0[2],mem[2],xmm0[3],mem[3]13501; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill13502; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload13503; AVX2-FP-NEXT: # xmm15 = mem[2,2,2,2]13504; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload13505; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm0[0,1,2],xmm15[3]13506; AVX2-FP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1513507; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1413508; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]13509; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13510; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm15 # 16-byte Folded Reload13511; AVX2-FP-NEXT: # xmm15 = xmm5[2],mem[2],xmm5[3],mem[3]13512; AVX2-FP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload13513; AVX2-FP-NEXT: # xmm13 = mem[2,2,2,2]13514; AVX2-FP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload13515; AVX2-FP-NEXT: # xmm13 = mem[0,1,2],xmm13[3]13516; AVX2-FP-NEXT: vblendps {{.*#+}} xmm13 = xmm15[0,1],xmm13[2,3]13517; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm14[4,5,6,7]13518; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13519; AVX2-FP-NEXT: vmovaps (%rsp), %xmm5 # 16-byte Reload13520; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm13 # 16-byte Folded Reload13521; AVX2-FP-NEXT: # xmm13 = xmm5[2],mem[2],xmm5[3],mem[3]13522; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13523; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm13 = xmm5[1],xmm13[1]13524; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13525; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm14 # 16-byte Folded Reload13526; AVX2-FP-NEXT: # xmm14 = xmm5[2],mem[2],xmm5[3],mem[3]13527; AVX2-FP-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm1413528; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload13529; AVX2-FP-NEXT: # xmm12 = mem[2,3,2,3]13530; AVX2-FP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1213531; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm14[6,7]13532; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]13533; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13534; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13535; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm12 = xmm5[2],xmm9[2],xmm5[3],xmm9[3]13536; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13537; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm11 = xmm5[1],xmm12[1]13538; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13539; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm12 # 16-byte Folded Reload13540; AVX2-FP-NEXT: # xmm12 = xmm5[2],mem[2],xmm5[3],mem[3]13541; AVX2-FP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1213542; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload13543; AVX2-FP-NEXT: # xmm13 = mem[2,3,2,3]13544; AVX2-FP-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm1313545; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]13546; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]13547; AVX2-FP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13548; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm11 # 16-byte Folded Reload13549; AVX2-FP-NEXT: # xmm11 = xmm7[2],mem[2],xmm7[3],mem[3]13550; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm10[1],xmm11[1]13551; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13552; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm11 # 16-byte Folded Reload13553; AVX2-FP-NEXT: # xmm11 = xmm5[2],mem[2],xmm5[3],mem[3]13554; AVX2-FP-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm1113555; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload13556; AVX2-FP-NEXT: # xmm10 = mem[2,3,2,3]13557; AVX2-FP-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm1013558; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm11[6,7]13559; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]13560; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13561; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13562; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm9 # 16-byte Folded Reload13563; AVX2-FP-NEXT: # xmm9 = xmm5[2],mem[2],xmm5[3],mem[3]13564; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm8[1],xmm9[1]13565; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13566; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm9 # 16-byte Folded Reload13567; AVX2-FP-NEXT: # xmm9 = xmm5[2],mem[2],xmm5[3],mem[3]13568; AVX2-FP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm913569; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload13570; AVX2-FP-NEXT: # xmm8 = mem[2,3,2,3]13571; AVX2-FP-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm813572; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm9[6,7]13573; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]13574; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13575; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload13576; AVX2-FP-NEXT: # xmm7 = xmm6[2],mem[2],xmm6[3],mem[3]13577; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload13578; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm7[1]13579; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload13580; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload13581; AVX2-FP-NEXT: # xmm7 = xmm6[2],mem[2],xmm6[3],mem[3]13582; AVX2-FP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm713583; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload13584; AVX2-FP-NEXT: # xmm6 = mem[2,3,2,3]13585; AVX2-FP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm613586; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]13587; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7]13588; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13589; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]13590; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload13591; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm5[1]13592; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload13593; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm5 # 16-byte Folded Reload13594; AVX2-FP-NEXT: # xmm5 = xmm4[2],mem[2],xmm4[3],mem[3]13595; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm513596; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload13597; AVX2-FP-NEXT: # xmm4 = mem[2,3,2,3]13598; AVX2-FP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm413599; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]13600; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]13601; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13602; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm3 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]13603; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13604; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]13605; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload13606; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload13607; AVX2-FP-NEXT: # xmm3 = xmm2[2],mem[2],xmm2[3],mem[3]13608; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm313609; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload13610; AVX2-FP-NEXT: # xmm2 = mem[2,3,2,3]13611; AVX2-FP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm213612; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]13613; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]13614; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13615; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload13616; AVX2-FP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]13617; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm113618; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload13619; AVX2-FP-NEXT: # xmm0 = mem[2,3,2,3]13620; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm013621; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]13622; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload13623; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload13624; AVX2-FP-NEXT: # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]13625; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm15[1],xmm1[1]13626; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13627; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13628; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm013629; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13630; AVX2-FP-NEXT: vmovaps (%rdi), %ymm113631; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13632; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]13633; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm013634; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm113635; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13636; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm213637; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13638; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]13639; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13640; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]13641; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13642; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm213643; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13644; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm313645; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13646; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm1513647; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm113648; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13649; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]13650; AVX2-FP-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13651; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]13652; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13653; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]13654; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13655; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13656; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm013657; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13658; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm113659; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13660; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]13661; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm013662; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm113663; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13664; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm213665; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13666; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]13667; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13668; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]13669; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13670; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm213671; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13672; AVX2-FP-NEXT: vmovaps 448(%rdi), %ymm313673; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13674; AVX2-FP-NEXT: vmovaps 416(%rdi), %ymm413675; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13676; AVX2-FP-NEXT: vmovaps 384(%rdi), %ymm113677; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13678; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[4],ymm4[4],ymm1[5],ymm4[5]13679; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]13680; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm12[0],ymm1[2],ymm12[2]13681; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13682; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13683; AVX2-FP-NEXT: vmovaps 544(%rdi), %ymm013684; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13685; AVX2-FP-NEXT: vmovaps 512(%rdi), %ymm113686; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13687; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]13688; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm013689; AVX2-FP-NEXT: vmovaps 608(%rdi), %ymm113690; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13691; AVX2-FP-NEXT: vmovaps 576(%rdi), %ymm213692; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13693; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]13694; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13695; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]13696; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13697; AVX2-FP-NEXT: vmovaps 736(%rdi), %ymm213698; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13699; AVX2-FP-NEXT: vmovaps 704(%rdi), %ymm313700; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13701; AVX2-FP-NEXT: vmovaps 672(%rdi), %ymm413702; AVX2-FP-NEXT: vmovaps 640(%rdi), %ymm113703; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13704; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[4],ymm4[4],ymm1[5],ymm4[5]13705; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13706; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]13707; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[2],ymm8[2]13708; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13709; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13710; AVX2-FP-NEXT: vmovaps 800(%rdi), %ymm013711; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13712; AVX2-FP-NEXT: vmovaps 768(%rdi), %ymm113713; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13714; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]13715; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm013716; AVX2-FP-NEXT: vmovaps 864(%rdi), %ymm113717; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13718; AVX2-FP-NEXT: vmovaps 832(%rdi), %ymm213719; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13720; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]13721; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13722; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]13723; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13724; AVX2-FP-NEXT: vmovaps 992(%rdi), %ymm213725; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13726; AVX2-FP-NEXT: vmovaps 960(%rdi), %ymm513727; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13728; AVX2-FP-NEXT: vmovaps 928(%rdi), %ymm313729; AVX2-FP-NEXT: vmovaps 896(%rdi), %ymm113730; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13731; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]13732; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13733; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm5 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[4],ymm2[4],ymm5[5],ymm2[5]13734; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm5[0],ymm1[2],ymm5[2]13735; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13736; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13737; AVX2-FP-NEXT: vmovaps 1056(%rdi), %ymm013738; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13739; AVX2-FP-NEXT: vmovaps 1024(%rdi), %ymm113740; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13741; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]13742; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm013743; AVX2-FP-NEXT: vmovaps 1120(%rdi), %ymm113744; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13745; AVX2-FP-NEXT: vmovaps 1088(%rdi), %ymm213746; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13747; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]13748; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13749; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]13750; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]13751; AVX2-FP-NEXT: vmovaps 1248(%rdi), %ymm113752; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13753; AVX2-FP-NEXT: vmovaps 1216(%rdi), %ymm713754; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13755; AVX2-FP-NEXT: vmovaps 1184(%rdi), %ymm213756; AVX2-FP-NEXT: vmovaps 1152(%rdi), %ymm613757; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13758; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[4],ymm2[4],ymm6[5],ymm2[5]13759; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13760; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[1],ymm1[1],ymm7[4],ymm1[4],ymm7[5],ymm1[5]13761; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]13762; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]13763; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13764; AVX2-FP-NEXT: vmovaps 1312(%rdi), %ymm013765; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13766; AVX2-FP-NEXT: vmovaps 1280(%rdi), %ymm613767; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13768; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]13769; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm613770; AVX2-FP-NEXT: vmovaps 1376(%rdi), %ymm013771; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13772; AVX2-FP-NEXT: vmovaps 1344(%rdi), %ymm713773; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13774; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[4],ymm0[4],ymm7[5],ymm0[5]13775; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13776; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[2,2,2,2]13777; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]13778; AVX2-FP-NEXT: vmovaps 1504(%rdi), %ymm613779; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13780; AVX2-FP-NEXT: vmovaps 1472(%rdi), %ymm1013781; AVX2-FP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13782; AVX2-FP-NEXT: vmovaps 1440(%rdi), %ymm013783; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill13784; AVX2-FP-NEXT: vmovaps 1408(%rdi), %ymm913785; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13786; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[1],ymm0[1],ymm9[4],ymm0[4],ymm9[5],ymm0[5]13787; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[4],ymm6[4],ymm10[5],ymm6[5]13788; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]13789; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]13790; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13791; AVX2-FP-NEXT: vmovaps 1568(%rdi), %ymm613792; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13793; AVX2-FP-NEXT: vmovaps 1536(%rdi), %ymm713794; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13795; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm7 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]13796; AVX2-FP-NEXT: vextractf128 $1, %ymm7, %xmm913797; AVX2-FP-NEXT: vmovaps 1632(%rdi), %ymm613798; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13799; AVX2-FP-NEXT: vmovaps 1600(%rdi), %ymm713800; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13801; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]13802; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13803; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm6[2,2,2,2]13804; AVX2-FP-NEXT: vblendps {{.*#+}} xmm10 = xmm9[0,1],xmm10[2,3]13805; AVX2-FP-NEXT: vmovaps 1760(%rdi), %ymm913806; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13807; AVX2-FP-NEXT: vmovaps 1728(%rdi), %ymm613808; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13809; AVX2-FP-NEXT: vmovaps 1696(%rdi), %ymm713810; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13811; AVX2-FP-NEXT: vmovaps 1664(%rdi), %ymm1113812; AVX2-FP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13813; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm11 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[4],ymm7[4],ymm11[5],ymm7[5]13814; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[4],ymm9[4],ymm6[5],ymm9[5]13815; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13816; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm6[0],ymm11[2],ymm6[2]13817; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]13818; AVX2-FP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13819; AVX2-FP-NEXT: vmovaps 1824(%rdi), %ymm613820; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13821; AVX2-FP-NEXT: vmovaps 1792(%rdi), %ymm713822; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13823; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm10 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]13824; AVX2-FP-NEXT: vextractf128 $1, %ymm10, %xmm1113825; AVX2-FP-NEXT: vmovaps 1888(%rdi), %ymm613826; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13827; AVX2-FP-NEXT: vmovaps 1856(%rdi), %ymm713828; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13829; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]13830; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13831; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm6[2,2,2,2]13832; AVX2-FP-NEXT: vblendps {{.*#+}} xmm13 = xmm11[0,1],xmm13[2,3]13833; AVX2-FP-NEXT: vmovaps 2016(%rdi), %ymm1113834; AVX2-FP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13835; AVX2-FP-NEXT: vmovaps 1984(%rdi), %ymm613836; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13837; AVX2-FP-NEXT: vmovaps 1952(%rdi), %ymm713838; AVX2-FP-NEXT: vmovaps 1920(%rdi), %ymm913839; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13840; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm14 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]13841; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13842; AVX2-FP-NEXT: vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm11[0],ymm6[1],ymm11[1],ymm6[4],ymm11[4],ymm6[5],ymm11[5]13843; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm14[0],ymm11[0],ymm14[2],ymm11[2]13844; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]13845; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13846; AVX2-FP-NEXT: vbroadcastss 148(%rdi), %ymm1313847; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm15[5],ymm13[6,7]13848; AVX2-FP-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload13849; AVX2-FP-NEXT: # ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]13850; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload13851; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm1413852; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload13853; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm9[1,1,1,1,5,5,5,5]13854; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload13855; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm6[1],ymm15[2,3,4],ymm6[5],ymm15[6,7]13856; AVX2-FP-NEXT: vextractf128 $1, %ymm15, %xmm1513857; AVX2-FP-NEXT: vblendps {{.*#+}} xmm14 = xmm15[0,1],xmm14[2,3]13858; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]13859; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13860; AVX2-FP-NEXT: vbroadcastss 404(%rdi), %ymm1313861; AVX2-FP-NEXT: vblendps $32, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload13862; AVX2-FP-NEXT: # ymm13 = ymm13[0,1,2,3,4],mem[5],ymm13[6,7]13863; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]13864; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload13865; AVX2-FP-NEXT: vextractf128 $1, %ymm10, %xmm1313866; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload13867; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm15[1,1,1,1,5,5,5,5]13868; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload13869; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm10[1],ymm14[2,3,4],ymm10[5],ymm14[6,7]13870; AVX2-FP-NEXT: vextractf128 $1, %ymm14, %xmm1413871; AVX2-FP-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]13872; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]13873; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13874; AVX2-FP-NEXT: vbroadcastss 660(%rdi), %ymm1213875; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm4[5],ymm12[6,7]13876; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm8[6,7]13877; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload13878; AVX2-FP-NEXT: vextractf128 $1, %ymm8, %xmm813879; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload13880; AVX2-FP-NEXT: vshufps {{.*#+}} ymm12 = ymm13[1,1,1,1,5,5,5,5]13881; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload13882; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3,4],ymm14[5],ymm12[6,7]13883; AVX2-FP-NEXT: vextractf128 $1, %ymm12, %xmm1213884; AVX2-FP-NEXT: vblendps {{.*#+}} xmm8 = xmm12[0,1],xmm8[2,3]13885; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]13886; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13887; AVX2-FP-NEXT: vbroadcastss 916(%rdi), %ymm413888; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]13889; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm5[6,7]13890; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload13891; AVX2-FP-NEXT: vextractf128 $1, %ymm4, %xmm413892; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload13893; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm8[1,1,1,1,5,5,5,5]13894; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload13895; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm12[1],ymm5[2,3,4],ymm12[5],ymm5[6,7]13896; AVX2-FP-NEXT: vextractf128 $1, %ymm5, %xmm513897; AVX2-FP-NEXT: vblendps {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]13898; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]13899; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13900; AVX2-FP-NEXT: vbroadcastss 1172(%rdi), %ymm313901; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]13902; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]13903; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload13904; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm213905; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload13906; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm4[1,1,1,1,5,5,5,5]13907; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload13908; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4],ymm5[5],ymm3[6,7]13909; AVX2-FP-NEXT: vextractf128 $1, %ymm3, %xmm313910; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]13911; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]13912; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13913; AVX2-FP-NEXT: vbroadcastss 1428(%rdi), %ymm113914; AVX2-FP-NEXT: vblendps $32, (%rsp), %ymm1, %ymm1 # 32-byte Folded Reload13915; AVX2-FP-NEXT: # ymm1 = ymm1[0,1,2,3,4],mem[5],ymm1[6,7]13916; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm0[6,7]13917; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13918; AVX2-FP-NEXT: vextractf128 $1, %ymm0, %xmm013919; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload13920; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm3[1,1,1,1,5,5,5,5]13921; AVX2-FP-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload13922; AVX2-FP-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]13923; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm213924; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]13925; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13926; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13927; AVX2-FP-NEXT: vbroadcastss 1684(%rdi), %ymm013928; AVX2-FP-NEXT: vblendps $32, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13929; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2,3,4],mem[5],ymm0[6,7]13930; AVX2-FP-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13931; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]13932; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13933; AVX2-FP-NEXT: vextractf128 $1, %ymm1, %xmm113934; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload13935; AVX2-FP-NEXT: # ymm2 = mem[1,1,1,1,5,5,5,5]13936; AVX2-FP-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload13937; AVX2-FP-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]13938; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm213939; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]13940; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13941; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13942; AVX2-FP-NEXT: vbroadcastss 1940(%rdi), %ymm013943; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]13944; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm11[6,7]13945; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13946; AVX2-FP-NEXT: vextractf128 $1, %ymm1, %xmm113947; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload13948; AVX2-FP-NEXT: # ymm2 = mem[1,1,1,1,5,5,5,5]13949; AVX2-FP-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload13950; AVX2-FP-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]13951; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm213952; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]13953; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13954; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13955; AVX2-FP-NEXT: vbroadcastss 248(%rdi), %ymm013956; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13957; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]13958; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13959; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload13960; AVX2-FP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]13961; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13962; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]13963; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13964; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13965; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload13966; AVX2-FP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]13967; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13968; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm113969; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm6[2,2,2,2]13970; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]13971; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]13972; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13973; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13974; AVX2-FP-NEXT: vbroadcastss 504(%rdi), %ymm013975; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13976; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]13977; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13978; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload13979; AVX2-FP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]13980; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13981; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[6],ymm10[6],ymm15[7],ymm10[7]13982; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13983; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13984; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload13985; AVX2-FP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]13986; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13987; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm113988; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm7[2,2,2,2]13989; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]13990; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]13991; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]13992; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13993; AVX2-FP-NEXT: vbroadcastss 760(%rdi), %ymm013994; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload13995; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]13996; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload13997; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload13998; AVX2-FP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]13999; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14000; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[6],ymm14[6],ymm13[7],ymm14[7]14001; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14002; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14003; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload14004; AVX2-FP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14005; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14006; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm114007; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm7[2,2,2,2]14008; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]14009; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]14010; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14011; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14012; AVX2-FP-NEXT: vbroadcastss 1016(%rdi), %ymm014013; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14014; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]14015; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14016; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload14017; AVX2-FP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14018; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14019; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[6],ymm12[6],ymm8[7],ymm12[7]14020; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14021; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14022; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload14023; AVX2-FP-NEXT: # ymm14 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14024; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm114025; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm14[2,2,2,2]14026; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]14027; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]14028; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14029; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14030; AVX2-FP-NEXT: vbroadcastss 1272(%rdi), %ymm014031; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14032; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]14033; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14034; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm12 # 32-byte Folded Reload14035; AVX2-FP-NEXT: # ymm12 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14036; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm13 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]14037; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14038; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm11 # 32-byte Folded Reload14039; AVX2-FP-NEXT: # ymm11 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14040; AVX2-FP-NEXT: vextractf128 $1, %ymm13, %xmm114041; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm11[2,2,2,2]14042; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]14043; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],ymm0[6,7]14044; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14045; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14046; AVX2-FP-NEXT: vbroadcastss 1528(%rdi), %ymm014047; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14048; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]14049; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14050; AVX2-FP-NEXT: vunpckhps (%rsp), %ymm1, %ymm10 # 32-byte Folded Reload14051; AVX2-FP-NEXT: # ymm10 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14052; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm9 # 32-byte Folded Reload14053; AVX2-FP-NEXT: # ymm9 = ymm3[2],mem[2],ymm3[3],mem[3],ymm3[6],mem[6],ymm3[7],mem[7]14054; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14055; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload14056; AVX2-FP-NEXT: # ymm8 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14057; AVX2-FP-NEXT: vextractf128 $1, %ymm9, %xmm114058; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm8[2,2,2,2]14059; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]14060; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],ymm0[6,7]14061; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14062; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14063; AVX2-FP-NEXT: vbroadcastss 1784(%rdi), %ymm014064; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14065; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]14066; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14067; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload14068; AVX2-FP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14069; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14070; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload14071; AVX2-FP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14072; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14073; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload14074; AVX2-FP-NEXT: # ymm5 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]14075; AVX2-FP-NEXT: vextractf128 $1, %ymm7, %xmm114076; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm5[2,2,2,2]14077; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]14078; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]14079; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14080; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14081; AVX2-FP-NEXT: vbroadcastss 2040(%rdi), %ymm014082; AVX2-FP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload14083; AVX2-FP-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm0[7]14084; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14085; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload14086; AVX2-FP-NEXT: # ymm4 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14087; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14088; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload14089; AVX2-FP-NEXT: # ymm3 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14090; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14091; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload14092; AVX2-FP-NEXT: # ymm2 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14093; AVX2-FP-NEXT: vextractf128 $1, %ymm3, %xmm014094; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm2[2,2,2,2]14095; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm15[2,3]14096; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]14097; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14098; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill14099; AVX2-FP-NEXT: vbroadcastss 220(%rdi), %ymm014100; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14101; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14102; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14103; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]14104; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14105; AVX2-FP-NEXT: vextractf128 $1, %ymm1, %xmm114106; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload14107; AVX2-FP-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]14108; AVX2-FP-NEXT: vextractf128 $1, %ymm15, %xmm1514109; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]14110; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14111; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14112; AVX2-FP-NEXT: vbroadcastss 476(%rdi), %ymm014113; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14114; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14115; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14116; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]14117; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14118; AVX2-FP-NEXT: vextractf128 $1, %ymm1, %xmm114119; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload14120; AVX2-FP-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]14121; AVX2-FP-NEXT: vextractf128 $1, %ymm15, %xmm1514122; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]14123; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14124; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14125; AVX2-FP-NEXT: vbroadcastss 732(%rdi), %ymm014126; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14127; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14128; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14129; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]14130; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14131; AVX2-FP-NEXT: vextractf128 $1, %ymm1, %xmm114132; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload14133; AVX2-FP-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]14134; AVX2-FP-NEXT: vextractf128 $1, %ymm15, %xmm1514135; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]14136; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm0[4,5,6,7]14137; AVX2-FP-NEXT: vbroadcastss 988(%rdi), %ymm014138; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14139; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14140; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload14141; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]14142; AVX2-FP-NEXT: vextractf128 $1, %ymm14, %xmm114143; AVX2-FP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload14144; AVX2-FP-NEXT: # ymm14 = mem[2,3,2,3,6,7,6,7]14145; AVX2-FP-NEXT: vextractf128 $1, %ymm14, %xmm1414146; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]14147; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm0[4,5,6,7]14148; AVX2-FP-NEXT: vbroadcastss 1244(%rdi), %ymm014149; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14150; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14151; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm12[1],ymm0[1],ymm12[3],ymm0[3]14152; AVX2-FP-NEXT: vextractf128 $1, %ymm11, %xmm114153; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm13[2,3,2,3,6,7,6,7]14154; AVX2-FP-NEXT: vextractf128 $1, %ymm11, %xmm1114155; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm11[0,1],xmm1[2,3]14156; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]14157; AVX2-FP-NEXT: vbroadcastss 1500(%rdi), %ymm014158; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14159; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14160; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm0[1],ymm10[3],ymm0[3]14161; AVX2-FP-NEXT: vextractf128 $1, %ymm8, %xmm814162; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]14163; AVX2-FP-NEXT: vextractf128 $1, %ymm9, %xmm914164; AVX2-FP-NEXT: vblendps {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]14165; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm0[4,5,6,7]14166; AVX2-FP-NEXT: vbroadcastss 1756(%rdi), %ymm014167; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14168; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14169; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm0[1],ymm6[3],ymm0[3]14170; AVX2-FP-NEXT: vextractf128 $1, %ymm5, %xmm514171; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm7[2,3,2,3,6,7,6,7]14172; AVX2-FP-NEXT: vextractf128 $1, %ymm6, %xmm614173; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]14174; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]14175; AVX2-FP-NEXT: vbroadcastss 2012(%rdi), %ymm014176; AVX2-FP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload14177; AVX2-FP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]14178; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm0[1],ymm4[3],ymm0[3]14179; AVX2-FP-NEXT: vextractf128 $1, %ymm2, %xmm214180; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]14181; AVX2-FP-NEXT: vextractf128 $1, %ymm3, %xmm314182; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]14183; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]14184; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14185; AVX2-FP-NEXT: vmovaps %ymm2, 192(%rsi)14186; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14187; AVX2-FP-NEXT: vmovaps %ymm2, 128(%rsi)14188; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14189; AVX2-FP-NEXT: vmovaps %ymm2, 64(%rsi)14190; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14191; AVX2-FP-NEXT: vmovaps %ymm2, (%rsi)14192; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14193; AVX2-FP-NEXT: vmovaps %ymm2, 224(%rsi)14194; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14195; AVX2-FP-NEXT: vmovaps %ymm2, 160(%rsi)14196; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14197; AVX2-FP-NEXT: vmovaps %ymm2, 96(%rsi)14198; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14199; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rsi)14200; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14201; AVX2-FP-NEXT: vmovaps %ymm2, 192(%rdx)14202; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14203; AVX2-FP-NEXT: vmovaps %ymm2, 128(%rdx)14204; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14205; AVX2-FP-NEXT: vmovaps %ymm2, 64(%rdx)14206; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14207; AVX2-FP-NEXT: vmovaps %ymm2, (%rdx)14208; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14209; AVX2-FP-NEXT: vmovaps %ymm2, 224(%rdx)14210; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14211; AVX2-FP-NEXT: vmovaps %ymm2, 160(%rdx)14212; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14213; AVX2-FP-NEXT: vmovaps %ymm2, 96(%rdx)14214; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14215; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rdx)14216; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14217; AVX2-FP-NEXT: vmovaps %ymm2, 192(%rcx)14218; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14219; AVX2-FP-NEXT: vmovaps %ymm2, 128(%rcx)14220; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14221; AVX2-FP-NEXT: vmovaps %ymm2, 64(%rcx)14222; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14223; AVX2-FP-NEXT: vmovaps %ymm2, (%rcx)14224; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14225; AVX2-FP-NEXT: vmovaps %ymm2, 224(%rcx)14226; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14227; AVX2-FP-NEXT: vmovaps %ymm2, 160(%rcx)14228; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14229; AVX2-FP-NEXT: vmovaps %ymm2, 96(%rcx)14230; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14231; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rcx)14232; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14233; AVX2-FP-NEXT: vmovaps %ymm2, 192(%r8)14234; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14235; AVX2-FP-NEXT: vmovaps %ymm2, 128(%r8)14236; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14237; AVX2-FP-NEXT: vmovaps %ymm2, 64(%r8)14238; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14239; AVX2-FP-NEXT: vmovaps %ymm2, (%r8)14240; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14241; AVX2-FP-NEXT: vmovaps %ymm2, 224(%r8)14242; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14243; AVX2-FP-NEXT: vmovaps %ymm2, 160(%r8)14244; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14245; AVX2-FP-NEXT: vmovaps %ymm2, 96(%r8)14246; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14247; AVX2-FP-NEXT: vmovaps %ymm2, 32(%r8)14248; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14249; AVX2-FP-NEXT: vmovaps %ymm2, 224(%r9)14250; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14251; AVX2-FP-NEXT: vmovaps %ymm2, 192(%r9)14252; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14253; AVX2-FP-NEXT: vmovaps %ymm2, 160(%r9)14254; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14255; AVX2-FP-NEXT: vmovaps %ymm2, 128(%r9)14256; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14257; AVX2-FP-NEXT: vmovaps %ymm2, 96(%r9)14258; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14259; AVX2-FP-NEXT: vmovaps %ymm2, 64(%r9)14260; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14261; AVX2-FP-NEXT: vmovaps %ymm2, 32(%r9)14262; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14263; AVX2-FP-NEXT: vmovaps %ymm2, (%r9)14264; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax14265; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14266; AVX2-FP-NEXT: vmovaps %ymm2, 224(%rax)14267; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14268; AVX2-FP-NEXT: vmovaps %ymm2, 192(%rax)14269; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14270; AVX2-FP-NEXT: vmovaps %ymm2, 160(%rax)14271; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14272; AVX2-FP-NEXT: vmovaps %ymm2, 128(%rax)14273; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14274; AVX2-FP-NEXT: vmovaps %ymm2, 96(%rax)14275; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14276; AVX2-FP-NEXT: vmovaps %ymm2, 64(%rax)14277; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14278; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rax)14279; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14280; AVX2-FP-NEXT: vmovaps %ymm2, (%rax)14281; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax14282; AVX2-FP-NEXT: vmovups (%rsp), %ymm2 # 32-byte Reload14283; AVX2-FP-NEXT: vmovaps %ymm2, 224(%rax)14284; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14285; AVX2-FP-NEXT: vmovaps %ymm2, 192(%rax)14286; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14287; AVX2-FP-NEXT: vmovaps %ymm2, 160(%rax)14288; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14289; AVX2-FP-NEXT: vmovaps %ymm2, 128(%rax)14290; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14291; AVX2-FP-NEXT: vmovaps %ymm2, 96(%rax)14292; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14293; AVX2-FP-NEXT: vmovaps %ymm2, 64(%rax)14294; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14295; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rax)14296; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload14297; AVX2-FP-NEXT: vmovaps %ymm2, (%rax)14298; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax14299; AVX2-FP-NEXT: vmovaps %ymm0, 224(%rax)14300; AVX2-FP-NEXT: vmovaps %ymm5, 192(%rax)14301; AVX2-FP-NEXT: vmovaps %ymm8, 160(%rax)14302; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rax)14303; AVX2-FP-NEXT: vmovaps %ymm14, 96(%rax)14304; AVX2-FP-NEXT: vmovaps %ymm15, 64(%rax)14305; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14306; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rax)14307; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload14308; AVX2-FP-NEXT: vmovaps %ymm0, (%rax)14309; AVX2-FP-NEXT: addq $3528, %rsp # imm = 0xDC814310; AVX2-FP-NEXT: vzeroupper14311; AVX2-FP-NEXT: retq14312;14313; AVX2-FCP-LABEL: load_i32_stride8_vf64:14314; AVX2-FCP: # %bb.0:14315; AVX2-FCP-NEXT: subq $3528, %rsp # imm = 0xDC814316; AVX2-FCP-NEXT: vmovaps 288(%rdi), %xmm1014317; AVX2-FCP-NEXT: vmovaps 256(%rdi), %xmm014318; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14319; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]14320; AVX2-FCP-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14321; AVX2-FCP-NEXT: vmovaps 352(%rdi), %xmm914322; AVX2-FCP-NEXT: vbroadcastss %xmm9, %xmm114323; AVX2-FCP-NEXT: vmovaps 320(%rdi), %xmm214324; AVX2-FCP-NEXT: vmovaps %xmm2, (%rsp) # 16-byte Spill14325; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214326; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14327; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14328; AVX2-FCP-NEXT: vmovaps 416(%rdi), %xmm114329; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14330; AVX2-FCP-NEXT: vmovaps 384(%rdi), %xmm214331; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14332; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14333; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114334; AVX2-FCP-NEXT: vmovaps 480(%rdi), %xmm1314335; AVX2-FCP-NEXT: vbroadcastss %xmm13, %xmm214336; AVX2-FCP-NEXT: vmovaps 448(%rdi), %xmm314337; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14338; AVX2-FCP-NEXT: vbroadcastss %xmm3, %xmm314339; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]14340; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214341; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]14342; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14343; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14344; AVX2-FCP-NEXT: vmovaps 800(%rdi), %xmm014345; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14346; AVX2-FCP-NEXT: vmovaps 768(%rdi), %xmm114347; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14348; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14349; AVX2-FCP-NEXT: vmovaps 864(%rdi), %xmm1214350; AVX2-FCP-NEXT: vbroadcastss %xmm12, %xmm114351; AVX2-FCP-NEXT: vmovaps 832(%rdi), %xmm214352; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14353; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214354; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14355; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14356; AVX2-FCP-NEXT: vmovaps 992(%rdi), %xmm114357; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14358; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114359; AVX2-FCP-NEXT: vmovaps 960(%rdi), %xmm214360; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14361; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214362; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14363; AVX2-FCP-NEXT: vmovaps 928(%rdi), %xmm214364; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14365; AVX2-FCP-NEXT: vmovaps 896(%rdi), %xmm314366; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14367; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]14368; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214369; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114370; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14371; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14372; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14373; AVX2-FCP-NEXT: vmovaps 1376(%rdi), %xmm014374; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14375; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm014376; AVX2-FCP-NEXT: vmovaps 1344(%rdi), %xmm114377; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14378; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114379; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14380; AVX2-FCP-NEXT: vmovaps 1312(%rdi), %xmm114381; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14382; AVX2-FCP-NEXT: vmovaps 1280(%rdi), %xmm214383; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14384; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14385; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]14386; AVX2-FCP-NEXT: vmovaps 1504(%rdi), %xmm114387; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14388; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114389; AVX2-FCP-NEXT: vmovaps 1472(%rdi), %xmm214390; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14391; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214392; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14393; AVX2-FCP-NEXT: vmovaps 1440(%rdi), %xmm214394; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14395; AVX2-FCP-NEXT: vmovaps 1408(%rdi), %xmm314396; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14397; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]14398; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214399; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114400; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14401; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14402; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14403; AVX2-FCP-NEXT: vmovaps 1888(%rdi), %xmm014404; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14405; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm014406; AVX2-FCP-NEXT: vmovaps 1856(%rdi), %xmm114407; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14408; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114409; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14410; AVX2-FCP-NEXT: vmovaps 1824(%rdi), %xmm114411; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14412; AVX2-FCP-NEXT: vmovaps 1792(%rdi), %xmm214413; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14414; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14415; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm1[0,1],xmm0[2,3]14416; AVX2-FCP-NEXT: vmovaps 2016(%rdi), %xmm014417; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14418; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm114419; AVX2-FCP-NEXT: vmovaps 1984(%rdi), %xmm014420; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14421; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm214422; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14423; AVX2-FCP-NEXT: vmovaps 1952(%rdi), %xmm014424; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14425; AVX2-FCP-NEXT: vmovaps 1920(%rdi), %xmm214426; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14427; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]14428; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214429; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114430; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14431; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm1[4,5,6,7]14432; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14433; AVX2-FCP-NEXT: vmovaps 608(%rdi), %xmm014434; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14435; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm014436; AVX2-FCP-NEXT: vmovaps 576(%rdi), %xmm114437; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14438; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114439; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14440; AVX2-FCP-NEXT: vmovaps 544(%rdi), %xmm214441; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14442; AVX2-FCP-NEXT: vmovaps 512(%rdi), %xmm114443; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14444; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]14445; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]14446; AVX2-FCP-NEXT: vmovaps 736(%rdi), %xmm114447; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14448; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114449; AVX2-FCP-NEXT: vmovaps 704(%rdi), %xmm214450; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14451; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214452; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14453; AVX2-FCP-NEXT: vmovaps 672(%rdi), %xmm314454; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14455; AVX2-FCP-NEXT: vmovaps 640(%rdi), %xmm214456; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14457; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]14458; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214459; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114460; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14461; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14462; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14463; AVX2-FCP-NEXT: vmovaps 1120(%rdi), %xmm014464; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14465; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm014466; AVX2-FCP-NEXT: vmovaps 1088(%rdi), %xmm114467; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14468; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114469; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14470; AVX2-FCP-NEXT: vmovaps 1056(%rdi), %xmm214471; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14472; AVX2-FCP-NEXT: vmovaps 1024(%rdi), %xmm114473; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14474; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]14475; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]14476; AVX2-FCP-NEXT: vmovaps 1248(%rdi), %xmm114477; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14478; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114479; AVX2-FCP-NEXT: vmovaps 1216(%rdi), %xmm214480; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14481; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214482; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14483; AVX2-FCP-NEXT: vmovaps 1184(%rdi), %xmm314484; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14485; AVX2-FCP-NEXT: vmovaps 1152(%rdi), %xmm214486; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14487; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]14488; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214489; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114490; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14491; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14492; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14493; AVX2-FCP-NEXT: vmovaps 1632(%rdi), %xmm014494; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14495; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm014496; AVX2-FCP-NEXT: vmovaps 1600(%rdi), %xmm114497; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14498; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114499; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14500; AVX2-FCP-NEXT: vmovaps 1568(%rdi), %xmm214501; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14502; AVX2-FCP-NEXT: vmovaps 1536(%rdi), %xmm114503; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14504; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]14505; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]14506; AVX2-FCP-NEXT: vmovaps 1760(%rdi), %xmm114507; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14508; AVX2-FCP-NEXT: vbroadcastss %xmm1, %xmm114509; AVX2-FCP-NEXT: vmovaps 1728(%rdi), %xmm214510; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14511; AVX2-FCP-NEXT: vbroadcastss %xmm2, %xmm214512; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14513; AVX2-FCP-NEXT: vmovaps 1696(%rdi), %xmm314514; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14515; AVX2-FCP-NEXT: vmovaps 1664(%rdi), %xmm214516; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14517; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]14518; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214519; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114520; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14521; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14522; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14523; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm014524; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14525; AVX2-FCP-NEXT: vbroadcastss %xmm0, %xmm014526; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm1114527; AVX2-FCP-NEXT: vbroadcastss %xmm11, %xmm114528; AVX2-FCP-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14529; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]14530; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm214531; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14532; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm114533; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14534; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]14535; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114536; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014537; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4,5],ymm0[6,7]14538; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm814539; AVX2-FCP-NEXT: vbroadcastss %xmm8, %xmm114540; AVX2-FCP-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14541; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm714542; AVX2-FCP-NEXT: vbroadcastss %xmm7, %xmm214543; AVX2-FCP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14544; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]14545; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm514546; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm614547; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]14548; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14549; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14550; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm3[2,3]14551; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm4[4,5,6,7]14552; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14553; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14554; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14555; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm10[1],xmm0[2,3]14556; AVX2-FCP-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14557; AVX2-FCP-NEXT: vmovaps (%rsp), %xmm1 # 16-byte Reload14558; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]14559; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14560; AVX2-FCP-NEXT: vmovaps %xmm13, %xmm914561; AVX2-FCP-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14562; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload14563; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm10[0],xmm13[0],xmm10[1],xmm13[1]14564; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114565; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14566; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14567; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14568; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14569; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214570; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14571; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14572; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14573; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14574; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14575; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14576; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]14577; AVX2-FCP-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14578; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14579; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm12[0],xmm1[1],xmm12[1]14580; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14581; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload14582; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload14583; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]14584; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114585; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14586; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14587; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14588; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14589; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214590; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14591; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14592; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14593; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14594; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14595; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14596; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]14597; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14598; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14599; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14600; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14601; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload14602; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload14603; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]14604; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114605; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14606; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14607; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14608; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14609; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214610; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14611; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14612; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14613; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14614; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14615; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14616; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]14617; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14618; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14619; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14620; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14621; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14622; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14623; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14624; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114625; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14626; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14627; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14628; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14629; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214630; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14631; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14632; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14633; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm5[1,1,1,1]14634; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm6[1],xmm0[2,3]14635; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm8[0],xmm7[1],xmm8[1]14636; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14637; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload14638; AVX2-FCP-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1]14639; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114640; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14641; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14642; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14643; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14644; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214645; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14646; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14647; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14648; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14649; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14650; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14651; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]14652; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14653; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14654; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14655; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14656; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload14657; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload14658; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]14659; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114660; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14661; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14662; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14663; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14664; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214665; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14666; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14667; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14668; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14669; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14670; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14671; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]14672; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14673; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14674; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14675; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14676; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14677; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload14678; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm11[0],xmm5[0],xmm11[1],xmm5[1]14679; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114680; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14681; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14682; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14683; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14684; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214685; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14686; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14687; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14688; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14689; AVX2-FCP-NEXT: # xmm0 = mem[1,1,1,1]14690; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14691; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[1],xmm0[2,3]14692; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14693; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14694; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14695; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14696; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14697; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14698; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]14699; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114700; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14701; AVX2-FCP-NEXT: # xmm2 = mem[1,1,1,1]14702; AVX2-FCP-NEXT: vblendps $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload14703; AVX2-FCP-NEXT: # xmm2 = xmm2[0],mem[1],xmm2[2,3]14704; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214705; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]14706; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]14707; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14708; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14709; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm6 # 16-byte Folded Reload14710; AVX2-FCP-NEXT: # xmm6 = xmm0[2],mem[2],xmm0[3],mem[3]14711; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14712; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14713; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14714; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14715; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14716; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm9[2,2,2,2]14717; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]14718; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014719; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114720; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]14721; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload14722; AVX2-FCP-NEXT: # xmm1 = mem[2,2,2,2]14723; AVX2-FCP-NEXT: vblendps $7, (%rsp), %xmm1, %xmm1 # 16-byte Folded Reload14724; AVX2-FCP-NEXT: # xmm1 = mem[0,1,2],xmm1[3]14725; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]14726; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14727; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14728; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14729; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload14730; AVX2-FCP-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]14731; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14732; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14733; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14734; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14735; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14736; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm12[2,2,2,2]14737; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm13[0,1,2],xmm0[3]14738; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014739; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114740; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]14741; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload14742; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm9[2,2,2,2]14743; AVX2-FCP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14744; AVX2-FCP-NEXT: # xmm1 = mem[0,1,2],xmm1[3]14745; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]14746; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14747; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14748; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14749; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload14750; AVX2-FCP-NEXT: # xmm10 = xmm0[2],mem[2],xmm0[3],mem[3]14751; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14752; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14753; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14754; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14755; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm14[2,2,2,2]14756; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm15[0,1,2],xmm0[3]14757; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014758; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114759; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]14760; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload14761; AVX2-FCP-NEXT: # xmm1 = mem[2,2,2,2]14762; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload14763; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm7[0,1,2],xmm1[3]14764; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm10[0,1],xmm1[2,3]14765; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14766; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14767; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14768; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm8 # 16-byte Folded Reload14769; AVX2-FCP-NEXT: # xmm8 = xmm0[2],mem[2],xmm0[3],mem[3]14770; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14771; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14772; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14773; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14774; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14775; AVX2-FCP-NEXT: # xmm0 = mem[2,2,2,2]14776; AVX2-FCP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14777; AVX2-FCP-NEXT: # xmm0 = mem[0,1,2],xmm0[3]14778; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014779; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114780; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]14781; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload14782; AVX2-FCP-NEXT: # xmm1 = mem[2,2,2,2]14783; AVX2-FCP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14784; AVX2-FCP-NEXT: # xmm1 = mem[0,1,2],xmm1[3]14785; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]14786; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14787; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14788; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14789; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload14790; AVX2-FCP-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]14791; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14792; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14793; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14794; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14795; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14796; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm3[2,2,2,2]14797; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm4[0,1,2],xmm0[3]14798; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014799; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114800; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]14801; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload14802; AVX2-FCP-NEXT: # xmm1 = mem[2,2,2,2]14803; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload14804; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1,2],xmm1[3]14805; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]14806; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14807; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14808; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14809; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload14810; AVX2-FCP-NEXT: # xmm2 = xmm0[2],mem[2],xmm0[3],mem[3]14811; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14812; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14813; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14814; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14815; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14816; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm0 = xmm5[2,2,2,2]14817; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm11[0,1,2],xmm0[3]14818; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014819; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114820; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]14821; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload14822; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm4[2,2,2,2]14823; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload14824; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]14825; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]14826; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14827; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14828; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14829; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload14830; AVX2-FCP-NEXT: # xmm5 = xmm0[2],mem[2],xmm0[3],mem[3]14831; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14832; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14833; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14834; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14835; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14836; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14837; AVX2-FCP-NEXT: # xmm0 = mem[2,2,2,2]14838; AVX2-FCP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload14839; AVX2-FCP-NEXT: # xmm0 = mem[0,1,2],xmm0[3]14840; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014841; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1514842; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5],ymm0[6,7]14843; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload14844; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm15 = xmm2[2,2,2,2]14845; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14846; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm1[0,1,2],xmm15[3]14847; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm5[0,1],xmm15[2,3]14848; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]14849; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14850; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14851; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload14852; AVX2-FCP-NEXT: # xmm5 = xmm0[2],mem[2],xmm0[3],mem[3]14853; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill14854; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload14855; AVX2-FCP-NEXT: # xmm15 = mem[2,2,2,2]14856; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload14857; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm15 = xmm0[0,1,2],xmm15[3]14858; AVX2-FCP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm1514859; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1414860; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]14861; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14862; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm15 # 16-byte Folded Reload14863; AVX2-FCP-NEXT: # xmm15 = xmm5[2],mem[2],xmm5[3],mem[3]14864; AVX2-FCP-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload14865; AVX2-FCP-NEXT: # xmm13 = mem[2,2,2,2]14866; AVX2-FCP-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload14867; AVX2-FCP-NEXT: # xmm13 = mem[0,1,2],xmm13[3]14868; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm13 = xmm15[0,1],xmm13[2,3]14869; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm14[4,5,6,7]14870; AVX2-FCP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14871; AVX2-FCP-NEXT: vmovaps (%rsp), %xmm5 # 16-byte Reload14872; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm13 # 16-byte Folded Reload14873; AVX2-FCP-NEXT: # xmm13 = xmm5[2],mem[2],xmm5[3],mem[3]14874; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14875; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm13 = xmm5[1],xmm13[1]14876; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14877; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm14 # 16-byte Folded Reload14878; AVX2-FCP-NEXT: # xmm14 = xmm5[2],mem[2],xmm5[3],mem[3]14879; AVX2-FCP-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm1414880; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload14881; AVX2-FCP-NEXT: # xmm12 = mem[2,3,2,3]14882; AVX2-FCP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1214883; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm14[6,7]14884; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]14885; AVX2-FCP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14886; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14887; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm12 = xmm5[2],xmm9[2],xmm5[3],xmm9[3]14888; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14889; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm11 = xmm5[1],xmm12[1]14890; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14891; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm12 # 16-byte Folded Reload14892; AVX2-FCP-NEXT: # xmm12 = xmm5[2],mem[2],xmm5[3],mem[3]14893; AVX2-FCP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1214894; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload14895; AVX2-FCP-NEXT: # xmm13 = mem[2,3,2,3]14896; AVX2-FCP-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm1314897; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]14898; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]14899; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14900; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm11 # 16-byte Folded Reload14901; AVX2-FCP-NEXT: # xmm11 = xmm7[2],mem[2],xmm7[3],mem[3]14902; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm10[1],xmm11[1]14903; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14904; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm11 # 16-byte Folded Reload14905; AVX2-FCP-NEXT: # xmm11 = xmm5[2],mem[2],xmm5[3],mem[3]14906; AVX2-FCP-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm1114907; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload14908; AVX2-FCP-NEXT: # xmm10 = mem[2,3,2,3]14909; AVX2-FCP-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm1014910; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm11[6,7]14911; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]14912; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14913; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14914; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm9 # 16-byte Folded Reload14915; AVX2-FCP-NEXT: # xmm9 = xmm5[2],mem[2],xmm5[3],mem[3]14916; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm8[1],xmm9[1]14917; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14918; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm9 # 16-byte Folded Reload14919; AVX2-FCP-NEXT: # xmm9 = xmm5[2],mem[2],xmm5[3],mem[3]14920; AVX2-FCP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm914921; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload14922; AVX2-FCP-NEXT: # xmm8 = mem[2,3,2,3]14923; AVX2-FCP-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm814924; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm9[6,7]14925; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]14926; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14927; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload14928; AVX2-FCP-NEXT: # xmm7 = xmm6[2],mem[2],xmm6[3],mem[3]14929; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload14930; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm7[1]14931; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload14932; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload14933; AVX2-FCP-NEXT: # xmm7 = xmm6[2],mem[2],xmm6[3],mem[3]14934; AVX2-FCP-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm714935; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload14936; AVX2-FCP-NEXT: # xmm6 = mem[2,3,2,3]14937; AVX2-FCP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm614938; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]14939; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7]14940; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14941; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]14942; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload14943; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm5[1]14944; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload14945; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm5 # 16-byte Folded Reload14946; AVX2-FCP-NEXT: # xmm5 = xmm4[2],mem[2],xmm4[3],mem[3]14947; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm514948; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload14949; AVX2-FCP-NEXT: # xmm4 = mem[2,3,2,3]14950; AVX2-FCP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm414951; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]14952; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]14953; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14954; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm3 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]14955; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14956; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]14957; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload14958; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload14959; AVX2-FCP-NEXT: # xmm3 = xmm2[2],mem[2],xmm2[3],mem[3]14960; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm314961; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload14962; AVX2-FCP-NEXT: # xmm2 = mem[2,3,2,3]14963; AVX2-FCP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm214964; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]14965; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]14966; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14967; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload14968; AVX2-FCP-NEXT: # xmm1 = xmm0[2],mem[2],xmm0[3],mem[3]14969; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm114970; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload14971; AVX2-FCP-NEXT: # xmm0 = mem[2,3,2,3]14972; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm014973; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]14974; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload14975; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload14976; AVX2-FCP-NEXT: # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]14977; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm15[1],xmm1[1]14978; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]14979; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14980; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm014981; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14982; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm114983; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14984; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]14985; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm014986; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm114987; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14988; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm214989; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14990; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]14991; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14992; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]14993; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]14994; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm214995; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14996; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm314997; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill14998; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm1514999; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm115000; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15001; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[4],ymm15[4],ymm1[5],ymm15[5]15002; AVX2-FCP-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15003; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]15004; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15005; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]15006; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15007; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15008; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm015009; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15010; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm115011; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15012; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]15013; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm015014; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm115015; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15016; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm215017; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15018; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]15019; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15020; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]15021; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]15022; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm215023; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15024; AVX2-FCP-NEXT: vmovaps 448(%rdi), %ymm315025; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15026; AVX2-FCP-NEXT: vmovaps 416(%rdi), %ymm415027; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15028; AVX2-FCP-NEXT: vmovaps 384(%rdi), %ymm115029; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15030; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[4],ymm4[4],ymm1[5],ymm4[5]15031; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm12 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]15032; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm12[0],ymm1[2],ymm12[2]15033; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15034; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15035; AVX2-FCP-NEXT: vmovaps 544(%rdi), %ymm015036; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15037; AVX2-FCP-NEXT: vmovaps 512(%rdi), %ymm115038; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15039; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]15040; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm015041; AVX2-FCP-NEXT: vmovaps 608(%rdi), %ymm115042; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15043; AVX2-FCP-NEXT: vmovaps 576(%rdi), %ymm215044; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15045; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]15046; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15047; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]15048; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]15049; AVX2-FCP-NEXT: vmovaps 736(%rdi), %ymm215050; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15051; AVX2-FCP-NEXT: vmovaps 704(%rdi), %ymm315052; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15053; AVX2-FCP-NEXT: vmovaps 672(%rdi), %ymm415054; AVX2-FCP-NEXT: vmovaps 640(%rdi), %ymm115055; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15056; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[4],ymm4[4],ymm1[5],ymm4[5]15057; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15058; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]15059; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[2],ymm8[2]15060; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15061; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15062; AVX2-FCP-NEXT: vmovaps 800(%rdi), %ymm015063; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15064; AVX2-FCP-NEXT: vmovaps 768(%rdi), %ymm115065; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15066; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]15067; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm015068; AVX2-FCP-NEXT: vmovaps 864(%rdi), %ymm115069; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15070; AVX2-FCP-NEXT: vmovaps 832(%rdi), %ymm215071; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15072; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]15073; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15074; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]15075; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]15076; AVX2-FCP-NEXT: vmovaps 992(%rdi), %ymm215077; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15078; AVX2-FCP-NEXT: vmovaps 960(%rdi), %ymm515079; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15080; AVX2-FCP-NEXT: vmovaps 928(%rdi), %ymm315081; AVX2-FCP-NEXT: vmovaps 896(%rdi), %ymm115082; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15083; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]15084; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15085; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm5 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[4],ymm2[4],ymm5[5],ymm2[5]15086; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm5[0],ymm1[2],ymm5[2]15087; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15088; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15089; AVX2-FCP-NEXT: vmovaps 1056(%rdi), %ymm015090; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15091; AVX2-FCP-NEXT: vmovaps 1024(%rdi), %ymm115092; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15093; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]15094; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm015095; AVX2-FCP-NEXT: vmovaps 1120(%rdi), %ymm115096; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15097; AVX2-FCP-NEXT: vmovaps 1088(%rdi), %ymm215098; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15099; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5]15100; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15101; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2]15102; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]15103; AVX2-FCP-NEXT: vmovaps 1248(%rdi), %ymm115104; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15105; AVX2-FCP-NEXT: vmovaps 1216(%rdi), %ymm715106; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15107; AVX2-FCP-NEXT: vmovaps 1184(%rdi), %ymm215108; AVX2-FCP-NEXT: vmovaps 1152(%rdi), %ymm615109; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15110; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[4],ymm2[4],ymm6[5],ymm2[5]15111; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15112; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm1 = ymm7[0],ymm1[0],ymm7[1],ymm1[1],ymm7[4],ymm1[4],ymm7[5],ymm1[5]15113; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]15114; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]15115; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15116; AVX2-FCP-NEXT: vmovaps 1312(%rdi), %ymm015117; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15118; AVX2-FCP-NEXT: vmovaps 1280(%rdi), %ymm615119; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15120; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]15121; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm615122; AVX2-FCP-NEXT: vmovaps 1376(%rdi), %ymm015123; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15124; AVX2-FCP-NEXT: vmovaps 1344(%rdi), %ymm715125; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15126; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[4],ymm0[4],ymm7[5],ymm0[5]15127; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15128; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[2,2,2,2]15129; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]15130; AVX2-FCP-NEXT: vmovaps 1504(%rdi), %ymm615131; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15132; AVX2-FCP-NEXT: vmovaps 1472(%rdi), %ymm1015133; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15134; AVX2-FCP-NEXT: vmovaps 1440(%rdi), %ymm015135; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill15136; AVX2-FCP-NEXT: vmovaps 1408(%rdi), %ymm915137; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15138; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[1],ymm0[1],ymm9[4],ymm0[4],ymm9[5],ymm0[5]15139; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm0 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[4],ymm6[4],ymm10[5],ymm6[5]15140; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[2],ymm0[2]15141; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm9[4,5,6,7]15142; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15143; AVX2-FCP-NEXT: vmovaps 1568(%rdi), %ymm615144; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15145; AVX2-FCP-NEXT: vmovaps 1536(%rdi), %ymm715146; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15147; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm7 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]15148; AVX2-FCP-NEXT: vextractf128 $1, %ymm7, %xmm915149; AVX2-FCP-NEXT: vmovaps 1632(%rdi), %ymm615150; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15151; AVX2-FCP-NEXT: vmovaps 1600(%rdi), %ymm715152; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15153; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]15154; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15155; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm10 = ymm6[2,2,2,2]15156; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm10 = xmm9[0,1],xmm10[2,3]15157; AVX2-FCP-NEXT: vmovaps 1760(%rdi), %ymm915158; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15159; AVX2-FCP-NEXT: vmovaps 1728(%rdi), %ymm615160; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15161; AVX2-FCP-NEXT: vmovaps 1696(%rdi), %ymm715162; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15163; AVX2-FCP-NEXT: vmovaps 1664(%rdi), %ymm1115164; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15165; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm11 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[4],ymm7[4],ymm11[5],ymm7[5]15166; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[4],ymm9[4],ymm6[5],ymm9[5]15167; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15168; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm6[0],ymm11[2],ymm6[2]15169; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]15170; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15171; AVX2-FCP-NEXT: vmovaps 1824(%rdi), %ymm615172; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15173; AVX2-FCP-NEXT: vmovaps 1792(%rdi), %ymm715174; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15175; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm10 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]15176; AVX2-FCP-NEXT: vextractf128 $1, %ymm10, %xmm1115177; AVX2-FCP-NEXT: vmovaps 1888(%rdi), %ymm615178; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15179; AVX2-FCP-NEXT: vmovaps 1856(%rdi), %ymm715180; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15181; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]15182; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15183; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm13 = ymm6[2,2,2,2]15184; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm13 = xmm11[0,1],xmm13[2,3]15185; AVX2-FCP-NEXT: vmovaps 2016(%rdi), %ymm1115186; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15187; AVX2-FCP-NEXT: vmovaps 1984(%rdi), %ymm615188; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15189; AVX2-FCP-NEXT: vmovaps 1952(%rdi), %ymm715190; AVX2-FCP-NEXT: vmovaps 1920(%rdi), %ymm915191; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15192; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm14 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]15193; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15194; AVX2-FCP-NEXT: vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm11[0],ymm6[1],ymm11[1],ymm6[4],ymm11[4],ymm6[5],ymm11[5]15195; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm14[0],ymm11[0],ymm14[2],ymm11[2]15196; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]15197; AVX2-FCP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15198; AVX2-FCP-NEXT: vbroadcastss 148(%rdi), %ymm1315199; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm15[5],ymm13[6,7]15200; AVX2-FCP-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload15201; AVX2-FCP-NEXT: # ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]15202; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload15203; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm1415204; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload15205; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm15 = ymm9[1,1,1,1,5,5,5,5]15206; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload15207; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm6[1],ymm15[2,3,4],ymm6[5],ymm15[6,7]15208; AVX2-FCP-NEXT: vextractf128 $1, %ymm15, %xmm1515209; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm14 = xmm15[0,1],xmm14[2,3]15210; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]15211; AVX2-FCP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15212; AVX2-FCP-NEXT: vbroadcastss 404(%rdi), %ymm1315213; AVX2-FCP-NEXT: vblendps $32, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload15214; AVX2-FCP-NEXT: # ymm13 = ymm13[0,1,2,3,4],mem[5],ymm13[6,7]15215; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]15216; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload15217; AVX2-FCP-NEXT: vextractf128 $1, %ymm10, %xmm1315218; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload15219; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm14 = ymm15[1,1,1,1,5,5,5,5]15220; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload15221; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm10[1],ymm14[2,3,4],ymm10[5],ymm14[6,7]15222; AVX2-FCP-NEXT: vextractf128 $1, %ymm14, %xmm1415223; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]15224; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]15225; AVX2-FCP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15226; AVX2-FCP-NEXT: vbroadcastss 660(%rdi), %ymm1215227; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm4[5],ymm12[6,7]15228; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm8[6,7]15229; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload15230; AVX2-FCP-NEXT: vextractf128 $1, %ymm8, %xmm815231; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload15232; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm12 = ymm13[1,1,1,1,5,5,5,5]15233; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload15234; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3,4],ymm14[5],ymm12[6,7]15235; AVX2-FCP-NEXT: vextractf128 $1, %ymm12, %xmm1215236; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm8 = xmm12[0,1],xmm8[2,3]15237; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]15238; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15239; AVX2-FCP-NEXT: vbroadcastss 916(%rdi), %ymm415240; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]15241; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm5[6,7]15242; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload15243; AVX2-FCP-NEXT: vextractf128 $1, %ymm4, %xmm415244; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload15245; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm8[1,1,1,1,5,5,5,5]15246; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload15247; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm12[1],ymm5[2,3,4],ymm12[5],ymm5[6,7]15248; AVX2-FCP-NEXT: vextractf128 $1, %ymm5, %xmm515249; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]15250; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]15251; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15252; AVX2-FCP-NEXT: vbroadcastss 1172(%rdi), %ymm315253; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]15254; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]15255; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15256; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm215257; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload15258; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm4[1,1,1,1,5,5,5,5]15259; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload15260; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4],ymm5[5],ymm3[6,7]15261; AVX2-FCP-NEXT: vextractf128 $1, %ymm3, %xmm315262; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]15263; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]15264; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15265; AVX2-FCP-NEXT: vbroadcastss 1428(%rdi), %ymm115266; AVX2-FCP-NEXT: vblendps $32, (%rsp), %ymm1, %ymm1 # 32-byte Folded Reload15267; AVX2-FCP-NEXT: # ymm1 = ymm1[0,1,2,3,4],mem[5],ymm1[6,7]15268; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm0[6,7]15269; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload15270; AVX2-FCP-NEXT: vextractf128 $1, %ymm0, %xmm015271; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload15272; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm3[1,1,1,1,5,5,5,5]15273; AVX2-FCP-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload15274; AVX2-FCP-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]15275; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm215276; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]15277; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15278; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15279; AVX2-FCP-NEXT: vbroadcastss 1684(%rdi), %ymm015280; AVX2-FCP-NEXT: vblendps $32, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15281; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1,2,3,4],mem[5],ymm0[6,7]15282; AVX2-FCP-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15283; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]15284; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15285; AVX2-FCP-NEXT: vextractf128 $1, %ymm1, %xmm115286; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload15287; AVX2-FCP-NEXT: # ymm2 = mem[1,1,1,1,5,5,5,5]15288; AVX2-FCP-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload15289; AVX2-FCP-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]15290; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm215291; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]15292; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15293; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15294; AVX2-FCP-NEXT: vbroadcastss 1940(%rdi), %ymm015295; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm7[5],ymm0[6,7]15296; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm11[6,7]15297; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15298; AVX2-FCP-NEXT: vextractf128 $1, %ymm1, %xmm115299; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload15300; AVX2-FCP-NEXT: # ymm2 = mem[1,1,1,1,5,5,5,5]15301; AVX2-FCP-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload15302; AVX2-FCP-NEXT: # ymm2 = ymm2[0],mem[1],ymm2[2,3,4],mem[5],ymm2[6,7]15303; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm215304; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]15305; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15306; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15307; AVX2-FCP-NEXT: vbroadcastss 248(%rdi), %ymm015308; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15309; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15310; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15311; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload15312; AVX2-FCP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15313; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15314; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]15315; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15316; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15317; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload15318; AVX2-FCP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15319; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15320; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm115321; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm6[2,2,2,2]15322; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15323; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]15324; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15325; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15326; AVX2-FCP-NEXT: vbroadcastss 504(%rdi), %ymm015327; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15328; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15329; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15330; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload15331; AVX2-FCP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15332; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15333; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[6],ymm10[6],ymm15[7],ymm10[7]15334; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15335; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15336; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload15337; AVX2-FCP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15338; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15339; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm115340; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm7[2,2,2,2]15341; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15342; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]15343; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15344; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15345; AVX2-FCP-NEXT: vbroadcastss 760(%rdi), %ymm015346; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15347; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15348; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15349; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload15350; AVX2-FCP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15351; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15352; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[6],ymm14[6],ymm13[7],ymm14[7]15353; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15354; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15355; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload15356; AVX2-FCP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15357; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15358; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm115359; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm7[2,2,2,2]15360; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15361; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]15362; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15363; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15364; AVX2-FCP-NEXT: vbroadcastss 1016(%rdi), %ymm015365; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15366; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15367; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15368; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload15369; AVX2-FCP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15370; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15371; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm2 = ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[6],ymm12[6],ymm8[7],ymm12[7]15372; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15373; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15374; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload15375; AVX2-FCP-NEXT: # ymm14 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15376; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm115377; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm14[2,2,2,2]15378; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15379; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]15380; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15381; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15382; AVX2-FCP-NEXT: vbroadcastss 1272(%rdi), %ymm015383; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15384; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15385; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15386; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm12 # 32-byte Folded Reload15387; AVX2-FCP-NEXT: # ymm12 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15388; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm13 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]15389; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15390; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm11 # 32-byte Folded Reload15391; AVX2-FCP-NEXT: # ymm11 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15392; AVX2-FCP-NEXT: vextractf128 $1, %ymm13, %xmm115393; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm11[2,2,2,2]15394; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15395; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],ymm0[6,7]15396; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15397; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15398; AVX2-FCP-NEXT: vbroadcastss 1528(%rdi), %ymm015399; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15400; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15401; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15402; AVX2-FCP-NEXT: vunpckhps (%rsp), %ymm1, %ymm10 # 32-byte Folded Reload15403; AVX2-FCP-NEXT: # ymm10 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15404; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm9 # 32-byte Folded Reload15405; AVX2-FCP-NEXT: # ymm9 = ymm3[2],mem[2],ymm3[3],mem[3],ymm3[6],mem[6],ymm3[7],mem[7]15406; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15407; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload15408; AVX2-FCP-NEXT: # ymm8 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15409; AVX2-FCP-NEXT: vextractf128 $1, %ymm9, %xmm115410; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm8[2,2,2,2]15411; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15412; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],ymm0[6,7]15413; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15414; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15415; AVX2-FCP-NEXT: vbroadcastss 1784(%rdi), %ymm015416; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15417; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5,6],ymm0[7]15418; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15419; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload15420; AVX2-FCP-NEXT: # ymm6 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15421; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15422; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload15423; AVX2-FCP-NEXT: # ymm7 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15424; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15425; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload15426; AVX2-FCP-NEXT: # ymm5 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7]15427; AVX2-FCP-NEXT: vextractf128 $1, %ymm7, %xmm115428; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm5[2,2,2,2]15429; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3]15430; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]15431; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15432; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15433; AVX2-FCP-NEXT: vbroadcastss 2040(%rdi), %ymm015434; AVX2-FCP-NEXT: vblendps $127, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload15435; AVX2-FCP-NEXT: # ymm1 = mem[0,1,2,3,4,5,6],ymm0[7]15436; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload15437; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload15438; AVX2-FCP-NEXT: # ymm4 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15439; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload15440; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload15441; AVX2-FCP-NEXT: # ymm3 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15442; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload15443; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload15444; AVX2-FCP-NEXT: # ymm2 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15445; AVX2-FCP-NEXT: vextractf128 $1, %ymm3, %xmm015446; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm15 = ymm2[2,2,2,2]15447; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm15[2,3]15448; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]15449; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15450; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill15451; AVX2-FCP-NEXT: vbroadcastss 220(%rdi), %ymm015452; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15453; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15454; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15455; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]15456; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15457; AVX2-FCP-NEXT: vextractf128 $1, %ymm1, %xmm115458; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload15459; AVX2-FCP-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]15460; AVX2-FCP-NEXT: vextractf128 $1, %ymm15, %xmm1515461; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]15462; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15463; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15464; AVX2-FCP-NEXT: vbroadcastss 476(%rdi), %ymm015465; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15466; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15467; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15468; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]15469; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15470; AVX2-FCP-NEXT: vextractf128 $1, %ymm1, %xmm115471; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload15472; AVX2-FCP-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]15473; AVX2-FCP-NEXT: vextractf128 $1, %ymm15, %xmm1515474; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]15475; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]15476; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill15477; AVX2-FCP-NEXT: vbroadcastss 732(%rdi), %ymm015478; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15479; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15480; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15481; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]15482; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15483; AVX2-FCP-NEXT: vextractf128 $1, %ymm1, %xmm115484; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload15485; AVX2-FCP-NEXT: # ymm15 = mem[2,3,2,3,6,7,6,7]15486; AVX2-FCP-NEXT: vextractf128 $1, %ymm15, %xmm1515487; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]15488; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm0[4,5,6,7]15489; AVX2-FCP-NEXT: vbroadcastss 988(%rdi), %ymm015490; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15491; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15492; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload15493; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]15494; AVX2-FCP-NEXT: vextractf128 $1, %ymm14, %xmm115495; AVX2-FCP-NEXT: vpermilps $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload15496; AVX2-FCP-NEXT: # ymm14 = mem[2,3,2,3,6,7,6,7]15497; AVX2-FCP-NEXT: vextractf128 $1, %ymm14, %xmm1415498; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]15499; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm0[4,5,6,7]15500; AVX2-FCP-NEXT: vbroadcastss 1244(%rdi), %ymm015501; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15502; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15503; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm12[1],ymm0[1],ymm12[3],ymm0[3]15504; AVX2-FCP-NEXT: vextractf128 $1, %ymm11, %xmm115505; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm11 = ymm13[2,3,2,3,6,7,6,7]15506; AVX2-FCP-NEXT: vextractf128 $1, %ymm11, %xmm1115507; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm11[0,1],xmm1[2,3]15508; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]15509; AVX2-FCP-NEXT: vbroadcastss 1500(%rdi), %ymm015510; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15511; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15512; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm0[1],ymm10[3],ymm0[3]15513; AVX2-FCP-NEXT: vextractf128 $1, %ymm8, %xmm815514; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]15515; AVX2-FCP-NEXT: vextractf128 $1, %ymm9, %xmm915516; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]15517; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm0[4,5,6,7]15518; AVX2-FCP-NEXT: vbroadcastss 1756(%rdi), %ymm015519; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15520; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15521; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm0[1],ymm6[3],ymm0[3]15522; AVX2-FCP-NEXT: vextractf128 $1, %ymm5, %xmm515523; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm7[2,3,2,3,6,7,6,7]15524; AVX2-FCP-NEXT: vextractf128 $1, %ymm6, %xmm615525; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]15526; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]15527; AVX2-FCP-NEXT: vbroadcastss 2012(%rdi), %ymm015528; AVX2-FCP-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload15529; AVX2-FCP-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7]15530; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm0[1],ymm4[3],ymm0[3]15531; AVX2-FCP-NEXT: vextractf128 $1, %ymm2, %xmm215532; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]15533; AVX2-FCP-NEXT: vextractf128 $1, %ymm3, %xmm315534; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]15535; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]15536; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15537; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%rsi)15538; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15539; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%rsi)15540; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15541; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%rsi)15542; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15543; AVX2-FCP-NEXT: vmovaps %ymm2, (%rsi)15544; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15545; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%rsi)15546; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15547; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%rsi)15548; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15549; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%rsi)15550; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15551; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rsi)15552; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15553; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%rdx)15554; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15555; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%rdx)15556; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15557; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%rdx)15558; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15559; AVX2-FCP-NEXT: vmovaps %ymm2, (%rdx)15560; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15561; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%rdx)15562; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15563; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%rdx)15564; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15565; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%rdx)15566; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15567; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rdx)15568; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15569; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%rcx)15570; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15571; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%rcx)15572; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15573; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%rcx)15574; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15575; AVX2-FCP-NEXT: vmovaps %ymm2, (%rcx)15576; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15577; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%rcx)15578; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15579; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%rcx)15580; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15581; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%rcx)15582; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15583; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rcx)15584; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15585; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%r8)15586; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15587; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%r8)15588; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15589; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%r8)15590; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15591; AVX2-FCP-NEXT: vmovaps %ymm2, (%r8)15592; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15593; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%r8)15594; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15595; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%r8)15596; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15597; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%r8)15598; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15599; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%r8)15600; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15601; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%r9)15602; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15603; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%r9)15604; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15605; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%r9)15606; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15607; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%r9)15608; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15609; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%r9)15610; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15611; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%r9)15612; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15613; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%r9)15614; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15615; AVX2-FCP-NEXT: vmovaps %ymm2, (%r9)15616; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax15617; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15618; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%rax)15619; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15620; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%rax)15621; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15622; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%rax)15623; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15624; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%rax)15625; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15626; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%rax)15627; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15628; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%rax)15629; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15630; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rax)15631; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15632; AVX2-FCP-NEXT: vmovaps %ymm2, (%rax)15633; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax15634; AVX2-FCP-NEXT: vmovups (%rsp), %ymm2 # 32-byte Reload15635; AVX2-FCP-NEXT: vmovaps %ymm2, 224(%rax)15636; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15637; AVX2-FCP-NEXT: vmovaps %ymm2, 192(%rax)15638; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15639; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%rax)15640; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15641; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%rax)15642; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15643; AVX2-FCP-NEXT: vmovaps %ymm2, 96(%rax)15644; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15645; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%rax)15646; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15647; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rax)15648; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload15649; AVX2-FCP-NEXT: vmovaps %ymm2, (%rax)15650; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax15651; AVX2-FCP-NEXT: vmovaps %ymm0, 224(%rax)15652; AVX2-FCP-NEXT: vmovaps %ymm5, 192(%rax)15653; AVX2-FCP-NEXT: vmovaps %ymm8, 160(%rax)15654; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rax)15655; AVX2-FCP-NEXT: vmovaps %ymm14, 96(%rax)15656; AVX2-FCP-NEXT: vmovaps %ymm15, 64(%rax)15657; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload15658; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rax)15659; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload15660; AVX2-FCP-NEXT: vmovaps %ymm0, (%rax)15661; AVX2-FCP-NEXT: addq $3528, %rsp # imm = 0xDC815662; AVX2-FCP-NEXT: vzeroupper15663; AVX2-FCP-NEXT: retq15664;15665; AVX512-LABEL: load_i32_stride8_vf64:15666; AVX512: # %bb.0:15667; AVX512-NEXT: subq $3144, %rsp # imm = 0xC4815668; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm1115669; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15670; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm1815671; AVX512-NEXT: vmovdqa64 1600(%rdi), %zmm3115672; AVX512-NEXT: vmovaps 1536(%rdi), %zmm015673; AVX512-NEXT: vmovups %zmm0, (%rsp) # 64-byte Spill15674; AVX512-NEXT: vmovdqa64 1728(%rdi), %zmm2415675; AVX512-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15676; AVX512-NEXT: vmovaps 1664(%rdi), %zmm015677; AVX512-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15678; AVX512-NEXT: vmovdqa64 1856(%rdi), %zmm2115679; AVX512-NEXT: vmovdqa64 1792(%rdi), %zmm2615680; AVX512-NEXT: vmovdqa64 1984(%rdi), %zmm2215681; AVX512-NEXT: vmovdqa64 1920(%rdi), %zmm515682; AVX512-NEXT: vmovdqa64 1088(%rdi), %zmm1315683; AVX512-NEXT: vmovdqa64 1024(%rdi), %zmm315684; AVX512-NEXT: vmovdqa64 1216(%rdi), %zmm3015685; AVX512-NEXT: vmovdqa64 1152(%rdi), %zmm215686; AVX512-NEXT: vmovdqa64 1344(%rdi), %zmm2915687; AVX512-NEXT: vmovdqa64 1280(%rdi), %zmm2715688; AVX512-NEXT: vmovdqa64 1472(%rdi), %zmm2015689; AVX512-NEXT: vmovdqa64 1408(%rdi), %zmm1015690; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm2515691; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm715692; AVX512-NEXT: vmovdqa64 704(%rdi), %zmm915693; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15694; AVX512-NEXT: vmovdqa64 640(%rdi), %zmm1215695; AVX512-NEXT: vmovdqa64 832(%rdi), %zmm615696; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15697; AVX512-NEXT: vmovdqa64 768(%rdi), %zmm2815698; AVX512-NEXT: vmovdqa64 960(%rdi), %zmm2315699; AVX512-NEXT: vmovdqa64 896(%rdi), %zmm415700; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]15701; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15702; AVX512-NEXT: vmovdqa64 %zmm4, %zmm115703; AVX512-NEXT: vmovdqa64 %zmm4, %zmm1615704; AVX512-NEXT: vpermt2d %zmm23, %zmm0, %zmm115705; AVX512-NEXT: vmovdqa64 %zmm28, %zmm415706; AVX512-NEXT: vpermt2d %zmm6, %zmm0, %zmm415707; AVX512-NEXT: movb $-64, %al15708; AVX512-NEXT: kmovw %eax, %k115709; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15710; AVX512-NEXT: vmovdqa64 %zmm12, %zmm115711; AVX512-NEXT: vmovdqa64 %zmm12, %zmm1515712; AVX512-NEXT: vpermt2d %zmm9, %zmm0, %zmm115713; AVX512-NEXT: vmovdqa64 %zmm7, %zmm1215714; AVX512-NEXT: vmovdqa64 %zmm7, %zmm915715; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm1215716; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15717; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115718; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15719; AVX512-NEXT: vmovdqa64 %zmm10, %zmm115720; AVX512-NEXT: vmovdqa64 %zmm10, %zmm615721; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15722; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm115723; AVX512-NEXT: vmovdqa64 %zmm27, %zmm415724; AVX512-NEXT: vpermt2d %zmm29, %zmm0, %zmm415725; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15726; AVX512-NEXT: vmovdqa64 %zmm2, %zmm115727; AVX512-NEXT: vmovdqa64 %zmm2, %zmm815728; AVX512-NEXT: vpermt2d %zmm30, %zmm0, %zmm115729; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1015730; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1215731; AVX512-NEXT: vpermt2d %zmm13, %zmm0, %zmm1215732; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15733; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115734; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15735; AVX512-NEXT: vmovdqa64 %zmm5, %zmm1415736; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15737; AVX512-NEXT: vmovdqa64 %zmm5, %zmm115738; AVX512-NEXT: vpermt2d %zmm22, %zmm0, %zmm115739; AVX512-NEXT: vmovdqa64 %zmm22, %zmm1715740; AVX512-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15741; AVX512-NEXT: vmovdqa64 %zmm26, %zmm415742; AVX512-NEXT: vpermt2d %zmm21, %zmm0, %zmm415743; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15744; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload15745; AVX512-NEXT: vmovdqa64 %zmm3, %zmm115746; AVX512-NEXT: vpermt2d %zmm24, %zmm0, %zmm115747; AVX512-NEXT: vmovdqu64 (%rsp), %zmm2 # 64-byte Reload15748; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1215749; AVX512-NEXT: vmovdqa64 %zmm31, %zmm2415750; AVX512-NEXT: vpermt2d %zmm31, %zmm0, %zmm1215751; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15752; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm515753; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15754; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115755; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15756; AVX512-NEXT: vmovdqa64 %zmm5, %zmm115757; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm115758; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm2215759; AVX512-NEXT: vmovdqa64 %zmm22, %zmm415760; AVX512-NEXT: vpermt2d %zmm11, %zmm0, %zmm415761; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15762; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm115763; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15764; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm515765; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15766; AVX512-NEXT: vpermt2d %zmm5, %zmm0, %zmm115767; AVX512-NEXT: vmovdqa64 (%rdi), %zmm715768; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm1915769; AVX512-NEXT: vpermi2d %zmm19, %zmm7, %zmm015770; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15771; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15772; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm015773; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15774; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [1,9,17,25,1,9,17,25,1,9,17,25,1,9,17,25]15775; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15776; AVX512-NEXT: vmovdqa64 %zmm6, %zmm115777; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm115778; AVX512-NEXT: vmovdqa64 %zmm27, %zmm415779; AVX512-NEXT: vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15780; AVX512-NEXT: vpermt2d %zmm29, %zmm0, %zmm415781; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15782; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15783; AVX512-NEXT: vmovdqa64 %zmm8, %zmm115784; AVX512-NEXT: vmovdqa64 %zmm30, %zmm3115785; AVX512-NEXT: vpermt2d %zmm30, %zmm0, %zmm115786; AVX512-NEXT: vmovdqa64 %zmm10, %zmm1215787; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15788; AVX512-NEXT: vmovdqa64 %zmm13, %zmm3015789; AVX512-NEXT: vpermt2d %zmm13, %zmm0, %zmm1215790; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15791; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115792; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15793; AVX512-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15794; AVX512-NEXT: vmovdqa64 %zmm16, %zmm115795; AVX512-NEXT: vpermt2d %zmm23, %zmm0, %zmm115796; AVX512-NEXT: vmovdqa64 %zmm28, %zmm415797; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15798; AVX512-NEXT: vpermt2d %zmm5, %zmm0, %zmm415799; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15800; AVX512-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15801; AVX512-NEXT: vmovdqa64 %zmm15, %zmm115802; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload15803; AVX512-NEXT: vpermt2d %zmm13, %zmm0, %zmm115804; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15805; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1215806; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm1215807; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15808; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115809; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15810; AVX512-NEXT: vmovdqa64 %zmm14, %zmm115811; AVX512-NEXT: vpermt2d %zmm17, %zmm0, %zmm115812; AVX512-NEXT: vmovdqa64 %zmm26, %zmm415813; AVX512-NEXT: vpermt2d %zmm21, %zmm0, %zmm415814; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15815; AVX512-NEXT: vmovdqa64 %zmm3, %zmm115816; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload15817; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm115818; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1215819; AVX512-NEXT: vpermt2d %zmm24, %zmm0, %zmm1215820; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15821; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115822; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15823; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15824; AVX512-NEXT: vmovdqa64 %zmm6, %zmm115825; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm115826; AVX512-NEXT: vmovdqa64 %zmm22, %zmm415827; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15828; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm415829; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15830; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload15831; AVX512-NEXT: vmovdqa64 %zmm14, %zmm115832; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload15833; AVX512-NEXT: vpermt2d %zmm11, %zmm0, %zmm115834; AVX512-NEXT: vpermi2d %zmm19, %zmm7, %zmm015835; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15836; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm015837; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15838; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,18,26,2,10,18,26,2,10,18,26,2,10,18,26]15839; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15840; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload15841; AVX512-NEXT: vmovdqa64 %zmm17, %zmm115842; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm115843; AVX512-NEXT: vmovdqa64 %zmm27, %zmm415844; AVX512-NEXT: vpermt2d %zmm29, %zmm0, %zmm415845; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15846; AVX512-NEXT: vpermt2d %zmm31, %zmm0, %zmm815847; AVX512-NEXT: vpermt2d %zmm30, %zmm0, %zmm1015848; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm10[0,1,2,3],ymm8[4,5,6,7]15849; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115850; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15851; AVX512-NEXT: vmovdqa64 %zmm16, %zmm115852; AVX512-NEXT: vpermt2d %zmm23, %zmm0, %zmm115853; AVX512-NEXT: vmovdqa64 %zmm28, %zmm415854; AVX512-NEXT: vpermt2d %zmm5, %zmm0, %zmm415855; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15856; AVX512-NEXT: vpermt2d %zmm13, %zmm0, %zmm1515857; AVX512-NEXT: vmovdqa64 %zmm13, %zmm1615858; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm915859; AVX512-NEXT: vmovdqa64 %zmm25, %zmm2915860; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3],ymm15[4,5,6,7]15861; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115862; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15863; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15864; AVX512-NEXT: vmovdqa64 %zmm8, %zmm115865; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload15866; AVX512-NEXT: vpermt2d %zmm10, %zmm0, %zmm115867; AVX512-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15868; AVX512-NEXT: vmovdqa64 %zmm26, %zmm415869; AVX512-NEXT: vmovdqa64 %zmm21, %zmm715870; AVX512-NEXT: vpermt2d %zmm21, %zmm0, %zmm415871; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15872; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload15873; AVX512-NEXT: vmovdqa64 %zmm9, %zmm115874; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2515875; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm115876; AVX512-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload15877; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1215878; AVX512-NEXT: vmovdqa64 %zmm24, %zmm1515879; AVX512-NEXT: vpermt2d %zmm24, %zmm0, %zmm1215880; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15881; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115882; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15883; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1315884; AVX512-NEXT: vmovdqa64 %zmm6, %zmm115885; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm115886; AVX512-NEXT: vmovdqa64 %zmm22, %zmm415887; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm415888; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15889; AVX512-NEXT: vmovdqa64 %zmm14, %zmm615890; AVX512-NEXT: vmovdqa64 %zmm14, %zmm115891; AVX512-NEXT: vpermt2d %zmm11, %zmm0, %zmm115892; AVX512-NEXT: vmovdqa64 %zmm11, %zmm2415893; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload15894; AVX512-NEXT: vpermi2d %zmm19, %zmm14, %zmm015895; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15896; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm015897; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15898; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,19,27,3,11,19,27,3,11,19,27,3,11,19,27]15899; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15900; AVX512-NEXT: vmovdqa64 %zmm17, %zmm115901; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm115902; AVX512-NEXT: vmovdqa64 %zmm27, %zmm415903; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15904; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm415905; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15906; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15907; AVX512-NEXT: vmovdqa64 %zmm31, %zmm2115908; AVX512-NEXT: vpermt2d %zmm31, %zmm0, %zmm115909; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload15910; AVX512-NEXT: vpermt2d %zmm30, %zmm0, %zmm1215911; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15912; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115913; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15914; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload15915; AVX512-NEXT: vmovdqa64 %zmm11, %zmm115916; AVX512-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15917; AVX512-NEXT: vpermt2d %zmm23, %zmm0, %zmm115918; AVX512-NEXT: vmovdqa64 %zmm28, %zmm415919; AVX512-NEXT: vpermt2d %zmm5, %zmm0, %zmm415920; AVX512-NEXT: vmovdqa64 %zmm5, %zmm3115921; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15922; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15923; AVX512-NEXT: vpermt2d %zmm16, %zmm0, %zmm115924; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload15925; AVX512-NEXT: vpermt2d %zmm29, %zmm0, %zmm1215926; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15927; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115928; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15929; AVX512-NEXT: vmovdqa64 %zmm8, %zmm115930; AVX512-NEXT: vpermt2d %zmm10, %zmm0, %zmm115931; AVX512-NEXT: vmovdqa64 %zmm26, %zmm415932; AVX512-NEXT: vpermt2d %zmm7, %zmm0, %zmm415933; AVX512-NEXT: vmovdqa64 %zmm7, %zmm2615934; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15935; AVX512-NEXT: vmovdqa64 %zmm9, %zmm115936; AVX512-NEXT: vmovdqa64 %zmm25, %zmm915937; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm115938; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1215939; AVX512-NEXT: vpermt2d %zmm15, %zmm0, %zmm1215940; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15941; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115942; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15943; AVX512-NEXT: vmovdqa64 %zmm13, %zmm115944; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm115945; AVX512-NEXT: vmovdqa64 %zmm18, %zmm2515946; AVX512-NEXT: vmovdqa64 %zmm22, %zmm415947; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload15948; AVX512-NEXT: vpermt2d %zmm5, %zmm0, %zmm415949; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15950; AVX512-NEXT: vmovdqa64 %zmm6, %zmm115951; AVX512-NEXT: vmovdqa64 %zmm24, %zmm1815952; AVX512-NEXT: vpermt2d %zmm24, %zmm0, %zmm115953; AVX512-NEXT: vpermi2d %zmm19, %zmm14, %zmm015954; AVX512-NEXT: vmovdqa64 %zmm19, %zmm2415955; AVX512-NEXT: vmovdqa64 %zmm14, %zmm1315956; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]15957; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm015958; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15959; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,20,28,4,12,20,28,4,12,20,28,4,12,20,28]15960; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]15961; AVX512-NEXT: vmovdqa64 %zmm17, %zmm115962; AVX512-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15963; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm115964; AVX512-NEXT: vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15965; AVX512-NEXT: vmovdqa64 %zmm27, %zmm415966; AVX512-NEXT: vmovdqa64 %zmm2, %zmm315967; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm415968; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15969; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload15970; AVX512-NEXT: vmovdqa64 %zmm7, %zmm115971; AVX512-NEXT: vmovdqa64 %zmm21, %zmm615972; AVX512-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15973; AVX512-NEXT: vpermt2d %zmm21, %zmm0, %zmm115974; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload15975; AVX512-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15976; AVX512-NEXT: vpermt2d %zmm30, %zmm0, %zmm1215977; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15978; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115979; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15980; AVX512-NEXT: vmovdqa64 %zmm11, %zmm115981; AVX512-NEXT: vpermt2d %zmm23, %zmm0, %zmm115982; AVX512-NEXT: vmovdqa64 %zmm28, %zmm415983; AVX512-NEXT: vpermt2d %zmm31, %zmm0, %zmm415984; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}15985; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15986; AVX512-NEXT: vmovdqa64 %zmm8, %zmm115987; AVX512-NEXT: vpermt2d %zmm16, %zmm0, %zmm115988; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload15989; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1215990; AVX512-NEXT: vmovdqa64 %zmm29, %zmm1115991; AVX512-NEXT: vpermt2d %zmm29, %zmm0, %zmm1215992; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]15993; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm115994; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15995; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload15996; AVX512-NEXT: vmovdqa64 %zmm23, %zmm115997; AVX512-NEXT: vpermt2d %zmm10, %zmm0, %zmm115998; AVX512-NEXT: vmovdqa64 %zmm10, %zmm3115999; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload16000; AVX512-NEXT: vmovdqa64 %zmm10, %zmm416001; AVX512-NEXT: vmovdqa64 %zmm26, %zmm1916002; AVX512-NEXT: vpermt2d %zmm26, %zmm0, %zmm416003; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}16004; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16005; AVX512-NEXT: vpermt2d %zmm9, %zmm0, %zmm116006; AVX512-NEXT: vmovdqa64 %zmm9, %zmm2616007; AVX512-NEXT: vmovdqu64 (%rsp), %zmm12 # 64-byte Reload16008; AVX512-NEXT: vpermt2d %zmm15, %zmm0, %zmm1216009; AVX512-NEXT: vmovdqa64 %zmm15, %zmm2916010; AVX512-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16011; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]16012; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm116013; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16014; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload16015; AVX512-NEXT: vmovdqa64 %zmm21, %zmm116016; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm116017; AVX512-NEXT: vmovdqa64 %zmm22, %zmm416018; AVX512-NEXT: vmovdqa64 %zmm5, %zmm1416019; AVX512-NEXT: vpermt2d %zmm5, %zmm0, %zmm416020; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}16021; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload16022; AVX512-NEXT: vmovdqa64 %zmm9, %zmm116023; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm116024; AVX512-NEXT: vmovdqa64 %zmm13, %zmm516025; AVX512-NEXT: vmovdqa64 %zmm24, %zmm1316026; AVX512-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16027; AVX512-NEXT: vpermi2d %zmm24, %zmm5, %zmm016028; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]16029; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm016030; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16031; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,21,29,5,13,21,29,5,13,21,29,5,13,21,29]16032; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16033; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm1716034; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm2716035; AVX512-NEXT: vmovdqa64 %zmm17, %zmm27 {%k1}16036; AVX512-NEXT: vmovdqa64 %zmm7, %zmm116037; AVX512-NEXT: vpermt2d %zmm6, %zmm0, %zmm116038; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload16039; AVX512-NEXT: vpermt2d %zmm30, %zmm0, %zmm1216040; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]16041; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm27, %zmm116042; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16043; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload16044; AVX512-NEXT: vmovdqa64 %zmm24, %zmm116045; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload16046; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm116047; AVX512-NEXT: vmovdqa64 %zmm28, %zmm416048; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload16049; AVX512-NEXT: vpermt2d %zmm7, %zmm0, %zmm416050; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}16051; AVX512-NEXT: vmovdqa64 %zmm8, %zmm116052; AVX512-NEXT: vmovdqa64 %zmm8, %zmm1516053; AVX512-NEXT: vpermt2d %zmm16, %zmm0, %zmm116054; AVX512-NEXT: vmovdqa64 %zmm16, %zmm816055; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1216056; AVX512-NEXT: vpermt2d %zmm11, %zmm0, %zmm1216057; AVX512-NEXT: vmovdqa64 %zmm11, %zmm216058; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]16059; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm116060; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16061; AVX512-NEXT: vmovdqa64 %zmm23, %zmm116062; AVX512-NEXT: vpermt2d %zmm31, %zmm0, %zmm116063; AVX512-NEXT: vmovdqa64 %zmm31, %zmm1616064; AVX512-NEXT: vmovdqa64 %zmm10, %zmm416065; AVX512-NEXT: vmovdqa64 %zmm10, %zmm2716066; AVX512-NEXT: vpermt2d %zmm19, %zmm0, %zmm416067; AVX512-NEXT: vmovdqa64 %zmm19, %zmm2016068; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}16069; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload16070; AVX512-NEXT: vmovdqa64 %zmm10, %zmm116071; AVX512-NEXT: vpermt2d %zmm26, %zmm0, %zmm116072; AVX512-NEXT: vmovdqa64 %zmm26, %zmm1716073; AVX512-NEXT: vmovdqu64 (%rsp), %zmm6 # 64-byte Reload16074; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1216075; AVX512-NEXT: vpermt2d %zmm29, %zmm0, %zmm1216076; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]16077; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm116078; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16079; AVX512-NEXT: vmovdqa64 %zmm21, %zmm116080; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm116081; AVX512-NEXT: vmovdqa64 %zmm22, %zmm416082; AVX512-NEXT: vpermt2d %zmm14, %zmm0, %zmm416083; AVX512-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}16084; AVX512-NEXT: vmovdqa64 %zmm9, %zmm116085; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm116086; AVX512-NEXT: vpermi2d %zmm13, %zmm5, %zmm016087; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]16088; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm4, %zmm016089; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16090; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm1 = [6,14,22,30,6,14,22,30,6,14,22,30,6,14,22,30]16091; AVX512-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16092; AVX512-NEXT: vmovdqa64 %zmm24, %zmm416093; AVX512-NEXT: vmovdqa64 %zmm24, %zmm3016094; AVX512-NEXT: vpermt2d %zmm3, %zmm1, %zmm3016095; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [7,15,23,31,7,15,23,31,7,15,23,31,7,15,23,31]16096; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16097; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm416098; AVX512-NEXT: vmovdqa64 %zmm4, %zmm1916099; AVX512-NEXT: vmovdqa64 %zmm28, %zmm1116100; AVX512-NEXT: vpermt2d %zmm7, %zmm1, %zmm1116101; AVX512-NEXT: vpermt2d %zmm7, %zmm0, %zmm2816102; AVX512-NEXT: vmovdqa64 %zmm15, %zmm916103; AVX512-NEXT: vpermt2d %zmm8, %zmm1, %zmm916104; AVX512-NEXT: vpermt2d %zmm8, %zmm0, %zmm1516105; AVX512-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16106; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload16107; AVX512-NEXT: vmovdqa64 %zmm3, %zmm816108; AVX512-NEXT: vpermt2d %zmm2, %zmm1, %zmm816109; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm316110; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16111; AVX512-NEXT: vmovdqa64 %zmm23, %zmm3116112; AVX512-NEXT: vpermt2d %zmm16, %zmm1, %zmm3116113; AVX512-NEXT: vpermt2d %zmm16, %zmm0, %zmm2316114; AVX512-NEXT: vmovdqa64 %zmm23, %zmm1616115; AVX512-NEXT: vmovdqa64 %zmm27, %zmm2916116; AVX512-NEXT: vpermt2d %zmm20, %zmm1, %zmm2916117; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm2716118; AVX512-NEXT: vmovdqa64 %zmm27, %zmm2616119; AVX512-NEXT: vmovdqa64 %zmm10, %zmm216120; AVX512-NEXT: vmovdqa64 %zmm10, %zmm516121; AVX512-NEXT: vpermt2d %zmm17, %zmm1, %zmm516122; AVX512-NEXT: vpermt2d %zmm17, %zmm0, %zmm216123; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16124; AVX512-NEXT: vmovdqa64 %zmm6, %zmm416125; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16126; AVX512-NEXT: vpermt2d %zmm2, %zmm1, %zmm416127; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm616128; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1516129; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload16130; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2316131; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16132; AVX512-NEXT: vpermt2d %zmm2, %zmm1, %zmm2316133; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm316134; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1716135; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload16136; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2416137; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16138; AVX512-NEXT: vpermt2d %zmm2, %zmm1, %zmm2416139; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm316140; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2716141; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload16142; AVX512-NEXT: vmovdqa64 %zmm12, %zmm1316143; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload16144; AVX512-NEXT: vpermt2d %zmm3, %zmm1, %zmm1316145; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm1216146; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload16147; AVX512-NEXT: vmovdqa64 %zmm10, %zmm316148; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16149; AVX512-NEXT: vpermt2d %zmm6, %zmm1, %zmm316150; AVX512-NEXT: vpermt2d %zmm6, %zmm0, %zmm1016151; AVX512-NEXT: vmovdqa64 %zmm21, %zmm616152; AVX512-NEXT: vmovdqa64 %zmm21, %zmm2016153; AVX512-NEXT: vpermt2d %zmm25, %zmm1, %zmm2016154; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm616155; AVX512-NEXT: vmovdqa64 %zmm6, %zmm2116156; AVX512-NEXT: vmovdqa64 %zmm22, %zmm2516157; AVX512-NEXT: vpermt2d %zmm14, %zmm1, %zmm2516158; AVX512-NEXT: vpermt2d %zmm14, %zmm0, %zmm2216159; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload16160; AVX512-NEXT: vmovdqa64 %zmm7, %zmm1416161; AVX512-NEXT: vpermt2d %zmm18, %zmm1, %zmm1416162; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm716163; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload16164; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16165; AVX512-NEXT: vpermi2d %zmm6, %zmm2, %zmm116166; AVX512-NEXT: vpermt2d %zmm6, %zmm0, %zmm216167; AVX512-NEXT: vmovdqa64 %zmm2, %zmm616168; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]16169; AVX512-NEXT: vmovdqa64 %zmm23, %zmm24 {%k1}16170; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm016171; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm9[4,5,6,7]16172; AVX512-NEXT: vmovdqa64 %zmm30, %zmm11 {%k1}16173; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm11, %zmm216174; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm5[4,5,6,7]16175; AVX512-NEXT: vmovdqa64 %zmm31, %zmm29 {%k1}16176; AVX512-NEXT: vinserti64x4 $0, %ymm3, %zmm29, %zmm316177; AVX512-NEXT: vmovdqa64 %zmm20, %zmm25 {%k1}16178; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]16179; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm25, %zmm116180; AVX512-NEXT: vmovdqa64 %zmm19, %zmm28 {%k1}16181; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload16182; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload16183; AVX512-NEXT: # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]16184; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm28, %zmm416185; AVX512-NEXT: vmovdqa64 %zmm16, %zmm26 {%k1}16186; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm5 # 32-byte Folded Reload16187; AVX512-NEXT: # ymm5 = ymm15[0,1,2,3],mem[4,5,6,7]16188; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm26, %zmm516189; AVX512-NEXT: vmovdqa64 %zmm17, %zmm27 {%k1}16190; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm12[4,5,6,7]16191; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm27, %zmm916192; AVX512-NEXT: vmovdqa64 %zmm21, %zmm22 {%k1}16193; AVX512-NEXT: vpblendd {{.*#+}} ymm7 = ymm6[0,1,2,3],ymm7[4,5,6,7]16194; AVX512-NEXT: vinserti64x4 $0, %ymm7, %zmm22, %zmm716195; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16196; AVX512-NEXT: vmovaps %zmm8, 192(%rsi)16197; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16198; AVX512-NEXT: vmovaps %zmm8, 128(%rsi)16199; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload16200; AVX512-NEXT: vmovaps %zmm8, 64(%rsi)16201; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16202; AVX512-NEXT: vmovaps %zmm6, (%rsi)16203; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16204; AVX512-NEXT: vmovaps %zmm6, 192(%rdx)16205; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16206; AVX512-NEXT: vmovaps %zmm6, (%rdx)16207; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16208; AVX512-NEXT: vmovaps %zmm6, 64(%rdx)16209; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16210; AVX512-NEXT: vmovaps %zmm6, 128(%rdx)16211; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16212; AVX512-NEXT: vmovaps %zmm6, 192(%rcx)16213; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16214; AVX512-NEXT: vmovaps %zmm6, (%rcx)16215; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16216; AVX512-NEXT: vmovaps %zmm6, 64(%rcx)16217; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16218; AVX512-NEXT: vmovaps %zmm6, 128(%rcx)16219; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16220; AVX512-NEXT: vmovaps %zmm6, 192(%r8)16221; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16222; AVX512-NEXT: vmovaps %zmm6, (%r8)16223; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16224; AVX512-NEXT: vmovaps %zmm6, 64(%r8)16225; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16226; AVX512-NEXT: vmovaps %zmm6, 128(%r8)16227; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16228; AVX512-NEXT: vmovaps %zmm6, 192(%r9)16229; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16230; AVX512-NEXT: vmovaps %zmm6, (%r9)16231; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16232; AVX512-NEXT: vmovaps %zmm6, 64(%r9)16233; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16234; AVX512-NEXT: vmovaps %zmm6, 128(%r9)16235; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax16236; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16237; AVX512-NEXT: vmovaps %zmm6, 192(%rax)16238; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16239; AVX512-NEXT: vmovaps %zmm6, (%rax)16240; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16241; AVX512-NEXT: vmovaps %zmm6, 64(%rax)16242; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16243; AVX512-NEXT: vmovaps %zmm6, 128(%rax)16244; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax16245; AVX512-NEXT: vmovdqa64 %zmm3, 192(%rax)16246; AVX512-NEXT: vmovdqa64 %zmm1, (%rax)16247; AVX512-NEXT: vmovdqa64 %zmm2, 64(%rax)16248; AVX512-NEXT: vmovdqa64 %zmm0, 128(%rax)16249; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax16250; AVX512-NEXT: vmovdqa64 %zmm9, 128(%rax)16251; AVX512-NEXT: vmovdqa64 %zmm5, 192(%rax)16252; AVX512-NEXT: vmovdqa64 %zmm7, (%rax)16253; AVX512-NEXT: vmovdqa64 %zmm4, 64(%rax)16254; AVX512-NEXT: addq $3144, %rsp # imm = 0xC4816255; AVX512-NEXT: vzeroupper16256; AVX512-NEXT: retq16257;16258; AVX512-FCP-LABEL: load_i32_stride8_vf64:16259; AVX512-FCP: # %bb.0:16260; AVX512-FCP-NEXT: subq $3144, %rsp # imm = 0xC4816261; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm1116262; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16263; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm1816264; AVX512-FCP-NEXT: vmovdqa64 1600(%rdi), %zmm3116265; AVX512-FCP-NEXT: vmovaps 1536(%rdi), %zmm016266; AVX512-FCP-NEXT: vmovups %zmm0, (%rsp) # 64-byte Spill16267; AVX512-FCP-NEXT: vmovdqa64 1728(%rdi), %zmm2416268; AVX512-FCP-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16269; AVX512-FCP-NEXT: vmovaps 1664(%rdi), %zmm016270; AVX512-FCP-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16271; AVX512-FCP-NEXT: vmovdqa64 1856(%rdi), %zmm2116272; AVX512-FCP-NEXT: vmovdqa64 1792(%rdi), %zmm2616273; AVX512-FCP-NEXT: vmovdqa64 1984(%rdi), %zmm2216274; AVX512-FCP-NEXT: vmovdqa64 1920(%rdi), %zmm516275; AVX512-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm1316276; AVX512-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm316277; AVX512-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm3016278; AVX512-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm216279; AVX512-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm2916280; AVX512-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm2716281; AVX512-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm2016282; AVX512-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm1016283; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm2516284; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm716285; AVX512-FCP-NEXT: vmovdqa64 704(%rdi), %zmm916286; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16287; AVX512-FCP-NEXT: vmovdqa64 640(%rdi), %zmm1216288; AVX512-FCP-NEXT: vmovdqa64 832(%rdi), %zmm616289; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16290; AVX512-FCP-NEXT: vmovdqa64 768(%rdi), %zmm2816291; AVX512-FCP-NEXT: vmovdqa64 960(%rdi), %zmm2316292; AVX512-FCP-NEXT: vmovdqa64 896(%rdi), %zmm416293; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24]16294; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]16295; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm116296; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm1616297; AVX512-FCP-NEXT: vpermt2d %zmm23, %zmm0, %zmm116298; AVX512-FCP-NEXT: vmovdqa64 %zmm28, %zmm416299; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm416300; AVX512-FCP-NEXT: movb $-64, %al16301; AVX512-FCP-NEXT: kmovw %eax, %k116302; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm4 {%k1}16303; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm116304; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm1516305; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm0, %zmm116306; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm1216307; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm916308; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm0, %zmm1216309; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0,1,2,3],ymm1[4,5,6,7]16310; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm116311; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16312; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm116313; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm616314; AVX512-FCP-File truncated to the first 1 MiB.