14188 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_stride6_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {19; SSE-LABEL: load_i32_stride6_vf2:20; SSE: # %bb.0:21; SSE-NEXT: movdqa (%rdi), %xmm122; SSE-NEXT: movdqa 16(%rdi), %xmm023; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]24; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,1,1]25; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm1[2,3,2,3]26; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm1[3,3,3,3]27; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]28; SSE-NEXT: movdqa 32(%rdi), %xmm229; SSE-NEXT: movq %xmm1, (%rsi)30; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]31; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]32; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax33; SSE-NEXT: movq %xmm3, (%rdx)34; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]35; SSE-NEXT: movq %xmm4, (%rcx)36; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]37; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]38; SSE-NEXT: movq %xmm5, (%r8)39; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,3,2,3]40; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,1,1]41; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]42; SSE-NEXT: movq %xmm0, (%r9)43; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[3,3,3,3]44; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]45; SSE-NEXT: movq %xmm3, (%rax)46; SSE-NEXT: retq47;48; AVX-LABEL: load_i32_stride6_vf2:49; AVX: # %bb.0:50; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax51; AVX-NEXT: vmovaps (%rdi), %xmm052; AVX-NEXT: vmovaps 16(%rdi), %xmm153; AVX-NEXT: vmovaps 32(%rdi), %xmm254; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm1[2,2,3,3]55; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm0[0],xmm3[1],xmm0[2,3]56; AVX-NEXT: vmovlps %xmm3, (%rsi)57; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm0[0,1],xmm1[2,3]58; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,3,2,3]59; AVX-NEXT: vmovlps %xmm3, (%rdx)60; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm2[0,1],xmm0[2,3]61; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,0,2,3]62; AVX-NEXT: vmovlps %xmm3, (%rcx)63; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[3,3,3,3]64; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]65; AVX-NEXT: vmovlps %xmm0, (%r8)66; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm2[2,2,3,3]67; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]68; AVX-NEXT: vmovlps %xmm0, (%r9)69; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm2[2,3]70; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3,2,3]71; AVX-NEXT: vmovlps %xmm0, (%rax)72; AVX-NEXT: retq73;74; AVX2-LABEL: load_i32_stride6_vf2:75; AVX2: # %bb.0:76; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax77; AVX2-NEXT: vmovaps 32(%rdi), %ymm078; AVX2-NEXT: vmovaps (%rdi), %xmm179; AVX2-NEXT: vmovaps 16(%rdi), %xmm280; AVX2-NEXT: vmovaps 32(%rdi), %xmm381; AVX2-NEXT: vshufps {{.*#+}} xmm4 = xmm2[2,2,3,3]82; AVX2-NEXT: vblendps {{.*#+}} xmm4 = xmm1[0],xmm4[1],xmm1[2,3]83; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]84; AVX2-NEXT: vmovlps %xmm4, (%rsi)85; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm2[2,3]86; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,3,2,3]87; AVX2-NEXT: vmovlps %xmm2, (%rdx)88; AVX2-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm1[2,3]89; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[2,0,2,3]90; AVX2-NEXT: vmovlps %xmm2, (%rcx)91; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm1[3,3,3,3]92; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3]93; AVX2-NEXT: vmovlps %xmm1, (%r8)94; AVX2-NEXT: vmovsd {{.*#+}} xmm1 = [4,2,0,0]95; AVX2-NEXT: vpermps %ymm0, %ymm1, %ymm196; AVX2-NEXT: vmovlps %xmm1, (%r9)97; AVX2-NEXT: vmovsd {{.*#+}} xmm1 = [5,3,0,0]98; AVX2-NEXT: vpermps %ymm0, %ymm1, %ymm099; AVX2-NEXT: vmovlps %xmm0, (%rax)100; AVX2-NEXT: vzeroupper101; AVX2-NEXT: retq102;103; AVX2-FP-LABEL: load_i32_stride6_vf2:104; AVX2-FP: # %bb.0:105; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax106; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm0107; AVX2-FP-NEXT: vmovaps (%rdi), %xmm1108; AVX2-FP-NEXT: vmovaps 16(%rdi), %xmm2109; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm3110; AVX2-FP-NEXT: vshufps {{.*#+}} xmm4 = xmm2[2,2,3,3]111; AVX2-FP-NEXT: vblendps {{.*#+}} xmm4 = xmm1[0],xmm4[1],xmm1[2,3]112; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]113; AVX2-FP-NEXT: vmovlps %xmm4, (%rsi)114; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm2[2,3]115; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,3,2,3]116; AVX2-FP-NEXT: vmovlps %xmm2, (%rdx)117; AVX2-FP-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm1[2,3]118; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[2,0,2,3]119; AVX2-FP-NEXT: vmovlps %xmm2, (%rcx)120; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm1[3,3,3,3]121; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3]122; AVX2-FP-NEXT: vmovlps %xmm1, (%r8)123; AVX2-FP-NEXT: vmovsd {{.*#+}} xmm1 = [4,2,0,0]124; AVX2-FP-NEXT: vpermps %ymm0, %ymm1, %ymm1125; AVX2-FP-NEXT: vmovlps %xmm1, (%r9)126; AVX2-FP-NEXT: vmovsd {{.*#+}} xmm1 = [5,3,0,0]127; AVX2-FP-NEXT: vpermps %ymm0, %ymm1, %ymm0128; AVX2-FP-NEXT: vmovlps %xmm0, (%rax)129; AVX2-FP-NEXT: vzeroupper130; AVX2-FP-NEXT: retq131;132; AVX2-FCP-LABEL: load_i32_stride6_vf2:133; AVX2-FCP: # %bb.0:134; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax135; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm0136; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm1137; AVX2-FCP-NEXT: vmovaps 16(%rdi), %xmm2138; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm3139; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm4 = xmm2[2,2,3,3]140; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm4 = xmm1[0],xmm4[1],xmm1[2,3]141; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]142; AVX2-FCP-NEXT: vmovlps %xmm4, (%rsi)143; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm2[2,3]144; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,3,2,3]145; AVX2-FCP-NEXT: vmovlps %xmm2, (%rdx)146; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm2 = xmm3[0,1],xmm1[2,3]147; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[2,0,2,3]148; AVX2-FCP-NEXT: vmovlps %xmm2, (%rcx)149; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm1[3,3,3,3]150; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3]151; AVX2-FCP-NEXT: vmovlps %xmm1, (%r8)152; AVX2-FCP-NEXT: vmovsd {{.*#+}} xmm1 = [4,2,0,0]153; AVX2-FCP-NEXT: vpermps %ymm0, %ymm1, %ymm1154; AVX2-FCP-NEXT: vmovlps %xmm1, (%r9)155; AVX2-FCP-NEXT: vmovsd {{.*#+}} xmm1 = [5,3,0,0]156; AVX2-FCP-NEXT: vpermps %ymm0, %ymm1, %ymm0157; AVX2-FCP-NEXT: vmovlps %xmm0, (%rax)158; AVX2-FCP-NEXT: vzeroupper159; AVX2-FCP-NEXT: retq160;161; AVX512-LABEL: load_i32_stride6_vf2:162; AVX512: # %bb.0:163; AVX512-NEXT: pushq %rbx164; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax165; AVX512-NEXT: vmovdqa (%rdi), %xmm0166; AVX512-NEXT: vmovaps 16(%rdi), %xmm1167; AVX512-NEXT: vmovdqa 32(%rdi), %xmm2168; AVX512-NEXT: vextractps $2, %xmm1, %r10d169; AVX512-NEXT: vextractps $3, %xmm1, %r11d170; AVX512-NEXT: vmovd %xmm2, %ebx171; AVX512-NEXT: vpinsrd $1, %r10d, %xmm0, %xmm1172; AVX512-NEXT: vpbroadcastd 8(%rdi), %xmm3173; AVX512-NEXT: vmovaps 32(%rdi), %ymm4174; AVX512-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]175; AVX512-NEXT: vmovq %xmm1, (%rsi)176; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]177; AVX512-NEXT: vpinsrd $1, %r11d, %xmm1, %xmm1178; AVX512-NEXT: vmovq %xmm1, (%rdx)179; AVX512-NEXT: vpinsrd $1, %ebx, %xmm3, %xmm1180; AVX512-NEXT: vmovq %xmm1, (%rcx)181; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]182; AVX512-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]183; AVX512-NEXT: vmovq %xmm0, (%r8)184; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]185; AVX512-NEXT: vpermps %ymm4, %ymm0, %ymm0186; AVX512-NEXT: vmovlps %xmm0, (%r9)187; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]188; AVX512-NEXT: vpermps %ymm4, %ymm0, %ymm0189; AVX512-NEXT: vmovlps %xmm0, (%rax)190; AVX512-NEXT: popq %rbx191; AVX512-NEXT: vzeroupper192; AVX512-NEXT: retq193;194; AVX512-FCP-LABEL: load_i32_stride6_vf2:195; AVX512-FCP: # %bb.0:196; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax197; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,0,0]198; AVX512-FCP-NEXT: vmovaps (%rdi), %ymm1199; AVX512-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0200; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm2201; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm3202; AVX512-FCP-NEXT: vblendps {{.*#+}} ymm4 = mem[0,1,2,3],ymm1[4,5,6,7]203; AVX512-FCP-NEXT: vmovlps %xmm0, (%rsi)204; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [1,7,0,0]205; AVX512-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0206; AVX512-FCP-NEXT: vmovlps %xmm0, (%rdx)207; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm0 = [2,4,2,4]208; AVX512-FCP-NEXT: vpermi2d %xmm3, %xmm2, %xmm0209; AVX512-FCP-NEXT: vmovq %xmm0, (%rcx)210; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [7,1,0,0]211; AVX512-FCP-NEXT: vpermi2d %xmm2, %xmm3, %xmm0212; AVX512-FCP-NEXT: vmovq %xmm0, (%r8)213; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]214; AVX512-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0215; AVX512-FCP-NEXT: vmovlps %xmm0, (%r9)216; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]217; AVX512-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0218; AVX512-FCP-NEXT: vmovlps %xmm0, (%rax)219; AVX512-FCP-NEXT: vzeroupper220; AVX512-FCP-NEXT: retq221;222; AVX512DQ-LABEL: load_i32_stride6_vf2:223; AVX512DQ: # %bb.0:224; AVX512DQ-NEXT: pushq %rbx225; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax226; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0227; AVX512DQ-NEXT: vmovaps 16(%rdi), %xmm1228; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm2229; AVX512DQ-NEXT: vextractps $2, %xmm1, %r10d230; AVX512DQ-NEXT: vextractps $3, %xmm1, %r11d231; AVX512DQ-NEXT: vmovd %xmm2, %ebx232; AVX512DQ-NEXT: vpinsrd $1, %r10d, %xmm0, %xmm1233; AVX512DQ-NEXT: vpbroadcastd 8(%rdi), %xmm3234; AVX512DQ-NEXT: vmovaps 32(%rdi), %ymm4235; AVX512DQ-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]236; AVX512DQ-NEXT: vmovq %xmm1, (%rsi)237; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]238; AVX512DQ-NEXT: vpinsrd $1, %r11d, %xmm1, %xmm1239; AVX512DQ-NEXT: vmovq %xmm1, (%rdx)240; AVX512DQ-NEXT: vpinsrd $1, %ebx, %xmm3, %xmm1241; AVX512DQ-NEXT: vmovq %xmm1, (%rcx)242; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]243; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]244; AVX512DQ-NEXT: vmovq %xmm0, (%r8)245; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]246; AVX512DQ-NEXT: vpermps %ymm4, %ymm0, %ymm0247; AVX512DQ-NEXT: vmovlps %xmm0, (%r9)248; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]249; AVX512DQ-NEXT: vpermps %ymm4, %ymm0, %ymm0250; AVX512DQ-NEXT: vmovlps %xmm0, (%rax)251; AVX512DQ-NEXT: popq %rbx252; AVX512DQ-NEXT: vzeroupper253; AVX512DQ-NEXT: retq254;255; AVX512DQ-FCP-LABEL: load_i32_stride6_vf2:256; AVX512DQ-FCP: # %bb.0:257; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax258; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,0,0]259; AVX512DQ-FCP-NEXT: vmovaps (%rdi), %ymm1260; AVX512DQ-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0261; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm2262; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm3263; AVX512DQ-FCP-NEXT: vblendps {{.*#+}} ymm4 = mem[0,1,2,3],ymm1[4,5,6,7]264; AVX512DQ-FCP-NEXT: vmovlps %xmm0, (%rsi)265; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [1,7,0,0]266; AVX512DQ-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0267; AVX512DQ-FCP-NEXT: vmovlps %xmm0, (%rdx)268; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm0 = [2,4,2,4]269; AVX512DQ-FCP-NEXT: vpermi2d %xmm3, %xmm2, %xmm0270; AVX512DQ-FCP-NEXT: vmovq %xmm0, (%rcx)271; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [7,1,0,0]272; AVX512DQ-FCP-NEXT: vpermi2d %xmm2, %xmm3, %xmm0273; AVX512DQ-FCP-NEXT: vmovq %xmm0, (%r8)274; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]275; AVX512DQ-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0276; AVX512DQ-FCP-NEXT: vmovlps %xmm0, (%r9)277; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]278; AVX512DQ-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0279; AVX512DQ-FCP-NEXT: vmovlps %xmm0, (%rax)280; AVX512DQ-FCP-NEXT: vzeroupper281; AVX512DQ-FCP-NEXT: retq282;283; AVX512BW-LABEL: load_i32_stride6_vf2:284; AVX512BW: # %bb.0:285; AVX512BW-NEXT: pushq %rbx286; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax287; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0288; AVX512BW-NEXT: vmovaps 16(%rdi), %xmm1289; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm2290; AVX512BW-NEXT: vextractps $2, %xmm1, %r10d291; AVX512BW-NEXT: vextractps $3, %xmm1, %r11d292; AVX512BW-NEXT: vmovd %xmm2, %ebx293; AVX512BW-NEXT: vpinsrd $1, %r10d, %xmm0, %xmm1294; AVX512BW-NEXT: vpbroadcastd 8(%rdi), %xmm3295; AVX512BW-NEXT: vmovaps 32(%rdi), %ymm4296; AVX512BW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]297; AVX512BW-NEXT: vmovq %xmm1, (%rsi)298; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]299; AVX512BW-NEXT: vpinsrd $1, %r11d, %xmm1, %xmm1300; AVX512BW-NEXT: vmovq %xmm1, (%rdx)301; AVX512BW-NEXT: vpinsrd $1, %ebx, %xmm3, %xmm1302; AVX512BW-NEXT: vmovq %xmm1, (%rcx)303; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]304; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]305; AVX512BW-NEXT: vmovq %xmm0, (%r8)306; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]307; AVX512BW-NEXT: vpermps %ymm4, %ymm0, %ymm0308; AVX512BW-NEXT: vmovlps %xmm0, (%r9)309; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]310; AVX512BW-NEXT: vpermps %ymm4, %ymm0, %ymm0311; AVX512BW-NEXT: vmovlps %xmm0, (%rax)312; AVX512BW-NEXT: popq %rbx313; AVX512BW-NEXT: vzeroupper314; AVX512BW-NEXT: retq315;316; AVX512BW-FCP-LABEL: load_i32_stride6_vf2:317; AVX512BW-FCP: # %bb.0:318; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax319; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,0,0]320; AVX512BW-FCP-NEXT: vmovaps (%rdi), %ymm1321; AVX512BW-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0322; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm2323; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm3324; AVX512BW-FCP-NEXT: vblendps {{.*#+}} ymm4 = mem[0,1,2,3],ymm1[4,5,6,7]325; AVX512BW-FCP-NEXT: vmovlps %xmm0, (%rsi)326; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [1,7,0,0]327; AVX512BW-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0328; AVX512BW-FCP-NEXT: vmovlps %xmm0, (%rdx)329; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm0 = [2,4,2,4]330; AVX512BW-FCP-NEXT: vpermi2d %xmm3, %xmm2, %xmm0331; AVX512BW-FCP-NEXT: vmovq %xmm0, (%rcx)332; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [7,1,0,0]333; AVX512BW-FCP-NEXT: vpermi2d %xmm2, %xmm3, %xmm0334; AVX512BW-FCP-NEXT: vmovq %xmm0, (%r8)335; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]336; AVX512BW-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0337; AVX512BW-FCP-NEXT: vmovlps %xmm0, (%r9)338; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]339; AVX512BW-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0340; AVX512BW-FCP-NEXT: vmovlps %xmm0, (%rax)341; AVX512BW-FCP-NEXT: vzeroupper342; AVX512BW-FCP-NEXT: retq343;344; AVX512DQ-BW-LABEL: load_i32_stride6_vf2:345; AVX512DQ-BW: # %bb.0:346; AVX512DQ-BW-NEXT: pushq %rbx347; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax348; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0349; AVX512DQ-BW-NEXT: vmovaps 16(%rdi), %xmm1350; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm2351; AVX512DQ-BW-NEXT: vextractps $2, %xmm1, %r10d352; AVX512DQ-BW-NEXT: vextractps $3, %xmm1, %r11d353; AVX512DQ-BW-NEXT: vmovd %xmm2, %ebx354; AVX512DQ-BW-NEXT: vpinsrd $1, %r10d, %xmm0, %xmm1355; AVX512DQ-BW-NEXT: vpbroadcastd 8(%rdi), %xmm3356; AVX512DQ-BW-NEXT: vmovaps 32(%rdi), %ymm4357; AVX512DQ-BW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]358; AVX512DQ-BW-NEXT: vmovq %xmm1, (%rsi)359; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]360; AVX512DQ-BW-NEXT: vpinsrd $1, %r11d, %xmm1, %xmm1361; AVX512DQ-BW-NEXT: vmovq %xmm1, (%rdx)362; AVX512DQ-BW-NEXT: vpinsrd $1, %ebx, %xmm3, %xmm1363; AVX512DQ-BW-NEXT: vmovq %xmm1, (%rcx)364; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]365; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]366; AVX512DQ-BW-NEXT: vmovq %xmm0, (%r8)367; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]368; AVX512DQ-BW-NEXT: vpermps %ymm4, %ymm0, %ymm0369; AVX512DQ-BW-NEXT: vmovlps %xmm0, (%r9)370; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]371; AVX512DQ-BW-NEXT: vpermps %ymm4, %ymm0, %ymm0372; AVX512DQ-BW-NEXT: vmovlps %xmm0, (%rax)373; AVX512DQ-BW-NEXT: popq %rbx374; AVX512DQ-BW-NEXT: vzeroupper375; AVX512DQ-BW-NEXT: retq376;377; AVX512DQ-BW-FCP-LABEL: load_i32_stride6_vf2:378; AVX512DQ-BW-FCP: # %bb.0:379; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax380; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,0,0]381; AVX512DQ-BW-FCP-NEXT: vmovaps (%rdi), %ymm1382; AVX512DQ-BW-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0383; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm2384; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm3385; AVX512DQ-BW-FCP-NEXT: vblendps {{.*#+}} ymm4 = mem[0,1,2,3],ymm1[4,5,6,7]386; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm0, (%rsi)387; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [1,7,0,0]388; AVX512DQ-BW-FCP-NEXT: vpermps %ymm1, %ymm0, %ymm0389; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm0, (%rdx)390; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm0 = [2,4,2,4]391; AVX512DQ-BW-FCP-NEXT: vpermi2d %xmm3, %xmm2, %xmm0392; AVX512DQ-BW-FCP-NEXT: vmovq %xmm0, (%rcx)393; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [7,1,0,0]394; AVX512DQ-BW-FCP-NEXT: vpermi2d %xmm2, %xmm3, %xmm0395; AVX512DQ-BW-FCP-NEXT: vmovq %xmm0, (%r8)396; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4,2,0,0]397; AVX512DQ-BW-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0398; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm0, (%r9)399; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [5,3,0,0]400; AVX512DQ-BW-FCP-NEXT: vpermps %ymm4, %ymm0, %ymm0401; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm0, (%rax)402; AVX512DQ-BW-FCP-NEXT: vzeroupper403; AVX512DQ-BW-FCP-NEXT: retq404 %wide.vec = load <12 x i32>, ptr %in.vec, align 64405 %strided.vec0 = shufflevector <12 x i32> %wide.vec, <12 x i32> poison, <2 x i32> <i32 0, i32 6>406 %strided.vec1 = shufflevector <12 x i32> %wide.vec, <12 x i32> poison, <2 x i32> <i32 1, i32 7>407 %strided.vec2 = shufflevector <12 x i32> %wide.vec, <12 x i32> poison, <2 x i32> <i32 2, i32 8>408 %strided.vec3 = shufflevector <12 x i32> %wide.vec, <12 x i32> poison, <2 x i32> <i32 3, i32 9>409 %strided.vec4 = shufflevector <12 x i32> %wide.vec, <12 x i32> poison, <2 x i32> <i32 4, i32 10>410 %strided.vec5 = shufflevector <12 x i32> %wide.vec, <12 x i32> poison, <2 x i32> <i32 5, i32 11>411 store <2 x i32> %strided.vec0, ptr %out.vec0, align 64412 store <2 x i32> %strided.vec1, ptr %out.vec1, align 64413 store <2 x i32> %strided.vec2, ptr %out.vec2, align 64414 store <2 x i32> %strided.vec3, ptr %out.vec3, align 64415 store <2 x i32> %strided.vec4, ptr %out.vec4, align 64416 store <2 x i32> %strided.vec5, ptr %out.vec5, align 64417 ret void418}419 420define void @load_i32_stride6_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {421; SSE-LABEL: load_i32_stride6_vf4:422; SSE: # %bb.0:423; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax424; SSE-NEXT: movdqa 80(%rdi), %xmm1425; SSE-NEXT: movdqa 64(%rdi), %xmm0426; SSE-NEXT: movdqa (%rdi), %xmm6427; SSE-NEXT: movdqa 16(%rdi), %xmm2428; SSE-NEXT: movdqa 48(%rdi), %xmm3429; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[2,3,2,3]430; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm6[1,1,1,1]431; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm6[2,3,2,3]432; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm6[3,3,3,3]433; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]434; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm0[2,2,3,3]435; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]436; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]437; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm6[0],xmm4[1]438; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[3,3,3,3]439; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]440; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm0[2,3,2,3]441; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm3[2,2,3,3]442; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm3[2,3,2,3]443; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm10[0],xmm3[1],xmm10[1]444; SSE-NEXT: movdqa 32(%rdi), %xmm10445; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm7[0],xmm3[1]446; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1]447; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm1[0,0,1,1]448; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1]449; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm8[0],xmm5[1]450; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm10[1,1,1,1]451; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm7[0],xmm9[1],xmm7[1]452; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]453; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm9[0],xmm6[1]454; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm10[2,3,2,3]455; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm2[1,1,1,1]456; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1]457; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm1[2,2,3,3]458; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm0[0,0,1,1]459; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm7[0],xmm9[1],xmm7[1]460; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm2[0],xmm9[1]461; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[3,3,3,3]462; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1]463; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]464; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]465; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm8[0],xmm0[1]466; SSE-NEXT: movapd %xmm4, (%rsi)467; SSE-NEXT: movapd %xmm3, (%rdx)468; SSE-NEXT: movapd %xmm5, (%rcx)469; SSE-NEXT: movapd %xmm6, (%r8)470; SSE-NEXT: movapd %xmm9, (%r9)471; SSE-NEXT: movapd %xmm0, (%rax)472; SSE-NEXT: retq473;474; AVX-LABEL: load_i32_stride6_vf4:475; AVX: # %bb.0:476; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax477; AVX-NEXT: vmovaps 32(%rdi), %ymm0478; AVX-NEXT: vmovaps (%rdi), %ymm1479; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]480; AVX-NEXT: vextractf128 $1, %ymm2, %xmm3481; AVX-NEXT: vblendps {{.*#+}} xmm4 = xmm2[0,1],xmm3[2,3]482; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,2],xmm3[0,3]483; AVX-NEXT: vmovaps 64(%rdi), %xmm5484; AVX-NEXT: vinsertps {{.*#+}} xmm4 = xmm4[0,1,2],xmm5[2]485; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,0],xmm3[3,0]486; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[1,3]487; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1,2],xmm5[3]488; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]489; AVX-NEXT: vextractf128 $1, %ymm0, %xmm1490; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm0[2,0],xmm1[2,3]491; AVX-NEXT: vmovaps 80(%rdi), %xmm6492; AVX-NEXT: vinsertps {{.*#+}} xmm3 = xmm3[0,1,2],xmm6[0]493; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[3,3]494; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm6[1]495; AVX-NEXT: vmovaps 32(%rdi), %xmm1496; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm1[2,2,3,3]497; AVX-NEXT: vmovaps 16(%rdi), %xmm8498; AVX-NEXT: vblendps {{.*#+}} xmm7 = xmm8[0],xmm7[1],xmm8[2,3]499; AVX-NEXT: vblendps {{.*#+}} xmm9 = xmm5[0,1],xmm6[2,3]500; AVX-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,0,2]501; AVX-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]502; AVX-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero503; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5],xmm6[6,7]504; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]505; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,3,2,3]506; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm5[2,3]507; AVX-NEXT: vmovaps %xmm4, (%rsi)508; AVX-NEXT: vmovaps %xmm2, (%rdx)509; AVX-NEXT: vmovaps %xmm3, (%rcx)510; AVX-NEXT: vmovaps %xmm0, (%r8)511; AVX-NEXT: vmovaps %xmm7, (%r9)512; AVX-NEXT: vmovaps %xmm1, (%rax)513; AVX-NEXT: vzeroupper514; AVX-NEXT: retq515;516; AVX2-LABEL: load_i32_stride6_vf4:517; AVX2: # %bb.0:518; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax519; AVX2-NEXT: vmovdqa (%rdi), %ymm1520; AVX2-NEXT: vmovdqa 32(%rdi), %ymm2521; AVX2-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,4,0]522; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm2[4,5],ymm1[6,7]523; AVX2-NEXT: vpermd %ymm3, %ymm0, %ymm0524; AVX2-NEXT: vmovdqa 64(%rdi), %xmm4525; AVX2-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[2,2,2,2]526; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm5[3]527; AVX2-NEXT: vpmovsxbd {{.*#+}} xmm5 = [1,7,5,0]528; AVX2-NEXT: vpermd %ymm3, %ymm5, %ymm3529; AVX2-NEXT: vpblendd {{.*#+}} xmm3 = xmm3[0,1,2],xmm4[3]530; AVX2-NEXT: vmovdqa 80(%rdi), %xmm5531; AVX2-NEXT: vpbroadcastd %xmm5, %xmm6532; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]533; AVX2-NEXT: vpshufd {{.*#+}} ymm8 = ymm7[2,0,2,3,6,4,6,7]534; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,3,2,3]535; AVX2-NEXT: vpblendd {{.*#+}} xmm6 = xmm8[0,1,2],xmm6[3]536; AVX2-NEXT: vpbroadcastd 84(%rdi), %xmm8537; AVX2-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[3,1,3,3,7,5,7,7]538; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,3,2,3]539; AVX2-NEXT: vpblendd {{.*#+}} xmm7 = xmm7[0,1,2],xmm8[3]540; AVX2-NEXT: vpblendd {{.*#+}} xmm8 = xmm4[0,1],xmm5[2,3]541; AVX2-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,1,0,2]542; AVX2-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,2,0,0]543; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]544; AVX2-NEXT: vpermd %ymm1, %ymm9, %ymm2545; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm8[2,3]546; AVX2-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero547; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1,2],xmm5[3]548; AVX2-NEXT: vpmovsxbd {{.*#+}} xmm5 = [5,3,0,0]549; AVX2-NEXT: vpermd %ymm1, %ymm5, %ymm1550; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1],xmm4[2,3]551; AVX2-NEXT: vmovdqa %xmm0, (%rsi)552; AVX2-NEXT: vmovdqa %xmm3, (%rdx)553; AVX2-NEXT: vmovdqa %xmm6, (%rcx)554; AVX2-NEXT: vmovdqa %xmm7, (%r8)555; AVX2-NEXT: vmovdqa %xmm2, (%r9)556; AVX2-NEXT: vmovdqa %xmm1, (%rax)557; AVX2-NEXT: vzeroupper558; AVX2-NEXT: retq559;560; AVX2-FP-LABEL: load_i32_stride6_vf4:561; AVX2-FP: # %bb.0:562; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax563; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm1564; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm2565; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,4,0]566; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm2[4,5],ymm1[6,7]567; AVX2-FP-NEXT: vpermd %ymm3, %ymm0, %ymm0568; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm4569; AVX2-FP-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[2,2,2,2]570; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm5[3]571; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [1,7,5,0]572; AVX2-FP-NEXT: vpermd %ymm3, %ymm5, %ymm3573; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm3 = xmm3[0,1,2],xmm4[3]574; AVX2-FP-NEXT: vmovdqa 80(%rdi), %xmm5575; AVX2-FP-NEXT: vpbroadcastd %xmm5, %xmm6576; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]577; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm8 = ymm7[2,0,2,3,6,4,6,7]578; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,3,2,3]579; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm6 = xmm8[0,1,2],xmm6[3]580; AVX2-FP-NEXT: vpbroadcastd 84(%rdi), %xmm8581; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[3,1,3,3,7,5,7,7]582; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,3,2,3]583; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm7 = xmm7[0,1,2],xmm8[3]584; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm8 = xmm4[0,1],xmm5[2,3]585; AVX2-FP-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,1,0,2]586; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,2,0,0]587; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]588; AVX2-FP-NEXT: vpermd %ymm1, %ymm9, %ymm2589; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm8[2,3]590; AVX2-FP-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero591; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1,2],xmm5[3]592; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [5,3,0,0]593; AVX2-FP-NEXT: vpermd %ymm1, %ymm5, %ymm1594; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1],xmm4[2,3]595; AVX2-FP-NEXT: vmovdqa %xmm0, (%rsi)596; AVX2-FP-NEXT: vmovdqa %xmm3, (%rdx)597; AVX2-FP-NEXT: vmovdqa %xmm6, (%rcx)598; AVX2-FP-NEXT: vmovdqa %xmm7, (%r8)599; AVX2-FP-NEXT: vmovdqa %xmm2, (%r9)600; AVX2-FP-NEXT: vmovdqa %xmm1, (%rax)601; AVX2-FP-NEXT: vzeroupper602; AVX2-FP-NEXT: retq603;604; AVX2-FCP-LABEL: load_i32_stride6_vf4:605; AVX2-FCP: # %bb.0:606; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax607; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm1608; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm2609; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} xmm0 = [0,6,4,0]610; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm2[4,5],ymm1[6,7]611; AVX2-FCP-NEXT: vpermd %ymm3, %ymm0, %ymm0612; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm4613; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[2,2,2,2]614; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm5[3]615; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [1,7,5,0]616; AVX2-FCP-NEXT: vpermd %ymm3, %ymm5, %ymm3617; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm3[0,1,2],xmm4[3]618; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [2,0,6,7]619; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]620; AVX2-FCP-NEXT: vpermd %ymm6, %ymm5, %ymm5621; AVX2-FCP-NEXT: vmovdqa 80(%rdi), %xmm7622; AVX2-FCP-NEXT: vpbroadcastd %xmm7, %xmm8623; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm5 = xmm5[0,1,2],xmm8[3]624; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} xmm8 = [3,1,7,7]625; AVX2-FCP-NEXT: vpermd %ymm6, %ymm8, %ymm6626; AVX2-FCP-NEXT: vpbroadcastd 84(%rdi), %xmm8627; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm6[0,1,2],xmm8[3]628; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm8 = xmm4[0,1],xmm7[2,3]629; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,1,0,2]630; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} xmm9 = [4,2,0,0]631; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]632; AVX2-FCP-NEXT: vpermd %ymm1, %ymm9, %ymm2633; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm8[2,3]634; AVX2-FCP-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero635; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1,2],xmm7[3]636; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,3,0,0]637; AVX2-FCP-NEXT: vpermd %ymm1, %ymm7, %ymm1638; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1],xmm4[2,3]639; AVX2-FCP-NEXT: vmovdqa %xmm0, (%rsi)640; AVX2-FCP-NEXT: vmovdqa %xmm3, (%rdx)641; AVX2-FCP-NEXT: vmovdqa %xmm5, (%rcx)642; AVX2-FCP-NEXT: vmovdqa %xmm6, (%r8)643; AVX2-FCP-NEXT: vmovdqa %xmm2, (%r9)644; AVX2-FCP-NEXT: vmovdqa %xmm1, (%rax)645; AVX2-FCP-NEXT: vzeroupper646; AVX2-FCP-NEXT: retq647;648; AVX512-LABEL: load_i32_stride6_vf4:649; AVX512: # %bb.0:650; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax651; AVX512-NEXT: vmovdqa64 (%rdi), %zmm0652; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm1653; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]654; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm2655; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]656; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm3657; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]658; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm4659; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]660; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm5661; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]662; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm6663; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]664; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm7665; AVX512-NEXT: vmovdqa %xmm2, (%rsi)666; AVX512-NEXT: vmovdqa %xmm3, (%rdx)667; AVX512-NEXT: vmovdqa %xmm4, (%rcx)668; AVX512-NEXT: vmovdqa %xmm5, (%r8)669; AVX512-NEXT: vmovdqa %xmm6, (%r9)670; AVX512-NEXT: vmovdqa %xmm7, (%rax)671; AVX512-NEXT: vzeroupper672; AVX512-NEXT: retq673;674; AVX512-FCP-LABEL: load_i32_stride6_vf4:675; AVX512-FCP: # %bb.0:676; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax677; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm0678; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1679; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]680; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm2681; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]682; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm3683; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]684; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4685; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]686; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm5687; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]688; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm6689; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]690; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm7691; AVX512-FCP-NEXT: vmovdqa %xmm2, (%rsi)692; AVX512-FCP-NEXT: vmovdqa %xmm3, (%rdx)693; AVX512-FCP-NEXT: vmovdqa %xmm4, (%rcx)694; AVX512-FCP-NEXT: vmovdqa %xmm5, (%r8)695; AVX512-FCP-NEXT: vmovdqa %xmm6, (%r9)696; AVX512-FCP-NEXT: vmovdqa %xmm7, (%rax)697; AVX512-FCP-NEXT: vzeroupper698; AVX512-FCP-NEXT: retq699;700; AVX512DQ-LABEL: load_i32_stride6_vf4:701; AVX512DQ: # %bb.0:702; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax703; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm0704; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm1705; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]706; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm2707; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]708; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm3709; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]710; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm4711; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]712; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm5713; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]714; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm6715; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]716; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm7717; AVX512DQ-NEXT: vmovdqa %xmm2, (%rsi)718; AVX512DQ-NEXT: vmovdqa %xmm3, (%rdx)719; AVX512DQ-NEXT: vmovdqa %xmm4, (%rcx)720; AVX512DQ-NEXT: vmovdqa %xmm5, (%r8)721; AVX512DQ-NEXT: vmovdqa %xmm6, (%r9)722; AVX512DQ-NEXT: vmovdqa %xmm7, (%rax)723; AVX512DQ-NEXT: vzeroupper724; AVX512DQ-NEXT: retq725;726; AVX512DQ-FCP-LABEL: load_i32_stride6_vf4:727; AVX512DQ-FCP: # %bb.0:728; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax729; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm0730; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1731; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]732; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm2733; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]734; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm3735; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]736; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4737; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]738; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm5739; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]740; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm6741; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]742; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm7743; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, (%rsi)744; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, (%rdx)745; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, (%rcx)746; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, (%r8)747; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, (%r9)748; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, (%rax)749; AVX512DQ-FCP-NEXT: vzeroupper750; AVX512DQ-FCP-NEXT: retq751;752; AVX512BW-LABEL: load_i32_stride6_vf4:753; AVX512BW: # %bb.0:754; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax755; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm0756; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm1757; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]758; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm2759; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]760; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm3761; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]762; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm4763; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]764; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm5765; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]766; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm6767; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]768; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm7769; AVX512BW-NEXT: vmovdqa %xmm2, (%rsi)770; AVX512BW-NEXT: vmovdqa %xmm3, (%rdx)771; AVX512BW-NEXT: vmovdqa %xmm4, (%rcx)772; AVX512BW-NEXT: vmovdqa %xmm5, (%r8)773; AVX512BW-NEXT: vmovdqa %xmm6, (%r9)774; AVX512BW-NEXT: vmovdqa %xmm7, (%rax)775; AVX512BW-NEXT: vzeroupper776; AVX512BW-NEXT: retq777;778; AVX512BW-FCP-LABEL: load_i32_stride6_vf4:779; AVX512BW-FCP: # %bb.0:780; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax781; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm0782; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1783; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]784; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm2785; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]786; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm3787; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]788; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4789; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]790; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm5791; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]792; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm6793; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]794; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm7795; AVX512BW-FCP-NEXT: vmovdqa %xmm2, (%rsi)796; AVX512BW-FCP-NEXT: vmovdqa %xmm3, (%rdx)797; AVX512BW-FCP-NEXT: vmovdqa %xmm4, (%rcx)798; AVX512BW-FCP-NEXT: vmovdqa %xmm5, (%r8)799; AVX512BW-FCP-NEXT: vmovdqa %xmm6, (%r9)800; AVX512BW-FCP-NEXT: vmovdqa %xmm7, (%rax)801; AVX512BW-FCP-NEXT: vzeroupper802; AVX512BW-FCP-NEXT: retq803;804; AVX512DQ-BW-LABEL: load_i32_stride6_vf4:805; AVX512DQ-BW: # %bb.0:806; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax807; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm0808; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm1809; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]810; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm2811; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]812; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm3813; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]814; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm4815; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]816; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm5817; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]818; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm6819; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]820; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm7821; AVX512DQ-BW-NEXT: vmovdqa %xmm2, (%rsi)822; AVX512DQ-BW-NEXT: vmovdqa %xmm3, (%rdx)823; AVX512DQ-BW-NEXT: vmovdqa %xmm4, (%rcx)824; AVX512DQ-BW-NEXT: vmovdqa %xmm5, (%r8)825; AVX512DQ-BW-NEXT: vmovdqa %xmm6, (%r9)826; AVX512DQ-BW-NEXT: vmovdqa %xmm7, (%rax)827; AVX512DQ-BW-NEXT: vzeroupper828; AVX512DQ-BW-NEXT: retq829;830; AVX512DQ-BW-FCP-LABEL: load_i32_stride6_vf4:831; AVX512DQ-BW-FCP: # %bb.0:832; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax833; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm0834; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1835; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm2 = [0,6,12,18]836; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm2837; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm3 = [1,7,13,19]838; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm3839; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm4 = [2,8,14,20]840; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4841; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm5 = [3,9,15,21]842; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm5843; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm6 = [4,10,16,22]844; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm6845; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm7 = [5,11,17,23]846; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm7847; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm2, (%rsi)848; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm3, (%rdx)849; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm4, (%rcx)850; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm5, (%r8)851; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm6, (%r9)852; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm7, (%rax)853; AVX512DQ-BW-FCP-NEXT: vzeroupper854; AVX512DQ-BW-FCP-NEXT: retq855 %wide.vec = load <24 x i32>, ptr %in.vec, align 64856 %strided.vec0 = shufflevector <24 x i32> %wide.vec, <24 x i32> poison, <4 x i32> <i32 0, i32 6, i32 12, i32 18>857 %strided.vec1 = shufflevector <24 x i32> %wide.vec, <24 x i32> poison, <4 x i32> <i32 1, i32 7, i32 13, i32 19>858 %strided.vec2 = shufflevector <24 x i32> %wide.vec, <24 x i32> poison, <4 x i32> <i32 2, i32 8, i32 14, i32 20>859 %strided.vec3 = shufflevector <24 x i32> %wide.vec, <24 x i32> poison, <4 x i32> <i32 3, i32 9, i32 15, i32 21>860 %strided.vec4 = shufflevector <24 x i32> %wide.vec, <24 x i32> poison, <4 x i32> <i32 4, i32 10, i32 16, i32 22>861 %strided.vec5 = shufflevector <24 x i32> %wide.vec, <24 x i32> poison, <4 x i32> <i32 5, i32 11, i32 17, i32 23>862 store <4 x i32> %strided.vec0, ptr %out.vec0, align 64863 store <4 x i32> %strided.vec1, ptr %out.vec1, align 64864 store <4 x i32> %strided.vec2, ptr %out.vec2, align 64865 store <4 x i32> %strided.vec3, ptr %out.vec3, align 64866 store <4 x i32> %strided.vec4, ptr %out.vec4, align 64867 store <4 x i32> %strided.vec5, ptr %out.vec5, align 64868 ret void869}870 871define void @load_i32_stride6_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {872; SSE-LABEL: load_i32_stride6_vf8:873; SSE: # %bb.0:874; SSE-NEXT: movdqa 144(%rdi), %xmm4875; SSE-NEXT: movdqa 160(%rdi), %xmm2876; SSE-NEXT: movdqa 96(%rdi), %xmm6877; SSE-NEXT: movdqa 112(%rdi), %xmm3878; SSE-NEXT: movdqa 64(%rdi), %xmm5879; SSE-NEXT: movdqa (%rdi), %xmm10880; SSE-NEXT: movdqa 16(%rdi), %xmm11881; SSE-NEXT: movdqa 48(%rdi), %xmm8882; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[2,3,2,3]883; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill884; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[1,1,1,1]885; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm10[2,3,2,3]886; SSE-NEXT: movdqa %xmm10, %xmm7887; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]888; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3]889; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm8[0,0,1,1]890; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1]891; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm7[0],xmm9[1]892; SSE-NEXT: movapd %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill893; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]894; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm6[1,1,1,1]895; SSE-NEXT: movdqa %xmm6, %xmm9896; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1]897; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,2,3,3]898; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm4[0,0,1,1]899; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]900; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm9[0],xmm7[1]901; SSE-NEXT: movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill902; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[3,3,3,3]903; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]904; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,3,2,3]905; SSE-NEXT: movdqa %xmm5, %xmm9906; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm8[2,2,3,3]907; SSE-NEXT: movdqa %xmm8, %xmm11908; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm1[0],xmm11[1],xmm1[1]909; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm0[0],xmm11[1]910; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[3,3,3,3]911; SSE-NEXT: movdqa %xmm3, %xmm1912; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill913; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1]914; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]915; SSE-NEXT: movdqa %xmm4, %xmm12916; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm0[0],xmm12[1],xmm0[1]917; SSE-NEXT: movdqa 80(%rdi), %xmm14918; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm15[0],xmm12[1]919; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[0,0,1,1]920; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]921; SSE-NEXT: movdqa 32(%rdi), %xmm7922; SSE-NEXT: punpckldq {{.*#+}} xmm13 = xmm13[0],xmm7[0],xmm13[1],xmm7[1]923; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm13[0],xmm5[1]924; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill925; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm4[2,2,3,3]926; SSE-NEXT: movdqa 176(%rdi), %xmm15927; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[0,0,1,1]928; SSE-NEXT: punpckldq {{.*#+}} xmm13 = xmm13[0],xmm0[0],xmm13[1],xmm0[1]929; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]930; SSE-NEXT: movdqa 128(%rdi), %xmm5931; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]932; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm0[0],xmm13[1]933; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[3,3,3,3]934; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm7[1,1,1,1]935; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]936; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[2,3,2,3]937; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1]938; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm0[0],xmm8[1]939; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[3,3,3,3]940; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,1,1,1]941; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]942; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,3,2,3]943; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm15[0],xmm4[1],xmm15[1]944; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm0[0],xmm4[1]945; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]946; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload947; SSE-NEXT: movdqa %xmm3, %xmm10948; SSE-NEXT: punpckldq {{.*#+}} xmm10 = xmm10[0],xmm0[0],xmm10[1],xmm0[1]949; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[2,2,3,3]950; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm9[0,0,1,1]951; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]952; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm10[0],xmm6[1]953; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm5[2,3,2,3]954; SSE-NEXT: movdqa %xmm1, %xmm0955; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]956; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[2,2,3,3]957; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm2[0,0,1,1]958; SSE-NEXT: punpckldq {{.*#+}} xmm10 = xmm10[0],xmm1[0],xmm10[1],xmm1[1]959; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm0[0],xmm10[1]960; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]961; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[3,3,3,3]962; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]963; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[2,3,2,3]964; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1]965; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm0[0],xmm9[1]966; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload967; SSE-NEXT: # xmm0 = mem[1,1,1,1]968; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[3,3,3,3]969; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]970; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[2,3,2,3]971; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]972; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]973; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload974; SSE-NEXT: movaps %xmm0, 16(%rsi)975; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload976; SSE-NEXT: movaps %xmm0, (%rsi)977; SSE-NEXT: movapd %xmm12, 16(%rdx)978; SSE-NEXT: movapd %xmm11, (%rdx)979; SSE-NEXT: movapd %xmm13, 16(%rcx)980; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload981; SSE-NEXT: movaps %xmm0, (%rcx)982; SSE-NEXT: movapd %xmm4, 16(%r8)983; SSE-NEXT: movapd %xmm8, (%r8)984; SSE-NEXT: movapd %xmm10, 16(%r9)985; SSE-NEXT: movapd %xmm6, (%r9)986; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax987; SSE-NEXT: movapd %xmm2, 16(%rax)988; SSE-NEXT: movapd %xmm9, (%rax)989; SSE-NEXT: retq990;991; AVX-LABEL: load_i32_stride6_vf8:992; AVX: # %bb.0:993; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax994; AVX-NEXT: vmovapd 160(%rdi), %ymm3995; AVX-NEXT: vmovapd 128(%rdi), %ymm4996; AVX-NEXT: vmovaps 96(%rdi), %ymm0997; AVX-NEXT: vmovaps 64(%rdi), %ymm1998; AVX-NEXT: vmovaps 32(%rdi), %ymm6999; AVX-NEXT: vmovaps (%rdi), %ymm71000; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm6[4,5],ymm7[6,7]1001; AVX-NEXT: vextractf128 $1, %ymm5, %xmm81002; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm5[0,1],xmm8[2,3]1003; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm8[0,3]1004; AVX-NEXT: vinsertf128 $1, 96(%rdi), %ymm1, %ymm91005; AVX-NEXT: vshufps {{.*#+}} ymm10 = ymm0[2,0],ymm9[0,0],ymm0[6,4],ymm9[4,4]1006; AVX-NEXT: vshufps {{.*#+}} ymm10 = ymm10[2,0],ymm1[2,2],ymm10[6,4],ymm1[6,6]1007; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm10[3,4,5,6,7]1008; AVX-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm4[2,3],ymm3[0,1]1009; AVX-NEXT: vshufpd {{.*#+}} ymm11 = ymm10[0],ymm4[1],ymm10[3],ymm4[2]1010; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[0,1,2,0,4,5,6,4]1011; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm11[6,7]1012; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm5[1,0],xmm8[3,0]1013; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,2],xmm8[1,3]1014; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,0],ymm9[1,0],ymm0[7,4],ymm9[5,4]1015; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[2,0],ymm1[2,3],ymm8[6,4],ymm1[6,7]1016; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm8[3,4,5,6,7]1017; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm10[3,1],ymm4[1,3],ymm10[7,5],ymm4[5,7]1018; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,1,2,0,4,5,6,4]1019; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm8[6,7]1020; AVX-NEXT: vblendpd {{.*#+}} ymm8 = ymm4[0,1],ymm3[2],ymm4[3]1021; AVX-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm8[2,3,0,1]1022; AVX-NEXT: vshufps {{.*#+}} ymm10 = ymm9[0,0],ymm8[2,0],ymm9[4,4],ymm8[6,4]1023; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm6[0,1],ymm7[2,3],ymm6[4,5,6,7]1024; AVX-NEXT: vextractf128 $1, %ymm7, %xmm111025; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm7[2,0],xmm11[2,3]1026; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm0[2,1],ymm1[2,0],ymm0[6,5],ymm1[6,4]1027; AVX-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3,0,1]1028; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm12[3,4,5,6,7]1029; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4],ymm10[5,6,7]1030; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm9[0,1],ymm8[3,1],ymm9[4,5],ymm8[7,5]1031; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm7[3,1],xmm11[3,3]1032; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm0[3,1],ymm1[2,1],ymm0[7,5],ymm1[6,5]1033; AVX-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3,0,1]1034; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm9[3,4,5,6,7]1035; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4],ymm8[5,6,7]1036; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3]1037; AVX-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm3[2,3,0,1]1038; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm3[2,0],ymm4[0,0],ymm3[6,4],ymm4[4,4]1039; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm4[0,2],ymm8[2,0],ymm4[4,6],ymm8[6,4]1040; AVX-NEXT: vmovaps 32(%rdi), %xmm91041; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm9[2,2,3,3]1042; AVX-NEXT: vmovaps 16(%rdi), %xmm111043; AVX-NEXT: vblendps {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]1044; AVX-NEXT: vmovapd 80(%rdi), %xmm121045; AVX-NEXT: vshufpd {{.*#+}} ymm13 = ymm12[1],ymm1[0],ymm12[2],ymm1[3]1046; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm0[0,1],ymm13[2,0],ymm0[4,5],ymm13[6,4]1047; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm13[2,3,4,5,6,7]1048; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3,4],ymm8[5,6,7]1049; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,0],ymm4[1,0],ymm3[7,4],ymm4[5,4]1050; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm4[0,3],ymm3[2,0],ymm4[4,7],ymm3[6,4]1051; AVX-NEXT: vblendps {{.*#+}} xmm4 = xmm11[0,1],xmm9[2,3]1052; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm4[1,3,2,3]1053; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm12[3,1],ymm1[1,3],ymm12[7,5],ymm1[5,7]1054; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[2,0],ymm0[5,5],ymm1[6,4]1055; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3,4,5,6,7]1056; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]1057; AVX-NEXT: vmovaps %ymm2, (%rsi)1058; AVX-NEXT: vmovaps %ymm5, (%rdx)1059; AVX-NEXT: vmovaps %ymm6, (%rcx)1060; AVX-NEXT: vmovaps %ymm7, (%r8)1061; AVX-NEXT: vmovaps %ymm8, (%r9)1062; AVX-NEXT: vmovaps %ymm0, (%rax)1063; AVX-NEXT: vzeroupper1064; AVX-NEXT: retq1065;1066; AVX2-LABEL: load_i32_stride6_vf8:1067; AVX2: # %bb.0:1068; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax1069; AVX2-NEXT: vmovaps 128(%rdi), %ymm01070; AVX2-NEXT: vmovaps 160(%rdi), %ymm11071; AVX2-NEXT: vmovaps 96(%rdi), %ymm51072; AVX2-NEXT: vmovaps (%rdi), %ymm21073; AVX2-NEXT: vmovaps 32(%rdi), %ymm41074; AVX2-NEXT: vmovaps 64(%rdi), %ymm61075; AVX2-NEXT: vmovaps {{.*#+}} xmm3 = [0,6,4,u]1076; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3],ymm4[4,5],ymm2[6,7]1077; AVX2-NEXT: vpermps %ymm7, %ymm3, %ymm31078; AVX2-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm6[0,1],ymm5[0,1]1079; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm8[0,1,2,3,4,5],ymm5[6,7]1080; AVX2-NEXT: vshufps {{.*#+}} ymm8 = ymm9[0,2,2,2,4,6,6,6]1081; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm8[3,4,5,6,7]1082; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5,6,7]1083; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm8 = [4,2,4,2,4,2,4,2]1084; AVX2-NEXT: vpermps %ymm10, %ymm8, %ymm111085; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm11[6,7]1086; AVX2-NEXT: vmovaps {{.*#+}} xmm11 = [1,7,5,u]1087; AVX2-NEXT: vpermps %ymm7, %ymm11, %ymm71088; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm9[1,3,2,3,5,7,6,7]1089; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm9[3,4,5,6,7]1090; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm9 = [5,3,5,3,5,3,5,3]1091; AVX2-NEXT: vpermps %ymm10, %ymm9, %ymm101092; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm10[6,7]1093; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm6[0,1],ymm5[2,3],ymm6[4,5],ymm5[6,7]1094; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm10[2,0,2,3,6,4,6,7]1095; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,2,0,3]1096; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm4[0,1],ymm2[2,3],ymm4[4,5,6,7]1097; AVX2-NEXT: vshufps {{.*#+}} ymm12 = ymm11[2,0,2,3,6,4,6,7]1098; AVX2-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[0,3,2,3]1099; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm12[0,1,2],ymm10[3,4,5,6,7]1100; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]1101; AVX2-NEXT: vshufps {{.*#+}} ymm13 = ymm12[0,0,2,0,4,4,6,4]1102; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,0,3]1103; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4],ymm13[5,6,7]1104; AVX2-NEXT: vshufps {{.*#+}} ymm13 = ymm5[3,3,3,3,7,7,7,7]1105; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm6[1],ymm13[2,3,4],ymm6[5],ymm13[6,7]1106; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,2,0,3]1107; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm11[3,1,3,3,7,5,7,7]1108; AVX2-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,3,2,3]1109; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm13[3,4,5,6,7]1110; AVX2-NEXT: vshufps {{.*#+}} ymm12 = ymm12[0,1,3,1,4,5,7,5]1111; AVX2-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[0,1,0,3]1112; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4],ymm12[5,6,7]1113; AVX2-NEXT: vmovaps 80(%rdi), %xmm121114; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1115; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm5[0,1],ymm12[2,3],ymm5[4,5,6,7]1116; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm6[0,1,0,2,4,5,4,6]1117; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]1118; AVX2-NEXT: vpermps %ymm2, %ymm8, %ymm41119; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm6[2,3,4,5,6,7]1120; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1121; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [0,2,0,6,0,2,0,6]1122; AVX2-NEXT: # ymm1 = mem[0,1,0,1]1123; AVX2-NEXT: vpermps %ymm0, %ymm1, %ymm11124; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4],ymm1[5,6,7]1125; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm5[1,1,1,1,5,5,5,5]1126; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm12[3],ymm4[4,5,6,7]1127; AVX2-NEXT: vpermps %ymm2, %ymm9, %ymm21128; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2,3,4,5,6,7]1129; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [0,3,1,7,0,3,1,7]1130; AVX2-NEXT: # ymm4 = mem[0,1,0,1]1131; AVX2-NEXT: vpermps %ymm0, %ymm4, %ymm01132; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6,7]1133; AVX2-NEXT: vmovaps %ymm3, (%rsi)1134; AVX2-NEXT: vmovaps %ymm7, (%rdx)1135; AVX2-NEXT: vmovaps %ymm10, (%rcx)1136; AVX2-NEXT: vmovaps %ymm11, (%r8)1137; AVX2-NEXT: vmovaps %ymm1, (%r9)1138; AVX2-NEXT: vmovaps %ymm0, (%rax)1139; AVX2-NEXT: vzeroupper1140; AVX2-NEXT: retq1141;1142; AVX2-FP-LABEL: load_i32_stride6_vf8:1143; AVX2-FP: # %bb.0:1144; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax1145; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm01146; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm11147; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm51148; AVX2-FP-NEXT: vmovaps (%rdi), %ymm21149; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm41150; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm61151; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm3 = [0,6,4,u]1152; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3],ymm4[4,5],ymm2[6,7]1153; AVX2-FP-NEXT: vpermps %ymm7, %ymm3, %ymm31154; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm6[0,1],ymm5[0,1]1155; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm8[0,1,2,3,4,5],ymm5[6,7]1156; AVX2-FP-NEXT: vshufps {{.*#+}} ymm8 = ymm9[0,2,2,2,4,6,6,6]1157; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm8[3,4,5,6,7]1158; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5,6,7]1159; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm8 = [4,2,4,2,4,2,4,2]1160; AVX2-FP-NEXT: vpermps %ymm10, %ymm8, %ymm111161; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm11[6,7]1162; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm11 = [1,7,5,u]1163; AVX2-FP-NEXT: vpermps %ymm7, %ymm11, %ymm71164; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[1,3,2,3,5,7,6,7]1165; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm9[3,4,5,6,7]1166; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm9 = [5,3,5,3,5,3,5,3]1167; AVX2-FP-NEXT: vpermps %ymm10, %ymm9, %ymm101168; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm10[6,7]1169; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm6[0,1],ymm5[2,3],ymm6[4,5],ymm5[6,7]1170; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm10[2,0,2,3,6,4,6,7]1171; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,2,0,3]1172; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm4[0,1],ymm2[2,3],ymm4[4,5,6,7]1173; AVX2-FP-NEXT: vshufps {{.*#+}} ymm12 = ymm11[2,0,2,3,6,4,6,7]1174; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[0,3,2,3]1175; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm12[0,1,2],ymm10[3,4,5,6,7]1176; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]1177; AVX2-FP-NEXT: vshufps {{.*#+}} ymm13 = ymm12[0,0,2,0,4,4,6,4]1178; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,0,3]1179; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4],ymm13[5,6,7]1180; AVX2-FP-NEXT: vshufps {{.*#+}} ymm13 = ymm5[3,3,3,3,7,7,7,7]1181; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm6[1],ymm13[2,3,4],ymm6[5],ymm13[6,7]1182; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,2,0,3]1183; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm11[3,1,3,3,7,5,7,7]1184; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,3,2,3]1185; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm13[3,4,5,6,7]1186; AVX2-FP-NEXT: vshufps {{.*#+}} ymm12 = ymm12[0,1,3,1,4,5,7,5]1187; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[0,1,0,3]1188; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4],ymm12[5,6,7]1189; AVX2-FP-NEXT: vmovaps 80(%rdi), %xmm121190; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1191; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm5[0,1],ymm12[2,3],ymm5[4,5,6,7]1192; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm6[0,1,0,2,4,5,4,6]1193; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]1194; AVX2-FP-NEXT: vpermps %ymm2, %ymm8, %ymm41195; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm6[2,3,4,5,6,7]1196; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1197; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [0,2,0,6,0,2,0,6]1198; AVX2-FP-NEXT: # ymm1 = mem[0,1,0,1]1199; AVX2-FP-NEXT: vpermps %ymm0, %ymm1, %ymm11200; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4],ymm1[5,6,7]1201; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm5[1,1,1,1,5,5,5,5]1202; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm12[3],ymm4[4,5,6,7]1203; AVX2-FP-NEXT: vpermps %ymm2, %ymm9, %ymm21204; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2,3,4,5,6,7]1205; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [0,3,1,7,0,3,1,7]1206; AVX2-FP-NEXT: # ymm4 = mem[0,1,0,1]1207; AVX2-FP-NEXT: vpermps %ymm0, %ymm4, %ymm01208; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6,7]1209; AVX2-FP-NEXT: vmovaps %ymm3, (%rsi)1210; AVX2-FP-NEXT: vmovaps %ymm7, (%rdx)1211; AVX2-FP-NEXT: vmovaps %ymm10, (%rcx)1212; AVX2-FP-NEXT: vmovaps %ymm11, (%r8)1213; AVX2-FP-NEXT: vmovaps %ymm1, (%r9)1214; AVX2-FP-NEXT: vmovaps %ymm0, (%rax)1215; AVX2-FP-NEXT: vzeroupper1216; AVX2-FP-NEXT: retq1217;1218; AVX2-FCP-LABEL: load_i32_stride6_vf8:1219; AVX2-FCP: # %bb.0:1220; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1221; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm01222; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm11223; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm51224; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm21225; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm41226; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm61227; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm3 = [0,6,4,u]1228; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3],ymm4[4,5],ymm2[6,7]1229; AVX2-FCP-NEXT: vpermps %ymm7, %ymm3, %ymm31230; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm6[0,1],ymm5[0,1]1231; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm8[0,1,2,3,4,5],ymm5[6,7]1232; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm8 = ymm9[0,2,2,2,4,6,6,6]1233; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm8[3,4,5,6,7]1234; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5,6,7]1235; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm8 = [4,2,4,2,4,2,4,2]1236; AVX2-FCP-NEXT: vpermps %ymm10, %ymm8, %ymm111237; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm11[6,7]1238; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm11 = [1,7,5,u]1239; AVX2-FCP-NEXT: vpermps %ymm7, %ymm11, %ymm71240; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[1,3,2,3,5,7,6,7]1241; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm9[3,4,5,6,7]1242; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm9 = [5,3,5,3,5,3,5,3]1243; AVX2-FCP-NEXT: vpermps %ymm10, %ymm9, %ymm101244; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm10[6,7]1245; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm6[0,1],ymm5[2,3],ymm6[4,5],ymm5[6,7]1246; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm11 = [2,4,2,4,2,4,2,4]1247; AVX2-FCP-NEXT: vpermps %ymm10, %ymm11, %ymm101248; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm11 = [2,0,6,7]1249; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm4[0,1],ymm2[2,3],ymm4[4,5,6,7]1250; AVX2-FCP-NEXT: vpermps %ymm12, %ymm11, %ymm111251; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2],ymm10[3,4,5,6,7]1252; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]1253; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm13 = [0,0,6,4,0,0,6,4]1254; AVX2-FCP-NEXT: # ymm13 = mem[0,1,0,1]1255; AVX2-FCP-NEXT: vpermps %ymm11, %ymm13, %ymm131256; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4],ymm13[5,6,7]1257; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm13 = [0,1,7,5,0,1,7,5]1258; AVX2-FCP-NEXT: # ymm13 = mem[0,1,0,1]1259; AVX2-FCP-NEXT: vpermps %ymm11, %ymm13, %ymm111260; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm13 = mem[3,3,3,3]1261; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm6[1],ymm13[2,3,4],ymm6[5],ymm13[6,7]1262; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4],ymm13[5],ymm12[6,7]1263; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm13 = [3,1,7,5,0,u,u,u]1264; AVX2-FCP-NEXT: vpermps %ymm12, %ymm13, %ymm121265; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5,6,7]1266; AVX2-FCP-NEXT: vmovaps 80(%rdi), %xmm121267; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]1268; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm5[0,1],ymm12[2,3],ymm5[4,5,6,7]1269; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm6[0,1,0,2,4,5,4,6]1270; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]1271; AVX2-FCP-NEXT: vpermps %ymm2, %ymm8, %ymm41272; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm6[2,3,4,5,6,7]1273; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1274; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [0,2,0,6,0,2,0,6]1275; AVX2-FCP-NEXT: # ymm1 = mem[0,1,0,1]1276; AVX2-FCP-NEXT: vpermps %ymm0, %ymm1, %ymm11277; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4],ymm1[5,6,7]1278; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm5[1,1,1,1,5,5,5,5]1279; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm12[3],ymm4[4,5,6,7]1280; AVX2-FCP-NEXT: vpermps %ymm2, %ymm9, %ymm21281; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2,3,4,5,6,7]1282; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [0,3,1,7,0,3,1,7]1283; AVX2-FCP-NEXT: # ymm4 = mem[0,1,0,1]1284; AVX2-FCP-NEXT: vpermps %ymm0, %ymm4, %ymm01285; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6,7]1286; AVX2-FCP-NEXT: vmovaps %ymm3, (%rsi)1287; AVX2-FCP-NEXT: vmovaps %ymm7, (%rdx)1288; AVX2-FCP-NEXT: vmovaps %ymm10, (%rcx)1289; AVX2-FCP-NEXT: vmovaps %ymm11, (%r8)1290; AVX2-FCP-NEXT: vmovaps %ymm1, (%r9)1291; AVX2-FCP-NEXT: vmovaps %ymm0, (%rax)1292; AVX2-FCP-NEXT: vzeroupper1293; AVX2-FCP-NEXT: retq1294;1295; AVX512-LABEL: load_i32_stride6_vf8:1296; AVX512: # %bb.0:1297; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax1298; AVX512-NEXT: vmovdqa 128(%rdi), %ymm01299; AVX512-NEXT: vmovdqa 160(%rdi), %ymm11300; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1301; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1302; AVX512-NEXT: vmovdqa64 (%rdi), %zmm41303; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm51304; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm61305; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm31306; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1307; AVX512-NEXT: vpermi2d %ymm2, %ymm3, %ymm71308; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1309; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm31310; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1311; AVX512-NEXT: vpermi2d %ymm2, %ymm3, %ymm81312; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1313; AVX512-NEXT: vpermd %zmm6, %zmm2, %zmm21314; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1315; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm31316; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1317; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1318; AVX512-NEXT: vpermd %zmm6, %zmm3, %zmm31319; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1320; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm61321; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1322; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1323; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1324; AVX512-NEXT: vpermi2d %zmm4, %zmm5, %zmm11325; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1326; AVX512-NEXT: vpermi2d %ymm0, %ymm1, %ymm61327; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1328; AVX512-NEXT: vpermi2d %zmm4, %zmm5, %zmm11329; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1330; AVX512-NEXT: vpermi2d %ymm0, %ymm1, %ymm41331; AVX512-NEXT: vmovdqa %ymm7, (%rsi)1332; AVX512-NEXT: vmovdqa %ymm8, (%rdx)1333; AVX512-NEXT: vmovdqa %ymm2, (%rcx)1334; AVX512-NEXT: vmovdqa %ymm3, (%r8)1335; AVX512-NEXT: vmovdqa %ymm6, (%r9)1336; AVX512-NEXT: vmovdqa %ymm4, (%rax)1337; AVX512-NEXT: vzeroupper1338; AVX512-NEXT: retq1339;1340; AVX512-FCP-LABEL: load_i32_stride6_vf8:1341; AVX512-FCP: # %bb.0:1342; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1343; AVX512-FCP-NEXT: vmovdqa 128(%rdi), %ymm01344; AVX512-FCP-NEXT: vmovdqa 160(%rdi), %ymm11345; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1346; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1347; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm41348; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm51349; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm61350; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31351; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1352; AVX512-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm71353; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1354; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31355; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1356; AVX512-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm81357; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1358; AVX512-FCP-NEXT: vpermd %zmm6, %zmm2, %zmm21359; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1360; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31361; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1362; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1363; AVX512-FCP-NEXT: vpermd %zmm6, %zmm3, %zmm31364; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1365; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm61366; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1367; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1368; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1369; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11370; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1371; AVX512-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm61372; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1373; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11374; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1375; AVX512-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm41376; AVX512-FCP-NEXT: vmovdqa %ymm7, (%rsi)1377; AVX512-FCP-NEXT: vmovdqa %ymm8, (%rdx)1378; AVX512-FCP-NEXT: vmovdqa %ymm2, (%rcx)1379; AVX512-FCP-NEXT: vmovdqa %ymm3, (%r8)1380; AVX512-FCP-NEXT: vmovdqa %ymm6, (%r9)1381; AVX512-FCP-NEXT: vmovdqa %ymm4, (%rax)1382; AVX512-FCP-NEXT: vzeroupper1383; AVX512-FCP-NEXT: retq1384;1385; AVX512DQ-LABEL: load_i32_stride6_vf8:1386; AVX512DQ: # %bb.0:1387; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax1388; AVX512DQ-NEXT: vmovdqa 128(%rdi), %ymm01389; AVX512DQ-NEXT: vmovdqa 160(%rdi), %ymm11390; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1391; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1392; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm41393; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm51394; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm61395; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm31396; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1397; AVX512DQ-NEXT: vpermi2d %ymm2, %ymm3, %ymm71398; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1399; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm31400; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1401; AVX512DQ-NEXT: vpermi2d %ymm2, %ymm3, %ymm81402; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1403; AVX512DQ-NEXT: vpermd %zmm6, %zmm2, %zmm21404; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1405; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm31406; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1407; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1408; AVX512DQ-NEXT: vpermd %zmm6, %zmm3, %zmm31409; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1410; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm61411; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1412; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1413; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1414; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm5, %zmm11415; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1416; AVX512DQ-NEXT: vpermi2d %ymm0, %ymm1, %ymm61417; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1418; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm5, %zmm11419; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1420; AVX512DQ-NEXT: vpermi2d %ymm0, %ymm1, %ymm41421; AVX512DQ-NEXT: vmovdqa %ymm7, (%rsi)1422; AVX512DQ-NEXT: vmovdqa %ymm8, (%rdx)1423; AVX512DQ-NEXT: vmovdqa %ymm2, (%rcx)1424; AVX512DQ-NEXT: vmovdqa %ymm3, (%r8)1425; AVX512DQ-NEXT: vmovdqa %ymm6, (%r9)1426; AVX512DQ-NEXT: vmovdqa %ymm4, (%rax)1427; AVX512DQ-NEXT: vzeroupper1428; AVX512DQ-NEXT: retq1429;1430; AVX512DQ-FCP-LABEL: load_i32_stride6_vf8:1431; AVX512DQ-FCP: # %bb.0:1432; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1433; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdi), %ymm01434; AVX512DQ-FCP-NEXT: vmovdqa 160(%rdi), %ymm11435; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1436; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1437; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm41438; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm51439; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm61440; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31441; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1442; AVX512DQ-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm71443; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1444; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31445; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1446; AVX512DQ-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm81447; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1448; AVX512DQ-FCP-NEXT: vpermd %zmm6, %zmm2, %zmm21449; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1450; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31451; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1452; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1453; AVX512DQ-FCP-NEXT: vpermd %zmm6, %zmm3, %zmm31454; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1455; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm61456; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1457; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1458; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1459; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11460; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1461; AVX512DQ-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm61462; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1463; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11464; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1465; AVX512DQ-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm41466; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, (%rsi)1467; AVX512DQ-FCP-NEXT: vmovdqa %ymm8, (%rdx)1468; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, (%rcx)1469; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, (%r8)1470; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, (%r9)1471; AVX512DQ-FCP-NEXT: vmovdqa %ymm4, (%rax)1472; AVX512DQ-FCP-NEXT: vzeroupper1473; AVX512DQ-FCP-NEXT: retq1474;1475; AVX512BW-LABEL: load_i32_stride6_vf8:1476; AVX512BW: # %bb.0:1477; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax1478; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm01479; AVX512BW-NEXT: vmovdqa 160(%rdi), %ymm11480; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1481; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1482; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm41483; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm51484; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm61485; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm31486; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1487; AVX512BW-NEXT: vpermi2d %ymm2, %ymm3, %ymm71488; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1489; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm31490; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1491; AVX512BW-NEXT: vpermi2d %ymm2, %ymm3, %ymm81492; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1493; AVX512BW-NEXT: vpermd %zmm6, %zmm2, %zmm21494; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1495; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm31496; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1497; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1498; AVX512BW-NEXT: vpermd %zmm6, %zmm3, %zmm31499; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1500; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm61501; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1502; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1503; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1504; AVX512BW-NEXT: vpermi2d %zmm4, %zmm5, %zmm11505; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1506; AVX512BW-NEXT: vpermi2d %ymm0, %ymm1, %ymm61507; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1508; AVX512BW-NEXT: vpermi2d %zmm4, %zmm5, %zmm11509; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1510; AVX512BW-NEXT: vpermi2d %ymm0, %ymm1, %ymm41511; AVX512BW-NEXT: vmovdqa %ymm7, (%rsi)1512; AVX512BW-NEXT: vmovdqa %ymm8, (%rdx)1513; AVX512BW-NEXT: vmovdqa %ymm2, (%rcx)1514; AVX512BW-NEXT: vmovdqa %ymm3, (%r8)1515; AVX512BW-NEXT: vmovdqa %ymm6, (%r9)1516; AVX512BW-NEXT: vmovdqa %ymm4, (%rax)1517; AVX512BW-NEXT: vzeroupper1518; AVX512BW-NEXT: retq1519;1520; AVX512BW-FCP-LABEL: load_i32_stride6_vf8:1521; AVX512BW-FCP: # %bb.0:1522; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1523; AVX512BW-FCP-NEXT: vmovdqa 128(%rdi), %ymm01524; AVX512BW-FCP-NEXT: vmovdqa 160(%rdi), %ymm11525; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1526; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1527; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm41528; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm51529; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm61530; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31531; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1532; AVX512BW-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm71533; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1534; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31535; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1536; AVX512BW-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm81537; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1538; AVX512BW-FCP-NEXT: vpermd %zmm6, %zmm2, %zmm21539; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1540; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31541; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1542; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1543; AVX512BW-FCP-NEXT: vpermd %zmm6, %zmm3, %zmm31544; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1545; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm61546; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1547; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1548; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1549; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11550; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1551; AVX512BW-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm61552; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1553; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11554; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1555; AVX512BW-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm41556; AVX512BW-FCP-NEXT: vmovdqa %ymm7, (%rsi)1557; AVX512BW-FCP-NEXT: vmovdqa %ymm8, (%rdx)1558; AVX512BW-FCP-NEXT: vmovdqa %ymm2, (%rcx)1559; AVX512BW-FCP-NEXT: vmovdqa %ymm3, (%r8)1560; AVX512BW-FCP-NEXT: vmovdqa %ymm6, (%r9)1561; AVX512BW-FCP-NEXT: vmovdqa %ymm4, (%rax)1562; AVX512BW-FCP-NEXT: vzeroupper1563; AVX512BW-FCP-NEXT: retq1564;1565; AVX512DQ-BW-LABEL: load_i32_stride6_vf8:1566; AVX512DQ-BW: # %bb.0:1567; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax1568; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %ymm01569; AVX512DQ-BW-NEXT: vmovdqa 160(%rdi), %ymm11570; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1571; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1572; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm41573; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm51574; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm61575; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm31576; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1577; AVX512DQ-BW-NEXT: vpermi2d %ymm2, %ymm3, %ymm71578; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1579; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm31580; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1581; AVX512DQ-BW-NEXT: vpermi2d %ymm2, %ymm3, %ymm81582; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1583; AVX512DQ-BW-NEXT: vpermd %zmm6, %zmm2, %zmm21584; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1585; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm31586; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1587; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1588; AVX512DQ-BW-NEXT: vpermd %zmm6, %zmm3, %zmm31589; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1590; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm61591; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1592; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1593; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1594; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm5, %zmm11595; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1596; AVX512DQ-BW-NEXT: vpermi2d %ymm0, %ymm1, %ymm61597; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1598; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm5, %zmm11599; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1600; AVX512DQ-BW-NEXT: vpermi2d %ymm0, %ymm1, %ymm41601; AVX512DQ-BW-NEXT: vmovdqa %ymm7, (%rsi)1602; AVX512DQ-BW-NEXT: vmovdqa %ymm8, (%rdx)1603; AVX512DQ-BW-NEXT: vmovdqa %ymm2, (%rcx)1604; AVX512DQ-BW-NEXT: vmovdqa %ymm3, (%r8)1605; AVX512DQ-BW-NEXT: vmovdqa %ymm6, (%r9)1606; AVX512DQ-BW-NEXT: vmovdqa %ymm4, (%rax)1607; AVX512DQ-BW-NEXT: vzeroupper1608; AVX512DQ-BW-NEXT: retq1609;1610; AVX512DQ-BW-FCP-LABEL: load_i32_stride6_vf8:1611; AVX512DQ-BW-FCP: # %bb.0:1612; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax1613; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdi), %ymm01614; AVX512DQ-BW-FCP-NEXT: vmovdqa 160(%rdi), %ymm11615; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]1616; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]1617; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm41618; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm51619; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm61620; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31621; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,1,2,3,4,5,12,10]1622; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm71623; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,7,13,19,25,31,0,0]1624; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31625; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,2,3,4,5,13,11]1626; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm2, %ymm3, %ymm81627; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,2,8,0,0,6,12]1628; AVX512DQ-BW-FCP-NEXT: vpermd %zmm6, %zmm2, %zmm21629; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [2,8,14,20,26,0,0,0]1630; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm31631; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]1632; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,1,3,9,0,1,7,13]1633; AVX512DQ-BW-FCP-NEXT: vpermd %zmm6, %zmm3, %zmm31634; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,9,15,21,27,0,0,0]1635; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm4, %zmm61636; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4],ymm3[5,6,7]1637; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]1638; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [20,26,0,6,12,0,0,0]1639; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11640; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,1,2,3,4,10,8,14]1641; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm61642; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [21,27,1,7,13,0,0,0]1643; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm5, %zmm11644; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,2,3,4,11,9,15]1645; AVX512DQ-BW-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm41646; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm7, (%rsi)1647; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm8, (%rdx)1648; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm2, (%rcx)1649; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm3, (%r8)1650; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm6, (%r9)1651; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm4, (%rax)1652; AVX512DQ-BW-FCP-NEXT: vzeroupper1653; AVX512DQ-BW-FCP-NEXT: retq1654 %wide.vec = load <48 x i32>, ptr %in.vec, align 641655 %strided.vec0 = shufflevector <48 x i32> %wide.vec, <48 x i32> poison, <8 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42>1656 %strided.vec1 = shufflevector <48 x i32> %wide.vec, <48 x i32> poison, <8 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43>1657 %strided.vec2 = shufflevector <48 x i32> %wide.vec, <48 x i32> poison, <8 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44>1658 %strided.vec3 = shufflevector <48 x i32> %wide.vec, <48 x i32> poison, <8 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45>1659 %strided.vec4 = shufflevector <48 x i32> %wide.vec, <48 x i32> poison, <8 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46>1660 %strided.vec5 = shufflevector <48 x i32> %wide.vec, <48 x i32> poison, <8 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47>1661 store <8 x i32> %strided.vec0, ptr %out.vec0, align 641662 store <8 x i32> %strided.vec1, ptr %out.vec1, align 641663 store <8 x i32> %strided.vec2, ptr %out.vec2, align 641664 store <8 x i32> %strided.vec3, ptr %out.vec3, align 641665 store <8 x i32> %strided.vec4, ptr %out.vec4, align 641666 store <8 x i32> %strided.vec5, ptr %out.vec5, align 641667 ret void1668}1669 1670define void @load_i32_stride6_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {1671; SSE-LABEL: load_i32_stride6_vf16:1672; SSE: # %bb.0:1673; SSE-NEXT: subq $408, %rsp # imm = 0x1981674; SSE-NEXT: movdqa 240(%rdi), %xmm91675; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1676; SSE-NEXT: movdqa 256(%rdi), %xmm31677; SSE-NEXT: movdqa 192(%rdi), %xmm101678; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1679; SSE-NEXT: movdqa 208(%rdi), %xmm41680; SSE-NEXT: movdqa 336(%rdi), %xmm141681; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1682; SSE-NEXT: movdqa 352(%rdi), %xmm51683; SSE-NEXT: movdqa 288(%rdi), %xmm151684; SSE-NEXT: movdqa 304(%rdi), %xmm71685; SSE-NEXT: movdqa 64(%rdi), %xmm121686; SSE-NEXT: movdqa (%rdi), %xmm81687; SSE-NEXT: movdqa 16(%rdi), %xmm111688; SSE-NEXT: movdqa 48(%rdi), %xmm131689; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[2,3,2,3]1690; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1691; SSE-NEXT: movdqa %xmm8, %xmm11692; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1693; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1694; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,2,3,3]1695; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1696; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[0,0,1,1]1697; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]1698; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]1699; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1700; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]1701; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1702; SSE-NEXT: movdqa %xmm15, %xmm11703; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1704; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]1705; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1706; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm14[0,0,1,1]1707; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]1708; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]1709; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1710; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]1711; SSE-NEXT: movdqa %xmm4, %xmm141712; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1713; SSE-NEXT: movdqa %xmm10, %xmm11714; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1715; SSE-NEXT: movdqa %xmm3, %xmm21716; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1717; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,2,3,3]1718; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm9[0,0,1,1]1719; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]1720; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]1721; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1722; SSE-NEXT: movdqa 96(%rdi), %xmm41723; SSE-NEXT: movdqa 112(%rdi), %xmm101724; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[2,3,2,3]1725; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1726; SSE-NEXT: movdqa %xmm4, %xmm11727; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1728; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1729; SSE-NEXT: movdqa 144(%rdi), %xmm91730; SSE-NEXT: movdqa 160(%rdi), %xmm61731; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,2,3,3]1732; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1733; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm9[0,0,1,1]1734; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]1735; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]1736; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1737; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[1,1,1,1]1738; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[3,3,3,3]1739; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1740; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[2,3,2,3]1741; SSE-NEXT: movdqa %xmm13, %xmm31742; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]1743; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]1744; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1745; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[1,1,1,1]1746; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[3,3,3,3]1747; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1748; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,3,2,3]1749; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload1750; SSE-NEXT: movdqa %xmm5, %xmm31751; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]1752; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]1753; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1754; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload1755; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,1,1]1756; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[3,3,3,3]1757; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1758; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,3,2,3]1759; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload1760; SSE-NEXT: movdqa %xmm3, %xmm21761; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]1762; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]1763; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1764; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,1,1]1765; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[3,3,3,3]1766; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1767; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[2,3,2,3]1768; SSE-NEXT: movdqa %xmm9, %xmm141769; SSE-NEXT: movdqa %xmm9, %xmm21770; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]1771; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]1772; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1773; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm13[2,2,3,3]1774; SSE-NEXT: movdqa 80(%rdi), %xmm101775; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[0,0,1,1]1776; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1777; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload1778; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,3,2,3]1779; SSE-NEXT: movdqa 32(%rdi), %xmm61780; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]1781; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1782; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1783; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3]1784; SSE-NEXT: movdqa 368(%rdi), %xmm01785; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1786; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]1787; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1788; SSE-NEXT: movdqa %xmm15, %xmm91789; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[2,3,2,3]1790; SSE-NEXT: movdqa 320(%rdi), %xmm81791; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1]1792; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1793; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1794; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[2,2,3,3]1795; SSE-NEXT: movdqa 272(%rdi), %xmm151796; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[0,0,1,1]1797; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1798; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]1799; SSE-NEXT: movdqa 224(%rdi), %xmm31800; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]1801; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1802; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1803; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[2,2,3,3]1804; SSE-NEXT: movdqa 176(%rdi), %xmm111805; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[0,0,1,1]1806; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1807; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload1808; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,3,2,3]1809; SSE-NEXT: movdqa 128(%rdi), %xmm41810; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]1811; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1812; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1813; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[3,3,3,3]1814; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1815; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[1,1,1,1]1816; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1817; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm13[2,3,2,3]1818; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1819; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1]1820; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1821; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1822; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[3,3,3,3]1823; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[1,1,1,1]1824; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1825; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1826; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1827; SSE-NEXT: # xmm1 = mem[2,3,2,3]1828; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1829; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]1830; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1831; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1832; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[3,3,3,3]1833; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,1,1,1]1834; SSE-NEXT: movdqa %xmm3, %xmm91835; SSE-NEXT: movdqa %xmm3, (%rsp) # 16-byte Spill1836; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1837; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload1838; SSE-NEXT: # xmm12 = mem[2,3,2,3]1839; SSE-NEXT: movdqa %xmm15, %xmm31840; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1841; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm15[0],xmm12[1],xmm15[1]1842; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm0[0],xmm12[1]1843; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[3,3,3,3]1844; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1]1845; SSE-NEXT: movdqa %xmm4, %xmm71846; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1847; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1848; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm14[2,3,2,3]1849; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm11[0],xmm15[1],xmm11[1]1850; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1851; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]1852; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]1853; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload1854; SSE-NEXT: movdqa %xmm6, %xmm11855; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1856; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[2,2,3,3]1857; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload1858; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm10[0,0,1,1]1859; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]1860; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm1[0],xmm5[1]1861; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[2,3,2,3]1862; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload1863; SSE-NEXT: movdqa %xmm14, %xmm11864; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1865; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,3,3]1866; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload1867; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm8[0,0,1,1]1868; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]1869; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm1[0],xmm4[1]1870; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[2,3,2,3]1871; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1872; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1873; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[2,2,3,3]1874; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload1875; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm9[0,0,1,1]1876; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]1877; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]1878; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]1879; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload1880; SSE-NEXT: movdqa %xmm7, %xmm11881; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1882; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[2,2,3,3]1883; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload1884; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[0,0,1,1]1885; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]1886; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]1887; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,1,1]1888; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1889; SSE-NEXT: # xmm1 = mem[3,3,3,3]1890; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1891; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1892; SSE-NEXT: # xmm1 = mem[2,3,2,3]1893; SSE-NEXT: punpckldq {{.*#+}} xmm10 = xmm10[0],xmm1[0],xmm10[1],xmm1[1]1894; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm0[0],xmm10[1]1895; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[1,1,1,1]1896; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1897; SSE-NEXT: # xmm1 = mem[3,3,3,3]1898; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1899; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1900; SSE-NEXT: # xmm1 = mem[2,3,2,3]1901; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm1[0],xmm8[1],xmm1[1]1902; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm0[0],xmm8[1]1903; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1904; SSE-NEXT: # xmm0 = mem[1,1,1,1]1905; SSE-NEXT: pshufd $255, (%rsp), %xmm1 # 16-byte Folded Reload1906; SSE-NEXT: # xmm1 = mem[3,3,3,3]1907; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1908; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1909; SSE-NEXT: # xmm1 = mem[2,3,2,3]1910; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1]1911; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm0[0],xmm9[1]1912; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,1,1]1913; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1914; SSE-NEXT: # xmm1 = mem[3,3,3,3]1915; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]1916; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1917; SSE-NEXT: # xmm1 = mem[2,3,2,3]1918; SSE-NEXT: punpckldq {{.*#+}} xmm13 = xmm13[0],xmm1[0],xmm13[1],xmm1[1]1919; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm0[0],xmm13[1]1920; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1921; SSE-NEXT: movaps %xmm0, 16(%rsi)1922; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1923; SSE-NEXT: movaps %xmm0, 32(%rsi)1924; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1925; SSE-NEXT: movaps %xmm0, 48(%rsi)1926; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1927; SSE-NEXT: movaps %xmm0, (%rsi)1928; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1929; SSE-NEXT: movaps %xmm0, 16(%rdx)1930; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1931; SSE-NEXT: movaps %xmm0, 32(%rdx)1932; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1933; SSE-NEXT: movaps %xmm0, 48(%rdx)1934; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1935; SSE-NEXT: movaps %xmm0, (%rdx)1936; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1937; SSE-NEXT: movaps %xmm0, 16(%rcx)1938; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1939; SSE-NEXT: movaps %xmm0, 32(%rcx)1940; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1941; SSE-NEXT: movaps %xmm0, 48(%rcx)1942; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1943; SSE-NEXT: movaps %xmm0, (%rcx)1944; SSE-NEXT: movapd %xmm15, 16(%r8)1945; SSE-NEXT: movapd %xmm12, 32(%r8)1946; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1947; SSE-NEXT: movaps %xmm0, 48(%r8)1948; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1949; SSE-NEXT: movaps %xmm0, (%r8)1950; SSE-NEXT: movapd %xmm2, 16(%r9)1951; SSE-NEXT: movapd %xmm3, 32(%r9)1952; SSE-NEXT: movapd %xmm4, 48(%r9)1953; SSE-NEXT: movapd %xmm5, (%r9)1954; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax1955; SSE-NEXT: movapd %xmm13, 16(%rax)1956; SSE-NEXT: movapd %xmm9, 32(%rax)1957; SSE-NEXT: movapd %xmm8, 48(%rax)1958; SSE-NEXT: movapd %xmm10, (%rax)1959; SSE-NEXT: addq $408, %rsp # imm = 0x1981960; SSE-NEXT: retq1961;1962; AVX-LABEL: load_i32_stride6_vf16:1963; AVX: # %bb.0:1964; AVX-NEXT: subq $264, %rsp # imm = 0x1081965; AVX-NEXT: vmovaps 224(%rdi), %ymm51966; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1967; AVX-NEXT: vmovaps 192(%rdi), %ymm101968; AVX-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1969; AVX-NEXT: vmovapd 160(%rdi), %ymm01970; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1971; AVX-NEXT: vmovapd 128(%rdi), %ymm71972; AVX-NEXT: vmovaps 96(%rdi), %ymm91973; AVX-NEXT: vmovaps 64(%rdi), %ymm151974; AVX-NEXT: vmovaps 32(%rdi), %ymm31975; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1976; AVX-NEXT: vmovaps (%rdi), %ymm21977; AVX-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill1978; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5],ymm2[6,7]1979; AVX-NEXT: vextractf128 $1, %ymm2, %xmm61980; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm2[0,1],xmm6[2,3]1981; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,2],xmm6[0,3]1982; AVX-NEXT: vinsertf128 $1, 96(%rdi), %ymm15, %ymm81983; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm9[2,0],ymm8[0,0],ymm9[6,4],ymm8[4,4]1984; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,0],ymm15[2,2],ymm4[6,4],ymm15[6,6]1985; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3,4,5,6,7]1986; AVX-NEXT: vmovapd %ymm7, %ymm11987; AVX-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm0[0,1]1988; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm7[0],ymm1[1],ymm7[3],ymm1[2]1989; AVX-NEXT: vmovapd %ymm1, %ymm131990; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,2,0,4,5,6,4]1991; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm4[6,7]1992; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1993; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0,1,2,3],ymm5[4,5],ymm10[6,7]1994; AVX-NEXT: vextractf128 $1, %ymm1, %xmm01995; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm1[0,1],xmm0[2,3]1996; AVX-NEXT: vshufps {{.*#+}} xmm12 = xmm3[0,2],xmm0[0,3]1997; AVX-NEXT: vmovaps 288(%rdi), %ymm111998; AVX-NEXT: vmovaps 256(%rdi), %ymm101999; AVX-NEXT: vinsertf128 $1, 288(%rdi), %ymm10, %ymm52000; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm11[2,0],ymm5[0,0],ymm11[6,4],ymm5[4,4]2001; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm14[2,0],ymm10[2,2],ymm14[6,4],ymm10[6,6]2002; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm12[0,1,2],ymm14[3,4,5,6,7]2003; AVX-NEXT: vmovapd 352(%rdi), %ymm32004; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2005; AVX-NEXT: vmovapd 320(%rdi), %ymm122006; AVX-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm12[2,3],ymm3[0,1]2007; AVX-NEXT: vshufpd {{.*#+}} ymm14 = ymm3[0],ymm12[1],ymm3[3],ymm12[2]2008; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm14[0,1,2,0,4,5,6,4]2009; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm14[6,7]2010; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2011; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,0],xmm6[3,0]2012; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm6[1,3]2013; AVX-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2014; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm9[3,0],ymm8[1,0],ymm9[7,4],ymm8[5,4]2015; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,0],ymm15[2,3],ymm4[6,4],ymm15[6,7]2016; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3,4,5,6,7]2017; AVX-NEXT: vmovupd %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2018; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm7[3,1],ymm13[1,3],ymm7[7,5],ymm13[5,7]2019; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,2,0,4,5,6,4]2020; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]2021; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2022; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,0],xmm0[3,0]2023; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm1[0,2],xmm0[1,3]2024; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm11[3,0],ymm5[1,0],ymm11[7,4],ymm5[5,4]2025; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm10[2,3],ymm1[6,4],ymm10[6,7]2026; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]2027; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm3[3,1],ymm12[1,3],ymm3[7,5],ymm12[5,7]2028; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]2029; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]2030; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2031; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2032; AVX-NEXT: vblendps $12, (%rsp), %ymm0, %ymm2 # 32-byte Folded Reload2033; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]2034; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm9[2,1],ymm15[2,0],ymm9[6,5],ymm15[6,4]2035; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]2036; AVX-NEXT: vextractf128 $1, %ymm2, %xmm72037; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[2,0],xmm7[2,3]2038; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3,4,5,6,7]2039; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2040; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm13[0,1,2,3],ymm9[4,5],ymm13[6,7]2041; AVX-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm3[2,3,0,1]2042; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm4[0,0],ymm3[2,0],ymm4[4,4],ymm3[6,4]2043; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm5[5,6,7]2044; AVX-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill2045; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2046; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload2047; AVX-NEXT: # ymm8 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]2048; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm11[2,1],ymm10[2,0],ymm11[6,5],ymm10[6,4]2049; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]2050; AVX-NEXT: vextractf128 $1, %ymm8, %xmm142051; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm8[2,0],xmm14[2,3]2052; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3,4,5,6,7]2053; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload2054; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm12[0,1,2,3],ymm13[4,5],ymm12[6,7]2055; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm6[2,3,0,1]2056; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm0[0,0],ymm6[2,0],ymm0[4,4],ymm6[6,4]2057; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm5[5,6,7]2058; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2059; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm4[0,1],ymm3[3,1],ymm4[4,5],ymm3[7,5]2060; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,1],xmm7[3,3]2061; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2062; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm1[3,1],ymm15[2,1],ymm1[7,5],ymm15[6,5]2063; AVX-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3,0,1]2064; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3,4,5,6,7]2065; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6,7]2066; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2067; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm6[3,1],ymm0[4,5],ymm6[7,5]2068; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm8[3,1],xmm14[3,3]2069; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm11[3,1],ymm10[2,1],ymm11[7,5],ymm10[6,5]2070; AVX-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3,0,1]2071; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3,4,5,6,7]2072; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6,7]2073; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2074; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload2075; AVX-NEXT: # ymm9 = ymm9[0,1],mem[2,3],ymm9[4,5,6,7]2076; AVX-NEXT: vmovaps 32(%rdi), %xmm32077; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm3[2,2,3,3]2078; AVX-NEXT: vmovaps 16(%rdi), %xmm52079; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm5[0],xmm2[1],xmm5[2,3]2080; AVX-NEXT: vmovapd 80(%rdi), %xmm62081; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm6[1],ymm15[0],ymm6[2],ymm15[3]2082; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm1[0,1],ymm7[2,0],ymm1[4,5],ymm7[6,4]2083; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm7[2,3,4,5,6,7]2084; AVX-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm9[2,3,0,1]2085; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm9[2,0],ymm7[0,0],ymm9[6,4],ymm7[4,4]2086; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm7[0,2],ymm8[2,0],ymm7[4,6],ymm8[6,4]2087; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm8[5,6,7]2088; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2089; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm13[0,1],ymm12[2,3],ymm13[4,5,6,7]2090; AVX-NEXT: vmovaps 224(%rdi), %xmm122091; AVX-NEXT: vshufps {{.*#+}} xmm14 = xmm12[2,2,3,3]2092; AVX-NEXT: vmovaps 208(%rdi), %xmm02093; AVX-NEXT: vblendps {{.*#+}} xmm14 = xmm0[0],xmm14[1],xmm0[2,3]2094; AVX-NEXT: vmovapd 272(%rdi), %xmm12095; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm1[1],ymm10[0],ymm1[2],ymm10[3]2096; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm11[0,1],ymm4[2,0],ymm11[4,5],ymm4[6,4]2097; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm14[0,1],ymm4[2,3,4,5,6,7]2098; AVX-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm8[2,3,0,1]2099; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm8[2,0],ymm14[0,0],ymm8[6,4],ymm14[4,4]2100; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm14[0,2],ymm2[2,0],ymm14[4,6],ymm2[6,4]2101; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5,6,7]2102; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm9[3,0],ymm7[1,0],ymm9[7,4],ymm7[5,4]2103; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm7[0,3],ymm4[2,0],ymm7[4,7],ymm4[6,4]2104; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3]2105; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm6[3,1],ymm15[1,3],ymm6[7,5],ymm15[5,7]2106; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2107; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm6[1,1],ymm5[2,0],ymm6[5,5],ymm5[6,4]2108; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,3,2,3]2109; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm5[2,3,4,5,6,7]2110; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm4[5,6,7]2111; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm8[3,0],ymm14[1,0],ymm8[7,4],ymm14[5,4]2112; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm14[0,3],ymm4[2,0],ymm14[4,7],ymm4[6,4]2113; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm12[2,3]2114; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,1],ymm10[1,3],ymm1[7,5],ymm10[5,7]2115; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm11[1,1],ymm1[2,0],ymm11[5,5],ymm1[6,4]2116; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3,2,3]2117; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]2118; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5,6,7]2119; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2120; AVX-NEXT: vmovaps %ymm1, 32(%rsi)2121; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2122; AVX-NEXT: vmovaps %ymm1, (%rsi)2123; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2124; AVX-NEXT: vmovaps %ymm1, 32(%rdx)2125; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2126; AVX-NEXT: vmovaps %ymm1, (%rdx)2127; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2128; AVX-NEXT: vmovaps %ymm1, 32(%rcx)2129; AVX-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload2130; AVX-NEXT: vmovaps %ymm1, (%rcx)2131; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2132; AVX-NEXT: vmovaps %ymm1, 32(%r8)2133; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2134; AVX-NEXT: vmovaps %ymm1, (%r8)2135; AVX-NEXT: vmovaps %ymm2, 32(%r9)2136; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2137; AVX-NEXT: vmovaps %ymm1, (%r9)2138; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax2139; AVX-NEXT: vmovaps %ymm0, 32(%rax)2140; AVX-NEXT: vmovaps %ymm3, (%rax)2141; AVX-NEXT: addq $264, %rsp # imm = 0x1082142; AVX-NEXT: vzeroupper2143; AVX-NEXT: retq2144;2145; AVX2-LABEL: load_i32_stride6_vf16:2146; AVX2: # %bb.0:2147; AVX2-NEXT: subq $392, %rsp # imm = 0x1882148; AVX2-NEXT: vmovaps 288(%rdi), %ymm102149; AVX2-NEXT: vmovaps 224(%rdi), %ymm12150; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2151; AVX2-NEXT: vmovaps 192(%rdi), %ymm22152; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2153; AVX2-NEXT: vmovaps 128(%rdi), %ymm02154; AVX2-NEXT: vmovaps 160(%rdi), %ymm32155; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2156; AVX2-NEXT: vmovaps 96(%rdi), %ymm152157; AVX2-NEXT: vmovaps (%rdi), %ymm42158; AVX2-NEXT: vmovups %ymm4, (%rsp) # 32-byte Spill2159; AVX2-NEXT: vmovaps 32(%rdi), %ymm52160; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2161; AVX2-NEXT: vmovaps 64(%rdi), %ymm132162; AVX2-NEXT: vmovaps {{.*#+}} xmm6 = [0,6,4,u]2163; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm4[0,1,2,3],ymm5[4,5],ymm4[6,7]2164; AVX2-NEXT: vpermps %ymm8, %ymm6, %ymm72165; AVX2-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm13[0,1],ymm15[0,1]2166; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm15[6,7]2167; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm9[0,2,2,2,4,6,6,6]2168; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm7[0,1,2],ymm11[3,4,5,6,7]2169; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm3[0,1,2,3],ymm0[4,5,6,7]2170; AVX2-NEXT: vmovaps %ymm0, %ymm72171; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2172; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm12 = [4,2,4,2,4,2,4,2]2173; AVX2-NEXT: vpermps %ymm4, %ymm12, %ymm142174; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2175; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm14[6,7]2176; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2177; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm1[4,5],ymm2[6,7]2178; AVX2-NEXT: vpermps %ymm3, %ymm6, %ymm02179; AVX2-NEXT: vmovaps 256(%rdi), %ymm112180; AVX2-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2181; AVX2-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm11[0,1],ymm10[0,1]2182; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],ymm10[6,7]2183; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm1[0,2,2,2,4,6,6,6]2184; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2],ymm14[3,4,5,6,7]2185; AVX2-NEXT: vmovaps 320(%rdi), %ymm52186; AVX2-NEXT: vmovaps 352(%rdi), %ymm62187; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm5[4,5,6,7]2188; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2189; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2190; AVX2-NEXT: vpermps %ymm0, %ymm12, %ymm102191; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]2192; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2193; AVX2-NEXT: vmovaps {{.*#+}} xmm2 = [1,7,5,u]2194; AVX2-NEXT: vpermps %ymm8, %ymm2, %ymm82195; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm9[1,3,2,3,5,7,6,7]2196; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2],ymm9[3,4,5,6,7]2197; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm14 = [5,3,5,3,5,3,5,3]2198; AVX2-NEXT: vpermps %ymm4, %ymm14, %ymm42199; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]2200; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2201; AVX2-NEXT: vpermps %ymm3, %ymm2, %ymm22202; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,3,2,3,5,7,6,7]2203; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3,4,5,6,7]2204; AVX2-NEXT: vpermps %ymm0, %ymm14, %ymm02205; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]2206; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2207; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1],ymm15[2,3],ymm13[4,5],ymm15[6,7]2208; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]2209; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]2210; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload2211; AVX2-NEXT: vmovups (%rsp), %ymm9 # 32-byte Reload2212; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0,1],ymm9[2,3],ymm10[4,5,6,7]2213; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm1[2,0,2,3,6,4,6,7]2214; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]2215; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]2216; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload2217; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0,1,2,3],ymm12[4,5],ymm7[6,7]2218; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm2[0,0,2,0,4,4,6,4]2219; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]2220; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]2221; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2222; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2223; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1],ymm7[2,3],ymm11[4,5],ymm7[6,7]2224; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]2225; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]2226; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2227; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload2228; AVX2-NEXT: # ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5,6,7]2229; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm3[2,0,2,3,6,4,6,7]2230; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,3,2,3]2231; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3,4,5,6,7]2232; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm6[4,5],ymm5[6,7]2233; AVX2-NEXT: vshufps {{.*#+}} ymm8 = ymm4[0,0,2,0,4,4,6,4]2234; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2235; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm8[5,6,7]2236; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2237; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm15[3,3,3,3,7,7,7,7]2238; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2,3,4],ymm13[5],ymm0[6,7]2239; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]2240; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,1,3,3,7,5,7,7]2241; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,3,2,3]2242; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]2243; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,1,4,5,7,5]2244; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]2245; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]2246; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm7[3,3,3,3,7,7,7,7]2247; AVX2-NEXT: vmovaps %ymm7, %ymm62248; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2,3,4],ymm11[5],ymm1[6,7]2249; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,0,3]2250; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,1,3,3,7,5,7,7]2251; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]2252; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3,4,5,6,7]2253; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm4[0,1,3,1,4,5,7,5]2254; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]2255; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4],ymm2[5,6,7]2256; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm15[4,5,6,7]2257; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3],ymm9[4,5,6,7]2258; AVX2-NEXT: vmovaps 80(%rdi), %xmm42259; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm4[2,3],ymm2[4,5,6,7]2260; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,0,2,4,5,4,6]2261; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2262; AVX2-NEXT: vpermps %ymm3, %ymm7, %ymm52263; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,3,4,5,6,7]2264; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm5 # 32-byte Folded Reload2265; AVX2-NEXT: # ymm5 = ymm12[0,1],mem[2,3],ymm12[4,5,6,7]2266; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [0,2,0,6,0,2,0,6]2267; AVX2-NEXT: # ymm10 = mem[0,1,0,1]2268; AVX2-NEXT: vpermps %ymm5, %ymm10, %ymm132269; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm13[5,6,7]2270; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm11[0,1,2,3],ymm6[4,5,6,7]2271; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2272; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm11 # 32-byte Folded Reload2273; AVX2-NEXT: # ymm11 = ymm9[0,1,2,3],mem[4,5,6,7]2274; AVX2-NEXT: vmovaps 272(%rdi), %xmm132275; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm6[0,1],ymm13[2,3],ymm6[4,5,6,7]2276; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm15[0,1,0,2,4,5,4,6]2277; AVX2-NEXT: vpermps %ymm11, %ymm7, %ymm72278; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm15[2,3,4,5,6,7]2279; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2280; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm12 # 32-byte Folded Reload2281; AVX2-NEXT: # ymm12 = mem[0,1],ymm9[2,3],mem[4,5,6,7]2282; AVX2-NEXT: vpermps %ymm12, %ymm10, %ymm102283; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4],ymm10[5,6,7]2284; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[1,1,1,1,5,5,5,5]2285; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5,6,7]2286; AVX2-NEXT: vpermps %ymm3, %ymm14, %ymm32287; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3,4,5,6,7]2288; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,3,1,7,0,3,1,7]2289; AVX2-NEXT: # ymm3 = mem[0,1,0,1]2290; AVX2-NEXT: vpermps %ymm5, %ymm3, %ymm42291; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm4[5,6,7]2292; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm6[1,1,1,1,5,5,5,5]2293; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm13[3],ymm4[4,5,6,7]2294; AVX2-NEXT: vpermps %ymm11, %ymm14, %ymm52295; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]2296; AVX2-NEXT: vpermps %ymm12, %ymm3, %ymm32297; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5,6,7]2298; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2299; AVX2-NEXT: vmovaps %ymm4, 32(%rsi)2300; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2301; AVX2-NEXT: vmovaps %ymm4, (%rsi)2302; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2303; AVX2-NEXT: vmovaps %ymm4, 32(%rdx)2304; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2305; AVX2-NEXT: vmovaps %ymm4, (%rdx)2306; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2307; AVX2-NEXT: vmovaps %ymm4, 32(%rcx)2308; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2309; AVX2-NEXT: vmovaps %ymm4, (%rcx)2310; AVX2-NEXT: vmovaps %ymm8, 32(%r8)2311; AVX2-NEXT: vmovaps %ymm0, (%r8)2312; AVX2-NEXT: vmovaps %ymm7, 32(%r9)2313; AVX2-NEXT: vmovaps %ymm1, (%r9)2314; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax2315; AVX2-NEXT: vmovaps %ymm3, 32(%rax)2316; AVX2-NEXT: vmovaps %ymm2, (%rax)2317; AVX2-NEXT: addq $392, %rsp # imm = 0x1882318; AVX2-NEXT: vzeroupper2319; AVX2-NEXT: retq2320;2321; AVX2-FP-LABEL: load_i32_stride6_vf16:2322; AVX2-FP: # %bb.0:2323; AVX2-FP-NEXT: subq $392, %rsp # imm = 0x1882324; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm102325; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm12326; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2327; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm22328; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2329; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm02330; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm32331; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2332; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm152333; AVX2-FP-NEXT: vmovaps (%rdi), %ymm42334; AVX2-FP-NEXT: vmovups %ymm4, (%rsp) # 32-byte Spill2335; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm52336; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2337; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm132338; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm6 = [0,6,4,u]2339; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm4[0,1,2,3],ymm5[4,5],ymm4[6,7]2340; AVX2-FP-NEXT: vpermps %ymm8, %ymm6, %ymm72341; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm13[0,1],ymm15[0,1]2342; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm15[6,7]2343; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm9[0,2,2,2,4,6,6,6]2344; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm7[0,1,2],ymm11[3,4,5,6,7]2345; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm3[0,1,2,3],ymm0[4,5,6,7]2346; AVX2-FP-NEXT: vmovaps %ymm0, %ymm72347; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2348; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm12 = [4,2,4,2,4,2,4,2]2349; AVX2-FP-NEXT: vpermps %ymm4, %ymm12, %ymm142350; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2351; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm14[6,7]2352; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2353; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm1[4,5],ymm2[6,7]2354; AVX2-FP-NEXT: vpermps %ymm3, %ymm6, %ymm02355; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm112356; AVX2-FP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2357; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm11[0,1],ymm10[0,1]2358; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],ymm10[6,7]2359; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm1[0,2,2,2,4,6,6,6]2360; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2],ymm14[3,4,5,6,7]2361; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm52362; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm62363; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm5[4,5,6,7]2364; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2365; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2366; AVX2-FP-NEXT: vpermps %ymm0, %ymm12, %ymm102367; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]2368; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2369; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm2 = [1,7,5,u]2370; AVX2-FP-NEXT: vpermps %ymm8, %ymm2, %ymm82371; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[1,3,2,3,5,7,6,7]2372; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2],ymm9[3,4,5,6,7]2373; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm14 = [5,3,5,3,5,3,5,3]2374; AVX2-FP-NEXT: vpermps %ymm4, %ymm14, %ymm42375; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]2376; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2377; AVX2-FP-NEXT: vpermps %ymm3, %ymm2, %ymm22378; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,3,2,3,5,7,6,7]2379; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3,4,5,6,7]2380; AVX2-FP-NEXT: vpermps %ymm0, %ymm14, %ymm02381; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]2382; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2383; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1],ymm15[2,3],ymm13[4,5],ymm15[6,7]2384; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]2385; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]2386; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload2387; AVX2-FP-NEXT: vmovups (%rsp), %ymm9 # 32-byte Reload2388; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0,1],ymm9[2,3],ymm10[4,5,6,7]2389; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm1[2,0,2,3,6,4,6,7]2390; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]2391; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]2392; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload2393; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0,1,2,3],ymm12[4,5],ymm7[6,7]2394; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm2[0,0,2,0,4,4,6,4]2395; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]2396; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]2397; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2398; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2399; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1],ymm7[2,3],ymm11[4,5],ymm7[6,7]2400; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]2401; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]2402; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2403; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload2404; AVX2-FP-NEXT: # ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5,6,7]2405; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm3[2,0,2,3,6,4,6,7]2406; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,3,2,3]2407; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3,4,5,6,7]2408; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm6[4,5],ymm5[6,7]2409; AVX2-FP-NEXT: vshufps {{.*#+}} ymm8 = ymm4[0,0,2,0,4,4,6,4]2410; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]2411; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm8[5,6,7]2412; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2413; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm15[3,3,3,3,7,7,7,7]2414; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2,3,4],ymm13[5],ymm0[6,7]2415; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]2416; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,1,3,3,7,5,7,7]2417; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,3,2,3]2418; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]2419; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,1,4,5,7,5]2420; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]2421; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]2422; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm7[3,3,3,3,7,7,7,7]2423; AVX2-FP-NEXT: vmovaps %ymm7, %ymm62424; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2,3,4],ymm11[5],ymm1[6,7]2425; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,0,3]2426; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,1,3,3,7,5,7,7]2427; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]2428; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3,4,5,6,7]2429; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm4[0,1,3,1,4,5,7,5]2430; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]2431; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4],ymm2[5,6,7]2432; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm15[4,5,6,7]2433; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3],ymm9[4,5,6,7]2434; AVX2-FP-NEXT: vmovaps 80(%rdi), %xmm42435; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm4[2,3],ymm2[4,5,6,7]2436; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,0,2,4,5,4,6]2437; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2438; AVX2-FP-NEXT: vpermps %ymm3, %ymm7, %ymm52439; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,3,4,5,6,7]2440; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm5 # 32-byte Folded Reload2441; AVX2-FP-NEXT: # ymm5 = ymm12[0,1],mem[2,3],ymm12[4,5,6,7]2442; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [0,2,0,6,0,2,0,6]2443; AVX2-FP-NEXT: # ymm10 = mem[0,1,0,1]2444; AVX2-FP-NEXT: vpermps %ymm5, %ymm10, %ymm132445; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm13[5,6,7]2446; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm11[0,1,2,3],ymm6[4,5,6,7]2447; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2448; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm11 # 32-byte Folded Reload2449; AVX2-FP-NEXT: # ymm11 = ymm9[0,1,2,3],mem[4,5,6,7]2450; AVX2-FP-NEXT: vmovaps 272(%rdi), %xmm132451; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm6[0,1],ymm13[2,3],ymm6[4,5,6,7]2452; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm15[0,1,0,2,4,5,4,6]2453; AVX2-FP-NEXT: vpermps %ymm11, %ymm7, %ymm72454; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm15[2,3,4,5,6,7]2455; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2456; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm12 # 32-byte Folded Reload2457; AVX2-FP-NEXT: # ymm12 = mem[0,1],ymm9[2,3],mem[4,5,6,7]2458; AVX2-FP-NEXT: vpermps %ymm12, %ymm10, %ymm102459; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4],ymm10[5,6,7]2460; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[1,1,1,1,5,5,5,5]2461; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5,6,7]2462; AVX2-FP-NEXT: vpermps %ymm3, %ymm14, %ymm32463; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3,4,5,6,7]2464; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,3,1,7,0,3,1,7]2465; AVX2-FP-NEXT: # ymm3 = mem[0,1,0,1]2466; AVX2-FP-NEXT: vpermps %ymm5, %ymm3, %ymm42467; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm4[5,6,7]2468; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm6[1,1,1,1,5,5,5,5]2469; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm13[3],ymm4[4,5,6,7]2470; AVX2-FP-NEXT: vpermps %ymm11, %ymm14, %ymm52471; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]2472; AVX2-FP-NEXT: vpermps %ymm12, %ymm3, %ymm32473; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5,6,7]2474; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2475; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rsi)2476; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2477; AVX2-FP-NEXT: vmovaps %ymm4, (%rsi)2478; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2479; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rdx)2480; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2481; AVX2-FP-NEXT: vmovaps %ymm4, (%rdx)2482; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2483; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rcx)2484; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2485; AVX2-FP-NEXT: vmovaps %ymm4, (%rcx)2486; AVX2-FP-NEXT: vmovaps %ymm8, 32(%r8)2487; AVX2-FP-NEXT: vmovaps %ymm0, (%r8)2488; AVX2-FP-NEXT: vmovaps %ymm7, 32(%r9)2489; AVX2-FP-NEXT: vmovaps %ymm1, (%r9)2490; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax2491; AVX2-FP-NEXT: vmovaps %ymm3, 32(%rax)2492; AVX2-FP-NEXT: vmovaps %ymm2, (%rax)2493; AVX2-FP-NEXT: addq $392, %rsp # imm = 0x1882494; AVX2-FP-NEXT: vzeroupper2495; AVX2-FP-NEXT: retq2496;2497; AVX2-FCP-LABEL: load_i32_stride6_vf16:2498; AVX2-FCP: # %bb.0:2499; AVX2-FCP-NEXT: subq $360, %rsp # imm = 0x1682500; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm102501; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm62502; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2503; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm22504; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2505; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm02506; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2507; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm132508; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm12509; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm42510; AVX2-FCP-NEXT: vmovups %ymm4, (%rsp) # 32-byte Spill2511; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm52512; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2513; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm152514; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm12 = [0,6,4,u]2515; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm4[0,1,2,3],ymm5[4,5],ymm4[6,7]2516; AVX2-FCP-NEXT: vpermps %ymm8, %ymm12, %ymm72517; AVX2-FCP-NEXT: vmovaps %ymm1, %ymm52518; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm15[0,1],ymm1[0,1]2519; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm1[6,7]2520; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2521; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm11 = ymm9[0,2,2,2,4,6,6,6]2522; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm7[0,1,2],ymm11[3,4,5,6,7]2523; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm13[0,1,2,3],ymm0[4,5,6,7]2524; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm1 = [4,2,4,2,4,2,4,2]2525; AVX2-FCP-NEXT: vpermps %ymm4, %ymm1, %ymm142526; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2527; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm14[6,7]2528; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2529; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm6[4,5],ymm2[6,7]2530; AVX2-FCP-NEXT: vpermps %ymm3, %ymm12, %ymm122531; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm112532; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2533; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm11[0,1],ymm10[0,1]2534; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm14[0,1,2,3,4,5],ymm10[6,7]2535; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm14 = ymm6[0,2,2,2,4,6,6,6]2536; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm12[0,1,2],ymm14[3,4,5,6,7]2537; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm72538; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm122539; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm7[4,5,6,7]2540; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2541; AVX2-FCP-NEXT: vpermps %ymm0, %ymm1, %ymm102542; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]2543; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2544; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm2 = [1,7,5,u]2545; AVX2-FCP-NEXT: vpermps %ymm8, %ymm2, %ymm82546; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[1,3,2,3,5,7,6,7]2547; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2],ymm9[3,4,5,6,7]2548; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm14 = [5,3,5,3,5,3,5,3]2549; AVX2-FCP-NEXT: vpermps %ymm4, %ymm14, %ymm42550; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]2551; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2552; AVX2-FCP-NEXT: vpermps %ymm3, %ymm2, %ymm22553; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm6[1,3,2,3,5,7,6,7]2554; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3,4,5,6,7]2555; AVX2-FCP-NEXT: vpermps %ymm0, %ymm14, %ymm02556; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]2557; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2558; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1],ymm5[2,3],ymm15[4,5],ymm5[6,7]2559; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm1 = [2,0,6,4,2,0,6,7]2560; AVX2-FCP-NEXT: vpermps %ymm0, %ymm1, %ymm02561; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm2 = [2,0,6,7]2562; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload2563; AVX2-FCP-NEXT: vmovups (%rsp), %ymm9 # 32-byte Reload2564; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm9[2,3],ymm5[4,5,6,7]2565; AVX2-FCP-NEXT: vpermps %ymm3, %ymm2, %ymm42566; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3,4,5,6,7]2567; AVX2-FCP-NEXT: vblendps $207, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm4 # 32-byte Folded Reload2568; AVX2-FCP-NEXT: # ymm4 = mem[0,1,2,3],ymm13[4,5],mem[6,7]2569; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm8 = [0,0,6,4,0,0,6,4]2570; AVX2-FCP-NEXT: # ymm8 = mem[0,1,0,1]2571; AVX2-FCP-NEXT: vpermps %ymm4, %ymm8, %ymm102572; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm10[5,6,7]2573; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2574; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2575; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1],ymm6[2,3],ymm11[4,5],ymm6[6,7]2576; AVX2-FCP-NEXT: vpermps %ymm0, %ymm1, %ymm02577; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload2578; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload2579; AVX2-FCP-NEXT: # ymm1 = ymm1[0,1],mem[2,3],ymm1[4,5,6,7]2580; AVX2-FCP-NEXT: vpermps %ymm1, %ymm2, %ymm22581; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]2582; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0,1,2,3],ymm12[4,5],ymm7[6,7]2583; AVX2-FCP-NEXT: vpermps %ymm2, %ymm8, %ymm82584; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm8[5,6,7]2585; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2586; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm0 = mem[3,3,3,3]2587; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3,4],ymm15[5],ymm0[6,7]2588; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1,2,3,4],ymm0[5],ymm3[6,7]2589; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,1,7,5,0,1,7,5]2590; AVX2-FCP-NEXT: # ymm3 = mem[0,1,0,1]2591; AVX2-FCP-NEXT: vpermps %ymm4, %ymm3, %ymm42592; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm8 = [3,1,7,5,0,u,u,u]2593; AVX2-FCP-NEXT: vpermps %ymm0, %ymm8, %ymm02594; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5,6,7]2595; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm4 = mem[3,3,3,3]2596; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm11[1],ymm4[2,3,4],ymm11[5],ymm4[6,7]2597; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0],ymm1[1,2,3,4],ymm4[5],ymm1[6,7]2598; AVX2-FCP-NEXT: vpermps %ymm1, %ymm8, %ymm12599; AVX2-FCP-NEXT: vpermps %ymm2, %ymm3, %ymm22600; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4],ymm2[5,6,7]2601; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm2 # 32-byte Folded Reload2602; AVX2-FCP-NEXT: # ymm2 = ymm15[0,1,2,3],mem[4,5,6,7]2603; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm9[4,5,6,7]2604; AVX2-FCP-NEXT: vmovaps 80(%rdi), %xmm42605; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm4[2,3],ymm2[4,5,6,7]2606; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,0,2,4,5,4,6]2607; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload2608; AVX2-FCP-NEXT: vpermps %ymm3, %ymm7, %ymm52609; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,3,4,5,6,7]2610; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm5 # 32-byte Folded Reload2611; AVX2-FCP-NEXT: # ymm5 = ymm13[0,1],mem[2,3],ymm13[4,5,6,7]2612; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [0,2,0,6,0,2,0,6]2613; AVX2-FCP-NEXT: # ymm10 = mem[0,1,0,1]2614; AVX2-FCP-NEXT: vpermps %ymm5, %ymm10, %ymm132615; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm13[5,6,7]2616; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm11[0,1,2,3],ymm6[4,5,6,7]2617; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload2618; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm11 # 32-byte Folded Reload2619; AVX2-FCP-NEXT: # ymm11 = ymm9[0,1,2,3],mem[4,5,6,7]2620; AVX2-FCP-NEXT: vmovaps 272(%rdi), %xmm132621; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm6[0,1],ymm13[2,3],ymm6[4,5,6,7]2622; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm15 = ymm15[0,1,0,2,4,5,4,6]2623; AVX2-FCP-NEXT: vpermps %ymm11, %ymm7, %ymm72624; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm15[2,3,4,5,6,7]2625; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload2626; AVX2-FCP-NEXT: # ymm12 = ymm12[0,1],mem[2,3],ymm12[4,5,6,7]2627; AVX2-FCP-NEXT: vpermps %ymm12, %ymm10, %ymm102628; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4],ymm10[5,6,7]2629; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[1,1,1,1,5,5,5,5]2630; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5,6,7]2631; AVX2-FCP-NEXT: vpermps %ymm3, %ymm14, %ymm32632; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3,4,5,6,7]2633; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,3,1,7,0,3,1,7]2634; AVX2-FCP-NEXT: # ymm3 = mem[0,1,0,1]2635; AVX2-FCP-NEXT: vpermps %ymm5, %ymm3, %ymm42636; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm4[5,6,7]2637; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm6[1,1,1,1,5,5,5,5]2638; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm13[3],ymm4[4,5,6,7]2639; AVX2-FCP-NEXT: vpermps %ymm11, %ymm14, %ymm52640; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]2641; AVX2-FCP-NEXT: vpermps %ymm12, %ymm3, %ymm32642; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5,6,7]2643; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2644; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%rsi)2645; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2646; AVX2-FCP-NEXT: vmovaps %ymm4, (%rsi)2647; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2648; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%rdx)2649; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2650; AVX2-FCP-NEXT: vmovaps %ymm4, (%rdx)2651; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2652; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%rcx)2653; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload2654; AVX2-FCP-NEXT: vmovaps %ymm4, (%rcx)2655; AVX2-FCP-NEXT: vmovaps %ymm8, 32(%r8)2656; AVX2-FCP-NEXT: vmovaps %ymm0, (%r8)2657; AVX2-FCP-NEXT: vmovaps %ymm7, 32(%r9)2658; AVX2-FCP-NEXT: vmovaps %ymm1, (%r9)2659; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax2660; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%rax)2661; AVX2-FCP-NEXT: vmovaps %ymm2, (%rax)2662; AVX2-FCP-NEXT: addq $360, %rsp # imm = 0x1682663; AVX2-FCP-NEXT: vzeroupper2664; AVX2-FCP-NEXT: retq2665;2666; AVX512-LABEL: load_i32_stride6_vf16:2667; AVX512: # %bb.0:2668; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax2669; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm02670; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm12671; AVX512-NEXT: vmovdqa64 (%rdi), %zmm32672; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm42673; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm52674; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm62675; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]2676; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2677; AVX512-NEXT: vpermi2d %zmm5, %zmm6, %zmm72678; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]2679; AVX512-NEXT: vpermi2d %zmm4, %zmm3, %zmm22680; AVX512-NEXT: movb $56, %dil2681; AVX512-NEXT: kmovw %edi, %k22682; AVX512-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}2683; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]2684; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2685; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm72686; AVX512-NEXT: movw $-2048, %di # imm = 0xF8002687; AVX512-NEXT: kmovw %edi, %k12688; AVX512-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}2689; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]2690; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2691; AVX512-NEXT: vpermi2d %zmm5, %zmm6, %zmm82692; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]2693; AVX512-NEXT: vpermi2d %zmm4, %zmm3, %zmm72694; AVX512-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}2695; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]2696; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2697; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm82698; AVX512-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}2699; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]2700; AVX512-NEXT: vpermi2d %zmm4, %zmm3, %zmm82701; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]2702; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2703; AVX512-NEXT: vpermi2d %zmm6, %zmm5, %zmm92704; AVX512-NEXT: movw $31, %di2705; AVX512-NEXT: kmovw %edi, %k22706; AVX512-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}2707; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]2708; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2709; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm82710; AVX512-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2711; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]2712; AVX512-NEXT: vpermi2d %zmm4, %zmm3, %zmm82713; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]2714; AVX512-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]2715; AVX512-NEXT: vpermi2d %zmm6, %zmm5, %zmm102716; AVX512-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}2717; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]2718; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2719; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm82720; AVX512-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}2721; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]2722; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2723; AVX512-NEXT: vpermi2d %zmm6, %zmm5, %zmm82724; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]2725; AVX512-NEXT: vpermi2d %zmm3, %zmm4, %zmm112726; AVX512-NEXT: movw $992, %di # imm = 0x3E02727; AVX512-NEXT: kmovw %edi, %k12728; AVX512-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}2729; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]2730; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2731; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm82732; AVX512-NEXT: movb $-32, %dil2733; AVX512-NEXT: kmovw %edi, %k22734; AVX512-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}2735; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]2736; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2737; AVX512-NEXT: vpermi2d %zmm6, %zmm5, %zmm82738; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]2739; AVX512-NEXT: vpermi2d %zmm3, %zmm4, %zmm52740; AVX512-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}2741; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]2742; AVX512-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]2743; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm32744; AVX512-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}2745; AVX512-NEXT: vmovdqa64 %zmm2, (%rsi)2746; AVX512-NEXT: vmovdqa64 %zmm7, (%rdx)2747; AVX512-NEXT: vmovdqa64 %zmm9, (%rcx)2748; AVX512-NEXT: vmovdqa64 %zmm10, (%r8)2749; AVX512-NEXT: vmovdqa64 %zmm11, (%r9)2750; AVX512-NEXT: vmovdqa64 %zmm5, (%rax)2751; AVX512-NEXT: vzeroupper2752; AVX512-NEXT: retq2753;2754; AVX512-FCP-LABEL: load_i32_stride6_vf16:2755; AVX512-FCP: # %bb.0:2756; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax2757; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm02758; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm12759; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm32760; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm42761; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm52762; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm62763; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]2764; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2765; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm72766; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]2767; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm22768; AVX512-FCP-NEXT: movb $56, %dil2769; AVX512-FCP-NEXT: kmovw %edi, %k22770; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}2771; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]2772; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2773; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm72774; AVX512-FCP-NEXT: movw $-2048, %di # imm = 0xF8002775; AVX512-FCP-NEXT: kmovw %edi, %k12776; AVX512-FCP-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}2777; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]2778; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2779; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm82780; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]2781; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm72782; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}2783; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]2784; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2785; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm82786; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}2787; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]2788; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm82789; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]2790; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2791; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm92792; AVX512-FCP-NEXT: movw $31, %di2793; AVX512-FCP-NEXT: kmovw %edi, %k22794; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}2795; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]2796; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2797; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82798; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2799; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]2800; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm82801; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]2802; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]2803; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm102804; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}2805; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]2806; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2807; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82808; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}2809; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]2810; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2811; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm82812; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]2813; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm112814; AVX512-FCP-NEXT: movw $992, %di # imm = 0x3E02815; AVX512-FCP-NEXT: kmovw %edi, %k12816; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}2817; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]2818; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2819; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm82820; AVX512-FCP-NEXT: movb $-32, %dil2821; AVX512-FCP-NEXT: kmovw %edi, %k22822; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}2823; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]2824; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2825; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm82826; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]2827; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm52828; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}2829; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]2830; AVX512-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]2831; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm32832; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}2833; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%rsi)2834; AVX512-FCP-NEXT: vmovdqa64 %zmm7, (%rdx)2835; AVX512-FCP-NEXT: vmovdqa64 %zmm9, (%rcx)2836; AVX512-FCP-NEXT: vmovdqa64 %zmm10, (%r8)2837; AVX512-FCP-NEXT: vmovdqa64 %zmm11, (%r9)2838; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%rax)2839; AVX512-FCP-NEXT: vzeroupper2840; AVX512-FCP-NEXT: retq2841;2842; AVX512DQ-LABEL: load_i32_stride6_vf16:2843; AVX512DQ: # %bb.0:2844; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax2845; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm02846; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm12847; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm32848; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm42849; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm52850; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm62851; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]2852; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2853; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm6, %zmm72854; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]2855; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm3, %zmm22856; AVX512DQ-NEXT: movb $56, %dil2857; AVX512DQ-NEXT: kmovw %edi, %k22858; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}2859; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]2860; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2861; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm72862; AVX512DQ-NEXT: movw $-2048, %di # imm = 0xF8002863; AVX512DQ-NEXT: kmovw %edi, %k12864; AVX512DQ-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}2865; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]2866; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2867; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm6, %zmm82868; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]2869; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm3, %zmm72870; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}2871; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]2872; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2873; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm82874; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}2875; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]2876; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm3, %zmm82877; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]2878; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2879; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm5, %zmm92880; AVX512DQ-NEXT: movw $31, %di2881; AVX512DQ-NEXT: kmovw %edi, %k22882; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}2883; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]2884; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2885; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm82886; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2887; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]2888; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm3, %zmm82889; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]2890; AVX512DQ-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]2891; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm5, %zmm102892; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}2893; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]2894; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2895; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm82896; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}2897; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]2898; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2899; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm5, %zmm82900; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]2901; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm4, %zmm112902; AVX512DQ-NEXT: movw $992, %di # imm = 0x3E02903; AVX512DQ-NEXT: kmovw %edi, %k12904; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}2905; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]2906; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2907; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm82908; AVX512DQ-NEXT: movb $-32, %dil2909; AVX512DQ-NEXT: kmovw %edi, %k22910; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}2911; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]2912; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2913; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm5, %zmm82914; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]2915; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm4, %zmm52916; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}2917; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]2918; AVX512DQ-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]2919; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm32920; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}2921; AVX512DQ-NEXT: vmovdqa64 %zmm2, (%rsi)2922; AVX512DQ-NEXT: vmovdqa64 %zmm7, (%rdx)2923; AVX512DQ-NEXT: vmovdqa64 %zmm9, (%rcx)2924; AVX512DQ-NEXT: vmovdqa64 %zmm10, (%r8)2925; AVX512DQ-NEXT: vmovdqa64 %zmm11, (%r9)2926; AVX512DQ-NEXT: vmovdqa64 %zmm5, (%rax)2927; AVX512DQ-NEXT: vzeroupper2928; AVX512DQ-NEXT: retq2929;2930; AVX512DQ-FCP-LABEL: load_i32_stride6_vf16:2931; AVX512DQ-FCP: # %bb.0:2932; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax2933; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm02934; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm12935; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm32936; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm42937; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm52938; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm62939; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]2940; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2941; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm72942; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]2943; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm22944; AVX512DQ-FCP-NEXT: movb $56, %dil2945; AVX512DQ-FCP-NEXT: kmovw %edi, %k22946; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}2947; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]2948; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2949; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm72950; AVX512DQ-FCP-NEXT: movw $-2048, %di # imm = 0xF8002951; AVX512DQ-FCP-NEXT: kmovw %edi, %k12952; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}2953; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]2954; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2955; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm82956; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]2957; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm72958; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}2959; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]2960; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2961; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm82962; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}2963; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]2964; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm82965; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]2966; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2967; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm92968; AVX512DQ-FCP-NEXT: movw $31, %di2969; AVX512DQ-FCP-NEXT: kmovw %edi, %k22970; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}2971; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]2972; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2973; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82974; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2975; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]2976; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm82977; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]2978; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]2979; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm102980; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}2981; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]2982; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2983; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82984; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}2985; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]2986; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2987; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm82988; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]2989; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm112990; AVX512DQ-FCP-NEXT: movw $992, %di # imm = 0x3E02991; AVX512DQ-FCP-NEXT: kmovw %edi, %k12992; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}2993; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]2994; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2995; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm82996; AVX512DQ-FCP-NEXT: movb $-32, %dil2997; AVX512DQ-FCP-NEXT: kmovw %edi, %k22998; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}2999; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]3000; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3001; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm83002; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]3003; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm53004; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}3005; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]3006; AVX512DQ-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]3007; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm33008; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}3009; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%rsi)3010; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, (%rdx)3011; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, (%rcx)3012; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, (%r8)3013; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, (%r9)3014; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%rax)3015; AVX512DQ-FCP-NEXT: vzeroupper3016; AVX512DQ-FCP-NEXT: retq3017;3018; AVX512BW-LABEL: load_i32_stride6_vf16:3019; AVX512BW: # %bb.0:3020; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax3021; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm03022; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm13023; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm33024; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm43025; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm53026; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm63027; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]3028; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3029; AVX512BW-NEXT: vpermi2d %zmm5, %zmm6, %zmm73030; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]3031; AVX512BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm23032; AVX512BW-NEXT: movb $56, %dil3033; AVX512BW-NEXT: kmovd %edi, %k23034; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}3035; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]3036; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3037; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm73038; AVX512BW-NEXT: movw $-2048, %di # imm = 0xF8003039; AVX512BW-NEXT: kmovd %edi, %k13040; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}3041; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]3042; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3043; AVX512BW-NEXT: vpermi2d %zmm5, %zmm6, %zmm83044; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]3045; AVX512BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm73046; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}3047; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]3048; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3049; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm83050; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}3051; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]3052; AVX512BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm83053; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]3054; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]3055; AVX512BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm93056; AVX512BW-NEXT: movw $31, %di3057; AVX512BW-NEXT: kmovd %edi, %k23058; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}3059; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]3060; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3061; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm83062; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}3063; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]3064; AVX512BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm83065; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]3066; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]3067; AVX512BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm103068; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}3069; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]3070; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3071; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm83072; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}3073; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]3074; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3075; AVX512BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm83076; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]3077; AVX512BW-NEXT: vpermi2d %zmm3, %zmm4, %zmm113078; AVX512BW-NEXT: movw $992, %di # imm = 0x3E03079; AVX512BW-NEXT: kmovd %edi, %k13080; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}3081; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]3082; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3083; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm83084; AVX512BW-NEXT: movb $-32, %dil3085; AVX512BW-NEXT: kmovd %edi, %k23086; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}3087; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]3088; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3089; AVX512BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm83090; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]3091; AVX512BW-NEXT: vpermi2d %zmm3, %zmm4, %zmm53092; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}3093; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]3094; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]3095; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm33096; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}3097; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rsi)3098; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rdx)3099; AVX512BW-NEXT: vmovdqa64 %zmm9, (%rcx)3100; AVX512BW-NEXT: vmovdqa64 %zmm10, (%r8)3101; AVX512BW-NEXT: vmovdqa64 %zmm11, (%r9)3102; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rax)3103; AVX512BW-NEXT: vzeroupper3104; AVX512BW-NEXT: retq3105;3106; AVX512BW-FCP-LABEL: load_i32_stride6_vf16:3107; AVX512BW-FCP: # %bb.0:3108; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3109; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm03110; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm13111; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm33112; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm43113; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm53114; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm63115; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]3116; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3117; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm73118; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]3119; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm23120; AVX512BW-FCP-NEXT: movb $56, %dil3121; AVX512BW-FCP-NEXT: kmovd %edi, %k23122; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}3123; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]3124; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3125; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm73126; AVX512BW-FCP-NEXT: movw $-2048, %di # imm = 0xF8003127; AVX512BW-FCP-NEXT: kmovd %edi, %k13128; AVX512BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}3129; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]3130; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3131; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm83132; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]3133; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm73134; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}3135; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]3136; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3137; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm83138; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}3139; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]3140; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm83141; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]3142; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]3143; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm93144; AVX512BW-FCP-NEXT: movw $31, %di3145; AVX512BW-FCP-NEXT: kmovd %edi, %k23146; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}3147; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]3148; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3149; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm83150; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}3151; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]3152; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm83153; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]3154; AVX512BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]3155; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm103156; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}3157; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]3158; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3159; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm83160; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}3161; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]3162; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3163; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm83164; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]3165; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm113166; AVX512BW-FCP-NEXT: movw $992, %di # imm = 0x3E03167; AVX512BW-FCP-NEXT: kmovd %edi, %k13168; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}3169; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]3170; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3171; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm83172; AVX512BW-FCP-NEXT: movb $-32, %dil3173; AVX512BW-FCP-NEXT: kmovd %edi, %k23174; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}3175; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]3176; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3177; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm83178; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]3179; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm53180; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}3181; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]3182; AVX512BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]3183; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm33184; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}3185; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, (%rsi)3186; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, (%rdx)3187; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, (%rcx)3188; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, (%r8)3189; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, (%r9)3190; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, (%rax)3191; AVX512BW-FCP-NEXT: vzeroupper3192; AVX512BW-FCP-NEXT: retq3193;3194; AVX512DQ-BW-LABEL: load_i32_stride6_vf16:3195; AVX512DQ-BW: # %bb.0:3196; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax3197; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm03198; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm13199; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm33200; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm43201; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm53202; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm63203; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]3204; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3205; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm6, %zmm73206; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]3207; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm23208; AVX512DQ-BW-NEXT: movb $56, %dil3209; AVX512DQ-BW-NEXT: kmovd %edi, %k23210; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}3211; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]3212; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3213; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm73214; AVX512DQ-BW-NEXT: movw $-2048, %di # imm = 0xF8003215; AVX512DQ-BW-NEXT: kmovd %edi, %k13216; AVX512DQ-BW-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}3217; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]3218; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3219; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm6, %zmm83220; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]3221; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm73222; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}3223; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]3224; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3225; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm83226; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}3227; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]3228; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm83229; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]3230; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]3231; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm93232; AVX512DQ-BW-NEXT: movw $31, %di3233; AVX512DQ-BW-NEXT: kmovd %edi, %k23234; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}3235; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]3236; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3237; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm83238; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}3239; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]3240; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm3, %zmm83241; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]3242; AVX512DQ-BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]3243; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm103244; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}3245; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]3246; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3247; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm83248; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}3249; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]3250; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3251; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm83252; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]3253; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm4, %zmm113254; AVX512DQ-BW-NEXT: movw $992, %di # imm = 0x3E03255; AVX512DQ-BW-NEXT: kmovd %edi, %k13256; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}3257; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]3258; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3259; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm83260; AVX512DQ-BW-NEXT: movb $-32, %dil3261; AVX512DQ-BW-NEXT: kmovd %edi, %k23262; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}3263; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]3264; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3265; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm5, %zmm83266; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]3267; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm4, %zmm53268; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}3269; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]3270; AVX512DQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]3271; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm33272; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}3273; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rsi)3274; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rdx)3275; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, (%rcx)3276; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, (%r8)3277; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, (%r9)3278; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rax)3279; AVX512DQ-BW-NEXT: vzeroupper3280; AVX512DQ-BW-NEXT: retq3281;3282; AVX512DQ-BW-FCP-LABEL: load_i32_stride6_vf16:3283; AVX512DQ-BW-FCP: # %bb.0:3284; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax3285; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm03286; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm13287; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm33288; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm43289; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm53290; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm63291; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]3292; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3293; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm73294; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,6,12,18,24,30,0,0]3295; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm23296; AVX512DQ-BW-FCP-NEXT: movb $56, %dil3297; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k23298; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2 {%k2}3299; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]3300; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3301; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm73302; AVX512DQ-BW-FCP-NEXT: movw $-2048, %di # imm = 0xF8003303; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k13304; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm2 {%k1}3305; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]3306; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3307; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm6, %zmm83308; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,7,13,19,25,31,0,0]3309; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm73310; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k2}3311; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]3312; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3313; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm83314; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm7 {%k1}3315; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [2,8,14,20,26,0,0,0]3316; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm83317; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]3318; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]3319; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm93320; AVX512DQ-BW-FCP-NEXT: movw $31, %di3321; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k23322; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k2}3323; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]3324; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3325; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm83326; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}3327; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,9,15,21,27,0,0,0]3328; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm3, %zmm83329; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]3330; AVX512DQ-BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]3331; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm103332; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k2}3333; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]3334; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3335; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm83336; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm10 {%k1}3337; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]3338; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3339; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm83340; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [20,26,0,6,12,0,0,0]3341; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm113342; AVX512DQ-BW-FCP-NEXT: movw $992, %di # imm = 0x3E03343; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k13344; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}3345; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]3346; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3347; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm83348; AVX512DQ-BW-FCP-NEXT: movb $-32, %dil3349; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k23350; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm11 {%k2}3351; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]3352; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3353; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm5, %zmm83354; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [21,27,1,7,13,0,0,0]3355; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm53356; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm5 {%k1}3357; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]3358; AVX512DQ-BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]3359; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm33360; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm5 {%k2}3361; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, (%rsi)3362; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, (%rdx)3363; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, (%rcx)3364; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, (%r8)3365; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, (%r9)3366; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, (%rax)3367; AVX512DQ-BW-FCP-NEXT: vzeroupper3368; AVX512DQ-BW-FCP-NEXT: retq3369 %wide.vec = load <96 x i32>, ptr %in.vec, align 643370 %strided.vec0 = shufflevector <96 x i32> %wide.vec, <96 x i32> poison, <16 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90>3371 %strided.vec1 = shufflevector <96 x i32> %wide.vec, <96 x i32> poison, <16 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91>3372 %strided.vec2 = shufflevector <96 x i32> %wide.vec, <96 x i32> poison, <16 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92>3373 %strided.vec3 = shufflevector <96 x i32> %wide.vec, <96 x i32> poison, <16 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93>3374 %strided.vec4 = shufflevector <96 x i32> %wide.vec, <96 x i32> poison, <16 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94>3375 %strided.vec5 = shufflevector <96 x i32> %wide.vec, <96 x i32> poison, <16 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95>3376 store <16 x i32> %strided.vec0, ptr %out.vec0, align 643377 store <16 x i32> %strided.vec1, ptr %out.vec1, align 643378 store <16 x i32> %strided.vec2, ptr %out.vec2, align 643379 store <16 x i32> %strided.vec3, ptr %out.vec3, align 643380 store <16 x i32> %strided.vec4, ptr %out.vec4, align 643381 store <16 x i32> %strided.vec5, ptr %out.vec5, align 643382 ret void3383}3384 3385define void @load_i32_stride6_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {3386; SSE-LABEL: load_i32_stride6_vf32:3387; SSE: # %bb.0:3388; SSE-NEXT: subq $1032, %rsp # imm = 0x4083389; SSE-NEXT: movdqa 64(%rdi), %xmm53390; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3391; SSE-NEXT: movdqa (%rdi), %xmm123392; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3393; SSE-NEXT: movdqa 16(%rdi), %xmm133394; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3395; SSE-NEXT: movdqa 48(%rdi), %xmm93396; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3397; SSE-NEXT: movdqa 528(%rdi), %xmm73398; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3399; SSE-NEXT: movdqa 544(%rdi), %xmm33400; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3401; SSE-NEXT: movdqa 480(%rdi), %xmm83402; SSE-NEXT: movdqa 496(%rdi), %xmm43403; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3404; SSE-NEXT: movdqa 144(%rdi), %xmm103405; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3406; SSE-NEXT: movdqa 160(%rdi), %xmm23407; SSE-NEXT: movdqa 96(%rdi), %xmm13408; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3409; SSE-NEXT: movdqa 112(%rdi), %xmm113410; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[2,3,2,3]3411; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3412; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3413; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,3,3]3414; SSE-NEXT: movdqa %xmm2, %xmm63415; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3416; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[0,0,1,1]3417; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3418; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3419; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3420; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]3421; SSE-NEXT: movdqa %xmm8, %xmm13422; SSE-NEXT: movdqa %xmm8, %xmm43423; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3424; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3425; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,2,3,3]3426; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,0,1,1]3427; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3428; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3429; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3430; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[2,3,2,3]3431; SSE-NEXT: movdqa %xmm12, %xmm13432; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3433; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]3434; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm9[0,0,1,1]3435; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3436; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3437; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3438; SSE-NEXT: movdqa 384(%rdi), %xmm13439; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3440; SSE-NEXT: movdqa 400(%rdi), %xmm143441; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[2,3,2,3]3442; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3443; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3444; SSE-NEXT: movdqa 432(%rdi), %xmm33445; SSE-NEXT: movdqa %xmm3, (%rsp) # 16-byte Spill3446; SSE-NEXT: movdqa 448(%rdi), %xmm93447; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[2,2,3,3]3448; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3449; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,0,1,1]3450; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3451; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3452; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3453; SSE-NEXT: movdqa 288(%rdi), %xmm13454; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3455; SSE-NEXT: movdqa 304(%rdi), %xmm153456; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[2,3,2,3]3457; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3458; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3459; SSE-NEXT: movdqa 336(%rdi), %xmm33460; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3461; SSE-NEXT: movdqa 352(%rdi), %xmm123462; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,2,3,3]3463; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3464; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,0,1,1]3465; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3466; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3467; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3468; SSE-NEXT: movdqa 672(%rdi), %xmm13469; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3470; SSE-NEXT: movdqa 688(%rdi), %xmm83471; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[2,3,2,3]3472; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3473; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3474; SSE-NEXT: movdqa 720(%rdi), %xmm33475; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3476; SSE-NEXT: movdqa 736(%rdi), %xmm53477; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]3478; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3479; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,0,1,1]3480; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3481; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3482; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3483; SSE-NEXT: movdqa 192(%rdi), %xmm13484; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3485; SSE-NEXT: movdqa 208(%rdi), %xmm03486; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3487; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]3488; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3489; SSE-NEXT: movdqa 240(%rdi), %xmm23490; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3491; SSE-NEXT: movdqa 256(%rdi), %xmm03492; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3493; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]3494; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]3495; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3496; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3497; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3498; SSE-NEXT: movdqa 576(%rdi), %xmm73499; SSE-NEXT: movdqa 592(%rdi), %xmm133500; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[2,3,2,3]3501; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3502; SSE-NEXT: movdqa %xmm7, %xmm13503; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3504; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3505; SSE-NEXT: movdqa 624(%rdi), %xmm103506; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3507; SSE-NEXT: movdqa 640(%rdi), %xmm33508; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,2,3,3]3509; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3510; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[0,0,1,1]3511; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3512; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3513; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3514; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3515; SSE-NEXT: # xmm0 = mem[1,1,1,1]3516; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[3,3,3,3]3517; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3518; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[2,3,2,3]3519; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload3520; SSE-NEXT: movdqa %xmm10, %xmm23521; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3522; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]3523; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3524; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,1,1]3525; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3526; SSE-NEXT: # xmm1 = mem[3,3,3,3]3527; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3528; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3529; SSE-NEXT: # xmm1 = mem[2,3,2,3]3530; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload3531; SSE-NEXT: movdqa %xmm11, %xmm23532; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3533; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]3534; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3535; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload3536; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,1,1]3537; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3538; SSE-NEXT: # xmm1 = mem[3,3,3,3]3539; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3540; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3541; SSE-NEXT: # xmm1 = mem[2,3,2,3]3542; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload3543; SSE-NEXT: movdqa %xmm4, %xmm23544; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3545; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]3546; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3547; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3548; SSE-NEXT: # xmm0 = mem[1,1,1,1]3549; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[3,3,3,3]3550; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3551; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[2,3,2,3]3552; SSE-NEXT: movdqa (%rsp), %xmm2 # 16-byte Reload3553; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3554; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]3555; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3556; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload3557; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[1,1,1,1]3558; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[3,3,3,3]3559; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3560; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[2,3,2,3]3561; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload3562; SSE-NEXT: movdqa %xmm15, %xmm23563; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3564; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]3565; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3566; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3567; SSE-NEXT: # xmm0 = mem[1,1,1,1]3568; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[3,3,3,3]3569; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3570; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,3,2,3]3571; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload3572; SSE-NEXT: movdqa %xmm8, %xmm23573; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]3574; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]3575; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3576; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload3577; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[1,1,1,1]3578; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3579; SSE-NEXT: # xmm1 = mem[3,3,3,3]3580; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3581; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3582; SSE-NEXT: # xmm1 = mem[2,3,2,3]3583; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload3584; SSE-NEXT: movdqa %xmm14, %xmm53585; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]3586; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm0[0],xmm5[1]3587; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3588; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,1,1]3589; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm13[3,3,3,3]3590; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3591; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]3592; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload3593; SSE-NEXT: movdqa %xmm5, %xmm33594; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]3595; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]3596; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3597; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[2,2,3,3]3598; SSE-NEXT: movdqa 176(%rdi), %xmm03599; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3600; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3601; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3602; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload3603; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[2,3,2,3]3604; SSE-NEXT: movdqa 128(%rdi), %xmm23605; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3606; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]3607; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3608; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3609; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[2,2,3,3]3610; SSE-NEXT: movdqa 80(%rdi), %xmm03611; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3612; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3613; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3614; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]3615; SSE-NEXT: movdqa 32(%rdi), %xmm133616; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm13[0],xmm0[1],xmm13[1]3617; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3618; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3619; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[2,2,3,3]3620; SSE-NEXT: movdqa 368(%rdi), %xmm03621; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3622; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3623; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3624; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[2,3,2,3]3625; SSE-NEXT: movdqa 320(%rdi), %xmm23626; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3627; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]3628; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3629; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3630; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[2,2,3,3]3631; SSE-NEXT: movdqa 272(%rdi), %xmm03632; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3633; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3634; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3635; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,3,2,3]3636; SSE-NEXT: movdqa %xmm12, %xmm143637; SSE-NEXT: movdqa 224(%rdi), %xmm63638; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]3639; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3640; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3641; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3642; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[2,2,3,3]3643; SSE-NEXT: movdqa 560(%rdi), %xmm03644; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3645; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3646; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3647; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3648; SSE-NEXT: # xmm0 = mem[2,3,2,3]3649; SSE-NEXT: movdqa 512(%rdi), %xmm23650; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3651; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]3652; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3653; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3654; SSE-NEXT: movdqa (%rsp), %xmm11 # 16-byte Reload3655; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[2,2,3,3]3656; SSE-NEXT: movdqa 464(%rdi), %xmm153657; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[0,0,1,1]3658; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3659; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload3660; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]3661; SSE-NEXT: movdqa 416(%rdi), %xmm43662; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]3663; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3664; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3665; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3666; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[2,2,3,3]3667; SSE-NEXT: movdqa 752(%rdi), %xmm03668; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3669; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3670; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3671; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3672; SSE-NEXT: # xmm0 = mem[2,3,2,3]3673; SSE-NEXT: movdqa 704(%rdi), %xmm123674; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1]3675; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3676; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3677; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3678; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3]3679; SSE-NEXT: movdqa 656(%rdi), %xmm03680; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3681; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]3682; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3683; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload3684; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]3685; SSE-NEXT: movdqa 608(%rdi), %xmm53686; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]3687; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3688; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3689; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3690; SSE-NEXT: # xmm0 = mem[3,3,3,3]3691; SSE-NEXT: movdqa %xmm13, %xmm93692; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3693; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm13[1,1,1,1]3694; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3695; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3696; SSE-NEXT: # xmm1 = mem[2,3,2,3]3697; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload3698; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1]3699; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3700; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3701; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[3,3,3,3]3702; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload3703; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[1,1,1,1]3704; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3705; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3706; SSE-NEXT: # xmm1 = mem[2,3,2,3]3707; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload3708; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]3709; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3710; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3711; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[3,3,3,3]3712; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[1,1,1,1]3713; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3714; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3715; SSE-NEXT: # xmm1 = mem[2,3,2,3]3716; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload3717; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1]3718; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3719; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3720; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3721; SSE-NEXT: # xmm0 = mem[3,3,3,3]3722; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload3723; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[1,1,1,1]3724; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3725; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3726; SSE-NEXT: # xmm1 = mem[2,3,2,3]3727; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload3728; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1]3729; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3730; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3731; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[3,3,3,3]3732; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1]3733; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3734; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[2,3,2,3]3735; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3736; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1]3737; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3738; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3739; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3740; SSE-NEXT: # xmm0 = mem[3,3,3,3]3741; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload3742; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[1,1,1,1]3743; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3744; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3745; SSE-NEXT: # xmm1 = mem[2,3,2,3]3746; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload3747; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]3748; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3749; SSE-NEXT: movapd %xmm1, (%rsp) # 16-byte Spill3750; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[3,3,3,3]3751; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3752; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1]3753; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3754; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3755; SSE-NEXT: # xmm1 = mem[2,3,2,3]3756; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload3757; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]3758; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3759; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3760; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3761; SSE-NEXT: # xmm0 = mem[3,3,3,3]3762; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[1,1,1,1]3763; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3764; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3765; SSE-NEXT: # xmm1 = mem[2,3,2,3]3766; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload3767; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1]3768; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3769; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3770; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[2,3,2,3]3771; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload3772; SSE-NEXT: movdqa %xmm12, %xmm13773; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3774; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[2,2,3,3]3775; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload3776; SSE-NEXT: # xmm9 = mem[0,0,1,1]3777; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]3778; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm1[0],xmm9[1]3779; SSE-NEXT: movapd %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3780; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[2,3,2,3]3781; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload3782; SSE-NEXT: movdqa %xmm13, %xmm13783; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3784; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,2,3,3]3785; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload3786; SSE-NEXT: # xmm3 = mem[0,0,1,1]3787; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]3788; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]3789; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3790; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3791; SSE-NEXT: # xmm0 = mem[2,3,2,3]3792; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3793; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3794; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[2,2,3,3]3795; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload3796; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm14[0,0,1,1]3797; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]3798; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm1[0],xmm9[1]3799; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[2,3,2,3]3800; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3801; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3802; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,2,3,3]3803; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload3804; SSE-NEXT: # xmm8 = mem[0,0,1,1]3805; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1]3806; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm1[0],xmm8[1]3807; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3808; SSE-NEXT: # xmm0 = mem[2,3,2,3]3809; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload3810; SSE-NEXT: movdqa %xmm10, %xmm13811; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3812; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[2,2,3,3]3813; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload3814; SSE-NEXT: # xmm6 = mem[0,0,1,1]3815; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]3816; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm1[0],xmm6[1]3817; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]3818; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3819; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3820; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,2,3,3]3821; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload3822; SSE-NEXT: # xmm4 = mem[0,0,1,1]3823; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]3824; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm1[0],xmm4[1]3825; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,3,2,3]3826; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3827; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3828; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,2,3,3]3829; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload3830; SSE-NEXT: # xmm3 = mem[0,0,1,1]3831; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]3832; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]3833; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload3834; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,3,2,3]3835; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3836; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3837; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[2,2,3,3]3838; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload3839; SSE-NEXT: # xmm2 = mem[0,0,1,1]3840; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]3841; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]3842; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[1,1,1,1]3843; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3844; SSE-NEXT: # xmm1 = mem[3,3,3,3]3845; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3846; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3847; SSE-NEXT: # xmm1 = mem[2,3,2,3]3848; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload3849; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]3850; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]3851; SSE-NEXT: movapd %xmm15, %xmm73852; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[1,1,1,1]3853; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3854; SSE-NEXT: # xmm1 = mem[3,3,3,3]3855; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3856; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3857; SSE-NEXT: # xmm1 = mem[2,3,2,3]3858; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload3859; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]3860; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]3861; SSE-NEXT: movapd %xmm15, %xmm113862; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3863; SSE-NEXT: # xmm0 = mem[1,1,1,1]3864; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3865; SSE-NEXT: # xmm1 = mem[3,3,3,3]3866; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3867; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3868; SSE-NEXT: # xmm1 = mem[2,3,2,3]3869; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]3870; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]3871; SSE-NEXT: movapd %xmm14, %xmm123872; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3873; SSE-NEXT: # xmm0 = mem[1,1,1,1]3874; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3875; SSE-NEXT: # xmm1 = mem[3,3,3,3]3876; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3877; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3878; SSE-NEXT: # xmm1 = mem[2,3,2,3]3879; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload3880; SSE-NEXT: punpckldq {{.*#+}} xmm13 = xmm13[0],xmm1[0],xmm13[1],xmm1[1]3881; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm0[0],xmm13[1]3882; SSE-NEXT: movapd %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3883; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[1,1,1,1]3884; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3885; SSE-NEXT: # xmm1 = mem[3,3,3,3]3886; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3887; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3888; SSE-NEXT: # xmm1 = mem[2,3,2,3]3889; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload3890; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]3891; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]3892; SSE-NEXT: movapd %xmm15, %xmm103893; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3894; SSE-NEXT: # xmm0 = mem[1,1,1,1]3895; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3896; SSE-NEXT: # xmm1 = mem[3,3,3,3]3897; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3898; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3899; SSE-NEXT: # xmm1 = mem[2,3,2,3]3900; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload3901; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]3902; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]3903; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3904; SSE-NEXT: # xmm0 = mem[1,1,1,1]3905; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3906; SSE-NEXT: # xmm1 = mem[3,3,3,3]3907; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3908; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3909; SSE-NEXT: # xmm1 = mem[2,3,2,3]3910; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload3911; SSE-NEXT: punpckldq {{.*#+}} xmm13 = xmm13[0],xmm1[0],xmm13[1],xmm1[1]3912; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm0[0],xmm13[1]3913; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload3914; SSE-NEXT: # xmm0 = mem[1,1,1,1]3915; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[3,3,3,3]3916; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]3917; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload3918; SSE-NEXT: # xmm1 = mem[2,3,2,3]3919; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload3920; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]3921; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]3922; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3923; SSE-NEXT: movaps %xmm0, 96(%rsi)3924; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3925; SSE-NEXT: movaps %xmm0, 32(%rsi)3926; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3927; SSE-NEXT: movaps %xmm0, 112(%rsi)3928; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3929; SSE-NEXT: movaps %xmm0, 48(%rsi)3930; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3931; SSE-NEXT: movaps %xmm0, 64(%rsi)3932; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3933; SSE-NEXT: movaps %xmm0, (%rsi)3934; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3935; SSE-NEXT: movaps %xmm0, 80(%rsi)3936; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3937; SSE-NEXT: movaps %xmm0, 16(%rsi)3938; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3939; SSE-NEXT: movaps %xmm0, 96(%rdx)3940; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3941; SSE-NEXT: movaps %xmm0, 32(%rdx)3942; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3943; SSE-NEXT: movaps %xmm0, 112(%rdx)3944; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3945; SSE-NEXT: movaps %xmm0, 48(%rdx)3946; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3947; SSE-NEXT: movaps %xmm0, 64(%rdx)3948; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3949; SSE-NEXT: movaps %xmm0, (%rdx)3950; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3951; SSE-NEXT: movaps %xmm0, 80(%rdx)3952; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3953; SSE-NEXT: movaps %xmm0, 16(%rdx)3954; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3955; SSE-NEXT: movaps %xmm0, 96(%rcx)3956; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3957; SSE-NEXT: movaps %xmm0, 112(%rcx)3958; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3959; SSE-NEXT: movaps %xmm0, 64(%rcx)3960; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3961; SSE-NEXT: movaps %xmm0, 80(%rcx)3962; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3963; SSE-NEXT: movaps %xmm0, 32(%rcx)3964; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3965; SSE-NEXT: movaps %xmm0, 48(%rcx)3966; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3967; SSE-NEXT: movaps %xmm0, (%rcx)3968; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3969; SSE-NEXT: movaps %xmm0, 16(%rcx)3970; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3971; SSE-NEXT: movaps %xmm0, 112(%r8)3972; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3973; SSE-NEXT: movaps %xmm0, 96(%r8)3974; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload3975; SSE-NEXT: movaps %xmm0, 80(%r8)3976; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3977; SSE-NEXT: movaps %xmm0, 64(%r8)3978; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3979; SSE-NEXT: movaps %xmm0, 48(%r8)3980; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3981; SSE-NEXT: movaps %xmm0, 32(%r8)3982; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3983; SSE-NEXT: movaps %xmm0, 16(%r8)3984; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3985; SSE-NEXT: movaps %xmm0, (%r8)3986; SSE-NEXT: movapd %xmm2, 112(%r9)3987; SSE-NEXT: movapd %xmm3, 96(%r9)3988; SSE-NEXT: movapd %xmm4, 80(%r9)3989; SSE-NEXT: movapd %xmm6, 64(%r9)3990; SSE-NEXT: movapd %xmm8, 48(%r9)3991; SSE-NEXT: movapd %xmm9, 32(%r9)3992; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3993; SSE-NEXT: movaps %xmm0, 16(%r9)3994; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3995; SSE-NEXT: movaps %xmm0, (%r9)3996; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax3997; SSE-NEXT: movapd %xmm14, 112(%rax)3998; SSE-NEXT: movapd %xmm13, 96(%rax)3999; SSE-NEXT: movapd %xmm15, 80(%rax)4000; SSE-NEXT: movapd %xmm10, 64(%rax)4001; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4002; SSE-NEXT: movaps %xmm0, 48(%rax)4003; SSE-NEXT: movapd %xmm12, 32(%rax)4004; SSE-NEXT: movapd %xmm11, 16(%rax)4005; SSE-NEXT: movapd %xmm7, (%rax)4006; SSE-NEXT: addq $1032, %rsp # imm = 0x4084007; SSE-NEXT: retq4008;4009; AVX-LABEL: load_i32_stride6_vf32:4010; AVX: # %bb.0:4011; AVX-NEXT: subq $1048, %rsp # imm = 0x4184012; AVX-NEXT: vmovaps 416(%rdi), %ymm74013; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4014; AVX-NEXT: vmovaps 384(%rdi), %ymm94015; AVX-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4016; AVX-NEXT: vmovapd 160(%rdi), %ymm54017; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4018; AVX-NEXT: vmovapd 128(%rdi), %ymm64019; AVX-NEXT: vmovaps 96(%rdi), %ymm24020; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4021; AVX-NEXT: vmovaps 64(%rdi), %ymm34022; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4023; AVX-NEXT: vmovaps 32(%rdi), %ymm04024; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4025; AVX-NEXT: vmovaps (%rdi), %ymm14026; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4027; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]4028; AVX-NEXT: vextractf128 $1, %ymm8, %xmm44029; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm8[0,1],xmm4[2,3]4030; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,3]4031; AVX-NEXT: vinsertf128 $1, 96(%rdi), %ymm3, %ymm14032; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4033; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[0,0],ymm2[6,4],ymm1[4,4]4034; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm3[2,2],ymm1[6,4],ymm3[6,6]4035; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4036; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4037; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm6[2,3],ymm5[0,1]4038; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4039; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm6[1],ymm1[3],ymm6[2]4040; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]4041; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4042; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4043; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5],ymm9[6,7]4044; AVX-NEXT: vextractf128 $1, %ymm11, %xmm34045; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm11[0,1],xmm3[2,3]4046; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,3]4047; AVX-NEXT: vmovaps 480(%rdi), %ymm14048; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4049; AVX-NEXT: vmovaps 448(%rdi), %ymm24050; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4051; AVX-NEXT: vinsertf128 $1, 480(%rdi), %ymm2, %ymm94052; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm9[0,0],ymm1[6,4],ymm9[4,4]4053; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,2],ymm1[6,4],ymm2[6,6]4054; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4055; AVX-NEXT: vmovapd 544(%rdi), %ymm14056; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4057; AVX-NEXT: vmovapd 512(%rdi), %ymm24058; AVX-NEXT: vmovupd %ymm2, (%rsp) # 32-byte Spill4059; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[0,1]4060; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4061; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[3],ymm2[2]4062; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]4063; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4064; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4065; AVX-NEXT: vmovaps 224(%rdi), %ymm04066; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4067; AVX-NEXT: vmovaps 192(%rdi), %ymm14068; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4069; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]4070; AVX-NEXT: vextractf128 $1, %ymm10, %xmm14071; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm10[0,1],xmm1[2,3]4072; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,3]4073; AVX-NEXT: vmovaps 288(%rdi), %ymm24074; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4075; AVX-NEXT: vmovaps 256(%rdi), %ymm54076; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4077; AVX-NEXT: vinsertf128 $1, 288(%rdi), %ymm5, %ymm64078; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4079; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0],ymm6[0,0],ymm2[6,4],ymm6[4,4]4080; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0],ymm5[2,2],ymm2[6,4],ymm5[6,6]4081; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]4082; AVX-NEXT: vmovapd 352(%rdi), %ymm24083; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4084; AVX-NEXT: vmovapd 320(%rdi), %ymm54085; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4086; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm5[2,3],ymm2[0,1]4087; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4088; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[3],ymm5[2]4089; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,2,0,4,5,6,4]4090; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]4091; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4092; AVX-NEXT: vmovaps 608(%rdi), %ymm04093; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4094; AVX-NEXT: vmovaps 576(%rdi), %ymm24095; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4096; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3],ymm0[4,5],ymm2[6,7]4097; AVX-NEXT: vextractf128 $1, %ymm7, %xmm24098; AVX-NEXT: vblendps {{.*#+}} xmm6 = xmm7[0,1],xmm2[2,3]4099; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,2],xmm2[0,3]4100; AVX-NEXT: vmovaps 672(%rdi), %ymm54101; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4102; AVX-NEXT: vmovaps 640(%rdi), %ymm04103; AVX-NEXT: vinsertf128 $1, 672(%rdi), %ymm0, %ymm154104; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4105; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm5[2,0],ymm15[0,0],ymm5[6,4],ymm15[4,4]4106; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm14[2,0],ymm0[2,2],ymm14[6,4],ymm0[6,6]4107; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm6[0,1,2],ymm14[3,4,5,6,7]4108; AVX-NEXT: vmovapd 736(%rdi), %ymm04109; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4110; AVX-NEXT: vmovapd 704(%rdi), %ymm54111; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4112; AVX-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm5[2,3],ymm0[0,1]4113; AVX-NEXT: vshufpd {{.*#+}} ymm13 = ymm12[0],ymm5[1],ymm12[3],ymm5[2]4114; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[0,1,2,0,4,5,6,4]4115; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3,4,5],ymm13[6,7]4116; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4117; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm8[1,0],xmm4[3,0]4118; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm8[0,2],xmm4[1,3]4119; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4120; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm8 # 32-byte Folded Reload4121; AVX-NEXT: # ymm8 = ymm13[3,0],mem[1,0],ymm13[7,4],mem[5,4]4122; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload4123; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[2,0],ymm14[2,3],ymm8[6,4],ymm14[6,7]4124; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm8[3,4,5,6,7]4125; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload4126; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4127; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,1],ymm5[1,3],ymm0[7,5],ymm5[5,7]4128; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,1,2,0,4,5,6,4]4129; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm8[6,7]4130; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4131; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm11[1,0],xmm3[3,0]4132; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm4[0,2],xmm3[1,3]4133; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload4134; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm8[3,0],ymm9[1,0],ymm8[7,4],ymm9[5,4]4135; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload4136; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,0],ymm6[2,3],ymm4[6,4],ymm6[6,7]4137; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3,4,5,6,7]4138; AVX-NEXT: vmovups (%rsp), %ymm9 # 32-byte Reload4139; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4140; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm0[3,1],ymm9[1,3],ymm0[7,5],ymm9[5,7]4141; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,2,0,4,5,6,4]4142; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm4[6,7]4143; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4144; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm10[1,0],xmm1[3,0]4145; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm3[0,2],xmm1[1,3]4146; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4147; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload4148; AVX-NEXT: # ymm3 = ymm0[3,0],mem[1,0],ymm0[7,4],mem[5,4]4149; AVX-NEXT: vshufps $226, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload4150; AVX-NEXT: # ymm3 = ymm3[2,0],mem[2,3],ymm3[6,4],mem[6,7]4151; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm3[3,4,5,6,7]4152; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4153; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload4154; AVX-NEXT: # ymm3 = ymm0[3,1],mem[1,3],ymm0[7,5],mem[5,7]4155; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,1,2,0,4,5,6,4]4156; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm3[6,7]4157; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4158; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm7[1,0],xmm2[3,0]4159; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm1[0,2],xmm2[1,3]4160; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload4161; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm10[3,0],ymm15[1,0],ymm10[7,4],ymm15[5,4]4162; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload4163; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm11[2,3],ymm1[6,4],ymm11[6,7]4164; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4165; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload4166; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm12[3,1],ymm15[1,3],ymm12[7,5],ymm15[5,7]4167; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]4168; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4169; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4170; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4171; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload4172; AVX-NEXT: # ymm4 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]4173; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm13[2,1],ymm14[2,0],ymm13[6,5],ymm14[6,4]4174; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[2,3,0,1]4175; AVX-NEXT: vextractf128 $1, %ymm4, %xmm04176; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4177; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm4[2,0],xmm0[2,3]4178; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3,4,5,6,7]4179; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm3 # 32-byte Folded Reload4180; AVX-NEXT: # ymm3 = ymm5[0,1,2,3],mem[4,5],ymm5[6,7]4181; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm3[2,3,0,1]4182; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4183; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm0[0,0],ymm3[2,0],ymm0[4,4],ymm3[6,4]4184; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm5[5,6,7]4185; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4186; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4187; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4188; AVX-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]4189; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4190; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm8[2,1],ymm6[2,0],ymm8[6,5],ymm6[6,4]4191; AVX-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm5[2,3,0,1]4192; AVX-NEXT: vextractf128 $1, %ymm0, %xmm14193; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4194; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm0[2,0],xmm1[2,3]4195; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3,4,5,6,7]4196; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm2 # 32-byte Folded Reload4197; AVX-NEXT: # ymm2 = ymm9[0,1,2,3],mem[4,5],ymm9[6,7]4198; AVX-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm2[2,3,0,1]4199; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm6[0,0],ymm2[2,0],ymm6[4,4],ymm2[6,4]4200; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4],ymm7[5,6,7]4201; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4202; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4203; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload4204; AVX-NEXT: # ymm5 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]4205; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm10[2,1],ymm11[2,0],ymm10[6,5],ymm11[6,4]4206; AVX-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm7[2,3,0,1]4207; AVX-NEXT: vextractf128 $1, %ymm5, %xmm04208; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4209; AVX-NEXT: vshufps {{.*#+}} xmm11 = xmm5[2,0],xmm0[2,3]4210; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm11[0,1,2],ymm10[3,4,5,6,7]4211; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm10 # 32-byte Folded Reload4212; AVX-NEXT: # ymm10 = ymm15[0,1,2,3],mem[4,5],ymm15[6,7]4213; AVX-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm10[2,3,0,1]4214; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm11[0,0],ymm10[2,0],ymm11[4,4],ymm10[6,4]4215; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4],ymm13[5,6,7]4216; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4217; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4218; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload4219; AVX-NEXT: # ymm12 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]4220; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload4221; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload4222; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm9[2,1],ymm8[2,0],ymm9[6,5],ymm8[6,4]4223; AVX-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm13[2,3,0,1]4224; AVX-NEXT: vextractf128 $1, %ymm12, %xmm144225; AVX-NEXT: vshufps {{.*#+}} xmm15 = xmm12[2,0],xmm14[2,3]4226; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0,1,2],ymm13[3,4,5,6,7]4227; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4228; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload4229; AVX-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]4230; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3,0,1]4231; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm0[0,0],ymm15[2,0],ymm0[4,4],ymm15[6,4]4232; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm13[0,1,2,3,4],ymm1[5,6,7]4233; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4234; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4235; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1],ymm3[3,1],ymm1[4,5],ymm3[7,5]4236; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm3 # 16-byte Folded Reload4237; AVX-NEXT: # xmm3 = xmm4[3,1],mem[3,3]4238; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4239; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload4240; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm13[3,1],ymm7[2,1],ymm13[7,5],ymm7[6,5]4241; AVX-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3,0,1]4242; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3,4,5,6,7]4243; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4],ymm1[5,6,7]4244; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4245; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm6[0,1],ymm2[3,1],ymm6[4,5],ymm2[7,5]4246; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4247; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload4248; AVX-NEXT: # xmm2 = xmm2[3,1],mem[3,3]4249; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4250; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload4251; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm4[3,1],ymm6[2,1],ymm4[7,5],ymm6[6,5]4252; AVX-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3,0,1]4253; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3,4,5,6,7]4254; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6,7]4255; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4256; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm15[3,1],ymm0[4,5],ymm15[7,5]4257; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm12[3,1],xmm14[3,3]4258; AVX-NEXT: vmovaps %ymm9, %ymm34259; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm9[3,1],ymm8[2,1],ymm9[7,5],ymm8[6,5]4260; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,0,1]4261; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]4262; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]4263; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4264; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm11[0,1],ymm10[3,1],ymm11[4,5],ymm10[7,5]4265; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm1 # 16-byte Folded Reload4266; AVX-NEXT: # xmm1 = xmm5[3,1],mem[3,3]4267; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload4268; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload4269; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm11[3,1],ymm15[2,1],ymm11[7,5],ymm15[6,5]4270; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,0,1]4271; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]4272; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]4273; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4274; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload4275; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload4276; AVX-NEXT: # ymm2 = mem[0,1],ymm0[2,3],mem[4,5,6,7]4277; AVX-NEXT: vmovaps 416(%rdi), %xmm04278; AVX-NEXT: vmovaps %xmm0, (%rsp) # 16-byte Spill4279; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]4280; AVX-NEXT: vmovaps 400(%rdi), %xmm14281; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4282; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]4283; AVX-NEXT: vmovapd 464(%rdi), %xmm14284; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4285; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1],ymm6[0],ymm1[2],ymm6[3]4286; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,1],ymm1[2,0],ymm4[4,5],ymm1[6,4]4287; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]4288; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4289; AVX-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm2[2,3,0,1]4290; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm9[0,0],ymm2[6,4],ymm9[4,4]4291; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm9[0,2],ymm1[2,0],ymm9[4,6],ymm1[6,4]4292; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4293; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4294; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4295; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload4296; AVX-NEXT: # ymm4 = mem[0,1],ymm0[2,3],mem[4,5,6,7]4297; AVX-NEXT: vmovaps 32(%rdi), %xmm04298; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4299; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]4300; AVX-NEXT: vmovaps 16(%rdi), %xmm144301; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm14[0],xmm0[1],xmm14[2,3]4302; AVX-NEXT: vmovapd 80(%rdi), %xmm54303; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm5[1],ymm7[0],ymm5[2],ymm7[3]4304; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm13[0,1],ymm1[2,0],ymm13[4,5],ymm1[6,4]4305; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]4306; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm4[2,3,0,1]4307; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm4[2,0],ymm1[0,0],ymm4[6,4],ymm1[4,4]4308; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm1[0,2],ymm2[2,0],ymm1[4,6],ymm2[6,4]4309; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]4310; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4311; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4312; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload4313; AVX-NEXT: # ymm6 = mem[0,1],ymm0[2,3],mem[4,5,6,7]4314; AVX-NEXT: vmovaps 224(%rdi), %xmm04315; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4316; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]4317; AVX-NEXT: vmovaps 208(%rdi), %xmm24318; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4319; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2,3]4320; AVX-NEXT: vmovapd 272(%rdi), %xmm24321; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4322; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[1],ymm8[0],ymm2[2],ymm8[3]4323; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,0],ymm3[4,5],ymm2[6,4]4324; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3,4,5,6,7]4325; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm6[2,3,0,1]4326; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm6[2,0],ymm2[0,0],ymm6[6,4],ymm2[4,4]4327; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm2[0,2],ymm3[2,0],ymm2[4,6],ymm3[6,4]4328; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]4329; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4330; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4331; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload4332; AVX-NEXT: # ymm12 = mem[0,1],ymm0[2,3],mem[4,5,6,7]4333; AVX-NEXT: vmovaps 608(%rdi), %xmm134334; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm13[2,2,3,3]4335; AVX-NEXT: vmovaps 592(%rdi), %xmm84336; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm8[0],xmm0[1],xmm8[2,3]4337; AVX-NEXT: vmovapd 656(%rdi), %xmm104338; AVX-NEXT: vmovaps %ymm15, %ymm34339; AVX-NEXT: vshufpd {{.*#+}} ymm15 = ymm10[1],ymm15[0],ymm10[2],ymm15[3]4340; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm11[0,1],ymm15[2,0],ymm11[4,5],ymm15[6,4]4341; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1],ymm15[2,3,4,5,6,7]4342; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm12[2,3,0,1]4343; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm12[2,0],ymm0[0,0],ymm12[6,4],ymm0[4,4]4344; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm0[0,2],ymm11[2,0],ymm0[4,6],ymm11[6,4]4345; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4],ymm11[5,6,7]4346; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm4[3,0],ymm1[1,0],ymm4[7,4],ymm1[5,4]4347; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,3],ymm7[2,0],ymm1[4,7],ymm7[6,4]4348; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm7 # 16-byte Folded Reload4349; AVX-NEXT: # xmm7 = xmm14[0,1],mem[2,3]4350; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload4351; AVX-NEXT: # ymm5 = ymm5[3,1],mem[1,3],ymm5[7,5],mem[5,7]4352; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4353; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm4[1,1],ymm5[2,0],ymm4[5,5],ymm5[6,4]4354; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm7[1,3,2,3]4355; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1],ymm5[2,3,4,5,6,7]4356; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm5[0,1,2,3,4],ymm1[5,6,7]4357; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4358; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,0],ymm9[1,0],ymm1[7,4],ymm9[5,4]4359; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm9[0,3],ymm1[2,0],ymm9[4,7],ymm1[6,4]4360; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload4361; AVX-NEXT: vblendps $12, (%rsp), %xmm4, %xmm5 # 16-byte Folded Reload4362; AVX-NEXT: # xmm5 = xmm4[0,1],mem[2,3]4363; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4364; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm9 # 32-byte Folded Reload4365; AVX-NEXT: # ymm9 = ymm4[3,1],mem[1,3],ymm4[7,5],mem[5,7]4366; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4367; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm4[1,1],ymm9[2,0],ymm4[5,5],ymm9[6,4]4368; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm5[1,3,2,3]4369; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1],ymm9[2,3,4,5,6,7]4370; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm1[5,6,7]4371; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm6[3,0],ymm2[1,0],ymm6[7,4],ymm2[5,4]4372; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,3],ymm1[2,0],ymm2[4,7],ymm1[6,4]4373; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload4374; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload4375; AVX-NEXT: # xmm2 = xmm2[0,1],mem[2,3]4376; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4377; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm6 # 32-byte Folded Reload4378; AVX-NEXT: # ymm6 = ymm4[3,1],mem[1,3],ymm4[7,5],mem[5,7]4379; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4380; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm4[1,1],ymm6[2,0],ymm4[5,5],ymm6[6,4]4381; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,3,2,3]4382; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm6[2,3,4,5,6,7]4383; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6,7]4384; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm12[3,0],ymm0[1,0],ymm12[7,4],ymm0[5,4]4385; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,3],ymm2[2,0],ymm0[4,7],ymm2[6,4]4386; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm8[0,1],xmm13[2,3]4387; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm10[3,1],ymm3[1,3],ymm10[7,5],ymm3[5,7]4388; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4389; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm4[1,1],ymm3[2,0],ymm4[5,5],ymm3[6,4]4390; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,3,2,3]4391; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3,4,5,6,7]4392; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6,7]4393; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4394; AVX-NEXT: vmovaps %ymm2, 96(%rsi)4395; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4396; AVX-NEXT: vmovaps %ymm2, 32(%rsi)4397; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4398; AVX-NEXT: vmovaps %ymm2, 64(%rsi)4399; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4400; AVX-NEXT: vmovaps %ymm2, (%rsi)4401; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4402; AVX-NEXT: vmovaps %ymm2, 96(%rdx)4403; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4404; AVX-NEXT: vmovaps %ymm2, 32(%rdx)4405; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4406; AVX-NEXT: vmovaps %ymm2, 64(%rdx)4407; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4408; AVX-NEXT: vmovaps %ymm2, (%rdx)4409; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4410; AVX-NEXT: vmovaps %ymm2, 32(%rcx)4411; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4412; AVX-NEXT: vmovaps %ymm2, 96(%rcx)4413; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4414; AVX-NEXT: vmovaps %ymm2, 64(%rcx)4415; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4416; AVX-NEXT: vmovaps %ymm2, (%rcx)4417; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4418; AVX-NEXT: vmovaps %ymm2, 96(%r8)4419; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4420; AVX-NEXT: vmovaps %ymm2, 32(%r8)4421; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4422; AVX-NEXT: vmovaps %ymm2, 64(%r8)4423; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4424; AVX-NEXT: vmovaps %ymm2, (%r8)4425; AVX-NEXT: vmovaps %ymm11, 96(%r9)4426; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4427; AVX-NEXT: vmovaps %ymm2, 32(%r9)4428; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4429; AVX-NEXT: vmovaps %ymm2, (%r9)4430; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4431; AVX-NEXT: vmovaps %ymm2, 64(%r9)4432; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax4433; AVX-NEXT: vmovaps %ymm0, 96(%rax)4434; AVX-NEXT: vmovaps %ymm1, 32(%rax)4435; AVX-NEXT: vmovaps %ymm5, 64(%rax)4436; AVX-NEXT: vmovaps %ymm7, (%rax)4437; AVX-NEXT: addq $1048, %rsp # imm = 0x4184438; AVX-NEXT: vzeroupper4439; AVX-NEXT: retq4440;4441; AVX2-LABEL: load_i32_stride6_vf32:4442; AVX2: # %bb.0:4443; AVX2-NEXT: subq $1224, %rsp # imm = 0x4C84444; AVX2-NEXT: vmovaps 480(%rdi), %ymm94445; AVX2-NEXT: vmovaps 448(%rdi), %ymm114446; AVX2-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4447; AVX2-NEXT: vmovaps 416(%rdi), %ymm84448; AVX2-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4449; AVX2-NEXT: vmovaps 128(%rdi), %ymm24450; AVX2-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill4451; AVX2-NEXT: vmovaps 160(%rdi), %ymm34452; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4453; AVX2-NEXT: vmovaps 96(%rdi), %ymm44454; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4455; AVX2-NEXT: vmovaps (%rdi), %ymm04456; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4457; AVX2-NEXT: vmovaps 32(%rdi), %ymm14458; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4459; AVX2-NEXT: vmovaps 64(%rdi), %ymm54460; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4461; AVX2-NEXT: vmovaps {{.*#+}} xmm10 = [0,6,4,u]4462; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]4463; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4464; AVX2-NEXT: vpermps %ymm0, %ymm10, %ymm04465; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm5[0,1],ymm4[0,1]4466; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm1[0,1,2,3,4,5],ymm4[6,7]4467; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm5[0,2,2,2,4,6,6,6]4468; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4469; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm3[0,1,2,3],ymm2[4,5,6,7]4470; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm6 = [4,2,4,2,4,2,4,2]4471; AVX2-NEXT: vpermps %ymm7, %ymm6, %ymm14472; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4473; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4474; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4475; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm11[0,1],ymm9[0,1]4476; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3,4,5],ymm9[6,7]4477; AVX2-NEXT: vmovaps 384(%rdi), %ymm04478; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4479; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3],ymm8[4,5],ymm0[6,7]4480; AVX2-NEXT: vpermps %ymm3, %ymm10, %ymm04481; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,2,2,2,4,6,6,6]4482; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4483; AVX2-NEXT: vmovaps 512(%rdi), %ymm14484; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4485; AVX2-NEXT: vmovaps 544(%rdi), %ymm24486; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4487; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4,5,6,7]4488; AVX2-NEXT: vpermps %ymm2, %ymm6, %ymm14489; AVX2-NEXT: vmovaps %ymm6, %ymm94490; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4491; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4492; AVX2-NEXT: vmovaps 288(%rdi), %ymm14493; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4494; AVX2-NEXT: vmovaps 256(%rdi), %ymm04495; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4496; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]4497; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3,4,5],ymm1[6,7]4498; AVX2-NEXT: vmovaps 224(%rdi), %ymm04499; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4500; AVX2-NEXT: vmovaps 192(%rdi), %ymm64501; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4502; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm0[4,5],ymm6[6,7]4503; AVX2-NEXT: vpermps %ymm1, %ymm10, %ymm84504; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm13[0,2,2,2,4,6,6,6]4505; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm8[0,1,2],ymm11[3,4,5,6,7]4506; AVX2-NEXT: vmovaps 320(%rdi), %ymm64507; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4508; AVX2-NEXT: vmovaps 352(%rdi), %ymm84509; AVX2-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4510; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm6[4,5,6,7]4511; AVX2-NEXT: vpermps %ymm8, %ymm9, %ymm144512; AVX2-NEXT: vmovaps %ymm9, %ymm04513; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4514; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm14[6,7]4515; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4516; AVX2-NEXT: vmovaps 608(%rdi), %ymm64517; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4518; AVX2-NEXT: vmovaps 576(%rdi), %ymm94519; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4520; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm9[0,1,2,3],ymm6[4,5],ymm9[6,7]4521; AVX2-NEXT: vpermps %ymm14, %ymm10, %ymm104522; AVX2-NEXT: vmovaps 672(%rdi), %ymm64523; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4524; AVX2-NEXT: vmovaps 640(%rdi), %ymm94525; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4526; AVX2-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm9[0,1],ymm6[0,1]4527; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm6[6,7]4528; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm12[0,2,2,2,4,6,6,6]4529; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2],ymm15[3,4,5,6,7]4530; AVX2-NEXT: vmovaps 704(%rdi), %ymm64531; AVX2-NEXT: vmovaps 736(%rdi), %ymm114532; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm11[0,1,2,3],ymm6[4,5,6,7]4533; AVX2-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4534; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4535; AVX2-NEXT: vpermps %ymm15, %ymm0, %ymm94536; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]4537; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4538; AVX2-NEXT: vmovaps {{.*#+}} xmm9 = [1,7,5,u]4539; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm10 # 32-byte Folded Reload4540; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm5[1,3,2,3,5,7,6,7]4541; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0,1,2],ymm5[3,4,5,6,7]4542; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm10 = [5,3,5,3,5,3,5,3]4543; AVX2-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4544; AVX2-NEXT: vpermps %ymm7, %ymm10, %ymm74545; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]4546; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4547; AVX2-NEXT: vpermps %ymm3, %ymm9, %ymm34548; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm4[1,3,2,3,5,7,6,7]4549; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3,4,5,6,7]4550; AVX2-NEXT: vpermps %ymm2, %ymm10, %ymm24551; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]4552; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4553; AVX2-NEXT: vpermps %ymm1, %ymm9, %ymm04554; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,3,2,3,5,7,6,7]4555; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4556; AVX2-NEXT: vpermps %ymm8, %ymm10, %ymm14557; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4558; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4559; AVX2-NEXT: vpermps %ymm14, %ymm9, %ymm04560; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,3,2,3,5,7,6,7]4561; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4562; AVX2-NEXT: vpermps %ymm15, %ymm10, %ymm14563; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4564; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4565; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload4566; AVX2-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload4567; AVX2-NEXT: # ymm0 = mem[0,1],ymm14[2,3],mem[4,5],ymm14[6,7]4568; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]4569; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]4570; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4571; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload4572; AVX2-NEXT: # ymm3 = mem[0,1],ymm1[2,3],mem[4,5,6,7]4573; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,0,2,3,6,4,6,7]4574; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,3,2,3]4575; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]4576; AVX2-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload4577; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm4 # 32-byte Folded Reload4578; AVX2-NEXT: # ymm4 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]4579; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,0,2,0,4,4,6,4]4580; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]4581; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4582; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4583; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4584; AVX2-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4585; AVX2-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]4586; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]4587; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]4588; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4589; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload4590; AVX2-NEXT: # ymm15 = ymm1[0,1],mem[2,3],ymm1[4,5,6,7]4591; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm15[2,0,2,3,6,4,6,7]4592; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,3,2,3]4593; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]4594; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4595; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload4596; AVX2-NEXT: # ymm2 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]4597; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,0,2,0,4,4,6,4]4598; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]4599; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4600; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4601; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4602; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload4603; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1],ymm12[2,3],ymm13[4,5],ymm12[6,7]4604; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]4605; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[0,2,0,3]4606; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4607; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4608; AVX2-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5,6,7]4609; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm0[2,0,2,3,6,4,6,7]4610; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,3,2,3]4611; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm1[3,4,5,6,7]4612; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm11[4,5],ymm6[6,7]4613; AVX2-NEXT: vshufps {{.*#+}} ymm8 = ymm1[0,0,2,0,4,4,6,4]4614; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]4615; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4],ymm8[5,6,7]4616; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4617; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload4618; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload4619; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm6[0,1],ymm11[2,3],ymm6[4,5],ymm11[6,7]4620; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm7[2,0,2,3,6,4,6,7]4621; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,2,0,3]4622; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload4623; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload4624; AVX2-NEXT: # ymm8 = mem[0,1],ymm8[2,3],mem[4,5,6,7]4625; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm8[2,0,2,3,6,4,6,7]4626; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,3,2,3]4627; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1,2],ymm7[3,4,5,6,7]4628; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload4629; AVX2-NEXT: vblendps $207, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload4630; AVX2-NEXT: # ymm9 = mem[0,1,2,3],ymm9[4,5],mem[6,7]4631; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm9[0,0,2,0,4,4,6,4]4632; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,1,0,3]4633; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4],ymm10[5,6,7]4634; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4635; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm14[3,3,3,3,7,7,7,7]4636; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload4637; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm10[1],ymm7[2,3,4],ymm10[5],ymm7[6,7]4638; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,2,0,3]4639; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,1,3,3,7,5,7,7]4640; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]4641; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm7[3,4,5,6,7]4642; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,3,1,4,5,7,5]4643; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,3]4644; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm4[5,6,7]4645; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4646; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload4647; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm5[3,3,3,3,7,7,7,7]4648; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload4649; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm7[1],ymm3[2,3,4],ymm7[5],ymm3[6,7]4650; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,2,0,3]4651; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm15[3,1,3,3,7,5,7,7]4652; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,3,2,3]4653; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3,4,5,6,7]4654; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,3,1,4,5,7,5]4655; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]4656; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]4657; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4658; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]4659; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7]4660; AVX2-NEXT: vmovaps %ymm6, %ymm44661; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,0,3]4662; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm8[3,1,3,3,7,5,7,7]4663; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]4664; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3,4,5,6,7]4665; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm9[0,1,3,1,4,5,7,5]4666; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]4667; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6,7]4668; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4669; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]4670; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm13[1],ymm2[2,3,4],ymm13[5],ymm2[6,7]4671; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,0,3]4672; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,1,3,3,7,5,7,7]4673; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,3,2,3]4674; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]4675; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,3,1,4,5,7,5]4676; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]4677; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4678; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4679; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm5[4,5,6,7]4680; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4681; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4682; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload4683; AVX2-NEXT: # ymm3 = ymm0[0,1,2,3],mem[4,5,6,7]4684; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4685; AVX2-NEXT: vmovaps 464(%rdi), %xmm04686; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4687; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]4688; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]4689; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload4690; AVX2-NEXT: vpermps %ymm3, %ymm8, %ymm14691; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5,6,7]4692; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4693; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload4694; AVX2-NEXT: # ymm7 = mem[0,1],ymm1[2,3],mem[4,5,6,7]4695; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,2,0,6,0,2,0,6]4696; AVX2-NEXT: # ymm3 = mem[0,1,0,1]4697; AVX2-NEXT: vpermps %ymm7, %ymm3, %ymm14698; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4699; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4700; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm10[0,1,2,3],ymm14[4,5,6,7]4701; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4702; AVX2-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload4703; AVX2-NEXT: # ymm14 = mem[0,1,2,3],ymm0[4,5,6,7]4704; AVX2-NEXT: vmovaps 80(%rdi), %xmm54705; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7]4706; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]4707; AVX2-NEXT: vpermps %ymm14, %ymm8, %ymm14708; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5,6,7]4709; AVX2-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload4710; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload4711; AVX2-NEXT: # ymm2 = mem[0,1],ymm1[2,3],mem[4,5,6,7]4712; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm14713; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4714; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4715; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm4[0,1,2,3],ymm11[4,5,6,7]4716; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4717; AVX2-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload4718; AVX2-NEXT: # ymm12 = mem[0,1,2,3],ymm0[4,5,6,7]4719; AVX2-NEXT: vmovaps 272(%rdi), %xmm44720; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1],ymm4[2,3],ymm9[4,5,6,7]4721; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]4722; AVX2-NEXT: vpermps %ymm12, %ymm8, %ymm114723; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1],ymm0[2,3,4,5,6,7]4724; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4725; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload4726; AVX2-NEXT: # ymm1 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]4727; AVX2-NEXT: vpermps %ymm1, %ymm3, %ymm154728; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4],ymm15[5,6,7]4729; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill4730; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm11 # 32-byte Folded Reload4731; AVX2-NEXT: # ymm11 = ymm13[0,1,2,3],mem[4,5,6,7]4732; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4733; AVX2-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload4734; AVX2-NEXT: # ymm15 = mem[0,1,2,3],ymm0[4,5,6,7]4735; AVX2-NEXT: vmovaps 656(%rdi), %xmm04736; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1],ymm0[2,3],ymm11[4,5,6,7]4737; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm10[0,1,0,2,4,5,4,6]4738; AVX2-NEXT: vpermps %ymm15, %ymm8, %ymm84739; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm10[2,3,4,5,6,7]4740; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload4741; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload4742; AVX2-NEXT: # ymm10 = ymm10[0,1],mem[2,3],ymm10[4,5,6,7]4743; AVX2-NEXT: vpermps %ymm10, %ymm3, %ymm34744; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4],ymm3[5,6,7]4745; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm6[1,1,1,1,5,5,5,5]4746; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2],ymm5[3],ymm6[4,5,6,7]4747; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4748; AVX2-NEXT: vpermps %ymm14, %ymm13, %ymm64749; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3,4,5,6,7]4750; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm6 = [0,3,1,7,0,3,1,7]4751; AVX2-NEXT: # ymm6 = mem[0,1,0,1]4752; AVX2-NEXT: vpermps %ymm2, %ymm6, %ymm24753; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4],ymm2[5,6,7]4754; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload4755; AVX2-NEXT: # ymm5 = mem[1,1,1,1,5,5,5,5]4756; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload4757; AVX2-NEXT: # ymm5 = ymm5[0,1,2],mem[3],ymm5[4,5,6,7]4758; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm8 # 32-byte Folded Reload4759; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1],ymm5[2,3,4,5,6,7]4760; AVX2-NEXT: vpermps %ymm7, %ymm6, %ymm74761; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm7[5,6,7]4762; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm9[1,1,1,1,5,5,5,5]4763; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm7[0,1,2],ymm4[3],ymm7[4,5,6,7]4764; AVX2-NEXT: vpermps %ymm12, %ymm13, %ymm74765; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm7[0,1],ymm4[2,3,4,5,6,7]4766; AVX2-NEXT: vpermps %ymm1, %ymm6, %ymm14767; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4],ymm1[5,6,7]4768; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm11[1,1,1,1,5,5,5,5]4769; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6,7]4770; AVX2-NEXT: vpermps %ymm15, %ymm13, %ymm44771; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3,4,5,6,7]4772; AVX2-NEXT: vpermps %ymm10, %ymm6, %ymm44773; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5,6,7]4774; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4775; AVX2-NEXT: vmovaps %ymm4, 96(%rsi)4776; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4777; AVX2-NEXT: vmovaps %ymm4, 32(%rsi)4778; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4779; AVX2-NEXT: vmovaps %ymm4, 64(%rsi)4780; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4781; AVX2-NEXT: vmovaps %ymm4, (%rsi)4782; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4783; AVX2-NEXT: vmovaps %ymm4, 96(%rdx)4784; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4785; AVX2-NEXT: vmovaps %ymm4, 32(%rdx)4786; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4787; AVX2-NEXT: vmovaps %ymm4, 64(%rdx)4788; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4789; AVX2-NEXT: vmovaps %ymm4, (%rdx)4790; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4791; AVX2-NEXT: vmovaps %ymm4, 32(%rcx)4792; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4793; AVX2-NEXT: vmovaps %ymm4, 96(%rcx)4794; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4795; AVX2-NEXT: vmovaps %ymm4, 64(%rcx)4796; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4797; AVX2-NEXT: vmovaps %ymm4, (%rcx)4798; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4799; AVX2-NEXT: vmovaps %ymm4, 96(%r8)4800; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4801; AVX2-NEXT: vmovaps %ymm4, 32(%r8)4802; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4803; AVX2-NEXT: vmovaps %ymm4, 64(%r8)4804; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload4805; AVX2-NEXT: vmovaps %ymm4, (%r8)4806; AVX2-NEXT: vmovaps %ymm3, 96(%r9)4807; AVX2-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload4808; AVX2-NEXT: vmovaps %ymm3, 32(%r9)4809; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload4810; AVX2-NEXT: vmovaps %ymm3, (%r9)4811; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload4812; AVX2-NEXT: vmovaps %ymm3, 64(%r9)4813; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax4814; AVX2-NEXT: vmovaps %ymm0, 96(%rax)4815; AVX2-NEXT: vmovaps %ymm1, 32(%rax)4816; AVX2-NEXT: vmovaps %ymm5, 64(%rax)4817; AVX2-NEXT: vmovaps %ymm2, (%rax)4818; AVX2-NEXT: addq $1224, %rsp # imm = 0x4C84819; AVX2-NEXT: vzeroupper4820; AVX2-NEXT: retq4821;4822; AVX2-FP-LABEL: load_i32_stride6_vf32:4823; AVX2-FP: # %bb.0:4824; AVX2-FP-NEXT: subq $1224, %rsp # imm = 0x4C84825; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm94826; AVX2-FP-NEXT: vmovaps 448(%rdi), %ymm114827; AVX2-FP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4828; AVX2-FP-NEXT: vmovaps 416(%rdi), %ymm84829; AVX2-FP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4830; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm24831; AVX2-FP-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill4832; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm34833; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4834; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm44835; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4836; AVX2-FP-NEXT: vmovaps (%rdi), %ymm04837; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4838; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm14839; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4840; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm54841; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4842; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm10 = [0,6,4,u]4843; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]4844; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4845; AVX2-FP-NEXT: vpermps %ymm0, %ymm10, %ymm04846; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm5[0,1],ymm4[0,1]4847; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm1[0,1,2,3,4,5],ymm4[6,7]4848; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[0,2,2,2,4,6,6,6]4849; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4850; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm3[0,1,2,3],ymm2[4,5,6,7]4851; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm6 = [4,2,4,2,4,2,4,2]4852; AVX2-FP-NEXT: vpermps %ymm7, %ymm6, %ymm14853; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4854; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4855; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4856; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm11[0,1],ymm9[0,1]4857; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3,4,5],ymm9[6,7]4858; AVX2-FP-NEXT: vmovaps 384(%rdi), %ymm04859; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4860; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3],ymm8[4,5],ymm0[6,7]4861; AVX2-FP-NEXT: vpermps %ymm3, %ymm10, %ymm04862; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,2,2,2,4,6,6,6]4863; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4864; AVX2-FP-NEXT: vmovaps 512(%rdi), %ymm14865; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4866; AVX2-FP-NEXT: vmovaps 544(%rdi), %ymm24867; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4868; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4,5,6,7]4869; AVX2-FP-NEXT: vpermps %ymm2, %ymm6, %ymm14870; AVX2-FP-NEXT: vmovaps %ymm6, %ymm94871; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4872; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4873; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm14874; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4875; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm04876; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4877; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]4878; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3,4,5],ymm1[6,7]4879; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm04880; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4881; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm64882; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4883; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm0[4,5],ymm6[6,7]4884; AVX2-FP-NEXT: vpermps %ymm1, %ymm10, %ymm84885; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm13[0,2,2,2,4,6,6,6]4886; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm8[0,1,2],ymm11[3,4,5,6,7]4887; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm64888; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4889; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm84890; AVX2-FP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4891; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm6[4,5,6,7]4892; AVX2-FP-NEXT: vpermps %ymm8, %ymm9, %ymm144893; AVX2-FP-NEXT: vmovaps %ymm9, %ymm04894; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4895; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm14[6,7]4896; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4897; AVX2-FP-NEXT: vmovaps 608(%rdi), %ymm64898; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4899; AVX2-FP-NEXT: vmovaps 576(%rdi), %ymm94900; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4901; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm9[0,1,2,3],ymm6[4,5],ymm9[6,7]4902; AVX2-FP-NEXT: vpermps %ymm14, %ymm10, %ymm104903; AVX2-FP-NEXT: vmovaps 672(%rdi), %ymm64904; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4905; AVX2-FP-NEXT: vmovaps 640(%rdi), %ymm94906; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4907; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm9[0,1],ymm6[0,1]4908; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm6[6,7]4909; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm12[0,2,2,2,4,6,6,6]4910; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2],ymm15[3,4,5,6,7]4911; AVX2-FP-NEXT: vmovaps 704(%rdi), %ymm64912; AVX2-FP-NEXT: vmovaps 736(%rdi), %ymm114913; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm11[0,1,2,3],ymm6[4,5,6,7]4914; AVX2-FP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4915; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4916; AVX2-FP-NEXT: vpermps %ymm15, %ymm0, %ymm94917; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]4918; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4919; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm9 = [1,7,5,u]4920; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm10 # 32-byte Folded Reload4921; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm5[1,3,2,3,5,7,6,7]4922; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0,1,2],ymm5[3,4,5,6,7]4923; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm10 = [5,3,5,3,5,3,5,3]4924; AVX2-FP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4925; AVX2-FP-NEXT: vpermps %ymm7, %ymm10, %ymm74926; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]4927; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4928; AVX2-FP-NEXT: vpermps %ymm3, %ymm9, %ymm34929; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[1,3,2,3,5,7,6,7]4930; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3,4,5,6,7]4931; AVX2-FP-NEXT: vpermps %ymm2, %ymm10, %ymm24932; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]4933; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4934; AVX2-FP-NEXT: vpermps %ymm1, %ymm9, %ymm04935; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,3,2,3,5,7,6,7]4936; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4937; AVX2-FP-NEXT: vpermps %ymm8, %ymm10, %ymm14938; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4939; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4940; AVX2-FP-NEXT: vpermps %ymm14, %ymm9, %ymm04941; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,3,2,3,5,7,6,7]4942; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]4943; AVX2-FP-NEXT: vpermps %ymm15, %ymm10, %ymm14944; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]4945; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4946; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload4947; AVX2-FP-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload4948; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm14[2,3],mem[4,5],ymm14[6,7]4949; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]4950; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]4951; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4952; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload4953; AVX2-FP-NEXT: # ymm3 = mem[0,1],ymm1[2,3],mem[4,5,6,7]4954; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,0,2,3,6,4,6,7]4955; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,3,2,3]4956; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]4957; AVX2-FP-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload4958; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm4 # 32-byte Folded Reload4959; AVX2-FP-NEXT: # ymm4 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]4960; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,0,2,0,4,4,6,4]4961; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]4962; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4963; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4964; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4965; AVX2-FP-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4966; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]4967; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]4968; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]4969; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4970; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload4971; AVX2-FP-NEXT: # ymm15 = ymm1[0,1],mem[2,3],ymm1[4,5,6,7]4972; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm15[2,0,2,3,6,4,6,7]4973; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,3,2,3]4974; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]4975; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4976; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload4977; AVX2-FP-NEXT: # ymm2 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]4978; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,0,2,0,4,4,6,4]4979; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]4980; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]4981; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4982; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4983; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload4984; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1],ymm12[2,3],ymm13[4,5],ymm12[6,7]4985; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]4986; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[0,2,0,3]4987; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4988; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4989; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5,6,7]4990; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm0[2,0,2,3,6,4,6,7]4991; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,3,2,3]4992; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm1[3,4,5,6,7]4993; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm11[4,5],ymm6[6,7]4994; AVX2-FP-NEXT: vshufps {{.*#+}} ymm8 = ymm1[0,0,2,0,4,4,6,4]4995; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,1,0,3]4996; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4],ymm8[5,6,7]4997; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4998; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload4999; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload5000; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm6[0,1],ymm11[2,3],ymm6[4,5],ymm11[6,7]5001; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm7[2,0,2,3,6,4,6,7]5002; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,2,0,3]5003; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload5004; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload5005; AVX2-FP-NEXT: # ymm8 = mem[0,1],ymm8[2,3],mem[4,5,6,7]5006; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm8[2,0,2,3,6,4,6,7]5007; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,3,2,3]5008; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1,2],ymm7[3,4,5,6,7]5009; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload5010; AVX2-FP-NEXT: vblendps $207, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload5011; AVX2-FP-NEXT: # ymm9 = mem[0,1,2,3],ymm9[4,5],mem[6,7]5012; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm9[0,0,2,0,4,4,6,4]5013; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,1,0,3]5014; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4],ymm10[5,6,7]5015; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5016; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm14[3,3,3,3,7,7,7,7]5017; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload5018; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm10[1],ymm7[2,3,4],ymm10[5],ymm7[6,7]5019; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,2,0,3]5020; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,1,3,3,7,5,7,7]5021; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]5022; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm7[3,4,5,6,7]5023; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,3,1,4,5,7,5]5024; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,3]5025; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm4[5,6,7]5026; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5027; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5028; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm5[3,3,3,3,7,7,7,7]5029; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload5030; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm7[1],ymm3[2,3,4],ymm7[5],ymm3[6,7]5031; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,2,0,3]5032; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm15[3,1,3,3,7,5,7,7]5033; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,3,2,3]5034; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3,4,5,6,7]5035; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,3,1,4,5,7,5]5036; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]5037; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]5038; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5039; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]5040; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7]5041; AVX2-FP-NEXT: vmovaps %ymm6, %ymm45042; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,0,3]5043; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm8[3,1,3,3,7,5,7,7]5044; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]5045; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3,4,5,6,7]5046; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm9[0,1,3,1,4,5,7,5]5047; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]5048; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6,7]5049; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5050; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]5051; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm13[1],ymm2[2,3,4],ymm13[5],ymm2[6,7]5052; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,0,3]5053; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,1,3,3,7,5,7,7]5054; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,3,2,3]5055; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]5056; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,3,1,4,5,7,5]5057; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]5058; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5059; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5060; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm5[4,5,6,7]5061; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5062; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5063; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload5064; AVX2-FP-NEXT: # ymm3 = ymm0[0,1,2,3],mem[4,5,6,7]5065; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5066; AVX2-FP-NEXT: vmovaps 464(%rdi), %xmm05067; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5068; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]5069; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]5070; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload5071; AVX2-FP-NEXT: vpermps %ymm3, %ymm8, %ymm15072; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5,6,7]5073; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5074; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm7 # 32-byte Folded Reload5075; AVX2-FP-NEXT: # ymm7 = mem[0,1],ymm1[2,3],mem[4,5,6,7]5076; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,2,0,6,0,2,0,6]5077; AVX2-FP-NEXT: # ymm3 = mem[0,1,0,1]5078; AVX2-FP-NEXT: vpermps %ymm7, %ymm3, %ymm15079; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5080; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5081; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm10[0,1,2,3],ymm14[4,5,6,7]5082; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5083; AVX2-FP-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload5084; AVX2-FP-NEXT: # ymm14 = mem[0,1,2,3],ymm0[4,5,6,7]5085; AVX2-FP-NEXT: vmovaps 80(%rdi), %xmm55086; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7]5087; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]5088; AVX2-FP-NEXT: vpermps %ymm14, %ymm8, %ymm15089; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5,6,7]5090; AVX2-FP-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload5091; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload5092; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm1[2,3],mem[4,5,6,7]5093; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm15094; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5095; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5096; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm4[0,1,2,3],ymm11[4,5,6,7]5097; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5098; AVX2-FP-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload5099; AVX2-FP-NEXT: # ymm12 = mem[0,1,2,3],ymm0[4,5,6,7]5100; AVX2-FP-NEXT: vmovaps 272(%rdi), %xmm45101; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1],ymm4[2,3],ymm9[4,5,6,7]5102; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]5103; AVX2-FP-NEXT: vpermps %ymm12, %ymm8, %ymm115104; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1],ymm0[2,3,4,5,6,7]5105; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5106; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload5107; AVX2-FP-NEXT: # ymm1 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]5108; AVX2-FP-NEXT: vpermps %ymm1, %ymm3, %ymm155109; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4],ymm15[5,6,7]5110; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill5111; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm11 # 32-byte Folded Reload5112; AVX2-FP-NEXT: # ymm11 = ymm13[0,1,2,3],mem[4,5,6,7]5113; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5114; AVX2-FP-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload5115; AVX2-FP-NEXT: # ymm15 = mem[0,1,2,3],ymm0[4,5,6,7]5116; AVX2-FP-NEXT: vmovaps 656(%rdi), %xmm05117; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1],ymm0[2,3],ymm11[4,5,6,7]5118; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm10[0,1,0,2,4,5,4,6]5119; AVX2-FP-NEXT: vpermps %ymm15, %ymm8, %ymm85120; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm10[2,3,4,5,6,7]5121; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload5122; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload5123; AVX2-FP-NEXT: # ymm10 = ymm10[0,1],mem[2,3],ymm10[4,5,6,7]5124; AVX2-FP-NEXT: vpermps %ymm10, %ymm3, %ymm35125; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4],ymm3[5,6,7]5126; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm6[1,1,1,1,5,5,5,5]5127; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2],ymm5[3],ymm6[4,5,6,7]5128; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload5129; AVX2-FP-NEXT: vpermps %ymm14, %ymm13, %ymm65130; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3,4,5,6,7]5131; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm6 = [0,3,1,7,0,3,1,7]5132; AVX2-FP-NEXT: # ymm6 = mem[0,1,0,1]5133; AVX2-FP-NEXT: vpermps %ymm2, %ymm6, %ymm25134; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4],ymm2[5,6,7]5135; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload5136; AVX2-FP-NEXT: # ymm5 = mem[1,1,1,1,5,5,5,5]5137; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload5138; AVX2-FP-NEXT: # ymm5 = ymm5[0,1,2],mem[3],ymm5[4,5,6,7]5139; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm8 # 32-byte Folded Reload5140; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1],ymm5[2,3,4,5,6,7]5141; AVX2-FP-NEXT: vpermps %ymm7, %ymm6, %ymm75142; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm7[5,6,7]5143; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm9[1,1,1,1,5,5,5,5]5144; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm7[0,1,2],ymm4[3],ymm7[4,5,6,7]5145; AVX2-FP-NEXT: vpermps %ymm12, %ymm13, %ymm75146; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm7[0,1],ymm4[2,3,4,5,6,7]5147; AVX2-FP-NEXT: vpermps %ymm1, %ymm6, %ymm15148; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4],ymm1[5,6,7]5149; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm11[1,1,1,1,5,5,5,5]5150; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6,7]5151; AVX2-FP-NEXT: vpermps %ymm15, %ymm13, %ymm45152; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3,4,5,6,7]5153; AVX2-FP-NEXT: vpermps %ymm10, %ymm6, %ymm45154; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5,6,7]5155; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5156; AVX2-FP-NEXT: vmovaps %ymm4, 96(%rsi)5157; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5158; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rsi)5159; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5160; AVX2-FP-NEXT: vmovaps %ymm4, 64(%rsi)5161; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5162; AVX2-FP-NEXT: vmovaps %ymm4, (%rsi)5163; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5164; AVX2-FP-NEXT: vmovaps %ymm4, 96(%rdx)5165; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5166; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rdx)5167; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5168; AVX2-FP-NEXT: vmovaps %ymm4, 64(%rdx)5169; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5170; AVX2-FP-NEXT: vmovaps %ymm4, (%rdx)5171; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5172; AVX2-FP-NEXT: vmovaps %ymm4, 32(%rcx)5173; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5174; AVX2-FP-NEXT: vmovaps %ymm4, 96(%rcx)5175; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5176; AVX2-FP-NEXT: vmovaps %ymm4, 64(%rcx)5177; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5178; AVX2-FP-NEXT: vmovaps %ymm4, (%rcx)5179; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5180; AVX2-FP-NEXT: vmovaps %ymm4, 96(%r8)5181; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5182; AVX2-FP-NEXT: vmovaps %ymm4, 32(%r8)5183; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5184; AVX2-FP-NEXT: vmovaps %ymm4, 64(%r8)5185; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5186; AVX2-FP-NEXT: vmovaps %ymm4, (%r8)5187; AVX2-FP-NEXT: vmovaps %ymm3, 96(%r9)5188; AVX2-FP-NEXT: vmovups (%rsp), %ymm3 # 32-byte Reload5189; AVX2-FP-NEXT: vmovaps %ymm3, 32(%r9)5190; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5191; AVX2-FP-NEXT: vmovaps %ymm3, (%r9)5192; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5193; AVX2-FP-NEXT: vmovaps %ymm3, 64(%r9)5194; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax5195; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rax)5196; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rax)5197; AVX2-FP-NEXT: vmovaps %ymm5, 64(%rax)5198; AVX2-FP-NEXT: vmovaps %ymm2, (%rax)5199; AVX2-FP-NEXT: addq $1224, %rsp # imm = 0x4C85200; AVX2-FP-NEXT: vzeroupper5201; AVX2-FP-NEXT: retq5202;5203; AVX2-FCP-LABEL: load_i32_stride6_vf32:5204; AVX2-FCP: # %bb.0:5205; AVX2-FCP-NEXT: subq $1192, %rsp # imm = 0x4A85206; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm65207; AVX2-FCP-NEXT: vmovaps 448(%rdi), %ymm115208; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5209; AVX2-FCP-NEXT: vmovaps 416(%rdi), %ymm105210; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5211; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm25212; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5213; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm35214; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5215; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm45216; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5217; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm05218; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill5219; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm15220; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5221; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm55222; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5223; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm9 = [0,6,4,u]5224; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]5225; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5226; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm05227; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm5[0,1],ymm4[0,1]5228; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],ymm4[6,7]5229; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm8[0,2,2,2,4,6,6,6]5230; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]5231; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm3[0,1,2,3],ymm2[4,5,6,7]5232; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm5 = [4,2,4,2,4,2,4,2]5233; AVX2-FCP-NEXT: vpermps %ymm7, %ymm5, %ymm15234; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]5235; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5236; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5237; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm11[0,1],ymm6[0,1]5238; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]5239; AVX2-FCP-NEXT: vmovaps 384(%rdi), %ymm05240; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5241; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3],ymm10[4,5],ymm0[6,7]5242; AVX2-FCP-NEXT: vpermps %ymm4, %ymm9, %ymm05243; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm6[0,2,2,2,4,6,6,6]5244; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]5245; AVX2-FCP-NEXT: vmovaps 512(%rdi), %ymm15246; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5247; AVX2-FCP-NEXT: vmovaps 544(%rdi), %ymm35248; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5249; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm1[4,5,6,7]5250; AVX2-FCP-NEXT: vpermps %ymm3, %ymm5, %ymm15251; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]5252; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5253; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm15254; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5255; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm05256; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5257; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]5258; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]5259; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm05260; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5261; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm15262; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5263; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]5264; AVX2-FCP-NEXT: vpermps %ymm1, %ymm9, %ymm05265; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm14 = ymm2[0,2,2,2,4,6,6,6]5266; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2],ymm14[3,4,5,6,7]5267; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm105268; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5269; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm05270; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5271; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3],ymm10[4,5,6,7]5272; AVX2-FCP-NEXT: vpermps %ymm13, %ymm5, %ymm155273; AVX2-FCP-NEXT: vmovaps %ymm5, %ymm05274; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5275; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm14[0,1,2,3,4,5],ymm15[6,7]5276; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5277; AVX2-FCP-NEXT: vmovaps 608(%rdi), %ymm105278; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5279; AVX2-FCP-NEXT: vmovaps 576(%rdi), %ymm115280; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5281; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm11[0,1,2,3],ymm10[4,5],ymm11[6,7]5282; AVX2-FCP-NEXT: vpermps %ymm14, %ymm9, %ymm155283; AVX2-FCP-NEXT: vmovaps 672(%rdi), %ymm55284; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5285; AVX2-FCP-NEXT: vmovaps 640(%rdi), %ymm95286; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5287; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm9[0,1],ymm5[0,1]5288; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm5[6,7]5289; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm11 = ymm12[0,2,2,2,4,6,6,6]5290; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1,2],ymm11[3,4,5,6,7]5291; AVX2-FCP-NEXT: vmovaps 704(%rdi), %ymm95292; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5293; AVX2-FCP-NEXT: vmovaps 736(%rdi), %ymm105294; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5295; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm10[0,1,2,3],ymm9[4,5,6,7]5296; AVX2-FCP-NEXT: vpermps %ymm15, %ymm0, %ymm105297; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm10[6,7]5298; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5299; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm10 = [1,7,5,u]5300; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm11 # 32-byte Folded Reload5301; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm8 = ymm8[1,3,2,3,5,7,6,7]5302; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm11[0,1,2],ymm8[3,4,5,6,7]5303; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm5 = [5,3,5,3,5,3,5,3]5304; AVX2-FCP-NEXT: vpermps %ymm7, %ymm5, %ymm75305; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]5306; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5307; AVX2-FCP-NEXT: vpermps %ymm4, %ymm10, %ymm45308; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm6[1,3,2,3,5,7,6,7]5309; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm6[3,4,5,6,7]5310; AVX2-FCP-NEXT: vpermps %ymm3, %ymm5, %ymm35311; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]5312; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5313; AVX2-FCP-NEXT: vpermps %ymm1, %ymm10, %ymm15314; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[1,3,2,3,5,7,6,7]5315; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]5316; AVX2-FCP-NEXT: vpermps %ymm13, %ymm5, %ymm05317; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]5318; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5319; AVX2-FCP-NEXT: vpermps %ymm14, %ymm10, %ymm05320; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,3,2,3,5,7,6,7]5321; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]5322; AVX2-FCP-NEXT: vpermps %ymm15, %ymm5, %ymm15323; AVX2-FCP-NEXT: vmovaps %ymm5, %ymm135324; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]5325; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5326; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload5327; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload5328; AVX2-FCP-NEXT: # ymm0 = ymm15[0,1],mem[2,3],ymm15[4,5],mem[6,7]5329; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm7 = [2,0,6,4,2,0,6,7]5330; AVX2-FCP-NEXT: vpermps %ymm0, %ymm7, %ymm05331; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm8 = [2,0,6,7]5332; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5333; AVX2-FCP-NEXT: vblendps $12, (%rsp), %ymm1, %ymm3 # 32-byte Folded Reload5334; AVX2-FCP-NEXT: # ymm3 = ymm1[0,1],mem[2,3],ymm1[4,5,6,7]5335; AVX2-FCP-NEXT: vpermps %ymm3, %ymm8, %ymm15336; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]5337; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5338; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload5339; AVX2-FCP-NEXT: # ymm2 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]5340; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [0,0,6,4,0,0,6,4]5341; AVX2-FCP-NEXT: # ymm10 = mem[0,1,0,1]5342; AVX2-FCP-NEXT: vpermps %ymm2, %ymm10, %ymm15343; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5344; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5345; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5346; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload5347; AVX2-FCP-NEXT: # ymm0 = ymm4[0,1],mem[2,3],ymm4[4,5],mem[6,7]5348; AVX2-FCP-NEXT: vpermps %ymm0, %ymm7, %ymm05349; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5350; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload5351; AVX2-FCP-NEXT: # ymm14 = mem[0,1],ymm1[2,3],mem[4,5,6,7]5352; AVX2-FCP-NEXT: vpermps %ymm14, %ymm8, %ymm15353; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]5354; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5355; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload5356; AVX2-FCP-NEXT: # ymm6 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]5357; AVX2-FCP-NEXT: vpermps %ymm6, %ymm10, %ymm15358; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5359; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5360; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5361; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload5362; AVX2-FCP-NEXT: # ymm0 = ymm5[0,1],mem[2,3],ymm5[4,5],mem[6,7]5363; AVX2-FCP-NEXT: vpermps %ymm0, %ymm7, %ymm15364; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5365; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5366; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]5367; AVX2-FCP-NEXT: vpermps %ymm0, %ymm8, %ymm115368; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm1[3,4,5,6,7]5369; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5370; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload5371; AVX2-FCP-NEXT: # ymm1 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]5372; AVX2-FCP-NEXT: vpermps %ymm1, %ymm10, %ymm125373; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm12[5,6,7]5374; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5375; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload5376; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm11 # 32-byte Folded Reload5377; AVX2-FCP-NEXT: # ymm11 = ymm9[0,1],mem[2,3],ymm9[4,5],mem[6,7]5378; AVX2-FCP-NEXT: vpermps %ymm11, %ymm7, %ymm75379; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload5380; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload5381; AVX2-FCP-NEXT: # ymm11 = ymm11[0,1],mem[2,3],ymm11[4,5,6,7]5382; AVX2-FCP-NEXT: vpermps %ymm11, %ymm8, %ymm85383; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3,4,5,6,7]5384; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload5385; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload5386; AVX2-FCP-NEXT: # ymm8 = ymm8[0,1,2,3],mem[4,5],ymm8[6,7]5387; AVX2-FCP-NEXT: vpermps %ymm8, %ymm10, %ymm105388; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4],ymm10[5,6,7]5389; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5390; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm7 = mem[3,3,3,3]5391; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm15[1],ymm7[2,3,4],ymm15[5],ymm7[6,7]5392; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm7[0],ymm3[1,2,3,4],ymm7[5],ymm3[6,7]5393; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm7 = [0,1,7,5,0,1,7,5]5394; AVX2-FCP-NEXT: # ymm7 = mem[0,1,0,1]5395; AVX2-FCP-NEXT: vpermps %ymm2, %ymm7, %ymm25396; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm10 = [3,1,7,5,0,u,u,u]5397; AVX2-FCP-NEXT: vpermps %ymm3, %ymm10, %ymm35398; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]5399; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5400; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm2 = mem[3,3,3,3]5401; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3,4],ymm4[5],ymm2[6,7]5402; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1,2,3,4],ymm2[5],ymm14[6,7]5403; AVX2-FCP-NEXT: vpermps %ymm2, %ymm10, %ymm25404; AVX2-FCP-NEXT: vpermps %ymm6, %ymm7, %ymm35405; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6,7]5406; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5407; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm2 = mem[3,3,3,3]5408; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm9[1],ymm2[2,3,4],ymm9[5],ymm2[6,7]5409; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm11[1,2,3,4],ymm2[5],ymm11[6,7]5410; AVX2-FCP-NEXT: vpermps %ymm2, %ymm10, %ymm25411; AVX2-FCP-NEXT: vpermps %ymm8, %ymm7, %ymm35412; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6,7]5413; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5414; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm2 = mem[3,3,3,3]5415; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2,3,4],ymm5[5],ymm2[6,7]5416; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5417; AVX2-FCP-NEXT: vpermps %ymm0, %ymm10, %ymm05418; AVX2-FCP-NEXT: vpermps %ymm1, %ymm7, %ymm15419; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5420; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5421; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm1 # 32-byte Folded Reload5422; AVX2-FCP-NEXT: # ymm1 = ymm4[0,1,2,3],mem[4,5,6,7]5423; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5424; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5425; AVX2-FCP-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload5426; AVX2-FCP-NEXT: # ymm3 = mem[0,1,2,3],ymm0[4,5,6,7]5427; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5428; AVX2-FCP-NEXT: vmovaps 464(%rdi), %xmm05429; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5430; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7]5431; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]5432; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5433; AVX2-FCP-NEXT: vpermps %ymm3, %ymm5, %ymm15434; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5,6,7]5435; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5436; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload5437; AVX2-FCP-NEXT: # ymm6 = mem[0,1],ymm1[2,3],mem[4,5,6,7]5438; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [0,2,0,6,0,2,0,6]5439; AVX2-FCP-NEXT: # ymm4 = mem[0,1,0,1]5440; AVX2-FCP-NEXT: vpermps %ymm6, %ymm4, %ymm15441; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]5442; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5443; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm3 # 32-byte Folded Reload5444; AVX2-FCP-NEXT: # ymm3 = ymm15[0,1,2,3],mem[4,5,6,7]5445; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5446; AVX2-FCP-NEXT: vblendps $240, (%rsp), %ymm0, %ymm12 # 32-byte Folded Reload5447; AVX2-FCP-NEXT: # ymm12 = ymm0[0,1,2,3],mem[4,5,6,7]5448; AVX2-FCP-NEXT: vmovaps 80(%rdi), %xmm145449; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1],ymm14[2,3],ymm3[4,5,6,7]5450; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]5451; AVX2-FCP-NEXT: vpermps %ymm12, %ymm5, %ymm15452; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm0[2,3,4,5,6,7]5453; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5454; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload5455; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm0[2,3],mem[4,5,6,7]5456; AVX2-FCP-NEXT: vpermps %ymm2, %ymm4, %ymm85457; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm8[5,6,7]5458; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill5459; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm10 # 32-byte Folded Reload5460; AVX2-FCP-NEXT: # ymm10 = ymm9[0,1,2,3],mem[4,5,6,7]5461; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5462; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 32-byte Folded Reload5463; AVX2-FCP-NEXT: # ymm11 = ymm0[0,1,2,3],mem[4,5,6,7]5464; AVX2-FCP-NEXT: vmovaps 272(%rdi), %xmm15465; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1],ymm1[2,3],ymm10[4,5,6,7]5466; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,1,0,2,4,5,4,6]5467; AVX2-FCP-NEXT: vpermps %ymm11, %ymm5, %ymm95468; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1],ymm8[2,3,4,5,6,7]5469; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5470; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload5471; AVX2-FCP-NEXT: # ymm8 = mem[0,1],ymm0[2,3],mem[4,5,6,7]5472; AVX2-FCP-NEXT: vpermps %ymm8, %ymm4, %ymm155473; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3,4],ymm15[5,6,7]5474; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5475; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5476; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm9 # 32-byte Folded Reload5477; AVX2-FCP-NEXT: # ymm9 = ymm0[0,1,2,3],mem[4,5,6,7]5478; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5479; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload5480; AVX2-FCP-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5,6,7]5481; AVX2-FCP-NEXT: vmovaps 656(%rdi), %xmm05482; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1],ymm0[2,3],ymm9[4,5,6,7]5483; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm7[0,1,0,2,4,5,4,6]5484; AVX2-FCP-NEXT: vpermps %ymm15, %ymm5, %ymm55485; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1],ymm7[2,3,4,5,6,7]5486; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload5487; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload5488; AVX2-FCP-NEXT: # ymm7 = mem[0,1],ymm7[2,3],mem[4,5,6,7]5489; AVX2-FCP-NEXT: vpermps %ymm7, %ymm4, %ymm45490; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5,6,7]5491; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm3[1,1,1,1,5,5,5,5]5492; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm14[3],ymm3[4,5,6,7]5493; AVX2-FCP-NEXT: vpermps %ymm12, %ymm13, %ymm55494; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3,4,5,6,7]5495; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm5 = [0,3,1,7,0,3,1,7]5496; AVX2-FCP-NEXT: # ymm5 = mem[0,1,0,1]5497; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm25498; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5,6,7]5499; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload5500; AVX2-FCP-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]5501; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload5502; AVX2-FCP-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]5503; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm12 # 32-byte Folded Reload5504; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm12[0,1],ymm3[2,3,4,5,6,7]5505; AVX2-FCP-NEXT: vpermps %ymm6, %ymm5, %ymm65506; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]5507; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm10[1,1,1,1,5,5,5,5]5508; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3],ymm6[4,5,6,7]5509; AVX2-FCP-NEXT: vpermps %ymm11, %ymm13, %ymm65510; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0,1],ymm1[2,3,4,5,6,7]5511; AVX2-FCP-NEXT: vpermps %ymm8, %ymm5, %ymm65512; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm6[5,6,7]5513; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm9[1,1,1,1,5,5,5,5]5514; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3],ymm6[4,5,6,7]5515; AVX2-FCP-NEXT: vpermps %ymm15, %ymm13, %ymm65516; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1],ymm0[2,3,4,5,6,7]5517; AVX2-FCP-NEXT: vpermps %ymm7, %ymm5, %ymm55518; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm5[5,6,7]5519; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5520; AVX2-FCP-NEXT: vmovaps %ymm5, 96(%rsi)5521; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5522; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%rsi)5523; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5524; AVX2-FCP-NEXT: vmovaps %ymm5, 64(%rsi)5525; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5526; AVX2-FCP-NEXT: vmovaps %ymm5, (%rsi)5527; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5528; AVX2-FCP-NEXT: vmovaps %ymm5, 96(%rdx)5529; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5530; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%rdx)5531; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5532; AVX2-FCP-NEXT: vmovaps %ymm5, 64(%rdx)5533; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5534; AVX2-FCP-NEXT: vmovaps %ymm5, (%rdx)5535; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5536; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%rcx)5537; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5538; AVX2-FCP-NEXT: vmovaps %ymm5, 96(%rcx)5539; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5540; AVX2-FCP-NEXT: vmovaps %ymm5, 64(%rcx)5541; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5542; AVX2-FCP-NEXT: vmovaps %ymm5, (%rcx)5543; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5544; AVX2-FCP-NEXT: vmovaps %ymm5, 96(%r8)5545; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5546; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%r8)5547; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5548; AVX2-FCP-NEXT: vmovaps %ymm5, 64(%r8)5549; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload5550; AVX2-FCP-NEXT: vmovaps %ymm5, (%r8)5551; AVX2-FCP-NEXT: vmovaps %ymm4, 96(%r9)5552; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5553; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%r9)5554; AVX2-FCP-NEXT: vmovups (%rsp), %ymm4 # 32-byte Reload5555; AVX2-FCP-NEXT: vmovaps %ymm4, (%r9)5556; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload5557; AVX2-FCP-NEXT: vmovaps %ymm4, 64(%r9)5558; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax5559; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rax)5560; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rax)5561; AVX2-FCP-NEXT: vmovaps %ymm3, 64(%rax)5562; AVX2-FCP-NEXT: vmovaps %ymm2, (%rax)5563; AVX2-FCP-NEXT: addq $1192, %rsp # imm = 0x4A85564; AVX2-FCP-NEXT: vzeroupper5565; AVX2-FCP-NEXT: retq5566;5567; AVX512-LABEL: load_i32_stride6_vf32:5568; AVX512: # %bb.0:5569; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax5570; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm05571; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm35572; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm25573; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm55574; AVX512-NEXT: vmovdqa64 704(%rdi), %zmm65575; AVX512-NEXT: vmovdqa64 640(%rdi), %zmm45576; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm105577; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm95578; AVX512-NEXT: vmovdqa64 (%rdi), %zmm115579; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm15580; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm125581; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm135582; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]5583; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]5584; AVX512-NEXT: vmovdqa64 %zmm13, %zmm155585; AVX512-NEXT: vpermt2d %zmm12, %zmm14, %zmm155586; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]5587; AVX512-NEXT: vmovdqa64 %zmm11, %zmm85588; AVX512-NEXT: vpermt2d %zmm1, %zmm7, %zmm85589; AVX512-NEXT: movb $56, %dil5590; AVX512-NEXT: kmovw %edi, %k25591; AVX512-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}5592; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]5593; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5594; AVX512-NEXT: vmovdqa64 %zmm9, %zmm165595; AVX512-NEXT: vpermt2d %zmm10, %zmm15, %zmm165596; AVX512-NEXT: movw $-2048, %di # imm = 0xF8005597; AVX512-NEXT: kmovw %edi, %k15598; AVX512-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}5599; AVX512-NEXT: vpermi2d %zmm6, %zmm4, %zmm155600; AVX512-NEXT: vpermi2d %zmm2, %zmm5, %zmm145601; AVX512-NEXT: vpermi2d %zmm0, %zmm3, %zmm75602; AVX512-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}5603; AVX512-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}5604; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]5605; AVX512-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]5606; AVX512-NEXT: vmovdqa64 %zmm13, %zmm175607; AVX512-NEXT: vpermt2d %zmm12, %zmm16, %zmm175608; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]5609; AVX512-NEXT: vmovdqa64 %zmm11, %zmm155610; AVX512-NEXT: vpermt2d %zmm1, %zmm14, %zmm155611; AVX512-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}5612; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]5613; AVX512-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]5614; AVX512-NEXT: vmovdqa64 %zmm9, %zmm185615; AVX512-NEXT: vpermt2d %zmm10, %zmm17, %zmm185616; AVX512-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}5617; AVX512-NEXT: vpermi2d %zmm6, %zmm4, %zmm175618; AVX512-NEXT: vpermi2d %zmm2, %zmm5, %zmm165619; AVX512-NEXT: vpermi2d %zmm0, %zmm3, %zmm145620; AVX512-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}5621; AVX512-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}5622; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]5623; AVX512-NEXT: vmovdqa64 %zmm11, %zmm195624; AVX512-NEXT: vpermt2d %zmm1, %zmm18, %zmm195625; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]5626; AVX512-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]5627; AVX512-NEXT: vmovdqa64 %zmm12, %zmm165628; AVX512-NEXT: vpermt2d %zmm13, %zmm17, %zmm165629; AVX512-NEXT: movw $31, %di5630; AVX512-NEXT: kmovw %edi, %k25631; AVX512-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}5632; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]5633; AVX512-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]5634; AVX512-NEXT: vmovdqa64 %zmm10, %zmm205635; AVX512-NEXT: vpermt2d %zmm9, %zmm19, %zmm205636; AVX512-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}5637; AVX512-NEXT: vpermi2d %zmm4, %zmm6, %zmm195638; AVX512-NEXT: vpermi2d %zmm5, %zmm2, %zmm175639; AVX512-NEXT: vpermi2d %zmm0, %zmm3, %zmm185640; AVX512-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}5641; AVX512-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}5642; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]5643; AVX512-NEXT: vmovdqa64 %zmm11, %zmm215644; AVX512-NEXT: vpermt2d %zmm1, %zmm20, %zmm215645; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]5646; AVX512-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]5647; AVX512-NEXT: vmovdqa64 %zmm12, %zmm185648; AVX512-NEXT: vpermt2d %zmm13, %zmm19, %zmm185649; AVX512-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}5650; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]5651; AVX512-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]5652; AVX512-NEXT: vmovdqa64 %zmm10, %zmm225653; AVX512-NEXT: vpermt2d %zmm9, %zmm21, %zmm225654; AVX512-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}5655; AVX512-NEXT: vpermi2d %zmm4, %zmm6, %zmm215656; AVX512-NEXT: vpermi2d %zmm5, %zmm2, %zmm195657; AVX512-NEXT: vpermi2d %zmm0, %zmm3, %zmm205658; AVX512-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}5659; AVX512-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}5660; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]5661; AVX512-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]5662; AVX512-NEXT: vmovdqa64 %zmm12, %zmm215663; AVX512-NEXT: vpermt2d %zmm13, %zmm20, %zmm215664; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]5665; AVX512-NEXT: vmovdqa64 %zmm1, %zmm235666; AVX512-NEXT: vpermt2d %zmm11, %zmm22, %zmm235667; AVX512-NEXT: movw $992, %di # imm = 0x3E05668; AVX512-NEXT: kmovw %edi, %k15669; AVX512-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}5670; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]5671; AVX512-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]5672; AVX512-NEXT: vmovdqa64 %zmm9, %zmm245673; AVX512-NEXT: vpermt2d %zmm10, %zmm21, %zmm245674; AVX512-NEXT: movb $-32, %dil5675; AVX512-NEXT: kmovw %edi, %k25676; AVX512-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}5677; AVX512-NEXT: vpermi2d %zmm6, %zmm4, %zmm215678; AVX512-NEXT: vpermi2d %zmm5, %zmm2, %zmm205679; AVX512-NEXT: vpermi2d %zmm3, %zmm0, %zmm225680; AVX512-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}5681; AVX512-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}5682; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]5683; AVX512-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]5684; AVX512-NEXT: vpermt2d %zmm13, %zmm20, %zmm125685; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]5686; AVX512-NEXT: vpermt2d %zmm11, %zmm13, %zmm15687; AVX512-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}5688; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]5689; AVX512-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5690; AVX512-NEXT: vpermt2d %zmm10, %zmm11, %zmm95691; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}5692; AVX512-NEXT: vpermt2d %zmm6, %zmm11, %zmm45693; AVX512-NEXT: vpermt2d %zmm5, %zmm20, %zmm25694; AVX512-NEXT: vpermt2d %zmm3, %zmm13, %zmm05695; AVX512-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}5696; AVX512-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}5697; AVX512-NEXT: vmovdqa64 %zmm7, 64(%rsi)5698; AVX512-NEXT: vmovdqa64 %zmm8, (%rsi)5699; AVX512-NEXT: vmovdqa64 %zmm14, 64(%rdx)5700; AVX512-NEXT: vmovdqa64 %zmm15, (%rdx)5701; AVX512-NEXT: vmovdqa64 %zmm17, 64(%rcx)5702; AVX512-NEXT: vmovdqa64 %zmm16, (%rcx)5703; AVX512-NEXT: vmovdqa64 %zmm19, 64(%r8)5704; AVX512-NEXT: vmovdqa64 %zmm18, (%r8)5705; AVX512-NEXT: vmovdqa64 %zmm22, 64(%r9)5706; AVX512-NEXT: vmovdqa64 %zmm23, (%r9)5707; AVX512-NEXT: vmovdqa64 %zmm0, 64(%rax)5708; AVX512-NEXT: vmovdqa64 %zmm1, (%rax)5709; AVX512-NEXT: vzeroupper5710; AVX512-NEXT: retq5711;5712; AVX512-FCP-LABEL: load_i32_stride6_vf32:5713; AVX512-FCP: # %bb.0:5714; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax5715; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm05716; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm35717; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm25718; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm55719; AVX512-FCP-NEXT: vmovdqa64 704(%rdi), %zmm65720; AVX512-FCP-NEXT: vmovdqa64 640(%rdi), %zmm45721; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm105722; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm95723; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm115724; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm15725; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm125726; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm135727; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]5728; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]5729; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm155730; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm14, %zmm155731; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]5732; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm85733; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm85734; AVX512-FCP-NEXT: movb $56, %dil5735; AVX512-FCP-NEXT: kmovw %edi, %k25736; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}5737; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]5738; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5739; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm165740; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm15, %zmm165741; AVX512-FCP-NEXT: movw $-2048, %di # imm = 0xF8005742; AVX512-FCP-NEXT: kmovw %edi, %k15743; AVX512-FCP-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}5744; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm155745; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm145746; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm75747; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}5748; AVX512-FCP-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}5749; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]5750; AVX512-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]5751; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm175752; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm16, %zmm175753; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]5754; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm155755; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm14, %zmm155756; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}5757; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]5758; AVX512-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]5759; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm185760; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm17, %zmm185761; AVX512-FCP-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}5762; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm175763; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm165764; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm145765; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}5766; AVX512-FCP-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}5767; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]5768; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm195769; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm18, %zmm195770; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]5771; AVX512-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]5772; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm165773; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm17, %zmm165774; AVX512-FCP-NEXT: movw $31, %di5775; AVX512-FCP-NEXT: kmovw %edi, %k25776; AVX512-FCP-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}5777; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]5778; AVX512-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]5779; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm205780; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm19, %zmm205781; AVX512-FCP-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}5782; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm195783; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm175784; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm185785; AVX512-FCP-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}5786; AVX512-FCP-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}5787; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]5788; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm215789; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm20, %zmm215790; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]5791; AVX512-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]5792; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm185793; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm19, %zmm185794; AVX512-FCP-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}5795; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]5796; AVX512-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]5797; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm225798; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm21, %zmm225799; AVX512-FCP-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}5800; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm215801; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm195802; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm205803; AVX512-FCP-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}5804; AVX512-FCP-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}5805; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]5806; AVX512-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]5807; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm215808; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm215809; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]5810; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm235811; AVX512-FCP-NEXT: vpermt2d %zmm11, %zmm22, %zmm235812; AVX512-FCP-NEXT: movw $992, %di # imm = 0x3E05813; AVX512-FCP-NEXT: kmovw %edi, %k15814; AVX512-FCP-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}5815; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]5816; AVX512-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]5817; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm245818; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm21, %zmm245819; AVX512-FCP-NEXT: movb $-32, %dil5820; AVX512-FCP-NEXT: kmovw %edi, %k25821; AVX512-FCP-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}5822; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm215823; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm205824; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm0, %zmm225825; AVX512-FCP-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}5826; AVX512-FCP-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}5827; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]5828; AVX512-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]5829; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm125830; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]5831; AVX512-FCP-NEXT: vpermt2d %zmm11, %zmm13, %zmm15832; AVX512-FCP-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}5833; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]5834; AVX512-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5835; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm11, %zmm95836; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}5837; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm11, %zmm45838; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm25839; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm13, %zmm05840; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}5841; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}5842; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 64(%rsi)5843; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)5844; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)5845; AVX512-FCP-NEXT: vmovdqa64 %zmm15, (%rdx)5846; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)5847; AVX512-FCP-NEXT: vmovdqa64 %zmm16, (%rcx)5848; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)5849; AVX512-FCP-NEXT: vmovdqa64 %zmm18, (%r8)5850; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 64(%r9)5851; AVX512-FCP-NEXT: vmovdqa64 %zmm23, (%r9)5852; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)5853; AVX512-FCP-NEXT: vmovdqa64 %zmm1, (%rax)5854; AVX512-FCP-NEXT: vzeroupper5855; AVX512-FCP-NEXT: retq5856;5857; AVX512DQ-LABEL: load_i32_stride6_vf32:5858; AVX512DQ: # %bb.0:5859; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax5860; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm05861; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm35862; AVX512DQ-NEXT: vmovdqa64 512(%rdi), %zmm25863; AVX512DQ-NEXT: vmovdqa64 576(%rdi), %zmm55864; AVX512DQ-NEXT: vmovdqa64 704(%rdi), %zmm65865; AVX512DQ-NEXT: vmovdqa64 640(%rdi), %zmm45866; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm105867; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm95868; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm115869; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm15870; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm125871; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm135872; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]5873; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]5874; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm155875; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm14, %zmm155876; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]5877; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm85878; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm7, %zmm85879; AVX512DQ-NEXT: movb $56, %dil5880; AVX512DQ-NEXT: kmovw %edi, %k25881; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}5882; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]5883; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5884; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm165885; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm15, %zmm165886; AVX512DQ-NEXT: movw $-2048, %di # imm = 0xF8005887; AVX512DQ-NEXT: kmovw %edi, %k15888; AVX512DQ-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}5889; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm4, %zmm155890; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm5, %zmm145891; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm3, %zmm75892; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}5893; AVX512DQ-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}5894; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]5895; AVX512DQ-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]5896; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm175897; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm16, %zmm175898; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]5899; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm155900; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm14, %zmm155901; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}5902; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]5903; AVX512DQ-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]5904; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm185905; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm17, %zmm185906; AVX512DQ-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}5907; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm4, %zmm175908; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm5, %zmm165909; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm3, %zmm145910; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}5911; AVX512DQ-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}5912; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]5913; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm195914; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm18, %zmm195915; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]5916; AVX512DQ-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]5917; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm165918; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm17, %zmm165919; AVX512DQ-NEXT: movw $31, %di5920; AVX512DQ-NEXT: kmovw %edi, %k25921; AVX512DQ-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}5922; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]5923; AVX512DQ-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]5924; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm205925; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm19, %zmm205926; AVX512DQ-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}5927; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm6, %zmm195928; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm2, %zmm175929; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm3, %zmm185930; AVX512DQ-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}5931; AVX512DQ-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}5932; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]5933; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm215934; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm20, %zmm215935; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]5936; AVX512DQ-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]5937; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm185938; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm19, %zmm185939; AVX512DQ-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}5940; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]5941; AVX512DQ-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]5942; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm225943; AVX512DQ-NEXT: vpermt2d %zmm9, %zmm21, %zmm225944; AVX512DQ-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}5945; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm6, %zmm215946; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm2, %zmm195947; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm3, %zmm205948; AVX512DQ-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}5949; AVX512DQ-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}5950; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]5951; AVX512DQ-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]5952; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm215953; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm20, %zmm215954; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]5955; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm235956; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm22, %zmm235957; AVX512DQ-NEXT: movw $992, %di # imm = 0x3E05958; AVX512DQ-NEXT: kmovw %edi, %k15959; AVX512DQ-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}5960; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]5961; AVX512DQ-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]5962; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm245963; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm21, %zmm245964; AVX512DQ-NEXT: movb $-32, %dil5965; AVX512DQ-NEXT: kmovw %edi, %k25966; AVX512DQ-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}5967; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm4, %zmm215968; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm2, %zmm205969; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm0, %zmm225970; AVX512DQ-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}5971; AVX512DQ-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}5972; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]5973; AVX512DQ-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]5974; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm20, %zmm125975; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]5976; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm13, %zmm15977; AVX512DQ-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}5978; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]5979; AVX512DQ-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5980; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm11, %zmm95981; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}5982; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm11, %zmm45983; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm20, %zmm25984; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm13, %zmm05985; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}5986; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}5987; AVX512DQ-NEXT: vmovdqa64 %zmm7, 64(%rsi)5988; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rsi)5989; AVX512DQ-NEXT: vmovdqa64 %zmm14, 64(%rdx)5990; AVX512DQ-NEXT: vmovdqa64 %zmm15, (%rdx)5991; AVX512DQ-NEXT: vmovdqa64 %zmm17, 64(%rcx)5992; AVX512DQ-NEXT: vmovdqa64 %zmm16, (%rcx)5993; AVX512DQ-NEXT: vmovdqa64 %zmm19, 64(%r8)5994; AVX512DQ-NEXT: vmovdqa64 %zmm18, (%r8)5995; AVX512DQ-NEXT: vmovdqa64 %zmm22, 64(%r9)5996; AVX512DQ-NEXT: vmovdqa64 %zmm23, (%r9)5997; AVX512DQ-NEXT: vmovdqa64 %zmm0, 64(%rax)5998; AVX512DQ-NEXT: vmovdqa64 %zmm1, (%rax)5999; AVX512DQ-NEXT: vzeroupper6000; AVX512DQ-NEXT: retq6001;6002; AVX512DQ-FCP-LABEL: load_i32_stride6_vf32:6003; AVX512DQ-FCP: # %bb.0:6004; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax6005; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm06006; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm36007; AVX512DQ-FCP-NEXT: vmovdqa64 512(%rdi), %zmm26008; AVX512DQ-FCP-NEXT: vmovdqa64 576(%rdi), %zmm56009; AVX512DQ-FCP-NEXT: vmovdqa64 704(%rdi), %zmm66010; AVX512DQ-FCP-NEXT: vmovdqa64 640(%rdi), %zmm46011; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm106012; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm96013; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm116014; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm16015; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm126016; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm136017; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]6018; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]6019; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm156020; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm14, %zmm156021; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]6022; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm86023; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm86024; AVX512DQ-FCP-NEXT: movb $56, %dil6025; AVX512DQ-FCP-NEXT: kmovw %edi, %k26026; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}6027; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]6028; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6029; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm166030; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm15, %zmm166031; AVX512DQ-FCP-NEXT: movw $-2048, %di # imm = 0xF8006032; AVX512DQ-FCP-NEXT: kmovw %edi, %k16033; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}6034; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm156035; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm146036; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm76037; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}6038; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}6039; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]6040; AVX512DQ-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]6041; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm176042; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm16, %zmm176043; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]6044; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm156045; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm14, %zmm156046; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}6047; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]6048; AVX512DQ-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6049; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm186050; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm17, %zmm186051; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}6052; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm176053; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm166054; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm146055; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}6056; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}6057; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]6058; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm196059; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm18, %zmm196060; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]6061; AVX512DQ-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6062; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm166063; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm17, %zmm166064; AVX512DQ-FCP-NEXT: movw $31, %di6065; AVX512DQ-FCP-NEXT: kmovw %edi, %k26066; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}6067; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]6068; AVX512DQ-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6069; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm206070; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm19, %zmm206071; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}6072; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm196073; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm176074; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm186075; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}6076; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}6077; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]6078; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm216079; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm20, %zmm216080; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]6081; AVX512DQ-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6082; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm186083; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm19, %zmm186084; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}6085; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]6086; AVX512DQ-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6087; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm226088; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm21, %zmm226089; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}6090; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm216091; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm196092; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm206093; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}6094; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}6095; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]6096; AVX512DQ-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6097; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm216098; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm216099; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]6100; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm236101; AVX512DQ-FCP-NEXT: vpermt2d %zmm11, %zmm22, %zmm236102; AVX512DQ-FCP-NEXT: movw $992, %di # imm = 0x3E06103; AVX512DQ-FCP-NEXT: kmovw %edi, %k16104; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}6105; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]6106; AVX512DQ-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6107; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm246108; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm21, %zmm246109; AVX512DQ-FCP-NEXT: movb $-32, %dil6110; AVX512DQ-FCP-NEXT: kmovw %edi, %k26111; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}6112; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm216113; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm206114; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm0, %zmm226115; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}6116; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}6117; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]6118; AVX512DQ-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6119; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm126120; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]6121; AVX512DQ-FCP-NEXT: vpermt2d %zmm11, %zmm13, %zmm16122; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}6123; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]6124; AVX512DQ-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6125; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm11, %zmm96126; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}6127; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm11, %zmm46128; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm26129; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm13, %zmm06130; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}6131; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}6132; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 64(%rsi)6133; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)6134; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)6135; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, (%rdx)6136; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)6137; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, (%rcx)6138; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)6139; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, (%r8)6140; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 64(%r9)6141; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, (%r9)6142; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)6143; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, (%rax)6144; AVX512DQ-FCP-NEXT: vzeroupper6145; AVX512DQ-FCP-NEXT: retq6146;6147; AVX512BW-LABEL: load_i32_stride6_vf32:6148; AVX512BW: # %bb.0:6149; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax6150; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm06151; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm36152; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm26153; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm56154; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm66155; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm46156; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm106157; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm96158; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm116159; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm16160; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm126161; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm136162; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]6163; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]6164; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm156165; AVX512BW-NEXT: vpermt2d %zmm12, %zmm14, %zmm156166; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]6167; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm86168; AVX512BW-NEXT: vpermt2d %zmm1, %zmm7, %zmm86169; AVX512BW-NEXT: movb $56, %dil6170; AVX512BW-NEXT: kmovd %edi, %k26171; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}6172; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]6173; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6174; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm166175; AVX512BW-NEXT: vpermt2d %zmm10, %zmm15, %zmm166176; AVX512BW-NEXT: movw $-2048, %di # imm = 0xF8006177; AVX512BW-NEXT: kmovd %edi, %k16178; AVX512BW-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}6179; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm156180; AVX512BW-NEXT: vpermi2d %zmm2, %zmm5, %zmm146181; AVX512BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm76182; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}6183; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}6184; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]6185; AVX512BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]6186; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm176187; AVX512BW-NEXT: vpermt2d %zmm12, %zmm16, %zmm176188; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]6189; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm156190; AVX512BW-NEXT: vpermt2d %zmm1, %zmm14, %zmm156191; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}6192; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]6193; AVX512BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6194; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm186195; AVX512BW-NEXT: vpermt2d %zmm10, %zmm17, %zmm186196; AVX512BW-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}6197; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm176198; AVX512BW-NEXT: vpermi2d %zmm2, %zmm5, %zmm166199; AVX512BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm146200; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}6201; AVX512BW-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}6202; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]6203; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm196204; AVX512BW-NEXT: vpermt2d %zmm1, %zmm18, %zmm196205; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]6206; AVX512BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6207; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm166208; AVX512BW-NEXT: vpermt2d %zmm13, %zmm17, %zmm166209; AVX512BW-NEXT: movw $31, %di6210; AVX512BW-NEXT: kmovd %edi, %k26211; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}6212; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]6213; AVX512BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6214; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm206215; AVX512BW-NEXT: vpermt2d %zmm9, %zmm19, %zmm206216; AVX512BW-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}6217; AVX512BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm196218; AVX512BW-NEXT: vpermi2d %zmm5, %zmm2, %zmm176219; AVX512BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm186220; AVX512BW-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}6221; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}6222; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]6223; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm216224; AVX512BW-NEXT: vpermt2d %zmm1, %zmm20, %zmm216225; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]6226; AVX512BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6227; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm186228; AVX512BW-NEXT: vpermt2d %zmm13, %zmm19, %zmm186229; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}6230; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]6231; AVX512BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6232; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm226233; AVX512BW-NEXT: vpermt2d %zmm9, %zmm21, %zmm226234; AVX512BW-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}6235; AVX512BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm216236; AVX512BW-NEXT: vpermi2d %zmm5, %zmm2, %zmm196237; AVX512BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm206238; AVX512BW-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}6239; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}6240; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]6241; AVX512BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6242; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm216243; AVX512BW-NEXT: vpermt2d %zmm13, %zmm20, %zmm216244; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]6245; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm236246; AVX512BW-NEXT: vpermt2d %zmm11, %zmm22, %zmm236247; AVX512BW-NEXT: movw $992, %di # imm = 0x3E06248; AVX512BW-NEXT: kmovd %edi, %k16249; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}6250; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]6251; AVX512BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6252; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm246253; AVX512BW-NEXT: vpermt2d %zmm10, %zmm21, %zmm246254; AVX512BW-NEXT: movb $-32, %dil6255; AVX512BW-NEXT: kmovd %edi, %k26256; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}6257; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm216258; AVX512BW-NEXT: vpermi2d %zmm5, %zmm2, %zmm206259; AVX512BW-NEXT: vpermi2d %zmm3, %zmm0, %zmm226260; AVX512BW-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}6261; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}6262; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]6263; AVX512BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6264; AVX512BW-NEXT: vpermt2d %zmm13, %zmm20, %zmm126265; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]6266; AVX512BW-NEXT: vpermt2d %zmm11, %zmm13, %zmm16267; AVX512BW-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}6268; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]6269; AVX512BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6270; AVX512BW-NEXT: vpermt2d %zmm10, %zmm11, %zmm96271; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}6272; AVX512BW-NEXT: vpermt2d %zmm6, %zmm11, %zmm46273; AVX512BW-NEXT: vpermt2d %zmm5, %zmm20, %zmm26274; AVX512BW-NEXT: vpermt2d %zmm3, %zmm13, %zmm06275; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}6276; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}6277; AVX512BW-NEXT: vmovdqa64 %zmm7, 64(%rsi)6278; AVX512BW-NEXT: vmovdqa64 %zmm8, (%rsi)6279; AVX512BW-NEXT: vmovdqa64 %zmm14, 64(%rdx)6280; AVX512BW-NEXT: vmovdqa64 %zmm15, (%rdx)6281; AVX512BW-NEXT: vmovdqa64 %zmm17, 64(%rcx)6282; AVX512BW-NEXT: vmovdqa64 %zmm16, (%rcx)6283; AVX512BW-NEXT: vmovdqa64 %zmm19, 64(%r8)6284; AVX512BW-NEXT: vmovdqa64 %zmm18, (%r8)6285; AVX512BW-NEXT: vmovdqa64 %zmm22, 64(%r9)6286; AVX512BW-NEXT: vmovdqa64 %zmm23, (%r9)6287; AVX512BW-NEXT: vmovdqa64 %zmm0, 64(%rax)6288; AVX512BW-NEXT: vmovdqa64 %zmm1, (%rax)6289; AVX512BW-NEXT: vzeroupper6290; AVX512BW-NEXT: retq6291;6292; AVX512BW-FCP-LABEL: load_i32_stride6_vf32:6293; AVX512BW-FCP: # %bb.0:6294; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax6295; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm06296; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm36297; AVX512BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm26298; AVX512BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm56299; AVX512BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm66300; AVX512BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm46301; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm106302; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm96303; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm116304; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm16305; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm126306; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm136307; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]6308; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]6309; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm156310; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm14, %zmm156311; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]6312; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm86313; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm86314; AVX512BW-FCP-NEXT: movb $56, %dil6315; AVX512BW-FCP-NEXT: kmovd %edi, %k26316; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}6317; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]6318; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6319; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm166320; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm15, %zmm166321; AVX512BW-FCP-NEXT: movw $-2048, %di # imm = 0xF8006322; AVX512BW-FCP-NEXT: kmovd %edi, %k16323; AVX512BW-FCP-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}6324; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm156325; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm146326; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm76327; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}6328; AVX512BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}6329; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]6330; AVX512BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]6331; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm176332; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm16, %zmm176333; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]6334; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm156335; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm14, %zmm156336; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}6337; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]6338; AVX512BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6339; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm186340; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm17, %zmm186341; AVX512BW-FCP-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}6342; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm176343; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm166344; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm146345; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}6346; AVX512BW-FCP-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}6347; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]6348; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm196349; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm18, %zmm196350; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]6351; AVX512BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6352; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm166353; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm17, %zmm166354; AVX512BW-FCP-NEXT: movw $31, %di6355; AVX512BW-FCP-NEXT: kmovd %edi, %k26356; AVX512BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}6357; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]6358; AVX512BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6359; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm206360; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm19, %zmm206361; AVX512BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}6362; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm196363; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm176364; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm186365; AVX512BW-FCP-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}6366; AVX512BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}6367; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]6368; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm216369; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm20, %zmm216370; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]6371; AVX512BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6372; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm186373; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm19, %zmm186374; AVX512BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}6375; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]6376; AVX512BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6377; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm226378; AVX512BW-FCP-NEXT: vpermt2d %zmm9, %zmm21, %zmm226379; AVX512BW-FCP-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}6380; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm216381; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm196382; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm206383; AVX512BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}6384; AVX512BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}6385; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]6386; AVX512BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6387; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm216388; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm216389; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]6390; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm236391; AVX512BW-FCP-NEXT: vpermt2d %zmm11, %zmm22, %zmm236392; AVX512BW-FCP-NEXT: movw $992, %di # imm = 0x3E06393; AVX512BW-FCP-NEXT: kmovd %edi, %k16394; AVX512BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}6395; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]6396; AVX512BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6397; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm246398; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm21, %zmm246399; AVX512BW-FCP-NEXT: movb $-32, %dil6400; AVX512BW-FCP-NEXT: kmovd %edi, %k26401; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}6402; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm216403; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm206404; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm0, %zmm226405; AVX512BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}6406; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}6407; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]6408; AVX512BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6409; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm126410; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]6411; AVX512BW-FCP-NEXT: vpermt2d %zmm11, %zmm13, %zmm16412; AVX512BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}6413; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]6414; AVX512BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6415; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm11, %zmm96416; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}6417; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm11, %zmm46418; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm26419; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm13, %zmm06420; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}6421; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}6422; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 64(%rsi)6423; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)6424; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)6425; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, (%rdx)6426; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)6427; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, (%rcx)6428; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)6429; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, (%r8)6430; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 64(%r9)6431; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, (%r9)6432; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)6433; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, (%rax)6434; AVX512BW-FCP-NEXT: vzeroupper6435; AVX512BW-FCP-NEXT: retq6436;6437; AVX512DQ-BW-LABEL: load_i32_stride6_vf32:6438; AVX512DQ-BW: # %bb.0:6439; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax6440; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm06441; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm36442; AVX512DQ-BW-NEXT: vmovdqa64 512(%rdi), %zmm26443; AVX512DQ-BW-NEXT: vmovdqa64 576(%rdi), %zmm56444; AVX512DQ-BW-NEXT: vmovdqa64 704(%rdi), %zmm66445; AVX512DQ-BW-NEXT: vmovdqa64 640(%rdi), %zmm46446; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm106447; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm96448; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm116449; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm16450; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm126451; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm136452; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]6453; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]6454; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm156455; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm14, %zmm156456; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]6457; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm86458; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm7, %zmm86459; AVX512DQ-BW-NEXT: movb $56, %dil6460; AVX512DQ-BW-NEXT: kmovd %edi, %k26461; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}6462; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]6463; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6464; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm166465; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm15, %zmm166466; AVX512DQ-BW-NEXT: movw $-2048, %di # imm = 0xF8006467; AVX512DQ-BW-NEXT: kmovd %edi, %k16468; AVX512DQ-BW-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}6469; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm156470; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm5, %zmm146471; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm76472; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}6473; AVX512DQ-BW-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}6474; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]6475; AVX512DQ-BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]6476; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm176477; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm16, %zmm176478; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]6479; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm156480; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm14, %zmm156481; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}6482; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]6483; AVX512DQ-BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6484; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm186485; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm17, %zmm186486; AVX512DQ-BW-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}6487; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm176488; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm5, %zmm166489; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm146490; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}6491; AVX512DQ-BW-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}6492; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]6493; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm196494; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm18, %zmm196495; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]6496; AVX512DQ-BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6497; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm166498; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm17, %zmm166499; AVX512DQ-BW-NEXT: movw $31, %di6500; AVX512DQ-BW-NEXT: kmovd %edi, %k26501; AVX512DQ-BW-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}6502; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]6503; AVX512DQ-BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6504; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm206505; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm19, %zmm206506; AVX512DQ-BW-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}6507; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm196508; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm2, %zmm176509; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm186510; AVX512DQ-BW-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}6511; AVX512DQ-BW-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}6512; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]6513; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm216514; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm20, %zmm216515; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]6516; AVX512DQ-BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6517; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm186518; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm19, %zmm186519; AVX512DQ-BW-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}6520; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]6521; AVX512DQ-BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6522; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm226523; AVX512DQ-BW-NEXT: vpermt2d %zmm9, %zmm21, %zmm226524; AVX512DQ-BW-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}6525; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm216526; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm2, %zmm196527; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm206528; AVX512DQ-BW-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}6529; AVX512DQ-BW-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}6530; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]6531; AVX512DQ-BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6532; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm216533; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm20, %zmm216534; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]6535; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm236536; AVX512DQ-BW-NEXT: vpermt2d %zmm11, %zmm22, %zmm236537; AVX512DQ-BW-NEXT: movw $992, %di # imm = 0x3E06538; AVX512DQ-BW-NEXT: kmovd %edi, %k16539; AVX512DQ-BW-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}6540; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]6541; AVX512DQ-BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6542; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm246543; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm21, %zmm246544; AVX512DQ-BW-NEXT: movb $-32, %dil6545; AVX512DQ-BW-NEXT: kmovd %edi, %k26546; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}6547; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm216548; AVX512DQ-BW-NEXT: vpermi2d %zmm5, %zmm2, %zmm206549; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm0, %zmm226550; AVX512DQ-BW-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}6551; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}6552; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]6553; AVX512DQ-BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6554; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm20, %zmm126555; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]6556; AVX512DQ-BW-NEXT: vpermt2d %zmm11, %zmm13, %zmm16557; AVX512DQ-BW-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}6558; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]6559; AVX512DQ-BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6560; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm11, %zmm96561; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}6562; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm11, %zmm46563; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm20, %zmm26564; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm13, %zmm06565; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}6566; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}6567; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 64(%rsi)6568; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%rsi)6569; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 64(%rdx)6570; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, (%rdx)6571; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 64(%rcx)6572; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, (%rcx)6573; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 64(%r8)6574; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, (%r8)6575; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 64(%r9)6576; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, (%r9)6577; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 64(%rax)6578; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, (%rax)6579; AVX512DQ-BW-NEXT: vzeroupper6580; AVX512DQ-BW-NEXT: retq6581;6582; AVX512DQ-BW-FCP-LABEL: load_i32_stride6_vf32:6583; AVX512DQ-BW-FCP: # %bb.0:6584; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax6585; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm06586; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm36587; AVX512DQ-BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm26588; AVX512DQ-BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm56589; AVX512DQ-BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm66590; AVX512DQ-BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm46591; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm106592; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm96593; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm116594; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm16595; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm126596; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm136597; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]6598; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]6599; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm156600; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm14, %zmm156601; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,6,12,18,24,30,0,0]6602; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm86603; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm86604; AVX512DQ-BW-FCP-NEXT: movb $56, %dil6605; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k26606; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm8 {%k2}6607; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]6608; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6609; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm166610; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm15, %zmm166611; AVX512DQ-BW-FCP-NEXT: movw $-2048, %di # imm = 0xF8006612; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k16613; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm16, %zmm8 {%k1}6614; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm156615; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm146616; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm76617; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k2}6618; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm7 {%k1}6619; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]6620; AVX512DQ-BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]6621; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm176622; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm16, %zmm176623; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [1,7,13,19,25,31,0,0]6624; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm156625; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm14, %zmm156626; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm15 {%k2}6627; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]6628; AVX512DQ-BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6629; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm186630; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm17, %zmm186631; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm18, %zmm15 {%k1}6632; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm176633; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm5, %zmm166634; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm146635; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm14 {%k2}6636; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm17, %zmm14 {%k1}6637; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [2,8,14,20,26,0,0,0]6638; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm196639; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm18, %zmm196640; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]6641; AVX512DQ-BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]6642; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm166643; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm17, %zmm166644; AVX512DQ-BW-FCP-NEXT: movw $31, %di6645; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k26646; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm16 {%k2}6647; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]6648; AVX512DQ-BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6649; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm206650; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm19, %zmm206651; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm16 {%k1}6652; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm196653; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm176654; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm186655; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm18, %zmm17 {%k2}6656; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm17 {%k1}6657; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [3,9,15,21,27,0,0,0]6658; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm216659; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm20, %zmm216660; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]6661; AVX512DQ-BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]6662; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm186663; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm19, %zmm186664; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2}6665; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]6666; AVX512DQ-BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6667; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm226668; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm9, %zmm21, %zmm226669; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm22, %zmm18 {%k1}6670; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm216671; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm196672; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm3, %zmm206673; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm19 {%k2}6674; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm19 {%k1}6675; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]6676; AVX512DQ-BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6677; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm216678; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm216679; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [20,26,0,6,12,0,0,0]6680; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm236681; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm11, %zmm22, %zmm236682; AVX512DQ-BW-FCP-NEXT: movw $992, %di # imm = 0x3E06683; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k16684; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm23 {%k1}6685; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]6686; AVX512DQ-BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]6687; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm246688; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm21, %zmm246689; AVX512DQ-BW-FCP-NEXT: movb $-32, %dil6690; AVX512DQ-BW-FCP-NEXT: kmovd %edi, %k26691; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, %zmm23 {%k2}6692; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm216693; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm206694; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm0, %zmm226695; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm22 {%k1}6696; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, %zmm22 {%k2}6697; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]6698; AVX512DQ-BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]6699; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm126700; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [21,27,1,7,13,0,0,0]6701; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm11, %zmm13, %zmm16702; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm1 {%k1}6703; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]6704; AVX512DQ-BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6705; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm11, %zmm96706; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k2}6707; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm11, %zmm46708; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm26709; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm13, %zmm06710; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}6711; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k2}6712; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 64(%rsi)6713; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)6714; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)6715; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, (%rdx)6716; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)6717; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, (%rcx)6718; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)6719; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, (%r8)6720; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 64(%r9)6721; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, (%r9)6722; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)6723; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, (%rax)6724; AVX512DQ-BW-FCP-NEXT: vzeroupper6725; AVX512DQ-BW-FCP-NEXT: retq6726 %wide.vec = load <192 x i32>, ptr %in.vec, align 646727 %strided.vec0 = shufflevector <192 x i32> %wide.vec, <192 x i32> poison, <32 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90, i32 96, i32 102, i32 108, i32 114, i32 120, i32 126, i32 132, i32 138, i32 144, i32 150, i32 156, i32 162, i32 168, i32 174, i32 180, i32 186>6728 %strided.vec1 = shufflevector <192 x i32> %wide.vec, <192 x i32> poison, <32 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91, i32 97, i32 103, i32 109, i32 115, i32 121, i32 127, i32 133, i32 139, i32 145, i32 151, i32 157, i32 163, i32 169, i32 175, i32 181, i32 187>6729 %strided.vec2 = shufflevector <192 x i32> %wide.vec, <192 x i32> poison, <32 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92, i32 98, i32 104, i32 110, i32 116, i32 122, i32 128, i32 134, i32 140, i32 146, i32 152, i32 158, i32 164, i32 170, i32 176, i32 182, i32 188>6730 %strided.vec3 = shufflevector <192 x i32> %wide.vec, <192 x i32> poison, <32 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93, i32 99, i32 105, i32 111, i32 117, i32 123, i32 129, i32 135, i32 141, i32 147, i32 153, i32 159, i32 165, i32 171, i32 177, i32 183, i32 189>6731 %strided.vec4 = shufflevector <192 x i32> %wide.vec, <192 x i32> poison, <32 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94, i32 100, i32 106, i32 112, i32 118, i32 124, i32 130, i32 136, i32 142, i32 148, i32 154, i32 160, i32 166, i32 172, i32 178, i32 184, i32 190>6732 %strided.vec5 = shufflevector <192 x i32> %wide.vec, <192 x i32> poison, <32 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95, i32 101, i32 107, i32 113, i32 119, i32 125, i32 131, i32 137, i32 143, i32 149, i32 155, i32 161, i32 167, i32 173, i32 179, i32 185, i32 191>6733 store <32 x i32> %strided.vec0, ptr %out.vec0, align 646734 store <32 x i32> %strided.vec1, ptr %out.vec1, align 646735 store <32 x i32> %strided.vec2, ptr %out.vec2, align 646736 store <32 x i32> %strided.vec3, ptr %out.vec3, align 646737 store <32 x i32> %strided.vec4, ptr %out.vec4, align 646738 store <32 x i32> %strided.vec5, ptr %out.vec5, align 646739 ret void6740}6741 6742define void @load_i32_stride6_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5) nounwind {6743; SSE-LABEL: load_i32_stride6_vf64:6744; SSE: # %bb.0:6745; SSE-NEXT: subq $2184, %rsp # imm = 0x8886746; SSE-NEXT: movdqa 912(%rdi), %xmm76747; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6748; SSE-NEXT: movdqa 928(%rdi), %xmm36749; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6750; SSE-NEXT: movdqa 864(%rdi), %xmm86751; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6752; SSE-NEXT: movdqa 880(%rdi), %xmm46753; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6754; SSE-NEXT: movdqa 528(%rdi), %xmm96755; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6756; SSE-NEXT: movdqa 544(%rdi), %xmm56757; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6758; SSE-NEXT: movdqa 480(%rdi), %xmm106759; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6760; SSE-NEXT: movdqa 496(%rdi), %xmm66761; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6762; SSE-NEXT: movdqa 144(%rdi), %xmm116763; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6764; SSE-NEXT: movdqa 160(%rdi), %xmm26765; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6766; SSE-NEXT: movdqa 96(%rdi), %xmm16767; SSE-NEXT: movdqa %xmm1, (%rsp) # 16-byte Spill6768; SSE-NEXT: movdqa 112(%rdi), %xmm06769; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6770; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6771; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6772; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,3,3]6773; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm11[0,0,1,1]6774; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]6775; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]6776; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6777; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]6778; SSE-NEXT: movdqa %xmm10, %xmm16779; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6780; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]6781; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm9[0,0,1,1]6782; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]6783; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]6784; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6785; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]6786; SSE-NEXT: movdqa %xmm8, %xmm16787; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6788; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,2,3,3]6789; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,0,1,1]6790; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]6791; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]6792; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6793; SSE-NEXT: movdqa 1248(%rdi), %xmm16794; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6795; SSE-NEXT: movdqa 1264(%rdi), %xmm06796; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6797; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6798; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6799; SSE-NEXT: movdqa 1296(%rdi), %xmm36800; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6801; SSE-NEXT: movdqa 1312(%rdi), %xmm06802; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6803; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6804; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,0,1,1]6805; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]6806; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]6807; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6808; SSE-NEXT: movdqa (%rdi), %xmm16809; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6810; SSE-NEXT: movdqa 16(%rdi), %xmm06811; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6812; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6813; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6814; SSE-NEXT: movdqa 64(%rdi), %xmm06815; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6816; SSE-NEXT: movdqa 48(%rdi), %xmm26817; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6818; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6819; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]6820; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6821; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6822; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6823; SSE-NEXT: movdqa 384(%rdi), %xmm16824; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6825; SSE-NEXT: movdqa 400(%rdi), %xmm06826; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6827; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6828; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6829; SSE-NEXT: movdqa 432(%rdi), %xmm26830; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6831; SSE-NEXT: movdqa 448(%rdi), %xmm06832; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6833; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6834; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]6835; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6836; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6837; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6838; SSE-NEXT: movdqa 768(%rdi), %xmm16839; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6840; SSE-NEXT: movdqa 784(%rdi), %xmm06841; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6842; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6843; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6844; SSE-NEXT: movdqa 816(%rdi), %xmm26845; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6846; SSE-NEXT: movdqa 832(%rdi), %xmm06847; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6848; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6849; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]6850; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6851; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6852; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6853; SSE-NEXT: movdqa 1152(%rdi), %xmm16854; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6855; SSE-NEXT: movdqa 1168(%rdi), %xmm06856; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6857; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6858; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6859; SSE-NEXT: movdqa 1200(%rdi), %xmm26860; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6861; SSE-NEXT: movdqa 1216(%rdi), %xmm06862; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6863; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6864; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]6865; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6866; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6867; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6868; SSE-NEXT: movdqa 288(%rdi), %xmm26869; SSE-NEXT: movdqa 304(%rdi), %xmm06870; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6871; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6872; SSE-NEXT: movdqa %xmm2, %xmm16873; SSE-NEXT: movdqa %xmm2, %xmm156874; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6875; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6876; SSE-NEXT: movdqa 336(%rdi), %xmm76877; SSE-NEXT: movdqa 352(%rdi), %xmm06878; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6879; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6880; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm7[0,0,1,1]6881; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6882; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6883; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6884; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6885; SSE-NEXT: movdqa 672(%rdi), %xmm16886; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6887; SSE-NEXT: movdqa 688(%rdi), %xmm06888; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6889; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6890; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6891; SSE-NEXT: movdqa 720(%rdi), %xmm36892; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6893; SSE-NEXT: movdqa 736(%rdi), %xmm06894; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6895; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6896; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]6897; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6898; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6899; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6900; SSE-NEXT: movdqa 1056(%rdi), %xmm16901; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6902; SSE-NEXT: movdqa 1072(%rdi), %xmm06903; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6904; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6905; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6906; SSE-NEXT: movdqa 1104(%rdi), %xmm26907; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6908; SSE-NEXT: movdqa 1120(%rdi), %xmm06909; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6910; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6911; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]6912; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6913; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6914; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6915; SSE-NEXT: movdqa 1440(%rdi), %xmm16916; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6917; SSE-NEXT: movdqa 1456(%rdi), %xmm06918; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6919; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]6920; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6921; SSE-NEXT: movdqa 1488(%rdi), %xmm26922; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6923; SSE-NEXT: movdqa 1504(%rdi), %xmm06924; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6925; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]6926; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]6927; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]6928; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]6929; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6930; SSE-NEXT: movdqa 192(%rdi), %xmm56931; SSE-NEXT: movdqa 208(%rdi), %xmm66932; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[2,3,2,3]6933; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6934; SSE-NEXT: movdqa %xmm5, %xmm36935; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6936; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]6937; SSE-NEXT: movdqa 240(%rdi), %xmm26938; SSE-NEXT: movdqa 256(%rdi), %xmm16939; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm1[2,2,3,3]6940; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6941; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm2[0,0,1,1]6942; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6943; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm4[0],xmm12[1],xmm4[1]6944; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm3[0],xmm12[1]6945; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6946; SSE-NEXT: movdqa 576(%rdi), %xmm106947; SSE-NEXT: movdqa 592(%rdi), %xmm146948; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm14[2,3,2,3]6949; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6950; SSE-NEXT: movdqa %xmm10, %xmm46951; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6952; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6953; SSE-NEXT: movdqa 624(%rdi), %xmm116954; SSE-NEXT: movdqa 640(%rdi), %xmm136955; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm13[2,2,3,3]6956; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6957; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm11[0,0,1,1]6958; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6959; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm3[0],xmm12[1],xmm3[1]6960; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm4[0],xmm12[1]6961; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6962; SSE-NEXT: movdqa 960(%rdi), %xmm46963; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6964; SSE-NEXT: movdqa 976(%rdi), %xmm06965; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6966; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]6967; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6968; SSE-NEXT: movdqa 1008(%rdi), %xmm86969; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6970; SSE-NEXT: movdqa 1024(%rdi), %xmm06971; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,2,3,3]6972; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6973; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm8[0,0,1,1]6974; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm3[0],xmm12[1],xmm3[1]6975; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm4[0],xmm12[1]6976; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6977; SSE-NEXT: movdqa 1344(%rdi), %xmm46978; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6979; SSE-NEXT: movdqa 1360(%rdi), %xmm36980; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6981; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,3,2,3]6982; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6983; SSE-NEXT: movdqa 1392(%rdi), %xmm96984; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6985; SSE-NEXT: movdqa 1408(%rdi), %xmm36986; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6987; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]6988; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm9[0,0,1,1]6989; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm3[0],xmm12[1],xmm3[1]6990; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm4[0],xmm12[1]6991; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6992; SSE-NEXT: pshufd $85, (%rsp), %xmm3 # 16-byte Folded Reload6993; SSE-NEXT: # xmm3 = mem[1,1,1,1]6994; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload6995; SSE-NEXT: # xmm4 = mem[3,3,3,3]6996; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]6997; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload6998; SSE-NEXT: # xmm4 = mem[2,3,2,3]6999; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload7000; SSE-NEXT: movdqa %xmm9, %xmm127001; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm4[0],xmm12[1],xmm4[1]7002; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm3[0],xmm12[1]7003; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7004; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7005; SSE-NEXT: # xmm3 = mem[1,1,1,1]7006; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7007; SSE-NEXT: # xmm4 = mem[3,3,3,3]7008; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7009; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7010; SSE-NEXT: # xmm4 = mem[2,3,2,3]7011; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload7012; SSE-NEXT: movdqa %xmm8, %xmm127013; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm4[0],xmm12[1],xmm4[1]7014; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm3[0],xmm12[1]7015; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7016; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm15[1,1,1,1]7017; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7018; SSE-NEXT: # xmm4 = mem[3,3,3,3]7019; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7020; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7021; SSE-NEXT: # xmm4 = mem[2,3,2,3]7022; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm4[0],xmm7[1],xmm4[1]7023; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm3[0],xmm7[1]7024; SSE-NEXT: movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7025; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm5[1,1,1,1]7026; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm6[3,3,3,3]7027; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7028; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm1[2,3,2,3]7029; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]7030; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm3[0],xmm2[1]7031; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7032; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7033; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm15[1,1,1,1]7034; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7035; SSE-NEXT: # xmm4 = mem[3,3,3,3]7036; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7037; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7038; SSE-NEXT: # xmm4 = mem[2,3,2,3]7039; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7040; SSE-NEXT: movdqa %xmm12, %xmm77041; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm4[0],xmm7[1],xmm4[1]7042; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm3[0],xmm7[1]7043; SSE-NEXT: movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7044; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload7045; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm7[1,1,1,1]7046; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7047; SSE-NEXT: # xmm4 = mem[3,3,3,3]7048; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7049; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7050; SSE-NEXT: # xmm4 = mem[2,3,2,3]7051; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload7052; SSE-NEXT: movdqa %xmm2, %xmm57053; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]7054; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm3[0],xmm5[1]7055; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7056; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7057; SSE-NEXT: # xmm3 = mem[1,1,1,1]7058; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7059; SSE-NEXT: # xmm4 = mem[3,3,3,3]7060; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7061; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7062; SSE-NEXT: # xmm4 = mem[2,3,2,3]7063; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload7064; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]7065; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm3[0],xmm5[1]7066; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7067; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm10[1,1,1,1]7068; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm14[3,3,3,3]7069; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7070; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm13[2,3,2,3]7071; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm4[0],xmm11[1],xmm4[1]7072; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm3[0],xmm11[1]7073; SSE-NEXT: movapd %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7074; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7075; SSE-NEXT: # xmm3 = mem[1,1,1,1]7076; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7077; SSE-NEXT: # xmm4 = mem[3,3,3,3]7078; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7079; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7080; SSE-NEXT: # xmm4 = mem[2,3,2,3]7081; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload7082; SSE-NEXT: movdqa %xmm11, %xmm17083; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]7084; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm3[0],xmm1[1]7085; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7086; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload7087; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,1,1]7088; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7089; SSE-NEXT: # xmm4 = mem[3,3,3,3]7090; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7091; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7092; SSE-NEXT: # xmm4 = mem[2,3,2,3]7093; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7094; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]7095; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm3[0],xmm1[1]7096; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7097; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7098; SSE-NEXT: # xmm3 = mem[1,1,1,1]7099; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7100; SSE-NEXT: # xmm4 = mem[3,3,3,3]7101; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7102; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7103; SSE-NEXT: # xmm4 = mem[2,3,2,3]7104; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload7105; SSE-NEXT: movdqa %xmm13, %xmm17106; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]7107; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm3[0],xmm1[1]7108; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7109; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7110; SSE-NEXT: # xmm3 = mem[1,1,1,1]7111; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7112; SSE-NEXT: # xmm4 = mem[3,3,3,3]7113; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7114; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]7115; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7116; SSE-NEXT: movdqa %xmm14, %xmm07117; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]7118; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7119; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7120; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7121; SSE-NEXT: # xmm3 = mem[1,1,1,1]7122; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7123; SSE-NEXT: # xmm4 = mem[3,3,3,3]7124; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7125; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7126; SSE-NEXT: # xmm4 = mem[2,3,2,3]7127; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7128; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]7129; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7130; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7131; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload7132; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm10[1,1,1,1]7133; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7134; SSE-NEXT: # xmm4 = mem[3,3,3,3]7135; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7136; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7137; SSE-NEXT: # xmm4 = mem[2,3,2,3]7138; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload7139; SSE-NEXT: movdqa %xmm5, %xmm07140; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]7141; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7142; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7143; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7144; SSE-NEXT: # xmm3 = mem[1,1,1,1]7145; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7146; SSE-NEXT: # xmm4 = mem[3,3,3,3]7147; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7148; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7149; SSE-NEXT: # xmm4 = mem[2,3,2,3]7150; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7151; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]7152; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7153; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7154; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7155; SSE-NEXT: # xmm3 = mem[1,1,1,1]7156; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7157; SSE-NEXT: # xmm4 = mem[3,3,3,3]7158; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7159; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7160; SSE-NEXT: # xmm4 = mem[2,3,2,3]7161; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7162; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]7163; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7164; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7165; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[2,2,3,3]7166; SSE-NEXT: movdqa 80(%rdi), %xmm17167; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7168; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7169; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7170; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7171; SSE-NEXT: # xmm3 = mem[2,3,2,3]7172; SSE-NEXT: movdqa 32(%rdi), %xmm17173; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7174; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7175; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7176; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7177; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[2,2,3,3]7178; SSE-NEXT: movdqa 176(%rdi), %xmm17179; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7180; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7181; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7182; SSE-NEXT: pshufd $238, (%rsp), %xmm3 # 16-byte Folded Reload7183; SSE-NEXT: # xmm3 = mem[2,3,2,3]7184; SSE-NEXT: movdqa 128(%rdi), %xmm17185; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7186; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7187; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7188; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7189; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7190; SSE-NEXT: # xmm0 = mem[2,2,3,3]7191; SSE-NEXT: movdqa 272(%rdi), %xmm17192; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7193; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7194; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7195; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7196; SSE-NEXT: # xmm3 = mem[2,3,2,3]7197; SSE-NEXT: movdqa 224(%rdi), %xmm17198; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7199; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7200; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7201; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7202; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7203; SSE-NEXT: # xmm0 = mem[2,2,3,3]7204; SSE-NEXT: movdqa 368(%rdi), %xmm17205; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7206; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7207; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7208; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7209; SSE-NEXT: # xmm3 = mem[2,3,2,3]7210; SSE-NEXT: movdqa 320(%rdi), %xmm17211; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7212; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7213; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7214; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7215; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,3,3]7216; SSE-NEXT: movdqa 464(%rdi), %xmm17217; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7218; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7219; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7220; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm7[2,3,2,3]7221; SSE-NEXT: movdqa 416(%rdi), %xmm17222; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7223; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7224; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7225; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7226; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,2,3,3]7227; SSE-NEXT: movdqa 560(%rdi), %xmm17228; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7229; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7230; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7231; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm15[2,3,2,3]7232; SSE-NEXT: movdqa 512(%rdi), %xmm27233; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7234; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]7235; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7236; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7237; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7238; SSE-NEXT: # xmm0 = mem[2,2,3,3]7239; SSE-NEXT: movdqa 656(%rdi), %xmm17240; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7241; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7242; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7243; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7244; SSE-NEXT: # xmm3 = mem[2,3,2,3]7245; SSE-NEXT: movdqa 608(%rdi), %xmm97246; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1]7247; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7248; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7249; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7250; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7251; SSE-NEXT: # xmm0 = mem[2,2,3,3]7252; SSE-NEXT: movdqa 752(%rdi), %xmm17253; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7254; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7255; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7256; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7257; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm12[2,3,2,3]7258; SSE-NEXT: movdqa 704(%rdi), %xmm27259; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7260; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]7261; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7262; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7263; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7264; SSE-NEXT: # xmm0 = mem[2,2,3,3]7265; SSE-NEXT: movdqa 848(%rdi), %xmm17266; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7267; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7268; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7269; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm6[2,3,2,3]7270; SSE-NEXT: movdqa 800(%rdi), %xmm27271; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7272; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]7273; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7274; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7275; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[2,2,3,3]7276; SSE-NEXT: movdqa 944(%rdi), %xmm17277; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7278; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7279; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7280; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7281; SSE-NEXT: # xmm3 = mem[2,3,2,3]7282; SSE-NEXT: movdqa 896(%rdi), %xmm87283; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1]7284; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7285; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7286; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7287; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[2,2,3,3]7288; SSE-NEXT: movdqa 1040(%rdi), %xmm17289; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7290; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7291; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7292; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload7293; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm7[2,3,2,3]7294; SSE-NEXT: movdqa 992(%rdi), %xmm17295; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7296; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7297; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7298; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7299; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[2,2,3,3]7300; SSE-NEXT: movdqa 1136(%rdi), %xmm17301; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7302; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7303; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7304; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7305; SSE-NEXT: # xmm3 = mem[2,3,2,3]7306; SSE-NEXT: movdqa 1088(%rdi), %xmm67307; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1]7308; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7309; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7310; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7311; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]7312; SSE-NEXT: movdqa 1232(%rdi), %xmm27313; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7314; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]7315; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7316; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm10[2,3,2,3]7317; SSE-NEXT: movdqa 1184(%rdi), %xmm17318; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7319; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7320; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7321; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7322; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7323; SSE-NEXT: # xmm0 = mem[2,2,3,3]7324; SSE-NEXT: movdqa 1328(%rdi), %xmm17325; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7326; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7327; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7328; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7329; SSE-NEXT: # xmm3 = mem[2,3,2,3]7330; SSE-NEXT: movdqa 1280(%rdi), %xmm17331; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7332; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7333; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7334; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7335; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7336; SSE-NEXT: # xmm0 = mem[2,2,3,3]7337; SSE-NEXT: movdqa 1424(%rdi), %xmm17338; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7339; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7340; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7341; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7342; SSE-NEXT: # xmm3 = mem[2,3,2,3]7343; SSE-NEXT: movdqa 1376(%rdi), %xmm17344; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7345; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7346; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7347; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7348; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7349; SSE-NEXT: # xmm0 = mem[2,2,3,3]7350; SSE-NEXT: movdqa 1520(%rdi), %xmm17351; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7352; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]7353; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7354; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7355; SSE-NEXT: # xmm3 = mem[2,3,2,3]7356; SSE-NEXT: movdqa 1472(%rdi), %xmm17357; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7358; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7359; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7360; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7361; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7362; SSE-NEXT: # xmm3 = mem[3,3,3,3]7363; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload7364; SSE-NEXT: # xmm4 = mem[1,1,1,1]7365; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]7366; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7367; SSE-NEXT: # xmm0 = mem[2,3,2,3]7368; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7369; SSE-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]7370; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]7371; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7372; SSE-NEXT: pshufd $255, (%rsp), %xmm2 # 16-byte Folded Reload7373; SSE-NEXT: # xmm2 = mem[3,3,3,3]7374; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7375; SSE-NEXT: # xmm3 = mem[1,1,1,1]7376; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]7377; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7378; SSE-NEXT: # xmm0 = mem[2,3,2,3]7379; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7380; SSE-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]7381; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm2[0],xmm0[1]7382; SSE-NEXT: movapd %xmm0, (%rsp) # 16-byte Spill7383; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7384; SSE-NEXT: # xmm0 = mem[3,3,3,3]7385; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7386; SSE-NEXT: # xmm2 = mem[1,1,1,1]7387; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]7388; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7389; SSE-NEXT: # xmm1 = mem[2,3,2,3]7390; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7391; SSE-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7392; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7393; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7394; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7395; SSE-NEXT: # xmm0 = mem[3,3,3,3]7396; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7397; SSE-NEXT: # xmm1 = mem[1,1,1,1]7398; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7399; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7400; SSE-NEXT: # xmm1 = mem[2,3,2,3]7401; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7402; SSE-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7403; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7404; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7405; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7406; SSE-NEXT: # xmm0 = mem[3,3,3,3]7407; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7408; SSE-NEXT: # xmm1 = mem[1,1,1,1]7409; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7410; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7411; SSE-NEXT: # xmm1 = mem[2,3,2,3]7412; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7413; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1]7414; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7415; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7416; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7417; SSE-NEXT: # xmm0 = mem[3,3,3,3]7418; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7419; SSE-NEXT: # xmm1 = mem[1,1,1,1]7420; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7421; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7422; SSE-NEXT: # xmm1 = mem[2,3,2,3]7423; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload7424; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]7425; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7426; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7427; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7428; SSE-NEXT: # xmm0 = mem[3,3,3,3]7429; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[1,1,1,1]7430; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7431; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7432; SSE-NEXT: # xmm1 = mem[2,3,2,3]7433; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7434; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1]7435; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7436; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7437; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[3,3,3,3]7438; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7439; SSE-NEXT: # xmm1 = mem[1,1,1,1]7440; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7441; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7442; SSE-NEXT: # xmm1 = mem[2,3,2,3]7443; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload7444; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1]7445; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7446; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7447; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7448; SSE-NEXT: # xmm0 = mem[3,3,3,3]7449; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7450; SSE-NEXT: # xmm1 = mem[1,1,1,1]7451; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7452; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7453; SSE-NEXT: # xmm1 = mem[2,3,2,3]7454; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload7455; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1]7456; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7457; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7458; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7459; SSE-NEXT: # xmm0 = mem[3,3,3,3]7460; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[1,1,1,1]7461; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7462; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7463; SSE-NEXT: # xmm1 = mem[2,3,2,3]7464; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload7465; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]7466; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7467; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7468; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[3,3,3,3]7469; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7470; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[1,1,1,1]7471; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7472; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7473; SSE-NEXT: # xmm1 = mem[2,3,2,3]7474; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload7475; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1]7476; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7477; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7478; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7479; SSE-NEXT: # xmm0 = mem[3,3,3,3]7480; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[1,1,1,1]7481; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7482; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7483; SSE-NEXT: # xmm1 = mem[2,3,2,3]7484; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload7485; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1]7486; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7487; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7488; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[3,3,3,3]7489; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload7490; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[1,1,1,1]7491; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7492; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7493; SSE-NEXT: # xmm1 = mem[2,3,2,3]7494; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7495; SSE-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7496; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7497; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7498; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7499; SSE-NEXT: # xmm0 = mem[3,3,3,3]7500; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7501; SSE-NEXT: # xmm1 = mem[1,1,1,1]7502; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7503; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7504; SSE-NEXT: # xmm1 = mem[2,3,2,3]7505; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload7506; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]7507; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7508; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7509; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7510; SSE-NEXT: # xmm0 = mem[3,3,3,3]7511; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload7512; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[1,1,1,1]7513; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7514; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7515; SSE-NEXT: # xmm1 = mem[2,3,2,3]7516; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7517; SSE-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7518; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7519; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7520; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7521; SSE-NEXT: # xmm0 = mem[3,3,3,3]7522; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7523; SSE-NEXT: # xmm1 = mem[1,1,1,1]7524; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7525; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7526; SSE-NEXT: # xmm1 = mem[2,3,2,3]7527; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload7528; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]7529; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7530; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7531; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7532; SSE-NEXT: # xmm0 = mem[2,3,2,3]7533; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7534; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7535; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7536; SSE-NEXT: # xmm0 = mem[2,2,3,3]7537; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7538; SSE-NEXT: # xmm2 = mem[0,0,1,1]7539; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7540; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7541; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7542; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7543; SSE-NEXT: # xmm0 = mem[2,3,2,3]7544; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7545; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7546; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7547; SSE-NEXT: # xmm3 = mem[2,2,3,3]7548; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7549; SSE-NEXT: # xmm2 = mem[0,0,1,1]7550; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]7551; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7552; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7553; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7554; SSE-NEXT: # xmm0 = mem[2,3,2,3]7555; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7556; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7557; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7558; SSE-NEXT: # xmm0 = mem[2,2,3,3]7559; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7560; SSE-NEXT: # xmm2 = mem[0,0,1,1]7561; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7562; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7563; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7564; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7565; SSE-NEXT: # xmm0 = mem[2,3,2,3]7566; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7567; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7568; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7569; SSE-NEXT: # xmm0 = mem[2,2,3,3]7570; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7571; SSE-NEXT: # xmm2 = mem[0,0,1,1]7572; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7573; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7574; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7575; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7576; SSE-NEXT: # xmm0 = mem[2,3,2,3]7577; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7578; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7579; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[2,2,3,3]7580; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload7581; SSE-NEXT: # xmm15 = mem[0,0,1,1]7582; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1]7583; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm1[0],xmm15[1]7584; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7585; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7586; SSE-NEXT: # xmm0 = mem[2,3,2,3]7587; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7588; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7589; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7590; SSE-NEXT: # xmm0 = mem[2,2,3,3]7591; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7592; SSE-NEXT: # xmm2 = mem[0,0,1,1]7593; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7594; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7595; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7596; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7597; SSE-NEXT: # xmm0 = mem[2,3,2,3]7598; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7599; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7600; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[2,2,3,3]7601; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7602; SSE-NEXT: # xmm2 = mem[0,0,1,1]7603; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7604; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7605; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7606; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7607; SSE-NEXT: # xmm0 = mem[2,3,2,3]7608; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7609; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7610; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[2,2,3,3]7611; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7612; SSE-NEXT: # xmm2 = mem[0,0,1,1]7613; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7614; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7615; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7616; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7617; SSE-NEXT: # xmm0 = mem[2,3,2,3]7618; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7619; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7620; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[2,2,3,3]7621; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload7622; SSE-NEXT: # xmm11 = mem[0,0,1,1]7623; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm0[0],xmm11[1],xmm0[1]7624; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm1[0],xmm11[1]7625; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7626; SSE-NEXT: # xmm0 = mem[2,3,2,3]7627; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7628; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7629; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[2,2,3,3]7630; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload7631; SSE-NEXT: # xmm9 = mem[0,0,1,1]7632; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]7633; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm1[0],xmm9[1]7634; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,3,2,3]7635; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7636; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7637; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[2,2,3,3]7638; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload7639; SSE-NEXT: # xmm8 = mem[0,0,1,1]7640; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1]7641; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm1[0],xmm8[1]7642; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7643; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[2,3,2,3]7644; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7645; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7646; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,2,3,3]7647; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload7648; SSE-NEXT: # xmm13 = mem[0,0,1,1]7649; SSE-NEXT: punpckldq {{.*#+}} xmm13 = xmm13[0],xmm0[0],xmm13[1],xmm0[1]7650; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm1[0],xmm13[1]7651; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[2,3,2,3]7652; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7653; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7654; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7655; SSE-NEXT: # xmm0 = mem[2,2,3,3]7656; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload7657; SSE-NEXT: # xmm7 = mem[0,0,1,1]7658; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm0[0],xmm7[1],xmm0[1]7659; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm1[0],xmm7[1]7660; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload7661; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[2,3,2,3]7662; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7663; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7664; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]7665; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload7666; SSE-NEXT: # xmm5 = mem[0,0,1,1]7667; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]7668; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm1[0],xmm5[1]7669; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[2,3,2,3]7670; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7671; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7672; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7673; SSE-NEXT: # xmm1 = mem[2,2,3,3]7674; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload7675; SSE-NEXT: # xmm3 = mem[0,0,1,1]7676; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]7677; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]7678; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload7679; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]7680; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7681; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]7682; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,2,3,3]7683; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload7684; SSE-NEXT: # xmm2 = mem[0,0,1,1]7685; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]7686; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]7687; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7688; SSE-NEXT: # xmm0 = mem[1,1,1,1]7689; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7690; SSE-NEXT: # xmm1 = mem[3,3,3,3]7691; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7692; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7693; SSE-NEXT: # xmm1 = mem[2,3,2,3]7694; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7695; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7696; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7697; SSE-NEXT: movapd %xmm15, %xmm47698; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7699; SSE-NEXT: # xmm0 = mem[1,1,1,1]7700; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7701; SSE-NEXT: # xmm1 = mem[3,3,3,3]7702; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7703; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7704; SSE-NEXT: # xmm1 = mem[2,3,2,3]7705; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7706; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7707; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7708; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7709; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7710; SSE-NEXT: # xmm0 = mem[1,1,1,1]7711; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7712; SSE-NEXT: # xmm1 = mem[3,3,3,3]7713; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7714; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7715; SSE-NEXT: # xmm1 = mem[2,3,2,3]7716; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7717; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7718; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7719; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7720; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7721; SSE-NEXT: # xmm0 = mem[1,1,1,1]7722; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7723; SSE-NEXT: # xmm1 = mem[3,3,3,3]7724; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7725; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7726; SSE-NEXT: # xmm1 = mem[2,3,2,3]7727; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7728; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7729; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7730; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7731; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7732; SSE-NEXT: # xmm0 = mem[1,1,1,1]7733; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7734; SSE-NEXT: # xmm1 = mem[3,3,3,3]7735; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7736; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7737; SSE-NEXT: # xmm1 = mem[2,3,2,3]7738; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7739; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7740; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7741; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7742; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7743; SSE-NEXT: # xmm0 = mem[1,1,1,1]7744; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7745; SSE-NEXT: # xmm1 = mem[3,3,3,3]7746; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7747; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7748; SSE-NEXT: # xmm1 = mem[2,3,2,3]7749; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7750; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7751; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7752; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7753; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7754; SSE-NEXT: # xmm0 = mem[1,1,1,1]7755; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7756; SSE-NEXT: # xmm1 = mem[3,3,3,3]7757; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7758; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7759; SSE-NEXT: # xmm1 = mem[2,3,2,3]7760; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7761; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]7762; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]7763; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7764; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7765; SSE-NEXT: # xmm0 = mem[1,1,1,1]7766; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7767; SSE-NEXT: # xmm1 = mem[3,3,3,3]7768; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7769; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7770; SSE-NEXT: # xmm1 = mem[2,3,2,3]7771; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7772; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]7773; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]7774; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7775; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7776; SSE-NEXT: # xmm0 = mem[1,1,1,1]7777; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7778; SSE-NEXT: # xmm1 = mem[3,3,3,3]7779; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7780; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7781; SSE-NEXT: # xmm1 = mem[2,3,2,3]7782; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7783; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]7784; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]7785; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7786; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7787; SSE-NEXT: # xmm0 = mem[1,1,1,1]7788; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7789; SSE-NEXT: # xmm1 = mem[3,3,3,3]7790; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7791; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7792; SSE-NEXT: # xmm1 = mem[2,3,2,3]7793; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7794; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]7795; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]7796; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7797; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7798; SSE-NEXT: # xmm0 = mem[1,1,1,1]7799; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7800; SSE-NEXT: # xmm1 = mem[3,3,3,3]7801; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7802; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7803; SSE-NEXT: # xmm1 = mem[2,3,2,3]7804; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7805; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]7806; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]7807; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7808; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7809; SSE-NEXT: # xmm0 = mem[1,1,1,1]7810; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[3,3,3,3]7811; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7812; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7813; SSE-NEXT: # xmm1 = mem[2,3,2,3]7814; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7815; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]7816; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm0[0],xmm12[1]7817; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7818; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7819; SSE-NEXT: # xmm0 = mem[1,1,1,1]7820; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7821; SSE-NEXT: # xmm1 = mem[3,3,3,3]7822; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7823; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7824; SSE-NEXT: # xmm1 = mem[2,3,2,3]7825; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7826; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]7827; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm0[0],xmm12[1]7828; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7829; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7830; SSE-NEXT: # xmm0 = mem[1,1,1,1]7831; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[3,3,3,3]7832; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7833; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7834; SSE-NEXT: # xmm1 = mem[2,3,2,3]7835; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload7836; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1]7837; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]7838; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7839; SSE-NEXT: # xmm0 = mem[1,1,1,1]7840; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7841; SSE-NEXT: # xmm1 = mem[3,3,3,3]7842; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7843; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7844; SSE-NEXT: # xmm1 = mem[2,3,2,3]7845; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload7846; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm1[0],xmm12[1],xmm1[1]7847; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm0[0],xmm12[1]7848; SSE-NEXT: pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload7849; SSE-NEXT: # xmm0 = mem[1,1,1,1]7850; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[3,3,3,3]7851; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]7852; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload7853; SSE-NEXT: # xmm1 = mem[2,3,2,3]7854; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload7855; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]7856; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]7857; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7858; SSE-NEXT: movaps %xmm0, 224(%rsi)7859; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7860; SSE-NEXT: movaps %xmm0, 160(%rsi)7861; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7862; SSE-NEXT: movaps %xmm0, 96(%rsi)7863; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7864; SSE-NEXT: movaps %xmm0, 32(%rsi)7865; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7866; SSE-NEXT: movaps %xmm0, 240(%rsi)7867; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7868; SSE-NEXT: movaps %xmm0, 176(%rsi)7869; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7870; SSE-NEXT: movaps %xmm0, 112(%rsi)7871; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7872; SSE-NEXT: movaps %xmm0, 48(%rsi)7873; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7874; SSE-NEXT: movaps %xmm0, 192(%rsi)7875; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7876; SSE-NEXT: movaps %xmm0, 128(%rsi)7877; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7878; SSE-NEXT: movaps %xmm0, 64(%rsi)7879; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7880; SSE-NEXT: movaps %xmm0, (%rsi)7881; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7882; SSE-NEXT: movaps %xmm0, 208(%rsi)7883; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7884; SSE-NEXT: movaps %xmm0, 144(%rsi)7885; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7886; SSE-NEXT: movaps %xmm0, 80(%rsi)7887; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7888; SSE-NEXT: movaps %xmm0, 16(%rsi)7889; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7890; SSE-NEXT: movaps %xmm0, 224(%rdx)7891; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7892; SSE-NEXT: movaps %xmm0, 240(%rdx)7893; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7894; SSE-NEXT: movaps %xmm0, 192(%rdx)7895; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7896; SSE-NEXT: movaps %xmm0, 208(%rdx)7897; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7898; SSE-NEXT: movaps %xmm0, 160(%rdx)7899; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7900; SSE-NEXT: movaps %xmm0, 176(%rdx)7901; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7902; SSE-NEXT: movaps %xmm0, 128(%rdx)7903; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7904; SSE-NEXT: movaps %xmm0, 144(%rdx)7905; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7906; SSE-NEXT: movaps %xmm0, 96(%rdx)7907; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7908; SSE-NEXT: movaps %xmm0, 112(%rdx)7909; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7910; SSE-NEXT: movaps %xmm0, 64(%rdx)7911; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7912; SSE-NEXT: movaps %xmm0, 80(%rdx)7913; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7914; SSE-NEXT: movaps %xmm0, 32(%rdx)7915; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7916; SSE-NEXT: movaps %xmm0, 48(%rdx)7917; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7918; SSE-NEXT: movaps %xmm0, (%rdx)7919; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7920; SSE-NEXT: movaps %xmm0, 16(%rdx)7921; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7922; SSE-NEXT: movaps %xmm0, 240(%rcx)7923; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7924; SSE-NEXT: movaps %xmm0, 224(%rcx)7925; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7926; SSE-NEXT: movaps %xmm0, 208(%rcx)7927; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7928; SSE-NEXT: movaps %xmm0, 192(%rcx)7929; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7930; SSE-NEXT: movaps %xmm0, 176(%rcx)7931; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7932; SSE-NEXT: movaps %xmm0, 160(%rcx)7933; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7934; SSE-NEXT: movaps %xmm0, 144(%rcx)7935; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7936; SSE-NEXT: movaps %xmm0, 128(%rcx)7937; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7938; SSE-NEXT: movaps %xmm0, 112(%rcx)7939; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7940; SSE-NEXT: movaps %xmm0, 96(%rcx)7941; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7942; SSE-NEXT: movaps %xmm0, 80(%rcx)7943; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7944; SSE-NEXT: movaps %xmm0, 64(%rcx)7945; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7946; SSE-NEXT: movaps %xmm0, 48(%rcx)7947; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7948; SSE-NEXT: movaps %xmm0, 32(%rcx)7949; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7950; SSE-NEXT: movaps %xmm0, 16(%rcx)7951; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7952; SSE-NEXT: movaps %xmm0, (%rcx)7953; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7954; SSE-NEXT: movaps %xmm0, 240(%r8)7955; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7956; SSE-NEXT: movaps %xmm0, 224(%r8)7957; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7958; SSE-NEXT: movaps %xmm0, 208(%r8)7959; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7960; SSE-NEXT: movaps %xmm0, 192(%r8)7961; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7962; SSE-NEXT: movaps %xmm0, 176(%r8)7963; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7964; SSE-NEXT: movaps %xmm0, 160(%r8)7965; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7966; SSE-NEXT: movaps %xmm0, 144(%r8)7967; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7968; SSE-NEXT: movaps %xmm0, 128(%r8)7969; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7970; SSE-NEXT: movaps %xmm0, 112(%r8)7971; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7972; SSE-NEXT: movaps %xmm0, 96(%r8)7973; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7974; SSE-NEXT: movaps %xmm0, 80(%r8)7975; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7976; SSE-NEXT: movaps %xmm0, 64(%r8)7977; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7978; SSE-NEXT: movaps %xmm0, 48(%r8)7979; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7980; SSE-NEXT: movaps %xmm0, 32(%r8)7981; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload7982; SSE-NEXT: movaps %xmm0, 16(%r8)7983; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7984; SSE-NEXT: movaps %xmm0, (%r8)7985; SSE-NEXT: movapd %xmm2, 240(%r9)7986; SSE-NEXT: movapd %xmm3, 224(%r9)7987; SSE-NEXT: movapd %xmm5, 208(%r9)7988; SSE-NEXT: movapd %xmm7, 192(%r9)7989; SSE-NEXT: movapd %xmm13, 176(%r9)7990; SSE-NEXT: movapd %xmm8, 160(%r9)7991; SSE-NEXT: movapd %xmm9, 144(%r9)7992; SSE-NEXT: movapd %xmm11, 128(%r9)7993; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7994; SSE-NEXT: movaps %xmm0, 112(%r9)7995; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7996; SSE-NEXT: movaps %xmm0, 96(%r9)7997; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7998; SSE-NEXT: movaps %xmm0, 80(%r9)7999; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8000; SSE-NEXT: movaps %xmm0, 64(%r9)8001; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8002; SSE-NEXT: movaps %xmm0, 48(%r9)8003; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8004; SSE-NEXT: movaps %xmm0, 32(%r9)8005; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8006; SSE-NEXT: movaps %xmm0, 16(%r9)8007; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8008; SSE-NEXT: movaps %xmm0, (%r9)8009; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax8010; SSE-NEXT: movapd %xmm14, 240(%rax)8011; SSE-NEXT: movapd %xmm12, 224(%rax)8012; SSE-NEXT: movapd %xmm15, 208(%rax)8013; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8014; SSE-NEXT: movaps %xmm0, 192(%rax)8015; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8016; SSE-NEXT: movaps %xmm0, 176(%rax)8017; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8018; SSE-NEXT: movaps %xmm0, 160(%rax)8019; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8020; SSE-NEXT: movaps %xmm0, 144(%rax)8021; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8022; SSE-NEXT: movaps %xmm0, 128(%rax)8023; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8024; SSE-NEXT: movaps %xmm0, 112(%rax)8025; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8026; SSE-NEXT: movaps %xmm0, 96(%rax)8027; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8028; SSE-NEXT: movaps %xmm0, 80(%rax)8029; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8030; SSE-NEXT: movaps %xmm0, 64(%rax)8031; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8032; SSE-NEXT: movaps %xmm0, 48(%rax)8033; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8034; SSE-NEXT: movaps %xmm0, 32(%rax)8035; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload8036; SSE-NEXT: movaps %xmm0, 16(%rax)8037; SSE-NEXT: movapd %xmm4, (%rax)8038; SSE-NEXT: addq $2184, %rsp # imm = 0x8888039; SSE-NEXT: retq8040;8041; AVX-LABEL: load_i32_stride6_vf64:8042; AVX: # %bb.0:8043; AVX-NEXT: subq $2616, %rsp # imm = 0xA388044; AVX-NEXT: vmovaps 608(%rdi), %ymm68045; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8046; AVX-NEXT: vmovaps 576(%rdi), %ymm78047; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8048; AVX-NEXT: vmovapd 352(%rdi), %ymm48049; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8050; AVX-NEXT: vmovapd 320(%rdi), %ymm58051; AVX-NEXT: vmovaps 288(%rdi), %ymm28052; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8053; AVX-NEXT: vmovaps 256(%rdi), %ymm38054; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8055; AVX-NEXT: vmovaps 224(%rdi), %ymm08056; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8057; AVX-NEXT: vmovaps 192(%rdi), %ymm18058; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8059; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]8060; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8061; AVX-NEXT: vextractf128 $1, %ymm0, %xmm138062; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm13[2,3]8063; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm13[0,3]8064; AVX-NEXT: vinsertf128 $1, 288(%rdi), %ymm3, %ymm18065; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8066; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[0,0],ymm2[6,4],ymm1[4,4]8067; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm3[2,2],ymm1[6,4],ymm3[6,6]8068; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8069; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8070; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm5[2,3],ymm4[0,1]8071; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8072; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[3],ymm5[2]8073; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8074; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8075; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8076; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm6[4,5],ymm7[6,7]8077; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8078; AVX-NEXT: vextractf128 $1, %ymm0, %xmm118079; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm11[2,3]8080; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm11[0,3]8081; AVX-NEXT: vmovaps 672(%rdi), %ymm28082; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8083; AVX-NEXT: vmovaps 640(%rdi), %ymm38084; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8085; AVX-NEXT: vinsertf128 $1, 672(%rdi), %ymm3, %ymm18086; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8087; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[0,0],ymm2[6,4],ymm1[4,4]8088; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm3[2,2],ymm1[6,4],ymm3[6,6]8089; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8090; AVX-NEXT: vmovapd 736(%rdi), %ymm18091; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8092; AVX-NEXT: vmovapd 704(%rdi), %ymm28093; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8094; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[0,1]8095; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8096; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[3],ymm2[2]8097; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8098; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8099; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8100; AVX-NEXT: vmovaps 992(%rdi), %ymm08101; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8102; AVX-NEXT: vmovaps 960(%rdi), %ymm18103; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8104; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]8105; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8106; AVX-NEXT: vextractf128 $1, %ymm0, %xmm88107; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm8[2,3]8108; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm8[0,3]8109; AVX-NEXT: vmovaps 1056(%rdi), %ymm28110; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8111; AVX-NEXT: vmovaps 1024(%rdi), %ymm128112; AVX-NEXT: vinsertf128 $1, 1056(%rdi), %ymm12, %ymm18113; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8114; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[0,0],ymm2[6,4],ymm1[4,4]8115; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm12[2,2],ymm1[6,4],ymm12[6,6]8116; AVX-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8117; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8118; AVX-NEXT: vmovapd 1120(%rdi), %ymm18119; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8120; AVX-NEXT: vmovapd 1088(%rdi), %ymm28121; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8122; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[0,1]8123; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8124; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[3],ymm2[2]8125; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8126; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8127; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8128; AVX-NEXT: vmovaps 1376(%rdi), %ymm08129; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8130; AVX-NEXT: vmovaps 1344(%rdi), %ymm18131; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8132; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]8133; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8134; AVX-NEXT: vextractf128 $1, %ymm0, %xmm68135; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm6[2,3]8136; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm6[0,3]8137; AVX-NEXT: vmovaps 1440(%rdi), %ymm28138; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8139; AVX-NEXT: vmovaps 1408(%rdi), %ymm38140; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8141; AVX-NEXT: vinsertf128 $1, 1440(%rdi), %ymm3, %ymm18142; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8143; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[0,0],ymm2[6,4],ymm1[4,4]8144; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm3[2,2],ymm1[6,4],ymm3[6,6]8145; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8146; AVX-NEXT: vmovapd 1504(%rdi), %ymm18147; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8148; AVX-NEXT: vmovapd 1472(%rdi), %ymm28149; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8150; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[0,1]8151; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8152; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[3],ymm2[2]8153; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8154; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8155; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8156; AVX-NEXT: vmovaps 32(%rdi), %ymm08157; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8158; AVX-NEXT: vmovaps (%rdi), %ymm18159; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8160; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]8161; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8162; AVX-NEXT: vextractf128 $1, %ymm0, %xmm48163; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]8164; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,3]8165; AVX-NEXT: vmovaps 96(%rdi), %ymm28166; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8167; AVX-NEXT: vmovaps 64(%rdi), %ymm38168; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8169; AVX-NEXT: vinsertf128 $1, 96(%rdi), %ymm3, %ymm18170; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8171; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm1[0,0],ymm2[6,4],ymm1[4,4]8172; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm3[2,2],ymm1[6,4],ymm3[6,6]8173; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8174; AVX-NEXT: vmovapd 160(%rdi), %ymm18175; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8176; AVX-NEXT: vmovapd 128(%rdi), %ymm28177; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8178; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[0,1]8179; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8180; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[3],ymm2[2]8181; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8182; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8183; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8184; AVX-NEXT: vmovaps 416(%rdi), %ymm08185; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8186; AVX-NEXT: vmovaps 384(%rdi), %ymm18187; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8188; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]8189; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8190; AVX-NEXT: vextractf128 $1, %ymm0, %xmm28191; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]8192; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,3]8193; AVX-NEXT: vmovaps 480(%rdi), %ymm18194; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8195; AVX-NEXT: vmovaps 448(%rdi), %ymm38196; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8197; AVX-NEXT: vinsertf128 $1, 480(%rdi), %ymm3, %ymm58198; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8199; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm5[0,0],ymm1[6,4],ymm5[4,4]8200; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm3[2,2],ymm1[6,4],ymm3[6,6]8201; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8202; AVX-NEXT: vmovapd 544(%rdi), %ymm18203; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8204; AVX-NEXT: vmovapd 512(%rdi), %ymm38205; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8206; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[0,1]8207; AVX-NEXT: vmovupd %ymm1, (%rsp) # 32-byte Spill8208; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[3],ymm3[2]8209; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8210; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8211; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8212; AVX-NEXT: vmovaps 800(%rdi), %ymm08213; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8214; AVX-NEXT: vmovaps 768(%rdi), %ymm18215; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8216; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]8217; AVX-NEXT: vextractf128 $1, %ymm10, %xmm18218; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm10[0,1],xmm1[2,3]8219; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,3]8220; AVX-NEXT: vmovaps 864(%rdi), %ymm38221; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8222; AVX-NEXT: vmovaps 832(%rdi), %ymm58223; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8224; AVX-NEXT: vinsertf128 $1, 864(%rdi), %ymm5, %ymm78225; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8226; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[2,0],ymm7[0,0],ymm3[6,4],ymm7[4,4]8227; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[2,0],ymm5[2,2],ymm3[6,4],ymm5[6,6]8228; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3,4,5,6,7]8229; AVX-NEXT: vmovapd 928(%rdi), %ymm38230; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8231; AVX-NEXT: vmovapd 896(%rdi), %ymm58232; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8233; AVX-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[0,1]8234; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8235; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[3],ymm5[2]8236; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,1,2,0,4,5,6,4]8237; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]8238; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8239; AVX-NEXT: vmovaps 1184(%rdi), %ymm08240; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8241; AVX-NEXT: vmovaps 1152(%rdi), %ymm38242; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8243; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm3[0,1,2,3],ymm0[4,5],ymm3[6,7]8244; AVX-NEXT: vextractf128 $1, %ymm5, %xmm08245; AVX-NEXT: vblendps {{.*#+}} xmm9 = xmm5[0,1],xmm0[2,3]8246; AVX-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,2],xmm0[0,3]8247; AVX-NEXT: vmovaps 1248(%rdi), %ymm38248; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8249; AVX-NEXT: vmovaps 1216(%rdi), %ymm78250; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8251; AVX-NEXT: vinsertf128 $1, 1248(%rdi), %ymm7, %ymm148252; AVX-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8253; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm3[2,0],ymm14[0,0],ymm3[6,4],ymm14[4,4]8254; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm15[2,0],ymm7[2,2],ymm15[6,4],ymm7[6,6]8255; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm9[0,1,2],ymm15[3,4,5,6,7]8256; AVX-NEXT: vmovapd 1312(%rdi), %ymm38257; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8258; AVX-NEXT: vmovapd 1280(%rdi), %ymm78259; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8260; AVX-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm7[2,3],ymm3[0,1]8261; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8262; AVX-NEXT: vshufpd {{.*#+}} ymm14 = ymm3[0],ymm7[1],ymm3[3],ymm7[2]8263; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm14[0,1,2,0,4,5,6,4]8264; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm15[0,1,2,3,4,5],ymm14[6,7]8265; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8266; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload8267; AVX-NEXT: vshufps {{.*#+}} xmm14 = xmm3[1,0],xmm13[3,0]8268; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm14[0,2],xmm13[1,3]8269; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload8270; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm14 # 32-byte Folded Reload8271; AVX-NEXT: # ymm14 = ymm3[3,0],mem[1,0],ymm3[7,4],mem[5,4]8272; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload8273; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm14[2,0],ymm3[2,3],ymm14[6,4],ymm3[6,7]8274; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2],ymm14[3,4,5,6,7]8275; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload8276; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm14 # 32-byte Folded Reload8277; AVX-NEXT: # ymm14 = ymm7[3,1],mem[1,3],ymm7[7,5],mem[5,7]8278; AVX-NEXT: vshufps {{.*#+}} ymm14 = ymm14[0,1,2,0,4,5,6,4]8279; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm13[0,1,2,3,4,5],ymm14[6,7]8280; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8281; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload8282; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm7[1,0],xmm11[3,0]8283; AVX-NEXT: vshufps {{.*#+}} xmm11 = xmm13[0,2],xmm11[1,3]8284; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8285; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm13 # 32-byte Folded Reload8286; AVX-NEXT: # ymm13 = ymm14[3,0],mem[1,0],ymm14[7,4],mem[5,4]8287; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload8288; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[2,0],ymm7[2,3],ymm13[6,4],ymm7[6,7]8289; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm13[3,4,5,6,7]8290; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload8291; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8292; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[3,1],ymm9[1,3],ymm13[7,5],ymm9[5,7]8293; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[0,1,2,0,4,5,6,4]8294; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],ymm13[6,7]8295; AVX-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8296; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload8297; AVX-NEXT: vshufps {{.*#+}} xmm11 = xmm11[1,0],xmm8[3,0]8298; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm11[0,2],xmm8[1,3]8299; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8300; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm11 # 32-byte Folded Reload8301; AVX-NEXT: # ymm11 = ymm13[3,0],mem[1,0],ymm13[7,4],mem[5,4]8302; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[2,0],ymm12[2,3],ymm11[6,4],ymm12[6,7]8303; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2],ymm11[3,4,5,6,7]8304; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload8305; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload8306; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[3,1],ymm12[1,3],ymm11[7,5],ymm12[5,7]8307; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[0,1,2,0,4,5,6,4]8308; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm11[6,7]8309; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8310; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload8311; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm8[1,0],xmm6[3,0]8312; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm8[0,2],xmm6[1,3]8313; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload8314; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm8 # 32-byte Folded Reload8315; AVX-NEXT: # ymm8 = ymm11[3,0],mem[1,0],ymm11[7,4],mem[5,4]8316; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload8317; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[2,0],ymm15[2,3],ymm8[6,4],ymm15[6,7]8318; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm8[3,4,5,6,7]8319; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload8320; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload8321; AVX-NEXT: # ymm8 = ymm8[3,1],mem[1,3],ymm8[7,5],mem[5,7]8322; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,1,2,0,4,5,6,4]8323; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm8[6,7]8324; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8325; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8326; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm6[1,0],xmm4[3,0]8327; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm6[0,2],xmm4[1,3]8328; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload8329; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm6 # 32-byte Folded Reload8330; AVX-NEXT: # ymm6 = ymm8[3,0],mem[1,0],ymm8[7,4],mem[5,4]8331; AVX-NEXT: vshufps $226, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload8332; AVX-NEXT: # ymm6 = ymm6[2,0],mem[2,3],ymm6[6,4],mem[6,7]8333; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm6[3,4,5,6,7]8334; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8335; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload8336; AVX-NEXT: # ymm6 = ymm6[3,1],mem[1,3],ymm6[7,5],mem[5,7]8337; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm6[0,1,2,0,4,5,6,4]8338; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm6[6,7]8339; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8340; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload8341; AVX-NEXT: vshufps {{.*#+}} xmm4 = xmm4[1,0],xmm2[3,0]8342; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm4[0,2],xmm2[1,3]8343; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload8344; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload8345; AVX-NEXT: # ymm4 = ymm4[3,0],mem[1,0],ymm4[7,4],mem[5,4]8346; AVX-NEXT: vshufps $226, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload8347; AVX-NEXT: # ymm4 = ymm4[2,0],mem[2,3],ymm4[6,4],mem[6,7]8348; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3,4,5,6,7]8349; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8350; AVX-NEXT: vmovups (%rsp), %ymm4 # 32-byte Reload8351; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,1],ymm6[1,3],ymm4[7,5],ymm6[5,7]8352; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,2,0,4,5,6,4]8353; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]8354; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8355; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm10[1,0],xmm1[3,0]8356; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,2],xmm1[1,3]8357; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload8358; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload8359; AVX-NEXT: # ymm2 = ymm2[3,0],mem[1,0],ymm2[7,4],mem[5,4]8360; AVX-NEXT: vshufps $226, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload8361; AVX-NEXT: # ymm2 = ymm2[2,0],mem[2,3],ymm2[6,4],mem[6,7]8362; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]8363; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload8364; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload8365; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,1],ymm4[1,3],ymm2[7,5],ymm4[5,7]8366; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,2,0,4,5,6,4]8367; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]8368; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8369; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm5[1,0],xmm0[3,0]8370; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm1[0,2],xmm0[1,3]8371; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8372; AVX-NEXT: vshufps $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload8373; AVX-NEXT: # ymm1 = ymm1[3,0],mem[1,0],ymm1[7,4],mem[5,4]8374; AVX-NEXT: vshufps $226, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload8375; AVX-NEXT: # ymm1 = ymm1[2,0],mem[2,3],ymm1[6,4],mem[6,7]8376; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6,7]8377; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8378; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload8379; AVX-NEXT: # ymm1 = ymm1[3,1],mem[1,3],ymm1[7,5],mem[5,7]8380; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,1,2,0,4,5,6,4]8381; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]8382; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8383; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8384; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8385; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8386; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8387; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8388; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,1],ymm3[2,0],ymm0[6,5],ymm3[6,4]8389; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8390; AVX-NEXT: vextractf128 $1, %ymm2, %xmm18391; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8392; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]8393; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8394; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8395; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload8396; AVX-NEXT: # ymm2 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]8397; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8398; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3,0,1]8399; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8400; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0],ymm2[2,0],ymm1[4,4],ymm2[6,4]8401; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8402; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8403; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8404; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8405; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8406; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8407; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm14[2,1],ymm7[2,0],ymm14[6,5],ymm7[6,4]8408; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8409; AVX-NEXT: vextractf128 $1, %ymm2, %xmm18410; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8411; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]8412; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8413; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm2 # 32-byte Folded Reload8414; AVX-NEXT: # ymm2 = ymm9[0,1,2,3],mem[4,5],ymm9[6,7]8415; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8416; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3,0,1]8417; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8418; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0],ymm2[2,0],ymm1[4,4],ymm2[6,4]8419; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8420; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8421; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8422; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8423; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8424; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8425; AVX-NEXT: vshufps $38, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload8426; AVX-NEXT: # ymm0 = ymm13[2,1],mem[2,0],ymm13[6,5],mem[6,4]8427; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8428; AVX-NEXT: vextractf128 $1, %ymm2, %xmm18429; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8430; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]8431; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8432; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm2 # 32-byte Folded Reload8433; AVX-NEXT: # ymm2 = ymm12[0,1,2,3],mem[4,5],ymm12[6,7]8434; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8435; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3,0,1]8436; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8437; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0],ymm2[2,0],ymm1[4,4],ymm2[6,4]8438; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8439; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8440; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8441; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8442; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8443; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8444; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm11[2,1],ymm15[2,0],ymm11[6,5],ymm15[6,4]8445; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8446; AVX-NEXT: vextractf128 $1, %ymm2, %xmm18447; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8448; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]8449; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8450; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8451; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm10 # 32-byte Folded Reload8452; AVX-NEXT: # ymm10 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]8453; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm10[2,3,0,1]8454; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8455; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0],ymm10[2,0],ymm1[4,4],ymm10[6,4]8456; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8457; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8458; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8459; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8460; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8461; AVX-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill8462; AVX-NEXT: vshufps $38, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm0 # 32-byte Folded Reload8463; AVX-NEXT: # ymm0 = ymm8[2,1],mem[2,0],ymm8[6,5],mem[6,4]8464; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8465; AVX-NEXT: vextractf128 $1, %ymm2, %xmm18466; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8467; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]8468; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8469; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8470; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm2 # 32-byte Folded Reload8471; AVX-NEXT: # ymm2 = ymm1[0,1,2,3],mem[4,5],ymm1[6,7]8472; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8473; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3,0,1]8474; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8475; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0],ymm2[2,0],ymm1[4,4],ymm2[6,4]8476; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8477; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8478; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8479; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8480; AVX-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8481; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8482; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8483; AVX-NEXT: vshufps $38, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8484; AVX-NEXT: # ymm0 = ymm0[2,1],mem[2,0],ymm0[6,5],mem[6,4]8485; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8486; AVX-NEXT: vextractf128 $1, %ymm2, %xmm18487; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8488; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]8489; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8490; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm9 # 32-byte Folded Reload8491; AVX-NEXT: # ymm9 = ymm6[0,1,2,3],mem[4,5],ymm6[6,7]8492; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm9[2,3,0,1]8493; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8494; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0],ymm9[2,0],ymm1[4,4],ymm9[6,4]8495; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8496; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8497; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8498; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload8499; AVX-NEXT: # ymm8 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8500; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8501; AVX-NEXT: vshufps $38, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8502; AVX-NEXT: # ymm0 = ymm0[2,1],mem[2,0],ymm0[6,5],mem[6,4]8503; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]8504; AVX-NEXT: vextractf128 $1, %ymm8, %xmm18505; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8506; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm8[2,0],xmm1[2,3]8507; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]8508; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm6 # 32-byte Folded Reload8509; AVX-NEXT: # ymm6 = ymm4[0,1,2,3],mem[4,5],ymm4[6,7]8510; AVX-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm6[2,3,0,1]8511; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm13[0,0],ymm6[2,0],ymm13[4,4],ymm6[6,4]8512; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]8513; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8514; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8515; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload8516; AVX-NEXT: # ymm5 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]8517; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload8518; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload8519; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm4[2,1],ymm3[2,0],ymm4[6,5],ymm3[6,4]8520; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[2,3,0,1]8521; AVX-NEXT: vextractf128 $1, %ymm5, %xmm78522; AVX-NEXT: vshufps {{.*#+}} xmm15 = xmm5[2,0],xmm7[2,3]8523; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2],ymm0[3,4,5,6,7]8524; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload8525; AVX-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8526; AVX-NEXT: # ymm15 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]8527; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm15[2,3,0,1]8528; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm2[0,0],ymm15[2,0],ymm2[4,4],ymm15[6,4]8529; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm11[5,6,7]8530; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8531; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8532; AVX-NEXT: vshufps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8533; AVX-NEXT: # ymm0 = ymm0[0,1],mem[3,1],ymm0[4,5],mem[7,5]8534; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8535; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm11 # 16-byte Folded Reload8536; AVX-NEXT: # xmm11 = xmm1[3,1],mem[3,3]8537; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8538; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8539; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm1[3,1],ymm14[2,1],ymm1[7,5],ymm14[6,5]8540; AVX-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3,0,1]8541; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm12[3,4,5,6,7]8542; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4],ymm0[5,6,7]8543; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8544; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8545; AVX-NEXT: vshufps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8546; AVX-NEXT: # ymm0 = ymm0[0,1],mem[3,1],ymm0[4,5],mem[7,5]8547; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8548; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm11 # 16-byte Folded Reload8549; AVX-NEXT: # xmm11 = xmm1[3,1],mem[3,3]8550; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload8551; AVX-NEXT: vshufps $103, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload8552; AVX-NEXT: # ymm12 = ymm12[3,1],mem[2,1],ymm12[7,5],mem[6,5]8553; AVX-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3,0,1]8554; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm12[3,4,5,6,7]8555; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4],ymm0[5,6,7]8556; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8557; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8558; AVX-NEXT: vshufps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload8559; AVX-NEXT: # ymm1 = ymm0[0,1],mem[3,1],ymm0[4,5],mem[7,5]8560; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8561; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm11 # 16-byte Folded Reload8562; AVX-NEXT: # xmm11 = xmm0[3,1],mem[3,3]8563; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload8564; AVX-NEXT: vshufps $103, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload8565; AVX-NEXT: # ymm12 = ymm12[3,1],mem[2,1],ymm12[7,5],mem[6,5]8566; AVX-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3,0,1]8567; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2],ymm12[3,4,5,6,7]8568; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3,4],ymm1[5,6,7]8569; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8570; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8571; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm10[3,1],ymm0[4,5],ymm10[7,5]8572; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8573; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm10 # 16-byte Folded Reload8574; AVX-NEXT: # xmm10 = xmm1[3,1],mem[3,3]8575; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload8576; AVX-NEXT: vshufps $103, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload8577; AVX-NEXT: # ymm11 = ymm11[3,1],mem[2,1],ymm11[7,5],mem[6,5]8578; AVX-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm11[2,3,0,1]8579; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2],ymm11[3,4,5,6,7]8580; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4],ymm0[5,6,7]8581; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8582; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,1],ymm15[3,1],ymm2[4,5],ymm15[7,5]8583; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm5[3,1],xmm7[3,3]8584; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm4[3,1],ymm3[2,1],ymm4[7,5],ymm3[6,5]8585; AVX-NEXT: vmovaps %ymm4, %ymm158586; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,0,1]8587; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]8588; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]8589; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8590; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm13[0,1],ymm6[3,1],ymm13[4,5],ymm6[7,5]8591; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm1 # 16-byte Folded Reload8592; AVX-NEXT: # xmm1 = xmm8[3,1],mem[3,3]8593; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload8594; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8595; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,1],ymm6[2,1],ymm3[7,5],ymm6[6,5]8596; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,0,1]8597; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]8598; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]8599; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8600; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8601; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm9[3,1],ymm0[4,5],ymm9[7,5]8602; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8603; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload8604; AVX-NEXT: # xmm1 = xmm1[3,1],mem[3,3]8605; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload8606; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload8607; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm4[3,1],ymm8[2,1],ymm4[7,5],ymm8[6,5]8608; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,0,1]8609; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]8610; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]8611; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8612; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8613; AVX-NEXT: vshufps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8614; AVX-NEXT: # ymm0 = ymm0[0,1],mem[3,1],ymm0[4,5],mem[7,5]8615; AVX-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload8616; AVX-NEXT: vshufps $247, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload8617; AVX-NEXT: # xmm1 = xmm1[3,1],mem[3,3]8618; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload8619; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload8620; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm5[3,1],ymm7[2,1],ymm5[7,5],ymm7[6,5]8621; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,0,1]8622; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4,5,6,7]8623; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]8624; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8625; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8626; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8627; AVX-NEXT: # ymm2 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8628; AVX-NEXT: vmovaps 32(%rdi), %xmm08629; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8630; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8631; AVX-NEXT: vmovaps 16(%rdi), %xmm18632; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8633; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8634; AVX-NEXT: vmovapd 80(%rdi), %xmm18635; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8636; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1],ymm7[0],ymm1[2],ymm7[3]8637; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,0],ymm5[4,5],ymm1[6,4]8638; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8639; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8640; AVX-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm2[2,3,0,1]8641; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm11[0,0],ymm2[6,4],ymm11[4,4]8642; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm11[0,2],ymm1[2,0],ymm11[4,6],ymm1[6,4]8643; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8644; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8645; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8646; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8647; AVX-NEXT: # ymm2 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8648; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8649; AVX-NEXT: vmovaps 224(%rdi), %xmm08650; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8651; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8652; AVX-NEXT: vmovaps 208(%rdi), %xmm18653; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8654; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8655; AVX-NEXT: vmovapd 272(%rdi), %xmm18656; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8657; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1],ymm14[0],ymm1[2],ymm14[3]8658; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload8659; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,0],ymm5[4,5],ymm1[6,4]8660; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8661; AVX-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm2[2,3,0,1]8662; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm7[0,0],ymm2[6,4],ymm7[4,4]8663; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm7[0,2],ymm1[2,0],ymm7[4,6],ymm1[6,4]8664; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8665; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8666; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8667; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload8668; AVX-NEXT: # ymm2 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8669; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8670; AVX-NEXT: vmovaps 416(%rdi), %xmm08671; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8672; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8673; AVX-NEXT: vmovaps 400(%rdi), %xmm18674; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8675; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8676; AVX-NEXT: vmovapd 464(%rdi), %xmm18677; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8678; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1],ymm8[0],ymm1[2],ymm8[3]8679; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,1],ymm1[2,0],ymm4[4,5],ymm1[6,4]8680; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8681; AVX-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm2[2,3,0,1]8682; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,0],ymm5[0,0],ymm2[6,4],ymm5[4,4]8683; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm5[0,2],ymm1[2,0],ymm5[4,6],ymm1[6,4]8684; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8685; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8686; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8687; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload8688; AVX-NEXT: # ymm12 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8689; AVX-NEXT: vmovaps 608(%rdi), %xmm08690; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8691; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8692; AVX-NEXT: vmovaps 592(%rdi), %xmm18693; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8694; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8695; AVX-NEXT: vmovapd 656(%rdi), %xmm18696; AVX-NEXT: vmovupd %ymm1, (%rsp) # 32-byte Spill8697; AVX-NEXT: vshufpd $9, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload8698; AVX-NEXT: # ymm1 = ymm1[1],mem[0],ymm1[2],mem[3]8699; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload8700; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,0],ymm2[4,5],ymm1[6,4]8701; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8702; AVX-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm12[2,3,0,1]8703; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm12[2,0],ymm4[0,0],ymm12[6,4],ymm4[4,4]8704; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm4[0,2],ymm1[2,0],ymm4[4,6],ymm1[6,4]8705; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8706; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8707; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8708; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload8709; AVX-NEXT: # ymm10 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8710; AVX-NEXT: vmovaps 800(%rdi), %xmm08711; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8712; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8713; AVX-NEXT: vmovaps 784(%rdi), %xmm18714; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8715; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8716; AVX-NEXT: vmovapd 848(%rdi), %xmm18717; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8718; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1],ymm6[0],ymm1[2],ymm6[3]8719; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,0],ymm3[4,5],ymm1[6,4]8720; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8721; AVX-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm10[2,3,0,1]8722; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm10[2,0],ymm3[0,0],ymm10[6,4],ymm3[4,4]8723; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm3[0,2],ymm1[2,0],ymm3[4,6],ymm1[6,4]8724; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8725; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8726; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8727; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm9 # 32-byte Folded Reload8728; AVX-NEXT: # ymm9 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8729; AVX-NEXT: vmovaps 992(%rdi), %xmm08730; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8731; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8732; AVX-NEXT: vmovaps 976(%rdi), %xmm18733; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8734; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8735; AVX-NEXT: vmovapd 1040(%rdi), %xmm18736; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8737; AVX-NEXT: vshufpd $9, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload8738; AVX-NEXT: # ymm1 = ymm1[1],mem[0],ymm1[2],mem[3]8739; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload8740; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,0],ymm2[4,5],ymm1[6,4]8741; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8742; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm9[2,3,0,1]8743; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm9[2,0],ymm2[0,0],ymm9[6,4],ymm2[4,4]8744; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,2],ymm1[2,0],ymm2[4,6],ymm1[6,4]8745; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]8746; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8747; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8748; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload8749; AVX-NEXT: # ymm8 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8750; AVX-NEXT: vmovaps 1184(%rdi), %xmm08751; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8752; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8753; AVX-NEXT: vmovaps 1168(%rdi), %xmm18754; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8755; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]8756; AVX-NEXT: vmovapd 1232(%rdi), %xmm18757; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8758; AVX-NEXT: vshufpd $9, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload8759; AVX-NEXT: # ymm1 = ymm1[1],mem[0],ymm1[2],mem[3]8760; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm15[0,1],ymm1[2,0],ymm15[4,5],ymm1[6,4]8761; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4,5,6,7]8762; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm8[2,3,0,1]8763; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm8[2,0],ymm1[0,0],ymm8[6,4],ymm1[4,4]8764; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm1[0,2],ymm6[2,0],ymm1[4,6],ymm6[6,4]8765; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5,6,7]8766; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8767; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8768; AVX-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload8769; AVX-NEXT: # ymm6 = mem[0,1],ymm0[2,3],mem[4,5,6,7]8770; AVX-NEXT: vmovaps 1376(%rdi), %xmm08771; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8772; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,3,3]8773; AVX-NEXT: vmovaps 1360(%rdi), %xmm138774; AVX-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8775; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm13[0],xmm0[1],xmm13[2,3]8776; AVX-NEXT: vmovapd 1424(%rdi), %xmm138777; AVX-NEXT: vmovupd %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8778; AVX-NEXT: vshufpd $9, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm15 # 32-byte Folded Reload8779; AVX-NEXT: # ymm15 = ymm13[1],mem[0],ymm13[2],mem[3]8780; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8781; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm13[0,1],ymm15[2,0],ymm13[4,5],ymm15[6,4]8782; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1],ymm15[2,3,4,5,6,7]8783; AVX-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm6[2,3,0,1]8784; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm6[2,0],ymm0[0,0],ymm6[6,4],ymm0[4,4]8785; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm0[0,2],ymm13[2,0],ymm0[4,6],ymm13[6,4]8786; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0,1,2,3,4],ymm13[5,6,7]8787; AVX-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8788; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8789; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[3,0],ymm11[1,0],ymm13[7,4],ymm11[5,4]8790; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[0,3],ymm13[2,0],ymm11[4,7],ymm13[6,4]8791; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload8792; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload8793; AVX-NEXT: # xmm13 = xmm13[0,1],mem[2,3]8794; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload8795; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload8796; AVX-NEXT: # ymm15 = ymm15[3,1],mem[1,3],ymm15[7,5],mem[5,7]8797; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8798; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm14[1,1],ymm15[2,0],ymm14[5,5],ymm15[6,4]8799; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm13[1,3,2,3]8800; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2,3,4,5,6,7]8801; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm13[0,1,2,3,4],ymm11[5,6,7]8802; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8803; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[3,0],ymm7[1,0],ymm13[7,4],ymm7[5,4]8804; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm7[0,3],ymm13[2,0],ymm7[4,7],ymm13[6,4]8805; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload8806; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload8807; AVX-NEXT: # xmm13 = xmm13[0,1],mem[2,3]8808; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8809; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm15 # 32-byte Folded Reload8810; AVX-NEXT: # ymm15 = ymm14[3,1],mem[1,3],ymm14[7,5],mem[5,7]8811; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8812; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm14[1,1],ymm15[2,0],ymm14[5,5],ymm15[6,4]8813; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm13[1,3,2,3]8814; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2,3,4,5,6,7]8815; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm13[0,1,2,3,4],ymm7[5,6,7]8816; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8817; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm13[3,0],ymm5[1,0],ymm13[7,4],ymm5[5,4]8818; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm5[0,3],ymm13[2,0],ymm5[4,7],ymm13[6,4]8819; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload8820; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload8821; AVX-NEXT: # xmm13 = xmm13[0,1],mem[2,3]8822; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8823; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm15 # 32-byte Folded Reload8824; AVX-NEXT: # ymm15 = ymm14[3,1],mem[1,3],ymm14[7,5],mem[5,7]8825; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload8826; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm14[1,1],ymm15[2,0],ymm14[5,5],ymm15[6,4]8827; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm13[1,3,2,3]8828; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2,3,4,5,6,7]8829; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm13[0,1,2,3,4],ymm5[5,6,7]8830; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm12[3,0],ymm4[1,0],ymm12[7,4],ymm4[5,4]8831; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,3],ymm12[2,0],ymm4[4,7],ymm12[6,4]8832; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload8833; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload8834; AVX-NEXT: # xmm12 = xmm12[0,1],mem[2,3]8835; AVX-NEXT: vmovups (%rsp), %ymm13 # 32-byte Reload8836; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload8837; AVX-NEXT: # ymm13 = ymm13[3,1],mem[1,3],ymm13[7,5],mem[5,7]8838; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload8839; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm15[1,1],ymm13[2,0],ymm15[5,5],ymm13[6,4]8840; AVX-NEXT: vshufps {{.*#+}} xmm12 = xmm12[1,3,2,3]8841; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1],ymm13[2,3,4,5,6,7]8842; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4],ymm4[5,6,7]8843; AVX-NEXT: vshufps {{.*#+}} ymm10 = ymm10[3,0],ymm3[1,0],ymm10[7,4],ymm3[5,4]8844; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,3],ymm10[2,0],ymm3[4,7],ymm10[6,4]8845; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload8846; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload8847; AVX-NEXT: # xmm10 = xmm10[0,1],mem[2,3]8848; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload8849; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload8850; AVX-NEXT: # ymm12 = ymm12[3,1],mem[1,3],ymm12[7,5],mem[5,7]8851; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload8852; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm13[1,1],ymm12[2,0],ymm13[5,5],ymm12[6,4]8853; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm10[1,3,2,3]8854; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm12[2,3,4,5,6,7]8855; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4],ymm3[5,6,7]8856; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm9[3,0],ymm2[1,0],ymm9[7,4],ymm2[5,4]8857; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,3],ymm9[2,0],ymm2[4,7],ymm9[6,4]8858; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload8859; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload8860; AVX-NEXT: # xmm9 = xmm9[0,1],mem[2,3]8861; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload8862; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload8863; AVX-NEXT: # ymm10 = ymm10[3,1],mem[1,3],ymm10[7,5],mem[5,7]8864; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload8865; AVX-NEXT: vshufps {{.*#+}} ymm10 = ymm12[1,1],ymm10[2,0],ymm12[5,5],ymm10[6,4]8866; AVX-NEXT: vshufps {{.*#+}} xmm9 = xmm9[1,3,2,3]8867; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1],ymm10[2,3,4,5,6,7]8868; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm9[0,1,2,3,4],ymm2[5,6,7]8869; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[3,0],ymm1[1,0],ymm8[7,4],ymm1[5,4]8870; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,3],ymm8[2,0],ymm1[4,7],ymm8[6,4]8871; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload8872; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload8873; AVX-NEXT: # xmm8 = xmm8[0,1],mem[2,3]8874; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload8875; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload8876; AVX-NEXT: # ymm9 = ymm9[3,1],mem[1,3],ymm9[7,5],mem[5,7]8877; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload8878; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm10[1,1],ymm9[2,0],ymm10[5,5],ymm9[6,4]8879; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm8[1,3,2,3]8880; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2,3,4,5,6,7]8881; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0,1,2,3,4],ymm1[5,6,7]8882; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm6[3,0],ymm0[1,0],ymm6[7,4],ymm0[5,4]8883; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,3],ymm6[2,0],ymm0[4,7],ymm6[6,4]8884; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload8885; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload8886; AVX-NEXT: # xmm6 = xmm6[0,1],mem[2,3]8887; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload8888; AVX-NEXT: vshufps $215, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload8889; AVX-NEXT: # ymm8 = ymm8[3,1],mem[1,3],ymm8[7,5],mem[5,7]8890; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload8891; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm9[1,1],ymm8[2,0],ymm9[5,5],ymm8[6,4]8892; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm6[1,3,2,3]8893; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm8[2,3,4,5,6,7]8894; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4],ymm0[5,6,7]8895; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8896; AVX-NEXT: vmovaps %ymm6, 192(%rsi)8897; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8898; AVX-NEXT: vmovaps %ymm6, 128(%rsi)8899; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8900; AVX-NEXT: vmovaps %ymm6, 64(%rsi)8901; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8902; AVX-NEXT: vmovaps %ymm6, (%rsi)8903; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8904; AVX-NEXT: vmovaps %ymm6, 224(%rsi)8905; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8906; AVX-NEXT: vmovaps %ymm6, 160(%rsi)8907; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8908; AVX-NEXT: vmovaps %ymm6, 96(%rsi)8909; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8910; AVX-NEXT: vmovaps %ymm6, 32(%rsi)8911; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8912; AVX-NEXT: vmovaps %ymm6, 192(%rdx)8913; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8914; AVX-NEXT: vmovaps %ymm6, 128(%rdx)8915; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8916; AVX-NEXT: vmovaps %ymm6, 64(%rdx)8917; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8918; AVX-NEXT: vmovaps %ymm6, (%rdx)8919; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8920; AVX-NEXT: vmovaps %ymm6, 224(%rdx)8921; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8922; AVX-NEXT: vmovaps %ymm6, 160(%rdx)8923; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8924; AVX-NEXT: vmovaps %ymm6, 96(%rdx)8925; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8926; AVX-NEXT: vmovaps %ymm6, 32(%rdx)8927; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8928; AVX-NEXT: vmovaps %ymm6, 192(%rcx)8929; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8930; AVX-NEXT: vmovaps %ymm6, 128(%rcx)8931; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8932; AVX-NEXT: vmovaps %ymm6, 64(%rcx)8933; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8934; AVX-NEXT: vmovaps %ymm6, (%rcx)8935; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8936; AVX-NEXT: vmovaps %ymm6, 224(%rcx)8937; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8938; AVX-NEXT: vmovaps %ymm6, 160(%rcx)8939; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8940; AVX-NEXT: vmovaps %ymm6, 96(%rcx)8941; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8942; AVX-NEXT: vmovaps %ymm6, 32(%rcx)8943; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8944; AVX-NEXT: vmovaps %ymm6, (%r8)8945; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8946; AVX-NEXT: vmovaps %ymm6, 64(%r8)8947; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8948; AVX-NEXT: vmovaps %ymm6, 128(%r8)8949; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8950; AVX-NEXT: vmovaps %ymm6, 192(%r8)8951; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8952; AVX-NEXT: vmovaps %ymm6, 224(%r8)8953; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8954; AVX-NEXT: vmovaps %ymm6, 160(%r8)8955; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8956; AVX-NEXT: vmovaps %ymm6, 96(%r8)8957; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8958; AVX-NEXT: vmovaps %ymm6, 32(%r8)8959; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8960; AVX-NEXT: vmovaps %ymm6, 224(%r9)8961; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8962; AVX-NEXT: vmovaps %ymm6, 192(%r9)8963; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8964; AVX-NEXT: vmovaps %ymm6, 160(%r9)8965; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8966; AVX-NEXT: vmovaps %ymm6, 128(%r9)8967; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8968; AVX-NEXT: vmovaps %ymm6, 96(%r9)8969; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8970; AVX-NEXT: vmovaps %ymm6, 64(%r9)8971; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8972; AVX-NEXT: vmovaps %ymm6, 32(%r9)8973; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload8974; AVX-NEXT: vmovaps %ymm6, (%r9)8975; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax8976; AVX-NEXT: vmovaps %ymm0, 224(%rax)8977; AVX-NEXT: vmovaps %ymm1, 192(%rax)8978; AVX-NEXT: vmovaps %ymm2, 160(%rax)8979; AVX-NEXT: vmovaps %ymm3, 128(%rax)8980; AVX-NEXT: vmovaps %ymm4, 96(%rax)8981; AVX-NEXT: vmovaps %ymm5, 64(%rax)8982; AVX-NEXT: vmovaps %ymm7, 32(%rax)8983; AVX-NEXT: vmovaps %ymm11, (%rax)8984; AVX-NEXT: addq $2616, %rsp # imm = 0xA388985; AVX-NEXT: vzeroupper8986; AVX-NEXT: retq8987;8988; AVX2-LABEL: load_i32_stride6_vf64:8989; AVX2: # %bb.0:8990; AVX2-NEXT: subq $2568, %rsp # imm = 0xA088991; AVX2-NEXT: vmovaps 672(%rdi), %ymm48992; AVX2-NEXT: vmovaps 640(%rdi), %ymm58993; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8994; AVX2-NEXT: vmovaps 608(%rdi), %ymm38995; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8996; AVX2-NEXT: vmovaps 320(%rdi), %ymm68997; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8998; AVX2-NEXT: vmovaps 352(%rdi), %ymm78999; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9000; AVX2-NEXT: vmovaps 288(%rdi), %ymm29001; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9002; AVX2-NEXT: vmovaps 256(%rdi), %ymm89003; AVX2-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9004; AVX2-NEXT: vmovaps 224(%rdi), %ymm09005; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9006; AVX2-NEXT: vmovaps 192(%rdi), %ymm19007; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9008; AVX2-NEXT: vmovaps {{.*#+}} xmm9 = [0,6,4,u]9009; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9010; AVX2-NEXT: vpermps %ymm14, %ymm9, %ymm09011; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm8[0,1],ymm2[0,1]9012; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2,3,4,5],ymm2[6,7]9013; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm15[0,2,2,2,4,6,6,6]9014; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9015; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm6[4,5,6,7]9016; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9017; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm6 = [4,2,4,2,4,2,4,2]9018; AVX2-NEXT: vpermps %ymm1, %ymm6, %ymm29019; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9020; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9021; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9022; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm5[0,1],ymm4[0,1]9023; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm4[6,7]9024; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9025; AVX2-NEXT: vmovaps 576(%rdi), %ymm09026; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9027; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5],ymm0[6,7]9028; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9029; AVX2-NEXT: vpermps %ymm0, %ymm9, %ymm09030; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm1[0,2,2,2,4,6,6,6]9031; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9032; AVX2-NEXT: vmovaps 704(%rdi), %ymm19033; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9034; AVX2-NEXT: vmovaps 736(%rdi), %ymm29035; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9036; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]9037; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9038; AVX2-NEXT: vpermps %ymm1, %ymm6, %ymm29039; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9040; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9041; AVX2-NEXT: vmovaps 1056(%rdi), %ymm19042; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9043; AVX2-NEXT: vmovaps 1024(%rdi), %ymm09044; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9045; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9046; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]9047; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9048; AVX2-NEXT: vmovaps 992(%rdi), %ymm09049; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9050; AVX2-NEXT: vmovaps 960(%rdi), %ymm19051; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9052; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9053; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9054; AVX2-NEXT: vpermps %ymm0, %ymm9, %ymm09055; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]9056; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9057; AVX2-NEXT: vmovaps 1088(%rdi), %ymm19058; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9059; AVX2-NEXT: vmovaps 1120(%rdi), %ymm29060; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9061; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]9062; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9063; AVX2-NEXT: vpermps %ymm1, %ymm6, %ymm29064; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9065; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9066; AVX2-NEXT: vmovaps 1440(%rdi), %ymm19067; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9068; AVX2-NEXT: vmovaps 1408(%rdi), %ymm09069; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9070; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9071; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]9072; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9073; AVX2-NEXT: vmovaps 1376(%rdi), %ymm09074; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9075; AVX2-NEXT: vmovaps 1344(%rdi), %ymm19076; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9077; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9078; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9079; AVX2-NEXT: vpermps %ymm0, %ymm9, %ymm09080; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]9081; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9082; AVX2-NEXT: vmovaps 1472(%rdi), %ymm19083; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9084; AVX2-NEXT: vmovaps 1504(%rdi), %ymm29085; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9086; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]9087; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9088; AVX2-NEXT: vpermps %ymm1, %ymm6, %ymm29089; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9090; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9091; AVX2-NEXT: vmovaps 96(%rdi), %ymm19092; AVX2-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill9093; AVX2-NEXT: vmovaps 64(%rdi), %ymm09094; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9095; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9096; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]9097; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9098; AVX2-NEXT: vmovaps (%rdi), %ymm09099; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9100; AVX2-NEXT: vmovaps 32(%rdi), %ymm19101; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9102; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]9103; AVX2-NEXT: vpermps %ymm13, %ymm9, %ymm09104; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]9105; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9106; AVX2-NEXT: vmovaps 128(%rdi), %ymm19107; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9108; AVX2-NEXT: vmovaps 160(%rdi), %ymm29109; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9110; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm2[0,1,2,3],ymm1[4,5,6,7]9111; AVX2-NEXT: vpermps %ymm12, %ymm6, %ymm29112; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9113; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9114; AVX2-NEXT: vmovaps 480(%rdi), %ymm19115; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9116; AVX2-NEXT: vmovaps 448(%rdi), %ymm09117; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9118; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9119; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3,4,5],ymm1[6,7]9120; AVX2-NEXT: vmovaps 416(%rdi), %ymm09121; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9122; AVX2-NEXT: vmovaps 384(%rdi), %ymm19123; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9124; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9125; AVX2-NEXT: vpermps %ymm10, %ymm9, %ymm09126; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm11[0,2,2,2,4,6,6,6]9127; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9128; AVX2-NEXT: vmovaps 512(%rdi), %ymm19129; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9130; AVX2-NEXT: vmovaps 544(%rdi), %ymm29131; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9132; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm2[0,1,2,3],ymm1[4,5,6,7]9133; AVX2-NEXT: vpermps %ymm8, %ymm6, %ymm29134; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9135; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9136; AVX2-NEXT: vmovaps 864(%rdi), %ymm19137; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9138; AVX2-NEXT: vmovaps 832(%rdi), %ymm09139; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9140; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9141; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],ymm1[6,7]9142; AVX2-NEXT: vmovaps 800(%rdi), %ymm09143; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9144; AVX2-NEXT: vmovaps 768(%rdi), %ymm19145; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9146; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9147; AVX2-NEXT: vpermps %ymm4, %ymm9, %ymm09148; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm7[0,2,2,2,4,6,6,6]9149; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9150; AVX2-NEXT: vmovaps 896(%rdi), %ymm19151; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9152; AVX2-NEXT: vmovaps 928(%rdi), %ymm29153; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9154; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm1[4,5,6,7]9155; AVX2-NEXT: vpermps %ymm3, %ymm6, %ymm29156; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9157; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9158; AVX2-NEXT: vmovaps 1184(%rdi), %ymm09159; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9160; AVX2-NEXT: vmovaps 1152(%rdi), %ymm19161; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9162; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9163; AVX2-NEXT: vpermps %ymm2, %ymm9, %ymm09164; AVX2-NEXT: vmovaps 1248(%rdi), %ymm19165; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9166; AVX2-NEXT: vmovaps 1216(%rdi), %ymm59167; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9168; AVX2-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm5[0,1],ymm1[0,1]9169; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm5[0,1,2,3,4,5],ymm1[6,7]9170; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm9[0,2,2,2,4,6,6,6]9171; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3,4,5,6,7]9172; AVX2-NEXT: vmovaps 1280(%rdi), %ymm19173; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9174; AVX2-NEXT: vmovaps 1312(%rdi), %ymm59175; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9176; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm1[4,5,6,7]9177; AVX2-NEXT: vpermps %ymm5, %ymm6, %ymm19178; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]9179; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9180; AVX2-NEXT: vmovaps {{.*#+}} xmm0 = [1,7,5,u]9181; AVX2-NEXT: vpermps %ymm14, %ymm0, %ymm19182; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm15[1,3,2,3,5,7,6,7]9183; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2],ymm15[3,4,5,6,7]9184; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm1 = [5,3,5,3,5,3,5,3]9185; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload9186; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]9187; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9188; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload9189; AVX2-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload9190; AVX2-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]9191; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]9192; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload9193; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]9194; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9195; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload9196; AVX2-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload9197; AVX2-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]9198; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]9199; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload9200; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]9201; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9202; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload9203; AVX2-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload9204; AVX2-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]9205; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]9206; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload9207; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]9208; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9209; AVX2-NEXT: vpermps %ymm13, %ymm0, %ymm139210; AVX2-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload9211; AVX2-NEXT: # ymm14 = mem[1,3,2,3,5,7,6,7]9212; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2],ymm14[3,4,5,6,7]9213; AVX2-NEXT: vpermps %ymm12, %ymm1, %ymm129214; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]9215; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9216; AVX2-NEXT: vpermps %ymm10, %ymm0, %ymm109217; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm11[1,3,2,3,5,7,6,7]9218; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2],ymm11[3,4,5,6,7]9219; AVX2-NEXT: vpermps %ymm8, %ymm1, %ymm89220; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3,4,5],ymm8[6,7]9221; AVX2-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9222; AVX2-NEXT: vpermps %ymm4, %ymm0, %ymm49223; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm7[1,3,2,3,5,7,6,7]9224; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm7[3,4,5,6,7]9225; AVX2-NEXT: vpermps %ymm3, %ymm1, %ymm39226; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]9227; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9228; AVX2-NEXT: vpermps %ymm2, %ymm0, %ymm09229; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm9[1,3,2,3,5,7,6,7]9230; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9231; AVX2-NEXT: vpermps %ymm5, %ymm1, %ymm29232; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9233; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9234; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload9235; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload9236; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1],ymm11[2,3],ymm12[4,5],ymm11[6,7]9237; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]9238; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9239; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9240; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9241; AVX2-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]9242; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9243; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9244; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9245; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9246; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9247; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9248; AVX2-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]9249; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9250; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]9251; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9252; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9253; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9254; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload9255; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload9256; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1],ymm7[2,3],ymm9[4,5],ymm7[6,7]9257; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]9258; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9259; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9260; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9261; AVX2-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]9262; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9263; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9264; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9265; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9266; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9267; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9268; AVX2-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]9269; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9270; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]9271; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9272; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9273; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9274; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload9275; AVX2-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload9276; AVX2-NEXT: # ymm0 = mem[0,1],ymm4[2,3],mem[4,5],ymm4[6,7]9277; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]9278; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9279; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9280; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9281; AVX2-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]9282; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9283; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9284; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9285; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9286; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9287; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9288; AVX2-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]9289; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9290; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]9291; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9292; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9293; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9294; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9295; AVX2-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9296; AVX2-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]9297; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]9298; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9299; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9300; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9301; AVX2-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]9302; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9303; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9304; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9305; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9306; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9307; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9308; AVX2-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]9309; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9310; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]9311; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9312; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9313; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9314; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload9315; AVX2-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9316; AVX2-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]9317; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]9318; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9319; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9320; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9321; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9322; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9323; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9324; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9325; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9326; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9327; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9328; AVX2-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]9329; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9330; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]9331; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9332; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9333; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9334; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9335; AVX2-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9336; AVX2-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]9337; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]9338; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm0[0,2,0,3]9339; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9340; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9341; AVX2-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]9342; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9343; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm0[2,0,2,3,6,4,6,7]9344; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,3,2,3]9345; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2],ymm2[3,4,5,6,7]9346; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9347; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9348; AVX2-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]9349; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9350; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm0[0,0,2,0,4,4,6,4]9351; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]9352; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm5[5,6,7]9353; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9354; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload9355; AVX2-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm2 # 32-byte Folded Reload9356; AVX2-NEXT: # ymm2 = ymm13[0,1],mem[2,3],ymm13[4,5],mem[6,7]9357; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9358; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,0,3]9359; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9360; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload9361; AVX2-NEXT: # ymm14 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]9362; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm14[2,0,2,3,6,4,6,7]9363; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,3,2,3]9364; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2],ymm2[3,4,5,6,7]9365; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9366; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload9367; AVX2-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]9368; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm15[0,0,2,0,4,4,6,4]9369; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]9370; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm5[5,6,7]9371; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9372; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload9373; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload9374; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm8[0,1],ymm10[2,3],ymm8[4,5],ymm10[6,7]9375; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]9376; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm2[0,2,0,3]9377; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9378; AVX2-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload9379; AVX2-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]9380; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm2[2,0,2,3,6,4,6,7]9381; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]9382; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm5[3,4,5,6,7]9383; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9384; AVX2-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload9385; AVX2-NEXT: # ymm5 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]9386; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm5[0,0,2,0,4,4,6,4]9387; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]9388; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4],ymm0[5,6,7]9389; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9390; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]9391; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3,4],ymm12[5],ymm0[6,7]9392; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9393; AVX2-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9394; AVX2-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]9395; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]9396; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]9397; AVX2-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9398; AVX2-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]9399; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]9400; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]9401; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9402; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm7[3,3,3,3,7,7,7,7]9403; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3,4],ymm9[5],ymm0[6,7]9404; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9405; AVX2-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9406; AVX2-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]9407; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]9408; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]9409; AVX2-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9410; AVX2-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]9411; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]9412; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]9413; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9414; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]9415; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload9416; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3,4],ymm9[5],ymm0[6,7]9417; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9418; AVX2-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9419; AVX2-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]9420; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]9421; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]9422; AVX2-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9423; AVX2-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]9424; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]9425; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]9426; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9427; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload9428; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm7[3,3,3,3,7,7,7,7]9429; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload9430; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7]9431; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9432; AVX2-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9433; AVX2-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]9434; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]9435; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]9436; AVX2-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9437; AVX2-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]9438; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]9439; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]9440; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9441; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm10[3,3,3,3,7,7,7,7]9442; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7]9443; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9444; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,1,3,3,7,5,7,7]9445; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9446; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9447; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm5[0,1,3,1,4,5,7,5]9448; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9449; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9450; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9451; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload9452; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]9453; AVX2-NEXT: vmovaps %ymm13, %ymm59454; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2,3,4],ymm13[5],ymm0[6,7]9455; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9456; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm14[3,1,3,3,7,5,7,7]9457; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9458; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9459; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm15[0,1,3,1,4,5,7,5]9460; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9461; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9462; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9463; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload9464; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm13[3,3,3,3,7,7,7,7]9465; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload9466; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3,4],ymm12[5],ymm0[6,7]9467; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9468; AVX2-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload9469; AVX2-NEXT: # ymm2 = mem[3,1,3,3,7,5,7,7]9470; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9471; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9472; AVX2-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload9473; AVX2-NEXT: # ymm2 = mem[0,1,3,1,4,5,7,5]9474; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9475; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9476; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9477; AVX2-NEXT: vmovups (%rsp), %ymm14 # 32-byte Reload9478; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm14[3,3,3,3,7,7,7,7]9479; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload9480; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]9481; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]9482; AVX2-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload9483; AVX2-NEXT: # ymm2 = mem[3,1,3,3,7,5,7,7]9484; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]9485; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]9486; AVX2-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload9487; AVX2-NEXT: # ymm2 = mem[0,1,3,1,4,5,7,5]9488; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]9489; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9490; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9491; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm14[4,5,6,7]9492; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9493; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9494; AVX2-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload9495; AVX2-NEXT: # ymm3 = mem[0,1,2,3],ymm0[4,5,6,7]9496; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9497; AVX2-NEXT: vmovaps 80(%rdi), %xmm09498; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9499; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]9500; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9501; AVX2-NEXT: vpermps %ymm3, %ymm6, %ymm29502; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9503; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9504; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9505; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9506; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9507; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,2,0,6,0,2,0,6]9508; AVX2-NEXT: # ymm3 = mem[0,1,0,1]9509; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm29510; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9511; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9512; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9513; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload9514; AVX2-NEXT: # ymm2 = ymm0[0,1,2,3],mem[4,5,6,7]9515; AVX2-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill9516; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9517; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload9518; AVX2-NEXT: # ymm14 = ymm0[0,1,2,3],mem[4,5,6,7]9519; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9520; AVX2-NEXT: vmovaps 272(%rdi), %xmm09521; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9522; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]9523; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9524; AVX2-NEXT: vpermps %ymm14, %ymm6, %ymm29525; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9526; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9527; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9528; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9529; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9530; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm29531; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9532; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9533; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm12[0,1,2,3],ymm13[4,5,6,7]9534; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9535; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9536; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload9537; AVX2-NEXT: # ymm12 = ymm0[0,1,2,3],mem[4,5,6,7]9538; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9539; AVX2-NEXT: vmovaps 464(%rdi), %xmm09540; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9541; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]9542; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9543; AVX2-NEXT: vpermps %ymm12, %ymm6, %ymm29544; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9545; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9546; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9547; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9548; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9549; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm29550; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9551; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9552; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9553; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload9554; AVX2-NEXT: # ymm2 = ymm0[0,1,2,3],mem[4,5,6,7]9555; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9556; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9557; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload9558; AVX2-NEXT: # ymm12 = ymm0[0,1,2,3],mem[4,5,6,7]9559; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9560; AVX2-NEXT: vmovaps 656(%rdi), %xmm09561; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9562; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]9563; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9564; AVX2-NEXT: vpermps %ymm12, %ymm6, %ymm29565; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9566; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9567; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9568; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9569; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9570; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm29571; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9572; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9573; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm11[4,5,6,7]9574; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9575; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9576; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload9577; AVX2-NEXT: # ymm5 = ymm0[0,1,2,3],mem[4,5,6,7]9578; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9579; AVX2-NEXT: vmovaps 848(%rdi), %xmm09580; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9581; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]9582; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9583; AVX2-NEXT: vpermps %ymm5, %ymm6, %ymm29584; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9585; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9586; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9587; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9588; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9589; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm29590; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9591; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9592; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm15 # 32-byte Folded Reload9593; AVX2-NEXT: # ymm15 = ymm9[0,1,2,3],mem[4,5,6,7]9594; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9595; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload9596; AVX2-NEXT: # ymm14 = ymm0[0,1,2,3],mem[4,5,6,7]9597; AVX2-NEXT: vmovaps 1040(%rdi), %xmm139598; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1],ymm13[2,3],ymm15[4,5,6,7]9599; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9600; AVX2-NEXT: vpermps %ymm14, %ymm6, %ymm29601; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9602; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9603; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm12 # 32-byte Folded Reload9604; AVX2-NEXT: # ymm12 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9605; AVX2-NEXT: vpermps %ymm12, %ymm3, %ymm29606; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9607; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9608; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm8[0,1,2,3],ymm10[4,5,6,7]9609; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9610; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload9611; AVX2-NEXT: # ymm10 = ymm0[0,1,2,3],mem[4,5,6,7]9612; AVX2-NEXT: vmovaps 1232(%rdi), %xmm99613; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1],ymm9[2,3],ymm11[4,5,6,7]9614; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9615; AVX2-NEXT: vpermps %ymm10, %ymm6, %ymm29616; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9617; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9618; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm8 # 32-byte Folded Reload9619; AVX2-NEXT: # ymm8 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9620; AVX2-NEXT: vpermps %ymm8, %ymm3, %ymm29621; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]9622; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9623; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm4[0,1,2,3],ymm7[4,5,6,7]9624; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9625; AVX2-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload9626; AVX2-NEXT: # ymm5 = ymm0[0,1,2,3],mem[4,5,6,7]9627; AVX2-NEXT: vmovaps 1424(%rdi), %xmm49628; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1],ymm4[2,3],ymm7[4,5,6,7]9629; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]9630; AVX2-NEXT: vpermps %ymm5, %ymm6, %ymm29631; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]9632; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9633; AVX2-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9634; AVX2-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]9635; AVX2-NEXT: vpermps %ymm2, %ymm3, %ymm39636; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]9637; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9638; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload9639; AVX2-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5]9640; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9641; AVX2-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]9642; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload9643; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm0[2,3,4,5,6,7]9644; AVX2-NEXT: vbroadcastf128 {{.*#+}} ymm0 = [0,3,1,7,0,3,1,7]9645; AVX2-NEXT: # ymm0 = mem[0,1,0,1]9646; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload9647; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]9648; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9649; AVX2-NEXT: vpermilps $85, (%rsp), %ymm3 # 32-byte Folded Reload9650; AVX2-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]9651; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload9652; AVX2-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]9653; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload9654; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3,4,5,6,7]9655; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload9656; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]9657; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9658; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9659; AVX2-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]9660; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload9661; AVX2-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]9662; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload9663; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3,4,5,6,7]9664; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload9665; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]9666; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9667; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload9668; AVX2-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]9669; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload9670; AVX2-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]9671; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload9672; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3,4,5,6,7]9673; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload9674; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]9675; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9676; AVX2-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload9677; AVX2-NEXT: # ymm6 = mem[1,1,1,1,5,5,5,5]9678; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload9679; AVX2-NEXT: # ymm6 = ymm6[0,1,2],mem[3],ymm6[4,5,6,7]9680; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload9681; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3,4,5,6,7]9682; AVX2-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload9683; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm3[0,1,2,3,4],ymm6[5,6,7]9684; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm15[1,1,1,1,5,5,5,5]9685; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm13[3],ymm3[4,5,6,7]9686; AVX2-NEXT: vpermps %ymm14, %ymm1, %ymm139687; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm13[0,1],ymm3[2,3,4,5,6,7]9688; AVX2-NEXT: vpermps %ymm12, %ymm0, %ymm129689; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm3[0,1,2,3,4],ymm12[5,6,7]9690; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm11[1,1,1,1,5,5,5,5]9691; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm9[3],ymm3[4,5,6,7]9692; AVX2-NEXT: vpermps %ymm10, %ymm1, %ymm99693; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm9[0,1],ymm3[2,3,4,5,6,7]9694; AVX2-NEXT: vpermps %ymm8, %ymm0, %ymm89695; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm3[0,1,2,3,4],ymm8[5,6,7]9696; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm7[1,1,1,1,5,5,5,5]9697; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5,6,7]9698; AVX2-NEXT: vpermps %ymm5, %ymm1, %ymm19699; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3,4,5,6,7]9700; AVX2-NEXT: vpermps %ymm2, %ymm0, %ymm09701; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]9702; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9703; AVX2-NEXT: vmovaps %ymm1, 192(%rsi)9704; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9705; AVX2-NEXT: vmovaps %ymm1, 128(%rsi)9706; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9707; AVX2-NEXT: vmovaps %ymm1, 64(%rsi)9708; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9709; AVX2-NEXT: vmovaps %ymm1, (%rsi)9710; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9711; AVX2-NEXT: vmovaps %ymm1, 224(%rsi)9712; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9713; AVX2-NEXT: vmovaps %ymm1, 160(%rsi)9714; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9715; AVX2-NEXT: vmovaps %ymm1, 96(%rsi)9716; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9717; AVX2-NEXT: vmovaps %ymm1, 32(%rsi)9718; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9719; AVX2-NEXT: vmovaps %ymm1, 192(%rdx)9720; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9721; AVX2-NEXT: vmovaps %ymm1, 128(%rdx)9722; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9723; AVX2-NEXT: vmovaps %ymm1, 64(%rdx)9724; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9725; AVX2-NEXT: vmovaps %ymm1, (%rdx)9726; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9727; AVX2-NEXT: vmovaps %ymm1, 224(%rdx)9728; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9729; AVX2-NEXT: vmovaps %ymm1, 160(%rdx)9730; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9731; AVX2-NEXT: vmovaps %ymm1, 96(%rdx)9732; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9733; AVX2-NEXT: vmovaps %ymm1, 32(%rdx)9734; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9735; AVX2-NEXT: vmovaps %ymm1, 192(%rcx)9736; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9737; AVX2-NEXT: vmovaps %ymm1, 128(%rcx)9738; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9739; AVX2-NEXT: vmovaps %ymm1, 64(%rcx)9740; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9741; AVX2-NEXT: vmovaps %ymm1, (%rcx)9742; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9743; AVX2-NEXT: vmovaps %ymm1, 224(%rcx)9744; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9745; AVX2-NEXT: vmovaps %ymm1, 160(%rcx)9746; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9747; AVX2-NEXT: vmovaps %ymm1, 96(%rcx)9748; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9749; AVX2-NEXT: vmovaps %ymm1, 32(%rcx)9750; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9751; AVX2-NEXT: vmovaps %ymm1, (%r8)9752; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9753; AVX2-NEXT: vmovaps %ymm1, 64(%r8)9754; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9755; AVX2-NEXT: vmovaps %ymm1, 128(%r8)9756; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9757; AVX2-NEXT: vmovaps %ymm1, 192(%r8)9758; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9759; AVX2-NEXT: vmovaps %ymm1, 224(%r8)9760; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9761; AVX2-NEXT: vmovaps %ymm1, 160(%r8)9762; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9763; AVX2-NEXT: vmovaps %ymm1, 96(%r8)9764; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9765; AVX2-NEXT: vmovaps %ymm1, 32(%r8)9766; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9767; AVX2-NEXT: vmovaps %ymm1, 224(%r9)9768; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9769; AVX2-NEXT: vmovaps %ymm1, 192(%r9)9770; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9771; AVX2-NEXT: vmovaps %ymm1, 160(%r9)9772; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9773; AVX2-NEXT: vmovaps %ymm1, 128(%r9)9774; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9775; AVX2-NEXT: vmovaps %ymm1, 96(%r9)9776; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9777; AVX2-NEXT: vmovaps %ymm1, 64(%r9)9778; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9779; AVX2-NEXT: vmovaps %ymm1, 32(%r9)9780; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9781; AVX2-NEXT: vmovaps %ymm1, (%r9)9782; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax9783; AVX2-NEXT: vmovaps %ymm0, 224(%rax)9784; AVX2-NEXT: vmovaps %ymm8, 192(%rax)9785; AVX2-NEXT: vmovaps %ymm12, 160(%rax)9786; AVX2-NEXT: vmovaps %ymm6, 128(%rax)9787; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9788; AVX2-NEXT: vmovaps %ymm0, 96(%rax)9789; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9790; AVX2-NEXT: vmovaps %ymm0, 64(%rax)9791; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9792; AVX2-NEXT: vmovaps %ymm0, 32(%rax)9793; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9794; AVX2-NEXT: vmovaps %ymm0, (%rax)9795; AVX2-NEXT: addq $2568, %rsp # imm = 0xA089796; AVX2-NEXT: vzeroupper9797; AVX2-NEXT: retq9798;9799; AVX2-FP-LABEL: load_i32_stride6_vf64:9800; AVX2-FP: # %bb.0:9801; AVX2-FP-NEXT: subq $2568, %rsp # imm = 0xA089802; AVX2-FP-NEXT: vmovaps 672(%rdi), %ymm49803; AVX2-FP-NEXT: vmovaps 640(%rdi), %ymm59804; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9805; AVX2-FP-NEXT: vmovaps 608(%rdi), %ymm39806; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9807; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm69808; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9809; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm79810; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9811; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm29812; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9813; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm89814; AVX2-FP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9815; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm09816; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9817; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm19818; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9819; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm9 = [0,6,4,u]9820; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9821; AVX2-FP-NEXT: vpermps %ymm14, %ymm9, %ymm09822; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm8[0,1],ymm2[0,1]9823; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2,3,4,5],ymm2[6,7]9824; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm15[0,2,2,2,4,6,6,6]9825; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9826; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm6[4,5,6,7]9827; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9828; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm6 = [4,2,4,2,4,2,4,2]9829; AVX2-FP-NEXT: vpermps %ymm1, %ymm6, %ymm29830; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9831; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9832; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9833; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm5[0,1],ymm4[0,1]9834; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm4[6,7]9835; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9836; AVX2-FP-NEXT: vmovaps 576(%rdi), %ymm09837; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9838; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5],ymm0[6,7]9839; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9840; AVX2-FP-NEXT: vpermps %ymm0, %ymm9, %ymm09841; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm1[0,2,2,2,4,6,6,6]9842; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9843; AVX2-FP-NEXT: vmovaps 704(%rdi), %ymm19844; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9845; AVX2-FP-NEXT: vmovaps 736(%rdi), %ymm29846; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9847; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]9848; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9849; AVX2-FP-NEXT: vpermps %ymm1, %ymm6, %ymm29850; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9851; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9852; AVX2-FP-NEXT: vmovaps 1056(%rdi), %ymm19853; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9854; AVX2-FP-NEXT: vmovaps 1024(%rdi), %ymm09855; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9856; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9857; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]9858; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9859; AVX2-FP-NEXT: vmovaps 992(%rdi), %ymm09860; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9861; AVX2-FP-NEXT: vmovaps 960(%rdi), %ymm19862; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9863; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9864; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9865; AVX2-FP-NEXT: vpermps %ymm0, %ymm9, %ymm09866; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]9867; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9868; AVX2-FP-NEXT: vmovaps 1088(%rdi), %ymm19869; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9870; AVX2-FP-NEXT: vmovaps 1120(%rdi), %ymm29871; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9872; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]9873; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9874; AVX2-FP-NEXT: vpermps %ymm1, %ymm6, %ymm29875; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9876; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9877; AVX2-FP-NEXT: vmovaps 1440(%rdi), %ymm19878; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9879; AVX2-FP-NEXT: vmovaps 1408(%rdi), %ymm09880; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9881; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9882; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]9883; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9884; AVX2-FP-NEXT: vmovaps 1376(%rdi), %ymm09885; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9886; AVX2-FP-NEXT: vmovaps 1344(%rdi), %ymm19887; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9888; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9889; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9890; AVX2-FP-NEXT: vpermps %ymm0, %ymm9, %ymm09891; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]9892; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9893; AVX2-FP-NEXT: vmovaps 1472(%rdi), %ymm19894; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9895; AVX2-FP-NEXT: vmovaps 1504(%rdi), %ymm29896; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9897; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]9898; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9899; AVX2-FP-NEXT: vpermps %ymm1, %ymm6, %ymm29900; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9901; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9902; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm19903; AVX2-FP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill9904; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm09905; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9906; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9907; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]9908; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9909; AVX2-FP-NEXT: vmovaps (%rdi), %ymm09910; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9911; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm19912; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9913; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]9914; AVX2-FP-NEXT: vpermps %ymm13, %ymm9, %ymm09915; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]9916; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9917; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm19918; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9919; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm29920; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9921; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm2[0,1,2,3],ymm1[4,5,6,7]9922; AVX2-FP-NEXT: vpermps %ymm12, %ymm6, %ymm29923; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9924; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9925; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm19926; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9927; AVX2-FP-NEXT: vmovaps 448(%rdi), %ymm09928; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9929; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9930; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3,4,5],ymm1[6,7]9931; AVX2-FP-NEXT: vmovaps 416(%rdi), %ymm09932; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9933; AVX2-FP-NEXT: vmovaps 384(%rdi), %ymm19934; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9935; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9936; AVX2-FP-NEXT: vpermps %ymm10, %ymm9, %ymm09937; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm11[0,2,2,2,4,6,6,6]9938; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9939; AVX2-FP-NEXT: vmovaps 512(%rdi), %ymm19940; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9941; AVX2-FP-NEXT: vmovaps 544(%rdi), %ymm29942; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9943; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm2[0,1,2,3],ymm1[4,5,6,7]9944; AVX2-FP-NEXT: vpermps %ymm8, %ymm6, %ymm29945; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9946; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9947; AVX2-FP-NEXT: vmovaps 864(%rdi), %ymm19948; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9949; AVX2-FP-NEXT: vmovaps 832(%rdi), %ymm09950; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9951; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]9952; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],ymm1[6,7]9953; AVX2-FP-NEXT: vmovaps 800(%rdi), %ymm09954; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9955; AVX2-FP-NEXT: vmovaps 768(%rdi), %ymm19956; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9957; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9958; AVX2-FP-NEXT: vpermps %ymm4, %ymm9, %ymm09959; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm7[0,2,2,2,4,6,6,6]9960; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]9961; AVX2-FP-NEXT: vmovaps 896(%rdi), %ymm19962; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9963; AVX2-FP-NEXT: vmovaps 928(%rdi), %ymm29964; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9965; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm1[4,5,6,7]9966; AVX2-FP-NEXT: vpermps %ymm3, %ymm6, %ymm29967; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]9968; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9969; AVX2-FP-NEXT: vmovaps 1184(%rdi), %ymm09970; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9971; AVX2-FP-NEXT: vmovaps 1152(%rdi), %ymm19972; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9973; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]9974; AVX2-FP-NEXT: vpermps %ymm2, %ymm9, %ymm09975; AVX2-FP-NEXT: vmovaps 1248(%rdi), %ymm19976; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9977; AVX2-FP-NEXT: vmovaps 1216(%rdi), %ymm59978; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9979; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm5[0,1],ymm1[0,1]9980; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm5[0,1,2,3,4,5],ymm1[6,7]9981; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm9[0,2,2,2,4,6,6,6]9982; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3,4,5,6,7]9983; AVX2-FP-NEXT: vmovaps 1280(%rdi), %ymm19984; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9985; AVX2-FP-NEXT: vmovaps 1312(%rdi), %ymm59986; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9987; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm1[4,5,6,7]9988; AVX2-FP-NEXT: vpermps %ymm5, %ymm6, %ymm19989; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]9990; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9991; AVX2-FP-NEXT: vmovaps {{.*#+}} xmm0 = [1,7,5,u]9992; AVX2-FP-NEXT: vpermps %ymm14, %ymm0, %ymm19993; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm15[1,3,2,3,5,7,6,7]9994; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2],ymm15[3,4,5,6,7]9995; AVX2-FP-NEXT: vbroadcastsd {{.*#+}} ymm1 = [5,3,5,3,5,3,5,3]9996; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload9997; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]9998; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9999; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10000; AVX2-FP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10001; AVX2-FP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10002; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]10003; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload10004; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]10005; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10006; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10007; AVX2-FP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10008; AVX2-FP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10009; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]10010; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload10011; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]10012; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10013; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10014; AVX2-FP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10015; AVX2-FP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10016; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]10017; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload10018; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]10019; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10020; AVX2-FP-NEXT: vpermps %ymm13, %ymm0, %ymm1310021; AVX2-FP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload10022; AVX2-FP-NEXT: # ymm14 = mem[1,3,2,3,5,7,6,7]10023; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2],ymm14[3,4,5,6,7]10024; AVX2-FP-NEXT: vpermps %ymm12, %ymm1, %ymm1210025; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]10026; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10027; AVX2-FP-NEXT: vpermps %ymm10, %ymm0, %ymm1010028; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm11[1,3,2,3,5,7,6,7]10029; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2],ymm11[3,4,5,6,7]10030; AVX2-FP-NEXT: vpermps %ymm8, %ymm1, %ymm810031; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3,4,5],ymm8[6,7]10032; AVX2-FP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10033; AVX2-FP-NEXT: vpermps %ymm4, %ymm0, %ymm410034; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm7[1,3,2,3,5,7,6,7]10035; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm7[3,4,5,6,7]10036; AVX2-FP-NEXT: vpermps %ymm3, %ymm1, %ymm310037; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]10038; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10039; AVX2-FP-NEXT: vpermps %ymm2, %ymm0, %ymm010040; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm9[1,3,2,3,5,7,6,7]10041; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10042; AVX2-FP-NEXT: vpermps %ymm5, %ymm1, %ymm210043; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10044; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10045; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload10046; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload10047; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1],ymm11[2,3],ymm12[4,5],ymm11[6,7]10048; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]10049; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10050; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10051; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10052; AVX2-FP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10053; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10054; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10055; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10056; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10057; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10058; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10059; AVX2-FP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10060; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10061; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]10062; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10063; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10064; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10065; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload10066; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload10067; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1],ymm7[2,3],ymm9[4,5],ymm7[6,7]10068; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]10069; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10070; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10071; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10072; AVX2-FP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10073; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10074; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10075; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10076; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10077; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10078; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10079; AVX2-FP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10080; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10081; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]10082; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10083; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10084; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10085; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload10086; AVX2-FP-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload10087; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm4[2,3],mem[4,5],ymm4[6,7]10088; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]10089; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10090; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10091; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10092; AVX2-FP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10093; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10094; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10095; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10096; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10097; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10098; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10099; AVX2-FP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10100; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10101; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]10102; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10103; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10104; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10105; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10106; AVX2-FP-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10107; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]10108; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]10109; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10110; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10111; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10112; AVX2-FP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10113; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10114; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10115; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10116; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10117; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10118; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10119; AVX2-FP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10120; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10121; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]10122; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10123; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10124; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10125; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload10126; AVX2-FP-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10127; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]10128; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]10129; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10130; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10131; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10132; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10133; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10134; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10135; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10136; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10137; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10138; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10139; AVX2-FP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10140; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10141; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,0,4,4,6,4]10142; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10143; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10144; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10145; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10146; AVX2-FP-NEXT: vblendps $51, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10147; AVX2-FP-NEXT: # ymm0 = mem[0,1],ymm0[2,3],mem[4,5],ymm0[6,7]10148; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,0,2,3,6,4,6,7]10149; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm0[0,2,0,3]10150; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10151; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10152; AVX2-FP-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]10153; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10154; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm0[2,0,2,3,6,4,6,7]10155; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,3,2,3]10156; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2],ymm2[3,4,5,6,7]10157; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10158; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10159; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]10160; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10161; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm0[0,0,2,0,4,4,6,4]10162; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]10163; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm5[5,6,7]10164; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10165; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload10166; AVX2-FP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm2 # 32-byte Folded Reload10167; AVX2-FP-NEXT: # ymm2 = ymm13[0,1],mem[2,3],ymm13[4,5],mem[6,7]10168; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10169; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,0,3]10170; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10171; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10172; AVX2-FP-NEXT: # ymm14 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]10173; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm14[2,0,2,3,6,4,6,7]10174; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,3,2,3]10175; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2],ymm2[3,4,5,6,7]10176; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10177; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm15 # 32-byte Folded Reload10178; AVX2-FP-NEXT: # ymm15 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]10179; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm15[0,0,2,0,4,4,6,4]10180; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,0,3]10181; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm5[5,6,7]10182; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10183; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload10184; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload10185; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm8[0,1],ymm10[2,3],ymm8[4,5],ymm10[6,7]10186; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0,2,3,6,4,6,7]10187; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm2[0,2,0,3]10188; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10189; AVX2-FP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload10190; AVX2-FP-NEXT: # ymm2 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7]10191; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm2[2,0,2,3,6,4,6,7]10192; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]10193; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm5[3,4,5,6,7]10194; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10195; AVX2-FP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload10196; AVX2-FP-NEXT: # ymm5 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]10197; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm5[0,0,2,0,4,4,6,4]10198; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]10199; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3,4],ymm0[5,6,7]10200; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10201; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]10202; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3,4],ymm12[5],ymm0[6,7]10203; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10204; AVX2-FP-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10205; AVX2-FP-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]10206; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]10207; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]10208; AVX2-FP-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10209; AVX2-FP-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]10210; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]10211; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]10212; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10213; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm7[3,3,3,3,7,7,7,7]10214; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3,4],ymm9[5],ymm0[6,7]10215; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10216; AVX2-FP-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10217; AVX2-FP-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]10218; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]10219; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]10220; AVX2-FP-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10221; AVX2-FP-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]10222; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]10223; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]10224; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10225; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]10226; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload10227; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3,4],ymm9[5],ymm0[6,7]10228; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10229; AVX2-FP-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10230; AVX2-FP-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]10231; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]10232; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]10233; AVX2-FP-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10234; AVX2-FP-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]10235; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]10236; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]10237; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10238; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload10239; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm7[3,3,3,3,7,7,7,7]10240; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload10241; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7]10242; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10243; AVX2-FP-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10244; AVX2-FP-NEXT: # ymm3 = mem[3,1,3,3,7,5,7,7]10245; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,3,2,3]10246; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7]10247; AVX2-FP-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10248; AVX2-FP-NEXT: # ymm3 = mem[0,1,3,1,4,5,7,5]10249; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,3]10250; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]10251; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10252; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm10[3,3,3,3,7,7,7,7]10253; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7]10254; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10255; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,1,3,3,7,5,7,7]10256; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10257; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10258; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm5[0,1,3,1,4,5,7,5]10259; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10260; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10261; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10262; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload10263; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]10264; AVX2-FP-NEXT: vmovaps %ymm13, %ymm510265; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2,3,4],ymm13[5],ymm0[6,7]10266; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10267; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm14[3,1,3,3,7,5,7,7]10268; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10269; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10270; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm15[0,1,3,1,4,5,7,5]10271; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10272; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10273; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10274; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload10275; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm13[3,3,3,3,7,7,7,7]10276; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload10277; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3,4],ymm12[5],ymm0[6,7]10278; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10279; AVX2-FP-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload10280; AVX2-FP-NEXT: # ymm2 = mem[3,1,3,3,7,5,7,7]10281; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10282; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10283; AVX2-FP-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload10284; AVX2-FP-NEXT: # ymm2 = mem[0,1,3,1,4,5,7,5]10285; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10286; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10287; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10288; AVX2-FP-NEXT: vmovups (%rsp), %ymm14 # 32-byte Reload10289; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm14[3,3,3,3,7,7,7,7]10290; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload10291; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]10292; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,0,3]10293; AVX2-FP-NEXT: vpermilps $247, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload10294; AVX2-FP-NEXT: # ymm2 = mem[3,1,3,3,7,5,7,7]10295; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,3,2,3]10296; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10297; AVX2-FP-NEXT: vpermilps $116, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload10298; AVX2-FP-NEXT: # ymm2 = mem[0,1,3,1,4,5,7,5]10299; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,0,3]10300; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10301; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10302; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm14[4,5,6,7]10303; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10304; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10305; AVX2-FP-NEXT: vblendps $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload10306; AVX2-FP-NEXT: # ymm3 = mem[0,1,2,3],ymm0[4,5,6,7]10307; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10308; AVX2-FP-NEXT: vmovaps 80(%rdi), %xmm010309; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10310; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]10311; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10312; AVX2-FP-NEXT: vpermps %ymm3, %ymm6, %ymm210313; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10314; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10315; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10316; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10317; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10318; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [0,2,0,6,0,2,0,6]10319; AVX2-FP-NEXT: # ymm3 = mem[0,1,0,1]10320; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm210321; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10322; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10323; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10324; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload10325; AVX2-FP-NEXT: # ymm2 = ymm0[0,1,2,3],mem[4,5,6,7]10326; AVX2-FP-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill10327; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10328; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10329; AVX2-FP-NEXT: # ymm14 = ymm0[0,1,2,3],mem[4,5,6,7]10330; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10331; AVX2-FP-NEXT: vmovaps 272(%rdi), %xmm010332; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10333; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]10334; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10335; AVX2-FP-NEXT: vpermps %ymm14, %ymm6, %ymm210336; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10337; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10338; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10339; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10340; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10341; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm210342; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10343; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10344; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm12[0,1,2,3],ymm13[4,5,6,7]10345; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10346; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10347; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload10348; AVX2-FP-NEXT: # ymm12 = ymm0[0,1,2,3],mem[4,5,6,7]10349; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10350; AVX2-FP-NEXT: vmovaps 464(%rdi), %xmm010351; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10352; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]10353; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10354; AVX2-FP-NEXT: vpermps %ymm12, %ymm6, %ymm210355; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10356; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10357; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10358; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10359; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10360; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm210361; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10362; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10363; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10364; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload10365; AVX2-FP-NEXT: # ymm2 = ymm0[0,1,2,3],mem[4,5,6,7]10366; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10367; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10368; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm12 # 32-byte Folded Reload10369; AVX2-FP-NEXT: # ymm12 = ymm0[0,1,2,3],mem[4,5,6,7]10370; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10371; AVX2-FP-NEXT: vmovaps 656(%rdi), %xmm010372; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10373; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]10374; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10375; AVX2-FP-NEXT: vpermps %ymm12, %ymm6, %ymm210376; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10377; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10378; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10379; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10380; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10381; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm210382; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10383; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10384; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm11[4,5,6,7]10385; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10386; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10387; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload10388; AVX2-FP-NEXT: # ymm5 = ymm0[0,1,2,3],mem[4,5,6,7]10389; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10390; AVX2-FP-NEXT: vmovaps 848(%rdi), %xmm010391; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10392; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]10393; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10394; AVX2-FP-NEXT: vpermps %ymm5, %ymm6, %ymm210395; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10396; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10397; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10398; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10399; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10400; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm210401; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10402; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10403; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm15 # 32-byte Folded Reload10404; AVX2-FP-NEXT: # ymm15 = ymm9[0,1,2,3],mem[4,5,6,7]10405; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10406; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10407; AVX2-FP-NEXT: # ymm14 = ymm0[0,1,2,3],mem[4,5,6,7]10408; AVX2-FP-NEXT: vmovaps 1040(%rdi), %xmm1310409; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1],ymm13[2,3],ymm15[4,5,6,7]10410; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10411; AVX2-FP-NEXT: vpermps %ymm14, %ymm6, %ymm210412; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10413; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10414; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm12 # 32-byte Folded Reload10415; AVX2-FP-NEXT: # ymm12 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10416; AVX2-FP-NEXT: vpermps %ymm12, %ymm3, %ymm210417; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10418; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10419; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm8[0,1,2,3],ymm10[4,5,6,7]10420; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10421; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload10422; AVX2-FP-NEXT: # ymm10 = ymm0[0,1,2,3],mem[4,5,6,7]10423; AVX2-FP-NEXT: vmovaps 1232(%rdi), %xmm910424; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1],ymm9[2,3],ymm11[4,5,6,7]10425; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10426; AVX2-FP-NEXT: vpermps %ymm10, %ymm6, %ymm210427; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10428; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10429; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm8 # 32-byte Folded Reload10430; AVX2-FP-NEXT: # ymm8 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10431; AVX2-FP-NEXT: vpermps %ymm8, %ymm3, %ymm210432; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10433; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10434; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm4[0,1,2,3],ymm7[4,5,6,7]10435; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10436; AVX2-FP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload10437; AVX2-FP-NEXT: # ymm5 = ymm0[0,1,2,3],mem[4,5,6,7]10438; AVX2-FP-NEXT: vmovaps 1424(%rdi), %xmm410439; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1],ymm4[2,3],ymm7[4,5,6,7]10440; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]10441; AVX2-FP-NEXT: vpermps %ymm5, %ymm6, %ymm210442; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]10443; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10444; AVX2-FP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10445; AVX2-FP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10446; AVX2-FP-NEXT: vpermps %ymm2, %ymm3, %ymm310447; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6,7]10448; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10449; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload10450; AVX2-FP-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5]10451; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10452; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]10453; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload10454; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm0[2,3,4,5,6,7]10455; AVX2-FP-NEXT: vbroadcastf128 {{.*#+}} ymm0 = [0,3,1,7,0,3,1,7]10456; AVX2-FP-NEXT: # ymm0 = mem[0,1,0,1]10457; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload10458; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]10459; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10460; AVX2-FP-NEXT: vpermilps $85, (%rsp), %ymm3 # 32-byte Folded Reload10461; AVX2-FP-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]10462; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload10463; AVX2-FP-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]10464; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload10465; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3,4,5,6,7]10466; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload10467; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]10468; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10469; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10470; AVX2-FP-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]10471; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload10472; AVX2-FP-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]10473; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload10474; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3,4,5,6,7]10475; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload10476; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]10477; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10478; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload10479; AVX2-FP-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5]10480; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload10481; AVX2-FP-NEXT: # ymm3 = ymm3[0,1,2],mem[3],ymm3[4,5,6,7]10482; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload10483; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3,4,5,6,7]10484; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload10485; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6,7]10486; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10487; AVX2-FP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload10488; AVX2-FP-NEXT: # ymm6 = mem[1,1,1,1,5,5,5,5]10489; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload10490; AVX2-FP-NEXT: # ymm6 = ymm6[0,1,2],mem[3],ymm6[4,5,6,7]10491; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm3 # 32-byte Folded Reload10492; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3,4,5,6,7]10493; AVX2-FP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload10494; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm3[0,1,2,3,4],ymm6[5,6,7]10495; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm15[1,1,1,1,5,5,5,5]10496; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm13[3],ymm3[4,5,6,7]10497; AVX2-FP-NEXT: vpermps %ymm14, %ymm1, %ymm1310498; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm13[0,1],ymm3[2,3,4,5,6,7]10499; AVX2-FP-NEXT: vpermps %ymm12, %ymm0, %ymm1210500; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm3[0,1,2,3,4],ymm12[5,6,7]10501; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm11[1,1,1,1,5,5,5,5]10502; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm9[3],ymm3[4,5,6,7]10503; AVX2-FP-NEXT: vpermps %ymm10, %ymm1, %ymm910504; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm9[0,1],ymm3[2,3,4,5,6,7]10505; AVX2-FP-NEXT: vpermps %ymm8, %ymm0, %ymm810506; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm3[0,1,2,3,4],ymm8[5,6,7]10507; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm7[1,1,1,1,5,5,5,5]10508; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5,6,7]10509; AVX2-FP-NEXT: vpermps %ymm5, %ymm1, %ymm110510; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3,4,5,6,7]10511; AVX2-FP-NEXT: vpermps %ymm2, %ymm0, %ymm010512; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]10513; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10514; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rsi)10515; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10516; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rsi)10517; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10518; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rsi)10519; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10520; AVX2-FP-NEXT: vmovaps %ymm1, (%rsi)10521; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10522; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rsi)10523; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10524; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rsi)10525; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10526; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rsi)10527; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10528; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rsi)10529; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10530; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rdx)10531; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10532; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rdx)10533; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10534; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rdx)10535; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10536; AVX2-FP-NEXT: vmovaps %ymm1, (%rdx)10537; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10538; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rdx)10539; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10540; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rdx)10541; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10542; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rdx)10543; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10544; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rdx)10545; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10546; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rcx)10547; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10548; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rcx)10549; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10550; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rcx)10551; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10552; AVX2-FP-NEXT: vmovaps %ymm1, (%rcx)10553; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10554; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rcx)10555; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10556; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rcx)10557; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10558; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rcx)10559; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10560; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rcx)10561; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10562; AVX2-FP-NEXT: vmovaps %ymm1, (%r8)10563; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10564; AVX2-FP-NEXT: vmovaps %ymm1, 64(%r8)10565; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10566; AVX2-FP-NEXT: vmovaps %ymm1, 128(%r8)10567; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10568; AVX2-FP-NEXT: vmovaps %ymm1, 192(%r8)10569; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10570; AVX2-FP-NEXT: vmovaps %ymm1, 224(%r8)10571; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10572; AVX2-FP-NEXT: vmovaps %ymm1, 160(%r8)10573; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10574; AVX2-FP-NEXT: vmovaps %ymm1, 96(%r8)10575; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10576; AVX2-FP-NEXT: vmovaps %ymm1, 32(%r8)10577; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10578; AVX2-FP-NEXT: vmovaps %ymm1, 224(%r9)10579; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10580; AVX2-FP-NEXT: vmovaps %ymm1, 192(%r9)10581; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10582; AVX2-FP-NEXT: vmovaps %ymm1, 160(%r9)10583; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10584; AVX2-FP-NEXT: vmovaps %ymm1, 128(%r9)10585; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10586; AVX2-FP-NEXT: vmovaps %ymm1, 96(%r9)10587; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10588; AVX2-FP-NEXT: vmovaps %ymm1, 64(%r9)10589; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10590; AVX2-FP-NEXT: vmovaps %ymm1, 32(%r9)10591; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10592; AVX2-FP-NEXT: vmovaps %ymm1, (%r9)10593; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax10594; AVX2-FP-NEXT: vmovaps %ymm0, 224(%rax)10595; AVX2-FP-NEXT: vmovaps %ymm8, 192(%rax)10596; AVX2-FP-NEXT: vmovaps %ymm12, 160(%rax)10597; AVX2-FP-NEXT: vmovaps %ymm6, 128(%rax)10598; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10599; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rax)10600; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10601; AVX2-FP-NEXT: vmovaps %ymm0, 64(%rax)10602; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10603; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rax)10604; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10605; AVX2-FP-NEXT: vmovaps %ymm0, (%rax)10606; AVX2-FP-NEXT: addq $2568, %rsp # imm = 0xA0810607; AVX2-FP-NEXT: vzeroupper10608; AVX2-FP-NEXT: retq10609;10610; AVX2-FCP-LABEL: load_i32_stride6_vf64:10611; AVX2-FCP: # %bb.0:10612; AVX2-FCP-NEXT: subq $2536, %rsp # imm = 0x9E810613; AVX2-FCP-NEXT: vmovaps 672(%rdi), %ymm510614; AVX2-FCP-NEXT: vmovaps 640(%rdi), %ymm610615; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10616; AVX2-FCP-NEXT: vmovaps 608(%rdi), %ymm310617; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10618; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm410619; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10620; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm710621; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10622; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm210623; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10624; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm810625; AVX2-FCP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10626; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm010627; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10628; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm110629; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10630; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm12 = [0,6,4,u]10631; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]10632; AVX2-FCP-NEXT: vpermps %ymm15, %ymm12, %ymm010633; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm8[0,1],ymm2[0,1]10634; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]10635; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10636; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm1[0,2,2,2,4,6,6,6]10637; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10638; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0,1,2,3],ymm4[4,5,6,7]10639; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10640; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm4 = [4,2,4,2,4,2,4,2]10641; AVX2-FCP-NEXT: vpermps %ymm2, %ymm4, %ymm210642; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10643; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10644; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10645; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm6[0,1],ymm5[0,1]10646; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm5[6,7]10647; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10648; AVX2-FCP-NEXT: vmovaps 576(%rdi), %ymm010649; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10650; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5],ymm0[6,7]10651; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10652; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm010653; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm1[0,2,2,2,4,6,6,6]10654; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10655; AVX2-FCP-NEXT: vmovaps 704(%rdi), %ymm110656; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10657; AVX2-FCP-NEXT: vmovaps 736(%rdi), %ymm210658; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10659; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]10660; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10661; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm210662; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10663; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10664; AVX2-FCP-NEXT: vmovaps 1056(%rdi), %ymm110665; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10666; AVX2-FCP-NEXT: vmovaps 1024(%rdi), %ymm010667; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10668; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]10669; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]10670; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10671; AVX2-FCP-NEXT: vmovaps 992(%rdi), %ymm010672; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10673; AVX2-FCP-NEXT: vmovaps 960(%rdi), %ymm110674; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10675; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]10676; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10677; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm010678; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]10679; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10680; AVX2-FCP-NEXT: vmovaps 1088(%rdi), %ymm110681; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10682; AVX2-FCP-NEXT: vmovaps 1120(%rdi), %ymm210683; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10684; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]10685; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10686; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm210687; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10688; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10689; AVX2-FCP-NEXT: vmovaps 1440(%rdi), %ymm110690; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10691; AVX2-FCP-NEXT: vmovaps 1408(%rdi), %ymm010692; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10693; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]10694; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]10695; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10696; AVX2-FCP-NEXT: vmovaps 1376(%rdi), %ymm010697; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10698; AVX2-FCP-NEXT: vmovaps 1344(%rdi), %ymm110699; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10700; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]10701; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10702; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm010703; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]10704; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10705; AVX2-FCP-NEXT: vmovaps 1472(%rdi), %ymm110706; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10707; AVX2-FCP-NEXT: vmovaps 1504(%rdi), %ymm210708; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10709; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]10710; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10711; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm210712; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10713; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10714; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm110715; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10716; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm010717; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10718; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]10719; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]10720; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10721; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm010722; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10723; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm110724; AVX2-FCP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill10725; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7]10726; AVX2-FCP-NEXT: vpermps %ymm13, %ymm12, %ymm010727; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,2,4,6,6,6]10728; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10729; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm110730; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10731; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm210732; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10733; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm2[0,1,2,3],ymm1[4,5,6,7]10734; AVX2-FCP-NEXT: vpermps %ymm11, %ymm4, %ymm210735; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10736; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10737; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm110738; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10739; AVX2-FCP-NEXT: vmovaps 448(%rdi), %ymm010740; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10741; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]10742; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],ymm1[6,7]10743; AVX2-FCP-NEXT: vmovaps 416(%rdi), %ymm010744; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10745; AVX2-FCP-NEXT: vmovaps 384(%rdi), %ymm110746; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10747; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]10748; AVX2-FCP-NEXT: vpermps %ymm9, %ymm12, %ymm010749; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm10[0,2,2,2,4,6,6,6]10750; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10751; AVX2-FCP-NEXT: vmovaps 512(%rdi), %ymm110752; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10753; AVX2-FCP-NEXT: vmovaps 544(%rdi), %ymm210754; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10755; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm2[0,1,2,3],ymm1[4,5,6,7]10756; AVX2-FCP-NEXT: vpermps %ymm8, %ymm4, %ymm210757; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10758; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10759; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10760; AVX2-FCP-NEXT: vmovaps 864(%rdi), %ymm110761; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10762; AVX2-FCP-NEXT: vmovaps 832(%rdi), %ymm010763; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10764; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1]10765; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],ymm1[6,7]10766; AVX2-FCP-NEXT: vmovaps 800(%rdi), %ymm010767; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10768; AVX2-FCP-NEXT: vmovaps 768(%rdi), %ymm110769; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10770; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]10771; AVX2-FCP-NEXT: vpermps %ymm6, %ymm12, %ymm010772; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm7[0,2,2,2,4,6,6,6]10773; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10774; AVX2-FCP-NEXT: vmovaps 896(%rdi), %ymm110775; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10776; AVX2-FCP-NEXT: vmovaps 928(%rdi), %ymm210777; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10778; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm2[0,1,2,3],ymm1[4,5,6,7]10779; AVX2-FCP-NEXT: vpermps %ymm5, %ymm4, %ymm210780; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10781; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10782; AVX2-FCP-NEXT: vmovaps 1184(%rdi), %ymm010783; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10784; AVX2-FCP-NEXT: vmovaps 1152(%rdi), %ymm110785; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10786; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7]10787; AVX2-FCP-NEXT: vpermps %ymm3, %ymm12, %ymm010788; AVX2-FCP-NEXT: vmovaps 1248(%rdi), %ymm110789; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10790; AVX2-FCP-NEXT: vmovaps 1216(%rdi), %ymm210791; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10792; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[0,1],ymm1[0,1]10793; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm2[0,1,2,3,4,5],ymm1[6,7]10794; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm12[0,2,2,2,4,6,6,6]10795; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4,5,6,7]10796; AVX2-FCP-NEXT: vmovaps 1280(%rdi), %ymm110797; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10798; AVX2-FCP-NEXT: vmovaps 1312(%rdi), %ymm210799; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10800; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4,5,6,7]10801; AVX2-FCP-NEXT: vpermps %ymm2, %ymm4, %ymm110802; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]10803; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10804; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm0 = [1,7,5,u]10805; AVX2-FCP-NEXT: vpermps %ymm15, %ymm0, %ymm110806; AVX2-FCP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10807; AVX2-FCP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10808; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0,1,2],ymm15[3,4,5,6,7]10809; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm1 = [5,3,5,3,5,3,5,3]10810; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload10811; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]10812; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10813; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10814; AVX2-FCP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10815; AVX2-FCP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10816; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]10817; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload10818; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]10819; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10820; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10821; AVX2-FCP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10822; AVX2-FCP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10823; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]10824; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload10825; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]10826; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10827; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload10828; AVX2-FCP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload10829; AVX2-FCP-NEXT: # ymm15 = mem[1,3,2,3,5,7,6,7]10830; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2],ymm15[3,4,5,6,7]10831; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm15 # 32-byte Folded Reload10832; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm14[0,1,2,3,4,5],ymm15[6,7]10833; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10834; AVX2-FCP-NEXT: vpermps %ymm13, %ymm0, %ymm1310835; AVX2-FCP-NEXT: vpermilps $237, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload10836; AVX2-FCP-NEXT: # ymm14 = mem[1,3,2,3,5,7,6,7]10837; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2],ymm14[3,4,5,6,7]10838; AVX2-FCP-NEXT: vpermps %ymm11, %ymm1, %ymm1110839; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm13[0,1,2,3,4,5],ymm11[6,7]10840; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10841; AVX2-FCP-NEXT: vpermps %ymm9, %ymm0, %ymm910842; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm10 = ymm10[1,3,2,3,5,7,6,7]10843; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2],ymm10[3,4,5,6,7]10844; AVX2-FCP-NEXT: vpermps %ymm8, %ymm1, %ymm810845; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5],ymm8[6,7]10846; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10847; AVX2-FCP-NEXT: vpermps %ymm6, %ymm0, %ymm610848; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm7[1,3,2,3,5,7,6,7]10849; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm7[3,4,5,6,7]10850; AVX2-FCP-NEXT: vpermps %ymm5, %ymm1, %ymm510851; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm6[0,1,2,3,4,5],ymm5[6,7]10852; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10853; AVX2-FCP-NEXT: vpermps %ymm3, %ymm0, %ymm010854; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm12[1,3,2,3,5,7,6,7]10855; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3,4,5,6,7]10856; AVX2-FCP-NEXT: vpermps %ymm2, %ymm1, %ymm210857; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]10858; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10859; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload10860; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload10861; AVX2-FCP-NEXT: # ymm0 = ymm7[0,1],mem[2,3],ymm7[4,5],mem[6,7]10862; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm13 = [2,0,6,4,2,0,6,7]10863; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010864; AVX2-FCP-NEXT: vmovaps {{.*#+}} xmm12 = [2,0,6,7]10865; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10866; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10867; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10868; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10869; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm210870; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10871; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10872; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10873; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10874; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10875; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm15 = [0,0,6,4,0,0,6,4]10876; AVX2-FCP-NEXT: # ymm15 = mem[0,1,0,1]10877; AVX2-FCP-NEXT: vpermps %ymm2, %ymm15, %ymm210878; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10879; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10880; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload10881; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm0 # 32-byte Folded Reload10882; AVX2-FCP-NEXT: # ymm0 = ymm6[0,1],mem[2,3],ymm6[4,5],mem[6,7]10883; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010884; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10885; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10886; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10887; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10888; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm210889; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10890; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10891; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10892; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10893; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10894; AVX2-FCP-NEXT: vpermps %ymm2, %ymm15, %ymm210895; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10896; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10897; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10898; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10899; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10900; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010901; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10902; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10903; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10904; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10905; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm210906; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10907; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10908; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm8 # 32-byte Folded Reload10909; AVX2-FCP-NEXT: # ymm8 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10910; AVX2-FCP-NEXT: vpermps %ymm8, %ymm15, %ymm210911; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10912; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10913; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload10914; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload10915; AVX2-FCP-NEXT: # ymm0 = ymm4[0,1],mem[2,3],ymm4[4,5],mem[6,7]10916; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010917; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10918; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm14 # 32-byte Folded Reload10919; AVX2-FCP-NEXT: # ymm14 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10920; AVX2-FCP-NEXT: vpermps %ymm14, %ymm12, %ymm210921; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10922; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10923; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm11 # 32-byte Folded Reload10924; AVX2-FCP-NEXT: # ymm11 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10925; AVX2-FCP-NEXT: vpermps %ymm11, %ymm15, %ymm210926; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10927; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10928; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10929; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10930; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10931; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010932; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10933; AVX2-FCP-NEXT: vblendps $243, (%rsp), %ymm2, %ymm2 # 32-byte Folded Reload10934; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]10935; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10936; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm210937; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10938; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10939; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10940; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10941; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10942; AVX2-FCP-NEXT: vpermps %ymm2, %ymm15, %ymm210943; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10944; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10945; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10946; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10947; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10948; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010949; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10950; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10951; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10952; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10953; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm210954; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10955; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10956; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10957; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10958; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10959; AVX2-FCP-NEXT: vpermps %ymm2, %ymm15, %ymm210960; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]10961; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10962; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload10963; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload10964; AVX2-FCP-NEXT: # ymm0 = ymm10[0,1],mem[2,3],ymm10[4,5],mem[6,7]10965; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010966; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10967; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm9 # 32-byte Folded Reload10968; AVX2-FCP-NEXT: # ymm9 = ymm2[0,1],mem[2,3],ymm2[4,5,6,7]10969; AVX2-FCP-NEXT: vpermps %ymm9, %ymm12, %ymm210970; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7]10971; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10972; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10973; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4,5],ymm2[6,7]10974; AVX2-FCP-NEXT: vpermps %ymm2, %ymm15, %ymm510975; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm5[5,6,7]10976; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10977; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload10978; AVX2-FCP-NEXT: vblendps $204, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload10979; AVX2-FCP-NEXT: # ymm0 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]10980; AVX2-FCP-NEXT: vpermps %ymm0, %ymm13, %ymm010981; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload10982; AVX2-FCP-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm13 # 32-byte Folded Reload10983; AVX2-FCP-NEXT: # ymm13 = ymm5[0,1],mem[2,3],ymm5[4,5,6,7]10984; AVX2-FCP-NEXT: vpermps %ymm13, %ymm12, %ymm510985; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm0[3,4,5,6,7]10986; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10987; AVX2-FCP-NEXT: vblendps $48, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10988; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5],ymm0[6,7]10989; AVX2-FCP-NEXT: vpermps %ymm0, %ymm15, %ymm1210990; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm12[5,6,7]10991; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10992; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm5 = mem[3,3,3,3]10993; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2,3,4],ymm7[5],ymm5[6,7]10994; AVX2-FCP-NEXT: vblendps $222, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm12 # 32-byte Folded Reload10995; AVX2-FCP-NEXT: # ymm12 = ymm5[0],mem[1,2,3,4],ymm5[5],mem[6,7]10996; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm5 = [3,1,7,5,0,u,u,u]10997; AVX2-FCP-NEXT: vpermps %ymm12, %ymm5, %ymm1510998; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm12 = [0,1,7,5,0,1,7,5]10999; AVX2-FCP-NEXT: # ymm12 = mem[0,1,0,1]11000; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm7 # 32-byte Folded Reload11001; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm15[0,1,2,3,4],ymm7[5,6,7]11002; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11003; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm7 = mem[3,3,3,3]11004; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm6[1],ymm7[2,3,4],ymm6[5],ymm7[6,7]11005; AVX2-FCP-NEXT: vblendps $222, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload11006; AVX2-FCP-NEXT: # ymm7 = ymm7[0],mem[1,2,3,4],ymm7[5],mem[6,7]11007; AVX2-FCP-NEXT: vpermps %ymm7, %ymm5, %ymm711008; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm15 # 32-byte Folded Reload11009; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4],ymm15[5,6,7]11010; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11011; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm7 = mem[3,3,3,3]11012; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload11013; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm15[1],ymm7[2,3,4],ymm15[5],ymm7[6,7]11014; AVX2-FCP-NEXT: vblendps $222, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload11015; AVX2-FCP-NEXT: # ymm7 = ymm7[0],mem[1,2,3,4],ymm7[5],mem[6,7]11016; AVX2-FCP-NEXT: vpermps %ymm7, %ymm5, %ymm711017; AVX2-FCP-NEXT: vpermps %ymm8, %ymm12, %ymm811018; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4],ymm8[5,6,7]11019; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11020; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm7 = mem[3,3,3,3]11021; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm4[1],ymm7[2,3,4],ymm4[5],ymm7[6,7]11022; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm14[1,2,3,4],ymm7[5],ymm14[6,7]11023; AVX2-FCP-NEXT: vpermps %ymm7, %ymm5, %ymm711024; AVX2-FCP-NEXT: vpermps %ymm11, %ymm12, %ymm611025; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm7[0,1,2,3,4],ymm6[5,6,7]11026; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11027; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm6 = mem[3,3,3,3]11028; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm3[1],ymm6[2,3,4],ymm3[5],ymm6[6,7]11029; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm13[1,2,3,4],ymm6[5],ymm13[6,7]11030; AVX2-FCP-NEXT: vpermps %ymm6, %ymm5, %ymm611031; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm011032; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4],ymm0[5,6,7]11033; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11034; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm0 = mem[3,3,3,3]11035; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[2,3,4],ymm10[5],ymm0[6,7]11036; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm9[1,2,3,4],ymm0[5],ymm9[6,7]11037; AVX2-FCP-NEXT: vpermps %ymm0, %ymm5, %ymm011038; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm211039; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11040; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11041; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm0 = mem[3,3,3,3]11042; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload11043; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2,3,4],ymm7[5],ymm0[6,7]11044; AVX2-FCP-NEXT: vblendps $222, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload11045; AVX2-FCP-NEXT: # ymm0 = ymm0[0],mem[1,2,3,4],ymm0[5],mem[6,7]11046; AVX2-FCP-NEXT: vpermps %ymm0, %ymm5, %ymm011047; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm2 # 32-byte Folded Reload11048; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11049; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11050; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm0 # 32-byte Folded Reload11051; AVX2-FCP-NEXT: vpermilps {{.*#+}} xmm2 = mem[3,3,3,3]11052; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload11053; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3,4],ymm4[5],ymm2[6,7]11054; AVX2-FCP-NEXT: vblendps $222, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11055; AVX2-FCP-NEXT: # ymm2 = ymm2[0],mem[1,2,3,4],ymm2[5],mem[6,7]11056; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm211057; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6,7]11058; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11059; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm2 # 32-byte Folded Reload11060; AVX2-FCP-NEXT: # ymm2 = ymm4[0,1,2,3],mem[4,5,6,7]11061; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11062; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11063; AVX2-FCP-NEXT: vblendps $15, (%rsp), %ymm0, %ymm5 # 32-byte Folded Reload11064; AVX2-FCP-NEXT: # ymm5 = mem[0,1,2,3],ymm0[4,5,6,7]11065; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11066; AVX2-FCP-NEXT: vmovaps 80(%rdi), %xmm011067; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill11068; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]11069; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11070; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload11071; AVX2-FCP-NEXT: vpermps %ymm5, %ymm4, %ymm211072; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11073; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11074; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11075; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11076; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11077; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm5 = [0,2,0,6,0,2,0,6]11078; AVX2-FCP-NEXT: # ymm5 = mem[0,1,0,1]11079; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm211080; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11081; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11082; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11083; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload11084; AVX2-FCP-NEXT: # ymm2 = ymm0[0,1,2,3],mem[4,5,6,7]11085; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11086; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11087; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload11088; AVX2-FCP-NEXT: # ymm8 = ymm0[0,1,2,3],mem[4,5,6,7]11089; AVX2-FCP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11090; AVX2-FCP-NEXT: vmovaps 272(%rdi), %xmm011091; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11092; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]11093; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11094; AVX2-FCP-NEXT: vpermps %ymm8, %ymm4, %ymm211095; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11096; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11097; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11098; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11099; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11100; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm211101; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11102; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11103; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm2 # 32-byte Folded Reload11104; AVX2-FCP-NEXT: # ymm2 = ymm7[0,1,2,3],mem[4,5,6,7]11105; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11106; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11107; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload11108; AVX2-FCP-NEXT: # ymm7 = ymm0[0,1,2,3],mem[4,5,6,7]11109; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11110; AVX2-FCP-NEXT: vmovaps 464(%rdi), %xmm011111; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11112; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]11113; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11114; AVX2-FCP-NEXT: vpermps %ymm7, %ymm4, %ymm211115; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11116; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11117; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11118; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11119; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11120; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm211121; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11122; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11123; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11124; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload11125; AVX2-FCP-NEXT: # ymm2 = ymm0[0,1,2,3],mem[4,5,6,7]11126; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11127; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11128; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload11129; AVX2-FCP-NEXT: # ymm7 = ymm0[0,1,2,3],mem[4,5,6,7]11130; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11131; AVX2-FCP-NEXT: vmovaps 656(%rdi), %xmm011132; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11133; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]11134; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11135; AVX2-FCP-NEXT: vpermps %ymm7, %ymm4, %ymm211136; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11137; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11138; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11139; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11140; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11141; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm211142; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11143; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11144; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm2 # 32-byte Folded Reload11145; AVX2-FCP-NEXT: # ymm2 = ymm10[0,1,2,3],mem[4,5,6,7]11146; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11147; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11148; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload11149; AVX2-FCP-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]11150; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11151; AVX2-FCP-NEXT: vmovaps 848(%rdi), %xmm011152; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11153; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]11154; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11155; AVX2-FCP-NEXT: vpermps %ymm6, %ymm4, %ymm211156; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11157; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11158; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11159; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11160; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11161; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm211162; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11163; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11164; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload11165; AVX2-FCP-NEXT: # ymm15 = ymm15[0,1,2,3],mem[4,5,6,7]11166; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11167; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm14 # 32-byte Folded Reload11168; AVX2-FCP-NEXT: # ymm14 = ymm0[0,1,2,3],mem[4,5,6,7]11169; AVX2-FCP-NEXT: vmovaps 1040(%rdi), %xmm1311170; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1],ymm13[2,3],ymm15[4,5,6,7]11171; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11172; AVX2-FCP-NEXT: vpermps %ymm14, %ymm4, %ymm211173; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11174; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11175; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm12 # 32-byte Folded Reload11176; AVX2-FCP-NEXT: # ymm12 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11177; AVX2-FCP-NEXT: vpermps %ymm12, %ymm5, %ymm211178; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11179; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11180; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload11181; AVX2-FCP-NEXT: # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]11182; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11183; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload11184; AVX2-FCP-NEXT: # ymm10 = ymm0[0,1,2,3],mem[4,5,6,7]11185; AVX2-FCP-NEXT: vmovaps 1232(%rdi), %xmm911186; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0,1],ymm9[2,3],ymm11[4,5,6,7]11187; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11188; AVX2-FCP-NEXT: vpermps %ymm10, %ymm4, %ymm211189; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11190; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11191; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm8 # 32-byte Folded Reload11192; AVX2-FCP-NEXT: # ymm8 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11193; AVX2-FCP-NEXT: vpermps %ymm8, %ymm5, %ymm211194; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]11195; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11196; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11197; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload11198; AVX2-FCP-NEXT: # ymm7 = ymm0[0,1,2,3],mem[4,5,6,7]11199; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11200; AVX2-FCP-NEXT: vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload11201; AVX2-FCP-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]11202; AVX2-FCP-NEXT: vmovaps 1424(%rdi), %xmm311203; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1],ymm3[2,3],ymm7[4,5,6,7]11204; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]11205; AVX2-FCP-NEXT: vpermps %ymm6, %ymm4, %ymm211206; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3,4,5,6,7]11207; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload11208; AVX2-FCP-NEXT: vblendps $243, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload11209; AVX2-FCP-NEXT: # ymm2 = mem[0,1],ymm2[2,3],mem[4,5,6,7]11210; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm411211; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5,6,7]11212; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11213; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload11214; AVX2-FCP-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5]11215; AVX2-FCP-NEXT: vblendps $8, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload11216; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]11217; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm4 # 32-byte Folded Reload11218; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm0[2,3,4,5,6,7]11219; AVX2-FCP-NEXT: vbroadcastf128 {{.*#+}} ymm0 = [0,3,1,7,0,3,1,7]11220; AVX2-FCP-NEXT: # ymm0 = mem[0,1,0,1]11221; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload11222; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5,6,7]11223; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11224; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload11225; AVX2-FCP-NEXT: # ymm4 = mem[1,1,1,1,5,5,5,5]11226; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload11227; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]11228; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload11229; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]11230; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload11231; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5,6,7]11232; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11233; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload11234; AVX2-FCP-NEXT: # ymm4 = mem[1,1,1,1,5,5,5,5]11235; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload11236; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]11237; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload11238; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]11239; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload11240; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5,6,7]11241; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11242; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload11243; AVX2-FCP-NEXT: # ymm4 = mem[1,1,1,1,5,5,5,5]11244; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload11245; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]11246; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload11247; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]11248; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload11249; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5,6,7]11250; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11251; AVX2-FCP-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload11252; AVX2-FCP-NEXT: # ymm4 = mem[1,1,1,1,5,5,5,5]11253; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload11254; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]11255; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload11256; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3,4,5,6,7]11257; AVX2-FCP-NEXT: vpermps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload11258; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5,6,7]11259; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm15[1,1,1,1,5,5,5,5]11260; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm13[3],ymm5[4,5,6,7]11261; AVX2-FCP-NEXT: vpermps %ymm14, %ymm1, %ymm1311262; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm13[0,1],ymm5[2,3,4,5,6,7]11263; AVX2-FCP-NEXT: vpermps %ymm12, %ymm0, %ymm1211264; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm5[0,1,2,3,4],ymm12[5,6,7]11265; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm11[1,1,1,1,5,5,5,5]11266; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm9[3],ymm5[4,5,6,7]11267; AVX2-FCP-NEXT: vpermps %ymm10, %ymm1, %ymm911268; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm9[0,1],ymm5[2,3,4,5,6,7]11269; AVX2-FCP-NEXT: vpermps %ymm8, %ymm0, %ymm811270; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm5[0,1,2,3,4],ymm8[5,6,7]11271; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm7[1,1,1,1,5,5,5,5]11272; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1,2],ymm3[3],ymm5[4,5,6,7]11273; AVX2-FCP-NEXT: vpermps %ymm6, %ymm1, %ymm111274; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3,4,5,6,7]11275; AVX2-FCP-NEXT: vpermps %ymm2, %ymm0, %ymm011276; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6,7]11277; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11278; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rsi)11279; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11280; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rsi)11281; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11282; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rsi)11283; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11284; AVX2-FCP-NEXT: vmovaps %ymm1, (%rsi)11285; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11286; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rsi)11287; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11288; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rsi)11289; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11290; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rsi)11291; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11292; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rsi)11293; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11294; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rdx)11295; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11296; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rdx)11297; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11298; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rdx)11299; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11300; AVX2-FCP-NEXT: vmovaps %ymm1, (%rdx)11301; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11302; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rdx)11303; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11304; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rdx)11305; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11306; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rdx)11307; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11308; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rdx)11309; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11310; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rcx)11311; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11312; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rcx)11313; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11314; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rcx)11315; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11316; AVX2-FCP-NEXT: vmovaps %ymm1, (%rcx)11317; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11318; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rcx)11319; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11320; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rcx)11321; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11322; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rcx)11323; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11324; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rcx)11325; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11326; AVX2-FCP-NEXT: vmovaps %ymm1, (%r8)11327; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11328; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%r8)11329; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11330; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%r8)11331; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11332; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%r8)11333; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11334; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%r8)11335; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11336; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%r8)11337; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11338; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%r8)11339; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11340; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%r8)11341; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11342; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%r9)11343; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11344; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%r9)11345; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11346; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%r9)11347; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11348; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%r9)11349; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11350; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%r9)11351; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11352; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%r9)11353; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11354; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%r9)11355; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload11356; AVX2-FCP-NEXT: vmovaps %ymm1, (%r9)11357; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax11358; AVX2-FCP-NEXT: vmovaps %ymm0, 224(%rax)11359; AVX2-FCP-NEXT: vmovaps %ymm8, 192(%rax)11360; AVX2-FCP-NEXT: vmovaps %ymm12, 160(%rax)11361; AVX2-FCP-NEXT: vmovaps %ymm4, 128(%rax)11362; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11363; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rax)11364; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11365; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%rax)11366; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11367; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rax)11368; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11369; AVX2-FCP-NEXT: vmovaps %ymm0, (%rax)11370; AVX2-FCP-NEXT: addq $2536, %rsp # imm = 0x9E811371; AVX2-FCP-NEXT: vzeroupper11372; AVX2-FCP-NEXT: retq11373;11374; AVX512-LABEL: load_i32_stride6_vf64:11375; AVX512: # %bb.0:11376; AVX512-NEXT: subq $2632, %rsp # imm = 0xA4811377; AVX512-NEXT: vmovdqa64 1472(%rdi), %zmm2111378; AVX512-NEXT: vmovdqa64 1408(%rdi), %zmm111379; AVX512-NEXT: vmovdqa64 1088(%rdi), %zmm2011380; AVX512-NEXT: vmovdqa64 1024(%rdi), %zmm011381; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm211382; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm1811383; AVX512-NEXT: vmovdqa64 704(%rdi), %zmm2711384; AVX512-NEXT: vmovdqa64 640(%rdi), %zmm311385; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm2511386; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm411387; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]11388; AVX512-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]11389; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]11390; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]11391; AVX512-NEXT: vmovdqa64 %zmm4, %zmm711392; AVX512-NEXT: vpermt2d %zmm25, %zmm6, %zmm711393; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11394; AVX512-NEXT: vmovdqa64 %zmm3, %zmm711395; AVX512-NEXT: vpermt2d %zmm27, %zmm6, %zmm711396; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11397; AVX512-NEXT: vmovdqa64 %zmm18, %zmm711398; AVX512-NEXT: vpermt2d %zmm2, %zmm5, %zmm711399; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11400; AVX512-NEXT: vmovdqa64 %zmm0, %zmm711401; AVX512-NEXT: vpermt2d %zmm20, %zmm6, %zmm711402; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11403; AVX512-NEXT: vpermi2d %zmm21, %zmm1, %zmm611404; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11405; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]11406; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]11407; AVX512-NEXT: vmovdqa64 %zmm0, %zmm811408; AVX512-NEXT: vmovdqa64 %zmm0, %zmm611409; AVX512-NEXT: vpermt2d %zmm20, %zmm7, %zmm811410; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11411; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]11412; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]11413; AVX512-NEXT: vmovdqa64 %zmm3, %zmm011414; AVX512-NEXT: vpermt2d %zmm27, %zmm7, %zmm011415; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11416; AVX512-NEXT: vmovdqa64 %zmm4, %zmm011417; AVX512-NEXT: vpermt2d %zmm25, %zmm7, %zmm011418; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11419; AVX512-NEXT: vpermi2d %zmm21, %zmm1, %zmm711420; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11421; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]11422; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11423; AVX512-NEXT: vmovdqa64 %zmm20, %zmm711424; AVX512-NEXT: vpermt2d %zmm6, %zmm0, %zmm711425; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11426; AVX512-NEXT: vmovdqa64 %zmm27, %zmm711427; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm711428; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11429; AVX512-NEXT: vmovdqa64 %zmm25, %zmm711430; AVX512-NEXT: vpermt2d %zmm4, %zmm0, %zmm711431; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11432; AVX512-NEXT: vpermi2d %zmm1, %zmm21, %zmm011433; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11434; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]11435; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11436; AVX512-NEXT: vmovdqa64 %zmm20, %zmm711437; AVX512-NEXT: vpermt2d %zmm6, %zmm0, %zmm711438; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11439; AVX512-NEXT: vmovdqa64 %zmm27, %zmm711440; AVX512-NEXT: vpermt2d %zmm3, %zmm0, %zmm711441; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11442; AVX512-NEXT: vmovdqa64 %zmm25, %zmm711443; AVX512-NEXT: vpermt2d %zmm4, %zmm0, %zmm711444; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11445; AVX512-NEXT: vpermi2d %zmm1, %zmm21, %zmm011446; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11447; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]11448; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]11449; AVX512-NEXT: vmovdqa64 %zmm3, %zmm011450; AVX512-NEXT: vpermt2d %zmm27, %zmm7, %zmm011451; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11452; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]11453; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11454; AVX512-NEXT: vpermt2d %zmm27, %zmm0, %zmm311455; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11456; AVX512-NEXT: vmovdqa64 %zmm4, %zmm311457; AVX512-NEXT: vpermt2d %zmm25, %zmm7, %zmm311458; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11459; AVX512-NEXT: vpermt2d %zmm25, %zmm0, %zmm411460; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11461; AVX512-NEXT: vmovdqa64 %zmm6, %zmm311462; AVX512-NEXT: vpermt2d %zmm20, %zmm7, %zmm311463; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11464; AVX512-NEXT: vpermi2d %zmm21, %zmm1, %zmm711465; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11466; AVX512-NEXT: vpermt2d %zmm21, %zmm0, %zmm111467; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11468; AVX512-NEXT: vmovdqa64 %zmm18, %zmm111469; AVX512-NEXT: vpermt2d %zmm2, %zmm8, %zmm111470; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11471; AVX512-NEXT: vpermt2d %zmm20, %zmm0, %zmm611472; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11473; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]11474; AVX512-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]11475; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011476; AVX512-NEXT: vpermt2d %zmm18, %zmm31, %zmm011477; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11478; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]11479; AVX512-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]11480; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011481; AVX512-NEXT: vpermt2d %zmm18, %zmm25, %zmm011482; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11483; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]11484; AVX512-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]11485; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011486; AVX512-NEXT: vpermt2d %zmm18, %zmm3, %zmm011487; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11488; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]11489; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11490; AVX512-NEXT: vpermt2d %zmm18, %zmm0, %zmm211491; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11492; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm2611493; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm111494; AVX512-NEXT: vmovdqa64 %zmm1, %zmm211495; AVX512-NEXT: vpermt2d %zmm26, %zmm5, %zmm211496; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11497; AVX512-NEXT: vmovdqa64 %zmm1, %zmm211498; AVX512-NEXT: vpermt2d %zmm26, %zmm8, %zmm211499; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11500; AVX512-NEXT: vmovdqa64 %zmm26, %zmm211501; AVX512-NEXT: vpermt2d %zmm1, %zmm31, %zmm211502; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11503; AVX512-NEXT: vmovdqa64 %zmm26, %zmm211504; AVX512-NEXT: vpermt2d %zmm1, %zmm25, %zmm211505; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11506; AVX512-NEXT: vmovdqa64 %zmm26, %zmm211507; AVX512-NEXT: vpermt2d %zmm1, %zmm3, %zmm211508; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11509; AVX512-NEXT: vpermt2d %zmm1, %zmm0, %zmm2611510; AVX512-NEXT: vmovdqa64 896(%rdi), %zmm2211511; AVX512-NEXT: vmovdqa64 960(%rdi), %zmm111512; AVX512-NEXT: vmovdqa64 %zmm1, %zmm211513; AVX512-NEXT: vpermt2d %zmm22, %zmm5, %zmm211514; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11515; AVX512-NEXT: vmovdqa64 1280(%rdi), %zmm1911516; AVX512-NEXT: vmovdqa64 1344(%rdi), %zmm211517; AVX512-NEXT: vpermi2d %zmm19, %zmm2, %zmm511518; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11519; AVX512-NEXT: vmovdqa64 %zmm1, %zmm411520; AVX512-NEXT: vpermt2d %zmm22, %zmm8, %zmm411521; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11522; AVX512-NEXT: vpermi2d %zmm19, %zmm2, %zmm811523; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11524; AVX512-NEXT: vmovdqa64 %zmm22, %zmm2811525; AVX512-NEXT: vpermt2d %zmm1, %zmm31, %zmm2811526; AVX512-NEXT: vpermi2d %zmm2, %zmm19, %zmm3111527; AVX512-NEXT: vmovdqa64 %zmm22, %zmm2911528; AVX512-NEXT: vpermt2d %zmm1, %zmm25, %zmm2911529; AVX512-NEXT: vpermi2d %zmm2, %zmm19, %zmm2511530; AVX512-NEXT: vmovdqa64 %zmm22, %zmm411531; AVX512-NEXT: vpermt2d %zmm1, %zmm3, %zmm411532; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11533; AVX512-NEXT: vpermi2d %zmm2, %zmm19, %zmm311534; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11535; AVX512-NEXT: vpermt2d %zmm2, %zmm0, %zmm1911536; AVX512-NEXT: vpermt2d %zmm1, %zmm0, %zmm2211537; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm211538; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm011539; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]11540; AVX512-NEXT: vmovdqa64 %zmm0, %zmm2011541; AVX512-NEXT: vpermt2d %zmm2, %zmm3, %zmm2011542; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]11543; AVX512-NEXT: vmovdqa64 %zmm0, %zmm2111544; AVX512-NEXT: vpermt2d %zmm2, %zmm4, %zmm2111545; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]11546; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111547; AVX512-NEXT: vpermt2d %zmm2, %zmm12, %zmm111548; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11549; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]11550; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111551; AVX512-NEXT: vpermt2d %zmm2, %zmm14, %zmm111552; AVX512-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill11553; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]11554; AVX512-NEXT: vmovdqa64 %zmm2, %zmm2311555; AVX512-NEXT: vpermt2d %zmm0, %zmm5, %zmm2311556; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]11557; AVX512-NEXT: vpermt2d %zmm0, %zmm24, %zmm211558; AVX512-NEXT: vmovdqa64 (%rdi), %zmm011559; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm611560; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1511561; AVX512-NEXT: vpermt2d %zmm6, %zmm3, %zmm1511562; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1611563; AVX512-NEXT: vpermt2d %zmm6, %zmm4, %zmm1611564; AVX512-NEXT: vmovdqa64 %zmm0, %zmm2711565; AVX512-NEXT: vpermt2d %zmm6, %zmm12, %zmm2711566; AVX512-NEXT: vmovdqa64 %zmm0, %zmm3011567; AVX512-NEXT: vpermt2d %zmm6, %zmm14, %zmm3011568; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1711569; AVX512-NEXT: vpermt2d %zmm0, %zmm5, %zmm1711570; AVX512-NEXT: vpermt2d %zmm0, %zmm24, %zmm611571; AVX512-NEXT: vmovdqa64 832(%rdi), %zmm1011572; AVX512-NEXT: vmovdqa64 768(%rdi), %zmm711573; AVX512-NEXT: vmovdqa64 %zmm7, %zmm811574; AVX512-NEXT: vpermt2d %zmm10, %zmm3, %zmm811575; AVX512-NEXT: vmovdqa64 1216(%rdi), %zmm111576; AVX512-NEXT: vmovdqa64 1152(%rdi), %zmm011577; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm311578; AVX512-NEXT: vmovdqa64 %zmm7, %zmm911579; AVX512-NEXT: vpermt2d %zmm10, %zmm4, %zmm911580; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm411581; AVX512-NEXT: vmovdqa64 %zmm7, %zmm1311582; AVX512-NEXT: vpermt2d %zmm10, %zmm12, %zmm1311583; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm1211584; AVX512-NEXT: vmovdqa64 %zmm7, %zmm1811585; AVX512-NEXT: vpermt2d %zmm10, %zmm14, %zmm1811586; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm1411587; AVX512-NEXT: vmovdqa64 %zmm10, %zmm1111588; AVX512-NEXT: vpermt2d %zmm7, %zmm5, %zmm1111589; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm511590; AVX512-NEXT: vpermt2d %zmm0, %zmm24, %zmm111591; AVX512-NEXT: vpermt2d %zmm7, %zmm24, %zmm1011592; AVX512-NEXT: movb $56, %al11593; AVX512-NEXT: kmovw %eax, %k211594; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11595; AVX512-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}11596; AVX512-NEXT: movw $-2048, %ax # imm = 0xF80011597; AVX512-NEXT: kmovw %eax, %k111598; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11599; AVX512-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}11600; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11601; AVX512-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}11602; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11603; AVX512-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}11604; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11605; AVX512-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}11606; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11607; AVX512-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}11608; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11609; AVX512-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}11610; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11611; AVX512-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}11612; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11613; AVX512-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}11614; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11615; AVX512-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}11616; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11617; AVX512-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}11618; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11619; AVX512-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}11620; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11621; AVX512-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}11622; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11623; AVX512-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}11624; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11625; AVX512-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}11626; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11627; AVX512-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}11628; AVX512-NEXT: movw $31, %ax11629; AVX512-NEXT: kmovw %eax, %k211630; AVX512-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}11631; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11632; AVX512-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}11633; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload11634; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11635; AVX512-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}11636; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11637; AVX512-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}11638; AVX512-NEXT: vmovdqa64 %zmm7, %zmm1311639; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload11640; AVX512-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}11641; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11642; AVX512-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}11643; AVX512-NEXT: vmovdqa64 %zmm7, %zmm2411644; AVX512-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}11645; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11646; AVX512-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}11647; AVX512-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}11648; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11649; AVX512-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}11650; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload11651; AVX512-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload11652; AVX512-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}11653; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11654; AVX512-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}11655; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload11656; AVX512-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}11657; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11658; AVX512-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}11659; AVX512-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}11660; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11661; AVX512-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}11662; AVX512-NEXT: movw $992, %ax # imm = 0x3E011663; AVX512-NEXT: kmovw %eax, %k111664; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11665; AVX512-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}11666; AVX512-NEXT: movb $-32, %al11667; AVX512-NEXT: kmovw %eax, %k211668; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11669; AVX512-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}11670; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11671; AVX512-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}11672; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11673; AVX512-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}11674; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11675; AVX512-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}11676; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11677; AVX512-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}11678; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11679; AVX512-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}11680; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11681; AVX512-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}11682; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11683; AVX512-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}11684; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11685; AVX512-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}11686; AVX512-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}11687; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11688; AVX512-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}11689; AVX512-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}11690; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11691; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}11692; AVX512-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}11693; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11694; AVX512-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}11695; AVX512-NEXT: vmovdqa64 %zmm3, 192(%rsi)11696; AVX512-NEXT: vmovdqa64 %zmm8, 128(%rsi)11697; AVX512-NEXT: vmovdqa64 %zmm20, 64(%rsi)11698; AVX512-NEXT: vmovdqa64 %zmm15, (%rsi)11699; AVX512-NEXT: vmovdqa64 %zmm4, 192(%rdx)11700; AVX512-NEXT: vmovdqa64 %zmm16, (%rdx)11701; AVX512-NEXT: vmovdqa64 %zmm21, 64(%rdx)11702; AVX512-NEXT: vmovdqa64 %zmm9, 128(%rdx)11703; AVX512-NEXT: vmovdqa64 %zmm31, 192(%rcx)11704; AVX512-NEXT: vmovdqa64 %zmm24, (%rcx)11705; AVX512-NEXT: vmovdqa64 %zmm13, 64(%rcx)11706; AVX512-NEXT: vmovdqa64 %zmm28, 128(%rcx)11707; AVX512-NEXT: vmovdqa64 %zmm25, 192(%r8)11708; AVX512-NEXT: vmovdqa64 %zmm12, (%r8)11709; AVX512-NEXT: vmovdqa64 %zmm7, 64(%r8)11710; AVX512-NEXT: vmovdqa64 %zmm29, 128(%r8)11711; AVX512-NEXT: vmovdqa64 %zmm5, 192(%r9)11712; AVX512-NEXT: vmovdqa64 %zmm17, (%r9)11713; AVX512-NEXT: vmovdqa64 %zmm23, 64(%r9)11714; AVX512-NEXT: vmovdqa64 %zmm11, 128(%r9)11715; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax11716; AVX512-NEXT: vmovdqa64 %zmm10, 128(%rax)11717; AVX512-NEXT: vmovdqa64 %zmm1, 192(%rax)11718; AVX512-NEXT: vmovdqa64 %zmm6, (%rax)11719; AVX512-NEXT: vmovdqa64 %zmm2, 64(%rax)11720; AVX512-NEXT: addq $2632, %rsp # imm = 0xA4811721; AVX512-NEXT: vzeroupper11722; AVX512-NEXT: retq11723;11724; AVX512-FCP-LABEL: load_i32_stride6_vf64:11725; AVX512-FCP: # %bb.0:11726; AVX512-FCP-NEXT: subq $2632, %rsp # imm = 0xA4811727; AVX512-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm2111728; AVX512-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm111729; AVX512-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm2011730; AVX512-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm011731; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm211732; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1811733; AVX512-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2711734; AVX512-FCP-NEXT: vmovdqa64 640(%rdi), %zmm311735; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm2511736; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm411737; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]11738; AVX512-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]11739; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]11740; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]11741; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm711742; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm6, %zmm711743; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11744; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm711745; AVX512-FCP-NEXT: vpermt2d %zmm27, %zmm6, %zmm711746; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11747; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm711748; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm711749; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11750; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm711751; AVX512-FCP-NEXT: vpermt2d %zmm20, %zmm6, %zmm711752; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11753; AVX512-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm611754; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11755; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]11756; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]11757; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm811758; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm611759; AVX512-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm811760; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11761; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]11762; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]11763; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm011764; AVX512-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm011765; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11766; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm011767; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm011768; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11769; AVX512-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm711770; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11771; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]11772; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11773; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm711774; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm711775; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11776; AVX512-FCP-NEXT: vmovdqa64 %zmm27, %zmm711777; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm711778; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11779; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm711780; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm711781; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11782; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm011783; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11784; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]11785; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11786; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm711787; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm711788; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11789; AVX512-FCP-NEXT: vmovdqa64 %zmm27, %zmm711790; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm711791; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11792; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm711793; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm711794; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11795; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm011796; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11797; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]11798; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]11799; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm011800; AVX512-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm011801; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11802; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]11803; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11804; AVX512-FCP-NEXT: vpermt2d %zmm27, %zmm0, %zmm311805; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11806; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm311807; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm311808; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11809; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm0, %zmm411810; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11811; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm311812; AVX512-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm311813; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11814; AVX512-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm711815; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11816; AVX512-FCP-NEXT: vpermt2d %zmm21, %zmm0, %zmm111817; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11818; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm111819; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm8, %zmm111820; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11821; AVX512-FCP-NEXT: vpermt2d %zmm20, %zmm0, %zmm611822; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11823; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]11824; AVX512-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]11825; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011826; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm31, %zmm011827; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11828; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]11829; AVX512-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]11830; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011831; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm25, %zmm011832; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11833; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]11834; AVX512-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]11835; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011836; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm3, %zmm011837; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11838; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]11839; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]11840; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm0, %zmm211841; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11842; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm2611843; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm111844; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm211845; AVX512-FCP-NEXT: vpermt2d %zmm26, %zmm5, %zmm211846; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11847; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm211848; AVX512-FCP-NEXT: vpermt2d %zmm26, %zmm8, %zmm211849; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11850; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm211851; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm211852; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11853; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm211854; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm211855; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11856; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm211857; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm211858; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11859; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2611860; AVX512-FCP-NEXT: vmovdqa64 896(%rdi), %zmm2211861; AVX512-FCP-NEXT: vmovdqa64 960(%rdi), %zmm111862; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm211863; AVX512-FCP-NEXT: vpermt2d %zmm22, %zmm5, %zmm211864; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11865; AVX512-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm1911866; AVX512-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm211867; AVX512-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm511868; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11869; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm411870; AVX512-FCP-NEXT: vpermt2d %zmm22, %zmm8, %zmm411871; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11872; AVX512-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm811873; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11874; AVX512-FCP-NEXT: vmovdqa64 %zmm22, %zmm2811875; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm2811876; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm3111877; AVX512-FCP-NEXT: vmovdqa64 %zmm22, %zmm2911878; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm2911879; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm2511880; AVX512-FCP-NEXT: vmovdqa64 %zmm22, %zmm411881; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm411882; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11883; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm311884; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11885; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm0, %zmm1911886; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2211887; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm211888; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm011889; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]11890; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm2011891; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm2011892; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]11893; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm2111894; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm4, %zmm2111895; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]11896; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111897; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm12, %zmm111898; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11899; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]11900; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111901; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm14, %zmm111902; AVX512-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill11903; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]11904; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm2311905; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm2311906; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]11907; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm211908; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm011909; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm611910; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1511911; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm3, %zmm1511912; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1611913; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm1611914; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm2711915; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm12, %zmm2711916; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm3011917; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm14, %zmm3011918; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm1711919; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm1711920; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm611921; AVX512-FCP-NEXT: vmovdqa64 832(%rdi), %zmm1011922; AVX512-FCP-NEXT: vmovdqa64 768(%rdi), %zmm711923; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm811924; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm3, %zmm811925; AVX512-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm111926; AVX512-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm011927; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm311928; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm911929; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm4, %zmm911930; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm411931; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm1311932; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm12, %zmm1311933; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1211934; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm1811935; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm14, %zmm1811936; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1411937; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm1111938; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm1111939; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm511940; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm111941; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm24, %zmm1011942; AVX512-FCP-NEXT: movb $56, %al11943; AVX512-FCP-NEXT: kmovw %eax, %k211944; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11945; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}11946; AVX512-FCP-NEXT: movw $-2048, %ax # imm = 0xF80011947; AVX512-FCP-NEXT: kmovw %eax, %k111948; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11949; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}11950; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11951; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}11952; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11953; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}11954; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11955; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}11956; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11957; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}11958; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11959; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}11960; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11961; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}11962; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11963; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}11964; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11965; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}11966; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11967; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}11968; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11969; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}11970; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11971; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}11972; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11973; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}11974; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11975; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}11976; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11977; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}11978; AVX512-FCP-NEXT: movw $31, %ax11979; AVX512-FCP-NEXT: kmovw %eax, %k211980; AVX512-FCP-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}11981; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11982; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}11983; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload11984; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11985; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}11986; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11987; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}11988; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm1311989; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload11990; AVX512-FCP-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}11991; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11992; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}11993; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm2411994; AVX512-FCP-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}11995; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11996; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}11997; AVX512-FCP-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}11998; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11999; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}12000; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12001; AVX512-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload12002; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}12003; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12004; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12005; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload12006; AVX512-FCP-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}12007; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12008; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}12009; AVX512-FCP-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}12010; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12011; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}12012; AVX512-FCP-NEXT: movw $992, %ax # imm = 0x3E012013; AVX512-FCP-NEXT: kmovw %eax, %k112014; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12015; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}12016; AVX512-FCP-NEXT: movb $-32, %al12017; AVX512-FCP-NEXT: kmovw %eax, %k212018; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12019; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}12020; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12021; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}12022; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12023; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}12024; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12025; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}12026; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12027; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}12028; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12029; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}12030; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12031; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}12032; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12033; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}12034; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12035; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}12036; AVX512-FCP-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}12037; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12038; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}12039; AVX512-FCP-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}12040; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12041; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}12042; AVX512-FCP-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}12043; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12044; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}12045; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 192(%rsi)12046; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 128(%rsi)12047; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 64(%rsi)12048; AVX512-FCP-NEXT: vmovdqa64 %zmm15, (%rsi)12049; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 192(%rdx)12050; AVX512-FCP-NEXT: vmovdqa64 %zmm16, (%rdx)12051; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 64(%rdx)12052; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 128(%rdx)12053; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 192(%rcx)12054; AVX512-FCP-NEXT: vmovdqa64 %zmm24, (%rcx)12055; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 64(%rcx)12056; AVX512-FCP-NEXT: vmovdqa64 %zmm28, 128(%rcx)12057; AVX512-FCP-NEXT: vmovdqa64 %zmm25, 192(%r8)12058; AVX512-FCP-NEXT: vmovdqa64 %zmm12, (%r8)12059; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 64(%r8)12060; AVX512-FCP-NEXT: vmovdqa64 %zmm29, 128(%r8)12061; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 192(%r9)12062; AVX512-FCP-NEXT: vmovdqa64 %zmm17, (%r9)12063; AVX512-FCP-NEXT: vmovdqa64 %zmm23, 64(%r9)12064; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 128(%r9)12065; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax12066; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)12067; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)12068; AVX512-FCP-NEXT: vmovdqa64 %zmm6, (%rax)12069; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)12070; AVX512-FCP-NEXT: addq $2632, %rsp # imm = 0xA4812071; AVX512-FCP-NEXT: vzeroupper12072; AVX512-FCP-NEXT: retq12073;12074; AVX512DQ-LABEL: load_i32_stride6_vf64:12075; AVX512DQ: # %bb.0:12076; AVX512DQ-NEXT: subq $2632, %rsp # imm = 0xA4812077; AVX512DQ-NEXT: vmovdqa64 1472(%rdi), %zmm2112078; AVX512DQ-NEXT: vmovdqa64 1408(%rdi), %zmm112079; AVX512DQ-NEXT: vmovdqa64 1088(%rdi), %zmm2012080; AVX512DQ-NEXT: vmovdqa64 1024(%rdi), %zmm012081; AVX512DQ-NEXT: vmovdqa64 512(%rdi), %zmm212082; AVX512DQ-NEXT: vmovdqa64 576(%rdi), %zmm1812083; AVX512DQ-NEXT: vmovdqa64 704(%rdi), %zmm2712084; AVX512DQ-NEXT: vmovdqa64 640(%rdi), %zmm312085; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm2512086; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm412087; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]12088; AVX512DQ-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]12089; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]12090; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12091; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm712092; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm6, %zmm712093; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12094; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm712095; AVX512DQ-NEXT: vpermt2d %zmm27, %zmm6, %zmm712096; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12097; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm712098; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm5, %zmm712099; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12100; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm712101; AVX512DQ-NEXT: vpermt2d %zmm20, %zmm6, %zmm712102; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12103; AVX512DQ-NEXT: vpermi2d %zmm21, %zmm1, %zmm612104; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12105; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]12106; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]12107; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm812108; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm612109; AVX512DQ-NEXT: vpermt2d %zmm20, %zmm7, %zmm812110; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12111; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]12112; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]12113; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm012114; AVX512DQ-NEXT: vpermt2d %zmm27, %zmm7, %zmm012115; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12116; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm012117; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm7, %zmm012118; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12119; AVX512DQ-NEXT: vpermi2d %zmm21, %zmm1, %zmm712120; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12121; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]12122; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12123; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm712124; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm0, %zmm712125; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12126; AVX512DQ-NEXT: vmovdqa64 %zmm27, %zmm712127; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm0, %zmm712128; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12129; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm712130; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm0, %zmm712131; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12132; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm21, %zmm012133; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12134; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]12135; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12136; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm712137; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm0, %zmm712138; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12139; AVX512DQ-NEXT: vmovdqa64 %zmm27, %zmm712140; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm0, %zmm712141; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12142; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm712143; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm0, %zmm712144; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12145; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm21, %zmm012146; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12147; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]12148; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]12149; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm012150; AVX512DQ-NEXT: vpermt2d %zmm27, %zmm7, %zmm012151; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12152; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]12153; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12154; AVX512DQ-NEXT: vpermt2d %zmm27, %zmm0, %zmm312155; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12156; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm312157; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm7, %zmm312158; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12159; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm0, %zmm412160; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12161; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm312162; AVX512DQ-NEXT: vpermt2d %zmm20, %zmm7, %zmm312163; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12164; AVX512DQ-NEXT: vpermi2d %zmm21, %zmm1, %zmm712165; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12166; AVX512DQ-NEXT: vpermt2d %zmm21, %zmm0, %zmm112167; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12168; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm112169; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm8, %zmm112170; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12171; AVX512DQ-NEXT: vpermt2d %zmm20, %zmm0, %zmm612172; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12173; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]12174; AVX512DQ-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]12175; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012176; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm31, %zmm012177; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12178; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]12179; AVX512DQ-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]12180; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012181; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm25, %zmm012182; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12183; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]12184; AVX512DQ-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]12185; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012186; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm3, %zmm012187; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12188; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]12189; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12190; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm0, %zmm212191; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12192; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm2612193; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm112194; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm212195; AVX512DQ-NEXT: vpermt2d %zmm26, %zmm5, %zmm212196; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12197; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm212198; AVX512DQ-NEXT: vpermt2d %zmm26, %zmm8, %zmm212199; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12200; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm212201; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm31, %zmm212202; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12203; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm212204; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm25, %zmm212205; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12206; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm212207; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm3, %zmm212208; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12209; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm0, %zmm2612210; AVX512DQ-NEXT: vmovdqa64 896(%rdi), %zmm2212211; AVX512DQ-NEXT: vmovdqa64 960(%rdi), %zmm112212; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm212213; AVX512DQ-NEXT: vpermt2d %zmm22, %zmm5, %zmm212214; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12215; AVX512DQ-NEXT: vmovdqa64 1280(%rdi), %zmm1912216; AVX512DQ-NEXT: vmovdqa64 1344(%rdi), %zmm212217; AVX512DQ-NEXT: vpermi2d %zmm19, %zmm2, %zmm512218; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12219; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm412220; AVX512DQ-NEXT: vpermt2d %zmm22, %zmm8, %zmm412221; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12222; AVX512DQ-NEXT: vpermi2d %zmm19, %zmm2, %zmm812223; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12224; AVX512DQ-NEXT: vmovdqa64 %zmm22, %zmm2812225; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm31, %zmm2812226; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm19, %zmm3112227; AVX512DQ-NEXT: vmovdqa64 %zmm22, %zmm2912228; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm25, %zmm2912229; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm19, %zmm2512230; AVX512DQ-NEXT: vmovdqa64 %zmm22, %zmm412231; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm3, %zmm412232; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12233; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm19, %zmm312234; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12235; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm0, %zmm1912236; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm0, %zmm2212237; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm212238; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm012239; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]12240; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm2012241; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm3, %zmm2012242; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]12243; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm2112244; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm4, %zmm2112245; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]12246; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm112247; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm12, %zmm112248; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12249; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]12250; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm112251; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm14, %zmm112252; AVX512DQ-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill12253; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]12254; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm2312255; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm5, %zmm2312256; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]12257; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm24, %zmm212258; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm012259; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm612260; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1512261; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm3, %zmm1512262; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1612263; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm4, %zmm1612264; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm2712265; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm12, %zmm2712266; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm3012267; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm14, %zmm3012268; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm1712269; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm5, %zmm1712270; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm24, %zmm612271; AVX512DQ-NEXT: vmovdqa64 832(%rdi), %zmm1012272; AVX512DQ-NEXT: vmovdqa64 768(%rdi), %zmm712273; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm812274; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm3, %zmm812275; AVX512DQ-NEXT: vmovdqa64 1216(%rdi), %zmm112276; AVX512DQ-NEXT: vmovdqa64 1152(%rdi), %zmm012277; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm312278; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm912279; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm4, %zmm912280; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm412281; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm1312282; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm12, %zmm1312283; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm1212284; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm1812285; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm14, %zmm1812286; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm1412287; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm1112288; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm5, %zmm1112289; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm512290; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm24, %zmm112291; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm24, %zmm1012292; AVX512DQ-NEXT: movb $56, %al12293; AVX512DQ-NEXT: kmovw %eax, %k212294; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12295; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}12296; AVX512DQ-NEXT: movw $-2048, %ax # imm = 0xF80012297; AVX512DQ-NEXT: kmovw %eax, %k112298; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12299; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}12300; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12301; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}12302; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12303; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}12304; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12305; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}12306; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12307; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}12308; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12309; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}12310; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12311; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}12312; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12313; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}12314; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12315; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}12316; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12317; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}12318; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12319; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}12320; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12321; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}12322; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12323; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}12324; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12325; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}12326; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12327; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}12328; AVX512DQ-NEXT: movw $31, %ax12329; AVX512DQ-NEXT: kmovw %eax, %k212330; AVX512DQ-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}12331; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12332; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}12333; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12334; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12335; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}12336; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12337; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12338; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm1312339; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12340; AVX512DQ-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}12341; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12342; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12343; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm2412344; AVX512DQ-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}12345; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12346; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}12347; AVX512DQ-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}12348; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12349; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}12350; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12351; AVX512DQ-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload12352; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}12353; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12354; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12355; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload12356; AVX512DQ-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}12357; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12358; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}12359; AVX512DQ-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}12360; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12361; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}12362; AVX512DQ-NEXT: movw $992, %ax # imm = 0x3E012363; AVX512DQ-NEXT: kmovw %eax, %k112364; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12365; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}12366; AVX512DQ-NEXT: movb $-32, %al12367; AVX512DQ-NEXT: kmovw %eax, %k212368; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12369; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}12370; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12371; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}12372; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12373; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}12374; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12375; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}12376; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12377; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}12378; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12379; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}12380; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12381; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}12382; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12383; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}12384; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12385; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}12386; AVX512DQ-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}12387; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12388; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}12389; AVX512DQ-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}12390; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12391; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}12392; AVX512DQ-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}12393; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12394; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}12395; AVX512DQ-NEXT: vmovdqa64 %zmm3, 192(%rsi)12396; AVX512DQ-NEXT: vmovdqa64 %zmm8, 128(%rsi)12397; AVX512DQ-NEXT: vmovdqa64 %zmm20, 64(%rsi)12398; AVX512DQ-NEXT: vmovdqa64 %zmm15, (%rsi)12399; AVX512DQ-NEXT: vmovdqa64 %zmm4, 192(%rdx)12400; AVX512DQ-NEXT: vmovdqa64 %zmm16, (%rdx)12401; AVX512DQ-NEXT: vmovdqa64 %zmm21, 64(%rdx)12402; AVX512DQ-NEXT: vmovdqa64 %zmm9, 128(%rdx)12403; AVX512DQ-NEXT: vmovdqa64 %zmm31, 192(%rcx)12404; AVX512DQ-NEXT: vmovdqa64 %zmm24, (%rcx)12405; AVX512DQ-NEXT: vmovdqa64 %zmm13, 64(%rcx)12406; AVX512DQ-NEXT: vmovdqa64 %zmm28, 128(%rcx)12407; AVX512DQ-NEXT: vmovdqa64 %zmm25, 192(%r8)12408; AVX512DQ-NEXT: vmovdqa64 %zmm12, (%r8)12409; AVX512DQ-NEXT: vmovdqa64 %zmm7, 64(%r8)12410; AVX512DQ-NEXT: vmovdqa64 %zmm29, 128(%r8)12411; AVX512DQ-NEXT: vmovdqa64 %zmm5, 192(%r9)12412; AVX512DQ-NEXT: vmovdqa64 %zmm17, (%r9)12413; AVX512DQ-NEXT: vmovdqa64 %zmm23, 64(%r9)12414; AVX512DQ-NEXT: vmovdqa64 %zmm11, 128(%r9)12415; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax12416; AVX512DQ-NEXT: vmovdqa64 %zmm10, 128(%rax)12417; AVX512DQ-NEXT: vmovdqa64 %zmm1, 192(%rax)12418; AVX512DQ-NEXT: vmovdqa64 %zmm6, (%rax)12419; AVX512DQ-NEXT: vmovdqa64 %zmm2, 64(%rax)12420; AVX512DQ-NEXT: addq $2632, %rsp # imm = 0xA4812421; AVX512DQ-NEXT: vzeroupper12422; AVX512DQ-NEXT: retq12423;12424; AVX512DQ-FCP-LABEL: load_i32_stride6_vf64:12425; AVX512DQ-FCP: # %bb.0:12426; AVX512DQ-FCP-NEXT: subq $2632, %rsp # imm = 0xA4812427; AVX512DQ-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm2112428; AVX512DQ-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm112429; AVX512DQ-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm2012430; AVX512DQ-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm012431; AVX512DQ-FCP-NEXT: vmovdqa64 512(%rdi), %zmm212432; AVX512DQ-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1812433; AVX512DQ-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2712434; AVX512DQ-FCP-NEXT: vmovdqa64 640(%rdi), %zmm312435; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm2512436; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm412437; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]12438; AVX512DQ-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]12439; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]12440; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12441; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm712442; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm6, %zmm712443; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12444; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm712445; AVX512DQ-FCP-NEXT: vpermt2d %zmm27, %zmm6, %zmm712446; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12447; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm712448; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm712449; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12450; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm712451; AVX512DQ-FCP-NEXT: vpermt2d %zmm20, %zmm6, %zmm712452; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12453; AVX512DQ-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm612454; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12455; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]12456; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]12457; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm812458; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm612459; AVX512DQ-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm812460; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12461; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]12462; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]12463; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm012464; AVX512DQ-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm012465; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12466; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm012467; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm012468; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12469; AVX512DQ-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm712470; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12471; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]12472; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12473; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm712474; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm712475; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12476; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, %zmm712477; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm712478; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12479; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm712480; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm712481; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12482; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm012483; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12484; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]12485; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12486; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm712487; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm712488; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12489; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, %zmm712490; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm712491; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12492; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm712493; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm712494; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12495; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm012496; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12497; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]12498; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]12499; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm012500; AVX512DQ-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm012501; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12502; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]12503; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12504; AVX512DQ-FCP-NEXT: vpermt2d %zmm27, %zmm0, %zmm312505; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12506; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm312507; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm312508; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12509; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm0, %zmm412510; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12511; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm312512; AVX512DQ-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm312513; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12514; AVX512DQ-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm712515; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12516; AVX512DQ-FCP-NEXT: vpermt2d %zmm21, %zmm0, %zmm112517; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12518; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm112519; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm8, %zmm112520; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12521; AVX512DQ-FCP-NEXT: vpermt2d %zmm20, %zmm0, %zmm612522; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12523; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]12524; AVX512DQ-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]12525; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012526; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm31, %zmm012527; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12528; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]12529; AVX512DQ-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]12530; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012531; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm25, %zmm012532; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12533; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]12534; AVX512DQ-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]12535; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012536; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm3, %zmm012537; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12538; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]12539; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12540; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm0, %zmm212541; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12542; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm2612543; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm112544; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm212545; AVX512DQ-FCP-NEXT: vpermt2d %zmm26, %zmm5, %zmm212546; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12547; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm212548; AVX512DQ-FCP-NEXT: vpermt2d %zmm26, %zmm8, %zmm212549; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12550; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm212551; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm212552; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12553; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm212554; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm212555; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12556; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm212557; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm212558; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12559; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2612560; AVX512DQ-FCP-NEXT: vmovdqa64 896(%rdi), %zmm2212561; AVX512DQ-FCP-NEXT: vmovdqa64 960(%rdi), %zmm112562; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm212563; AVX512DQ-FCP-NEXT: vpermt2d %zmm22, %zmm5, %zmm212564; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12565; AVX512DQ-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm1912566; AVX512DQ-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm212567; AVX512DQ-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm512568; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12569; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm412570; AVX512DQ-FCP-NEXT: vpermt2d %zmm22, %zmm8, %zmm412571; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12572; AVX512DQ-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm812573; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12574; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, %zmm2812575; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm2812576; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm3112577; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, %zmm2912578; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm2912579; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm2512580; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, %zmm412581; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm412582; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12583; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm312584; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12585; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm0, %zmm1912586; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2212587; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm212588; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm012589; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]12590; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2012591; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm2012592; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]12593; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2112594; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm4, %zmm2112595; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]12596; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112597; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm12, %zmm112598; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12599; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]12600; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112601; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm14, %zmm112602; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill12603; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]12604; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm2312605; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm2312606; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]12607; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm212608; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm012609; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm612610; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1512611; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm3, %zmm1512612; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1612613; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm1612614; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2712615; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm12, %zmm2712616; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm3012617; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm14, %zmm3012618; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm1712619; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm1712620; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm612621; AVX512DQ-FCP-NEXT: vmovdqa64 832(%rdi), %zmm1012622; AVX512DQ-FCP-NEXT: vmovdqa64 768(%rdi), %zmm712623; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm812624; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm3, %zmm812625; AVX512DQ-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm112626; AVX512DQ-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm012627; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm312628; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm912629; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm4, %zmm912630; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm412631; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm1312632; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm12, %zmm1312633; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1212634; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm1812635; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm14, %zmm1812636; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1412637; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm1112638; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm1112639; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm512640; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm112641; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm24, %zmm1012642; AVX512DQ-FCP-NEXT: movb $56, %al12643; AVX512DQ-FCP-NEXT: kmovw %eax, %k212644; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12645; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}12646; AVX512DQ-FCP-NEXT: movw $-2048, %ax # imm = 0xF80012647; AVX512DQ-FCP-NEXT: kmovw %eax, %k112648; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12649; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}12650; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12651; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}12652; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12653; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}12654; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12655; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}12656; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12657; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}12658; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12659; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}12660; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12661; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}12662; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12663; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}12664; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12665; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}12666; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12667; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}12668; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12669; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}12670; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12671; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}12672; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12673; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}12674; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12675; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}12676; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12677; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}12678; AVX512DQ-FCP-NEXT: movw $31, %ax12679; AVX512DQ-FCP-NEXT: kmovw %eax, %k212680; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}12681; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12682; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}12683; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12684; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12685; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}12686; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12687; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12688; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm1312689; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12690; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}12691; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12692; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12693; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm2412694; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}12695; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12696; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}12697; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}12698; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12699; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}12700; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload12701; AVX512DQ-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload12702; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}12703; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12704; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}12705; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload12706; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}12707; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12708; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}12709; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}12710; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12711; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}12712; AVX512DQ-FCP-NEXT: movw $992, %ax # imm = 0x3E012713; AVX512DQ-FCP-NEXT: kmovw %eax, %k112714; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12715; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}12716; AVX512DQ-FCP-NEXT: movb $-32, %al12717; AVX512DQ-FCP-NEXT: kmovw %eax, %k212718; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12719; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}12720; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12721; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}12722; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12723; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}12724; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12725; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}12726; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12727; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}12728; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12729; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}12730; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12731; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}12732; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12733; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}12734; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12735; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}12736; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}12737; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12738; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}12739; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}12740; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12741; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}12742; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}12743; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12744; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}12745; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 192(%rsi)12746; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 128(%rsi)12747; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 64(%rsi)12748; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, (%rsi)12749; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 192(%rdx)12750; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, (%rdx)12751; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 64(%rdx)12752; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 128(%rdx)12753; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 192(%rcx)12754; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, (%rcx)12755; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 64(%rcx)12756; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, 128(%rcx)12757; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, 192(%r8)12758; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, (%r8)12759; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 64(%r8)12760; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, 128(%r8)12761; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 192(%r9)12762; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, (%r9)12763; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, 64(%r9)12764; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 128(%r9)12765; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax12766; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)12767; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)12768; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, (%rax)12769; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)12770; AVX512DQ-FCP-NEXT: addq $2632, %rsp # imm = 0xA4812771; AVX512DQ-FCP-NEXT: vzeroupper12772; AVX512DQ-FCP-NEXT: retq12773;12774; AVX512BW-LABEL: load_i32_stride6_vf64:12775; AVX512BW: # %bb.0:12776; AVX512BW-NEXT: subq $2632, %rsp # imm = 0xA4812777; AVX512BW-NEXT: vmovdqa64 1472(%rdi), %zmm2112778; AVX512BW-NEXT: vmovdqa64 1408(%rdi), %zmm112779; AVX512BW-NEXT: vmovdqa64 1088(%rdi), %zmm2012780; AVX512BW-NEXT: vmovdqa64 1024(%rdi), %zmm012781; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm212782; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm1812783; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm2712784; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm312785; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm2512786; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm412787; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]12788; AVX512BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]12789; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]12790; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12791; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm712792; AVX512BW-NEXT: vpermt2d %zmm25, %zmm6, %zmm712793; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12794; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm712795; AVX512BW-NEXT: vpermt2d %zmm27, %zmm6, %zmm712796; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12797; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm712798; AVX512BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm712799; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12800; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm712801; AVX512BW-NEXT: vpermt2d %zmm20, %zmm6, %zmm712802; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12803; AVX512BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm612804; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12805; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]12806; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]12807; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm812808; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm612809; AVX512BW-NEXT: vpermt2d %zmm20, %zmm7, %zmm812810; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12811; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]12812; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]12813; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm012814; AVX512BW-NEXT: vpermt2d %zmm27, %zmm7, %zmm012815; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12816; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm012817; AVX512BW-NEXT: vpermt2d %zmm25, %zmm7, %zmm012818; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12819; AVX512BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm712820; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12821; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]12822; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12823; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm712824; AVX512BW-NEXT: vpermt2d %zmm6, %zmm0, %zmm712825; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12826; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm712827; AVX512BW-NEXT: vpermt2d %zmm3, %zmm0, %zmm712828; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12829; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm712830; AVX512BW-NEXT: vpermt2d %zmm4, %zmm0, %zmm712831; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12832; AVX512BW-NEXT: vpermi2d %zmm1, %zmm21, %zmm012833; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12834; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]12835; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12836; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm712837; AVX512BW-NEXT: vpermt2d %zmm6, %zmm0, %zmm712838; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12839; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm712840; AVX512BW-NEXT: vpermt2d %zmm3, %zmm0, %zmm712841; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12842; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm712843; AVX512BW-NEXT: vpermt2d %zmm4, %zmm0, %zmm712844; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12845; AVX512BW-NEXT: vpermi2d %zmm1, %zmm21, %zmm012846; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12847; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]12848; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]12849; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm012850; AVX512BW-NEXT: vpermt2d %zmm27, %zmm7, %zmm012851; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12852; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]12853; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12854; AVX512BW-NEXT: vpermt2d %zmm27, %zmm0, %zmm312855; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12856; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm312857; AVX512BW-NEXT: vpermt2d %zmm25, %zmm7, %zmm312858; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12859; AVX512BW-NEXT: vpermt2d %zmm25, %zmm0, %zmm412860; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12861; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm312862; AVX512BW-NEXT: vpermt2d %zmm20, %zmm7, %zmm312863; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12864; AVX512BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm712865; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12866; AVX512BW-NEXT: vpermt2d %zmm21, %zmm0, %zmm112867; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12868; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm112869; AVX512BW-NEXT: vpermt2d %zmm2, %zmm8, %zmm112870; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12871; AVX512BW-NEXT: vpermt2d %zmm20, %zmm0, %zmm612872; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12873; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]12874; AVX512BW-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]12875; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012876; AVX512BW-NEXT: vpermt2d %zmm18, %zmm31, %zmm012877; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12878; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]12879; AVX512BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]12880; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012881; AVX512BW-NEXT: vpermt2d %zmm18, %zmm25, %zmm012882; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12883; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]12884; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]12885; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012886; AVX512BW-NEXT: vpermt2d %zmm18, %zmm3, %zmm012887; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12888; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]12889; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]12890; AVX512BW-NEXT: vpermt2d %zmm18, %zmm0, %zmm212891; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12892; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm2612893; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm112894; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm212895; AVX512BW-NEXT: vpermt2d %zmm26, %zmm5, %zmm212896; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12897; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm212898; AVX512BW-NEXT: vpermt2d %zmm26, %zmm8, %zmm212899; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12900; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm212901; AVX512BW-NEXT: vpermt2d %zmm1, %zmm31, %zmm212902; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12903; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm212904; AVX512BW-NEXT: vpermt2d %zmm1, %zmm25, %zmm212905; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12906; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm212907; AVX512BW-NEXT: vpermt2d %zmm1, %zmm3, %zmm212908; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12909; AVX512BW-NEXT: vpermt2d %zmm1, %zmm0, %zmm2612910; AVX512BW-NEXT: vmovdqa64 896(%rdi), %zmm2212911; AVX512BW-NEXT: vmovdqa64 960(%rdi), %zmm112912; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm212913; AVX512BW-NEXT: vpermt2d %zmm22, %zmm5, %zmm212914; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12915; AVX512BW-NEXT: vmovdqa64 1280(%rdi), %zmm1912916; AVX512BW-NEXT: vmovdqa64 1344(%rdi), %zmm212917; AVX512BW-NEXT: vpermi2d %zmm19, %zmm2, %zmm512918; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12919; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm412920; AVX512BW-NEXT: vpermt2d %zmm22, %zmm8, %zmm412921; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12922; AVX512BW-NEXT: vpermi2d %zmm19, %zmm2, %zmm812923; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12924; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm2812925; AVX512BW-NEXT: vpermt2d %zmm1, %zmm31, %zmm2812926; AVX512BW-NEXT: vpermi2d %zmm2, %zmm19, %zmm3112927; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm2912928; AVX512BW-NEXT: vpermt2d %zmm1, %zmm25, %zmm2912929; AVX512BW-NEXT: vpermi2d %zmm2, %zmm19, %zmm2512930; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm412931; AVX512BW-NEXT: vpermt2d %zmm1, %zmm3, %zmm412932; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12933; AVX512BW-NEXT: vpermi2d %zmm2, %zmm19, %zmm312934; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12935; AVX512BW-NEXT: vpermt2d %zmm2, %zmm0, %zmm1912936; AVX512BW-NEXT: vpermt2d %zmm1, %zmm0, %zmm2212937; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm212938; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm012939; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]12940; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2012941; AVX512BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm2012942; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]12943; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2112944; AVX512BW-NEXT: vpermt2d %zmm2, %zmm4, %zmm2112945; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]12946; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112947; AVX512BW-NEXT: vpermt2d %zmm2, %zmm12, %zmm112948; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12949; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]12950; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112951; AVX512BW-NEXT: vpermt2d %zmm2, %zmm14, %zmm112952; AVX512BW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill12953; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]12954; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm2312955; AVX512BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm2312956; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]12957; AVX512BW-NEXT: vpermt2d %zmm0, %zmm24, %zmm212958; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm012959; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm612960; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1512961; AVX512BW-NEXT: vpermt2d %zmm6, %zmm3, %zmm1512962; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1612963; AVX512BW-NEXT: vpermt2d %zmm6, %zmm4, %zmm1612964; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2712965; AVX512BW-NEXT: vpermt2d %zmm6, %zmm12, %zmm2712966; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm3012967; AVX512BW-NEXT: vpermt2d %zmm6, %zmm14, %zmm3012968; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm1712969; AVX512BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm1712970; AVX512BW-NEXT: vpermt2d %zmm0, %zmm24, %zmm612971; AVX512BW-NEXT: vmovdqa64 832(%rdi), %zmm1012972; AVX512BW-NEXT: vmovdqa64 768(%rdi), %zmm712973; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm812974; AVX512BW-NEXT: vpermt2d %zmm10, %zmm3, %zmm812975; AVX512BW-NEXT: vmovdqa64 1216(%rdi), %zmm112976; AVX512BW-NEXT: vmovdqa64 1152(%rdi), %zmm012977; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm312978; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm912979; AVX512BW-NEXT: vpermt2d %zmm10, %zmm4, %zmm912980; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm412981; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm1312982; AVX512BW-NEXT: vpermt2d %zmm10, %zmm12, %zmm1312983; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm1212984; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm1812985; AVX512BW-NEXT: vpermt2d %zmm10, %zmm14, %zmm1812986; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm1412987; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm1112988; AVX512BW-NEXT: vpermt2d %zmm7, %zmm5, %zmm1112989; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm512990; AVX512BW-NEXT: vpermt2d %zmm0, %zmm24, %zmm112991; AVX512BW-NEXT: vpermt2d %zmm7, %zmm24, %zmm1012992; AVX512BW-NEXT: movb $56, %al12993; AVX512BW-NEXT: kmovd %eax, %k212994; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12995; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}12996; AVX512BW-NEXT: movw $-2048, %ax # imm = 0xF80012997; AVX512BW-NEXT: kmovd %eax, %k112998; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12999; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}13000; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13001; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}13002; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13003; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}13004; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13005; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}13006; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13007; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}13008; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13009; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}13010; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13011; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}13012; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13013; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}13014; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13015; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}13016; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13017; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}13018; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13019; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}13020; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13021; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}13022; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13023; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}13024; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13025; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}13026; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13027; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}13028; AVX512BW-NEXT: movw $31, %ax13029; AVX512BW-NEXT: kmovd %eax, %k213030; AVX512BW-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}13031; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13032; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}13033; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13034; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13035; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}13036; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13037; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13038; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm1313039; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13040; AVX512BW-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}13041; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13042; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13043; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm2413044; AVX512BW-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}13045; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13046; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}13047; AVX512BW-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}13048; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13049; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}13050; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13051; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload13052; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}13053; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13054; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13055; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload13056; AVX512BW-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}13057; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13058; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}13059; AVX512BW-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}13060; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13061; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}13062; AVX512BW-NEXT: movw $992, %ax # imm = 0x3E013063; AVX512BW-NEXT: kmovd %eax, %k113064; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13065; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}13066; AVX512BW-NEXT: movb $-32, %al13067; AVX512BW-NEXT: kmovd %eax, %k213068; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13069; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}13070; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13071; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}13072; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13073; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}13074; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13075; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}13076; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13077; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}13078; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13079; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}13080; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13081; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}13082; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13083; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}13084; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13085; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}13086; AVX512BW-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}13087; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13088; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}13089; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}13090; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13091; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}13092; AVX512BW-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}13093; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13094; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}13095; AVX512BW-NEXT: vmovdqa64 %zmm3, 192(%rsi)13096; AVX512BW-NEXT: vmovdqa64 %zmm8, 128(%rsi)13097; AVX512BW-NEXT: vmovdqa64 %zmm20, 64(%rsi)13098; AVX512BW-NEXT: vmovdqa64 %zmm15, (%rsi)13099; AVX512BW-NEXT: vmovdqa64 %zmm4, 192(%rdx)13100; AVX512BW-NEXT: vmovdqa64 %zmm16, (%rdx)13101; AVX512BW-NEXT: vmovdqa64 %zmm21, 64(%rdx)13102; AVX512BW-NEXT: vmovdqa64 %zmm9, 128(%rdx)13103; AVX512BW-NEXT: vmovdqa64 %zmm31, 192(%rcx)13104; AVX512BW-NEXT: vmovdqa64 %zmm24, (%rcx)13105; AVX512BW-NEXT: vmovdqa64 %zmm13, 64(%rcx)13106; AVX512BW-NEXT: vmovdqa64 %zmm28, 128(%rcx)13107; AVX512BW-NEXT: vmovdqa64 %zmm25, 192(%r8)13108; AVX512BW-NEXT: vmovdqa64 %zmm12, (%r8)13109; AVX512BW-NEXT: vmovdqa64 %zmm7, 64(%r8)13110; AVX512BW-NEXT: vmovdqa64 %zmm29, 128(%r8)13111; AVX512BW-NEXT: vmovdqa64 %zmm5, 192(%r9)13112; AVX512BW-NEXT: vmovdqa64 %zmm17, (%r9)13113; AVX512BW-NEXT: vmovdqa64 %zmm23, 64(%r9)13114; AVX512BW-NEXT: vmovdqa64 %zmm11, 128(%r9)13115; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax13116; AVX512BW-NEXT: vmovdqa64 %zmm10, 128(%rax)13117; AVX512BW-NEXT: vmovdqa64 %zmm1, 192(%rax)13118; AVX512BW-NEXT: vmovdqa64 %zmm6, (%rax)13119; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rax)13120; AVX512BW-NEXT: addq $2632, %rsp # imm = 0xA4813121; AVX512BW-NEXT: vzeroupper13122; AVX512BW-NEXT: retq13123;13124; AVX512BW-FCP-LABEL: load_i32_stride6_vf64:13125; AVX512BW-FCP: # %bb.0:13126; AVX512BW-FCP-NEXT: subq $2632, %rsp # imm = 0xA4813127; AVX512BW-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm2113128; AVX512BW-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm113129; AVX512BW-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm2013130; AVX512BW-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm013131; AVX512BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm213132; AVX512BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1813133; AVX512BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2713134; AVX512BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm313135; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm2513136; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm413137; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]13138; AVX512BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]13139; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]13140; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13141; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm713142; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm6, %zmm713143; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13144; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm713145; AVX512BW-FCP-NEXT: vpermt2d %zmm27, %zmm6, %zmm713146; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13147; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm713148; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm713149; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13150; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm713151; AVX512BW-FCP-NEXT: vpermt2d %zmm20, %zmm6, %zmm713152; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13153; AVX512BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm613154; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13155; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]13156; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]13157; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm813158; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm613159; AVX512BW-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm813160; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13161; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]13162; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]13163; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm013164; AVX512BW-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm013165; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13166; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm013167; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm013168; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13169; AVX512BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm713170; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13171; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]13172; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13173; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm713174; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm713175; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13176; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, %zmm713177; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm713178; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13179; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm713180; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm713181; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13182; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm013183; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13184; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]13185; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13186; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm713187; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm713188; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13189; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, %zmm713190; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm713191; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13192; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm713193; AVX512BW-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm713194; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13195; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm013196; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13197; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]13198; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]13199; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm013200; AVX512BW-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm013201; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13202; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]13203; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13204; AVX512BW-FCP-NEXT: vpermt2d %zmm27, %zmm0, %zmm313205; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13206; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm313207; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm313208; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13209; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm0, %zmm413210; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13211; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm313212; AVX512BW-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm313213; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13214; AVX512BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm713215; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13216; AVX512BW-FCP-NEXT: vpermt2d %zmm21, %zmm0, %zmm113217; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13218; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm113219; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm8, %zmm113220; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13221; AVX512BW-FCP-NEXT: vpermt2d %zmm20, %zmm0, %zmm613222; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13223; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]13224; AVX512BW-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]13225; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013226; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm31, %zmm013227; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13228; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]13229; AVX512BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]13230; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013231; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm25, %zmm013232; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13233; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]13234; AVX512BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]13235; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013236; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm3, %zmm013237; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13238; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]13239; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13240; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm0, %zmm213241; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13242; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm2613243; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm113244; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213245; AVX512BW-FCP-NEXT: vpermt2d %zmm26, %zmm5, %zmm213246; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13247; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213248; AVX512BW-FCP-NEXT: vpermt2d %zmm26, %zmm8, %zmm213249; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13250; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm213251; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm213252; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13253; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm213254; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm213255; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13256; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm213257; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm213258; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13259; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2613260; AVX512BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm2213261; AVX512BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm113262; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213263; AVX512BW-FCP-NEXT: vpermt2d %zmm22, %zmm5, %zmm213264; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13265; AVX512BW-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm1913266; AVX512BW-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm213267; AVX512BW-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm513268; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13269; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm413270; AVX512BW-FCP-NEXT: vpermt2d %zmm22, %zmm8, %zmm413271; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13272; AVX512BW-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm813273; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13274; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, %zmm2813275; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm2813276; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm3113277; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, %zmm2913278; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm2913279; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm2513280; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, %zmm413281; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm413282; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13283; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm313284; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13285; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm0, %zmm1913286; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2213287; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm213288; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm013289; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]13290; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2013291; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm2013292; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]13293; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2113294; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm4, %zmm2113295; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]13296; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113297; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm12, %zmm113298; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13299; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]13300; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113301; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm14, %zmm113302; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill13303; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]13304; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm2313305; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm2313306; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]13307; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm213308; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm013309; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm613310; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1513311; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm3, %zmm1513312; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1613313; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm1613314; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2713315; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm12, %zmm2713316; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3013317; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm14, %zmm3013318; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1713319; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm1713320; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm613321; AVX512BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm1013322; AVX512BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm713323; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm813324; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm3, %zmm813325; AVX512BW-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm113326; AVX512BW-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm013327; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm313328; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm913329; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm4, %zmm913330; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm413331; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm1313332; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm12, %zmm1313333; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1213334; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm1813335; AVX512BW-FCP-NEXT: vpermt2d %zmm10, %zmm14, %zmm1813336; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1413337; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm1113338; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm1113339; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm513340; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm113341; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm24, %zmm1013342; AVX512BW-FCP-NEXT: movb $56, %al13343; AVX512BW-FCP-NEXT: kmovd %eax, %k213344; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13345; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}13346; AVX512BW-FCP-NEXT: movw $-2048, %ax # imm = 0xF80013347; AVX512BW-FCP-NEXT: kmovd %eax, %k113348; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13349; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}13350; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13351; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}13352; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13353; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}13354; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13355; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}13356; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13357; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}13358; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13359; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}13360; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13361; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}13362; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13363; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}13364; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13365; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}13366; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13367; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}13368; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13369; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}13370; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13371; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}13372; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13373; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}13374; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13375; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}13376; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13377; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}13378; AVX512BW-FCP-NEXT: movw $31, %ax13379; AVX512BW-FCP-NEXT: kmovd %eax, %k213380; AVX512BW-FCP-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}13381; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13382; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}13383; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13384; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13385; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}13386; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13387; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13388; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm1313389; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13390; AVX512BW-FCP-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}13391; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13392; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13393; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2413394; AVX512BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}13395; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13396; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}13397; AVX512BW-FCP-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}13398; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13399; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}13400; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13401; AVX512BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload13402; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}13403; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13404; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13405; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload13406; AVX512BW-FCP-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}13407; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13408; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}13409; AVX512BW-FCP-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}13410; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13411; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}13412; AVX512BW-FCP-NEXT: movw $992, %ax # imm = 0x3E013413; AVX512BW-FCP-NEXT: kmovd %eax, %k113414; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13415; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}13416; AVX512BW-FCP-NEXT: movb $-32, %al13417; AVX512BW-FCP-NEXT: kmovd %eax, %k213418; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13419; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}13420; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13421; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}13422; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13423; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}13424; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13425; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}13426; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13427; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}13428; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13429; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}13430; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13431; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}13432; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13433; AVX512BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}13434; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13435; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}13436; AVX512BW-FCP-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}13437; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13438; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}13439; AVX512BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}13440; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13441; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}13442; AVX512BW-FCP-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}13443; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13444; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}13445; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 192(%rsi)13446; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 128(%rsi)13447; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 64(%rsi)13448; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, (%rsi)13449; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 192(%rdx)13450; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, (%rdx)13451; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 64(%rdx)13452; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 128(%rdx)13453; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, 192(%rcx)13454; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, (%rcx)13455; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 64(%rcx)13456; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, 128(%rcx)13457; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, 192(%r8)13458; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, (%r8)13459; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 64(%r8)13460; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 128(%r8)13461; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 192(%r9)13462; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, (%r9)13463; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, 64(%r9)13464; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, 128(%r9)13465; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax13466; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)13467; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)13468; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, (%rax)13469; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)13470; AVX512BW-FCP-NEXT: addq $2632, %rsp # imm = 0xA4813471; AVX512BW-FCP-NEXT: vzeroupper13472; AVX512BW-FCP-NEXT: retq13473;13474; AVX512DQ-BW-LABEL: load_i32_stride6_vf64:13475; AVX512DQ-BW: # %bb.0:13476; AVX512DQ-BW-NEXT: subq $2632, %rsp # imm = 0xA4813477; AVX512DQ-BW-NEXT: vmovdqa64 1472(%rdi), %zmm2113478; AVX512DQ-BW-NEXT: vmovdqa64 1408(%rdi), %zmm113479; AVX512DQ-BW-NEXT: vmovdqa64 1088(%rdi), %zmm2013480; AVX512DQ-BW-NEXT: vmovdqa64 1024(%rdi), %zmm013481; AVX512DQ-BW-NEXT: vmovdqa64 512(%rdi), %zmm213482; AVX512DQ-BW-NEXT: vmovdqa64 576(%rdi), %zmm1813483; AVX512DQ-BW-NEXT: vmovdqa64 704(%rdi), %zmm2713484; AVX512DQ-BW-NEXT: vmovdqa64 640(%rdi), %zmm313485; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm2513486; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm413487; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]13488; AVX512DQ-BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]13489; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]13490; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13491; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm713492; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm6, %zmm713493; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13494; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm713495; AVX512DQ-BW-NEXT: vpermt2d %zmm27, %zmm6, %zmm713496; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13497; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm713498; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm713499; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13500; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm713501; AVX512DQ-BW-NEXT: vpermt2d %zmm20, %zmm6, %zmm713502; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13503; AVX512DQ-BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm613504; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13505; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]13506; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]13507; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm813508; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm613509; AVX512DQ-BW-NEXT: vpermt2d %zmm20, %zmm7, %zmm813510; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13511; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]13512; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]13513; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm013514; AVX512DQ-BW-NEXT: vpermt2d %zmm27, %zmm7, %zmm013515; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13516; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm013517; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm7, %zmm013518; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13519; AVX512DQ-BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm713520; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13521; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]13522; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13523; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm713524; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm0, %zmm713525; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13526; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, %zmm713527; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm0, %zmm713528; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13529; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm713530; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm0, %zmm713531; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13532; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm21, %zmm013533; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13534; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]13535; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13536; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm713537; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm0, %zmm713538; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13539; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, %zmm713540; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm0, %zmm713541; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13542; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm713543; AVX512DQ-BW-NEXT: vpermt2d %zmm4, %zmm0, %zmm713544; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13545; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm21, %zmm013546; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13547; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]13548; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]13549; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm013550; AVX512DQ-BW-NEXT: vpermt2d %zmm27, %zmm7, %zmm013551; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13552; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]13553; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13554; AVX512DQ-BW-NEXT: vpermt2d %zmm27, %zmm0, %zmm313555; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13556; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm313557; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm7, %zmm313558; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13559; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm0, %zmm413560; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13561; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm313562; AVX512DQ-BW-NEXT: vpermt2d %zmm20, %zmm7, %zmm313563; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13564; AVX512DQ-BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm713565; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13566; AVX512DQ-BW-NEXT: vpermt2d %zmm21, %zmm0, %zmm113567; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13568; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm113569; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm8, %zmm113570; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13571; AVX512DQ-BW-NEXT: vpermt2d %zmm20, %zmm0, %zmm613572; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13573; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]13574; AVX512DQ-BW-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]13575; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013576; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm31, %zmm013577; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13578; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]13579; AVX512DQ-BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]13580; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013581; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm25, %zmm013582; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13583; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]13584; AVX512DQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]13585; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013586; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm3, %zmm013587; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13588; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]13589; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13590; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm0, %zmm213591; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13592; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm2613593; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm113594; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm213595; AVX512DQ-BW-NEXT: vpermt2d %zmm26, %zmm5, %zmm213596; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13597; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm213598; AVX512DQ-BW-NEXT: vpermt2d %zmm26, %zmm8, %zmm213599; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13600; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm213601; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm31, %zmm213602; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13603; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm213604; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm25, %zmm213605; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13606; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm213607; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm3, %zmm213608; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13609; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm0, %zmm2613610; AVX512DQ-BW-NEXT: vmovdqa64 896(%rdi), %zmm2213611; AVX512DQ-BW-NEXT: vmovdqa64 960(%rdi), %zmm113612; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm213613; AVX512DQ-BW-NEXT: vpermt2d %zmm22, %zmm5, %zmm213614; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13615; AVX512DQ-BW-NEXT: vmovdqa64 1280(%rdi), %zmm1913616; AVX512DQ-BW-NEXT: vmovdqa64 1344(%rdi), %zmm213617; AVX512DQ-BW-NEXT: vpermi2d %zmm19, %zmm2, %zmm513618; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13619; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm413620; AVX512DQ-BW-NEXT: vpermt2d %zmm22, %zmm8, %zmm413621; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13622; AVX512DQ-BW-NEXT: vpermi2d %zmm19, %zmm2, %zmm813623; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13624; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, %zmm2813625; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm31, %zmm2813626; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm19, %zmm3113627; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, %zmm2913628; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm25, %zmm2913629; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm19, %zmm2513630; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, %zmm413631; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm3, %zmm413632; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13633; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm19, %zmm313634; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13635; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm0, %zmm1913636; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm0, %zmm2213637; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm213638; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm013639; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]13640; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm2013641; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm2013642; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]13643; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm2113644; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm4, %zmm2113645; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]13646; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113647; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm12, %zmm113648; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13649; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]13650; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113651; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm14, %zmm113652; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill13653; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]13654; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm2313655; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm2313656; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]13657; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm24, %zmm213658; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm013659; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm613660; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1513661; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm3, %zmm1513662; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1613663; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm4, %zmm1613664; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm2713665; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm12, %zmm2713666; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm3013667; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm14, %zmm3013668; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm1713669; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm1713670; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm24, %zmm613671; AVX512DQ-BW-NEXT: vmovdqa64 832(%rdi), %zmm1013672; AVX512DQ-BW-NEXT: vmovdqa64 768(%rdi), %zmm713673; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm813674; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm3, %zmm813675; AVX512DQ-BW-NEXT: vmovdqa64 1216(%rdi), %zmm113676; AVX512DQ-BW-NEXT: vmovdqa64 1152(%rdi), %zmm013677; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm313678; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm913679; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm4, %zmm913680; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm413681; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm1313682; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm12, %zmm1313683; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm1213684; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm1813685; AVX512DQ-BW-NEXT: vpermt2d %zmm10, %zmm14, %zmm1813686; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm1413687; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm1113688; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm5, %zmm1113689; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm513690; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm24, %zmm113691; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm24, %zmm1013692; AVX512DQ-BW-NEXT: movb $56, %al13693; AVX512DQ-BW-NEXT: kmovd %eax, %k213694; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13695; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}13696; AVX512DQ-BW-NEXT: movw $-2048, %ax # imm = 0xF80013697; AVX512DQ-BW-NEXT: kmovd %eax, %k113698; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13699; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}13700; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13701; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}13702; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13703; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}13704; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13705; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}13706; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13707; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}13708; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13709; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}13710; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13711; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}13712; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13713; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}13714; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13715; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}13716; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13717; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}13718; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13719; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}13720; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13721; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}13722; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13723; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}13724; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13725; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}13726; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13727; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}13728; AVX512DQ-BW-NEXT: movw $31, %ax13729; AVX512DQ-BW-NEXT: kmovd %eax, %k213730; AVX512DQ-BW-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}13731; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13732; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}13733; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13734; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13735; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}13736; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13737; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13738; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm1313739; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13740; AVX512DQ-BW-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}13741; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13742; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13743; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm2413744; AVX512DQ-BW-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}13745; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13746; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}13747; AVX512DQ-BW-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}13748; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13749; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}13750; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload13751; AVX512DQ-BW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload13752; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}13753; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13754; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}13755; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload13756; AVX512DQ-BW-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}13757; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13758; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}13759; AVX512DQ-BW-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}13760; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13761; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}13762; AVX512DQ-BW-NEXT: movw $992, %ax # imm = 0x3E013763; AVX512DQ-BW-NEXT: kmovd %eax, %k113764; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13765; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}13766; AVX512DQ-BW-NEXT: movb $-32, %al13767; AVX512DQ-BW-NEXT: kmovd %eax, %k213768; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13769; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}13770; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13771; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}13772; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13773; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}13774; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13775; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}13776; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13777; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}13778; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13779; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}13780; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13781; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}13782; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13783; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}13784; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13785; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}13786; AVX512DQ-BW-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}13787; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13788; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}13789; AVX512DQ-BW-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}13790; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13791; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}13792; AVX512DQ-BW-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}13793; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13794; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}13795; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 192(%rsi)13796; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 128(%rsi)13797; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 64(%rsi)13798; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, (%rsi)13799; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 192(%rdx)13800; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, (%rdx)13801; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 64(%rdx)13802; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 128(%rdx)13803; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, 192(%rcx)13804; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, (%rcx)13805; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, 64(%rcx)13806; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, 128(%rcx)13807; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, 192(%r8)13808; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, (%r8)13809; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 64(%r8)13810; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, 128(%r8)13811; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 192(%r9)13812; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, (%r9)13813; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, 64(%r9)13814; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, 128(%r9)13815; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax13816; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 128(%rax)13817; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 192(%rax)13818; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, (%rax)13819; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rax)13820; AVX512DQ-BW-NEXT: addq $2632, %rsp # imm = 0xA4813821; AVX512DQ-BW-NEXT: vzeroupper13822; AVX512DQ-BW-NEXT: retq13823;13824; AVX512DQ-BW-FCP-LABEL: load_i32_stride6_vf64:13825; AVX512DQ-BW-FCP: # %bb.0:13826; AVX512DQ-BW-FCP-NEXT: subq $2632, %rsp # imm = 0xA4813827; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm2113828; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm113829; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm2013830; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm013831; AVX512DQ-BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm213832; AVX512DQ-BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1813833; AVX512DQ-BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2713834; AVX512DQ-BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm313835; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm2513836; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm413837; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,6,12,0,0,0,20,26,0,6,12,0,0,0,20,26]13838; AVX512DQ-BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]13839; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,0,0,2,8,14,20,26,0,0,0,2,8,14,20,26]13840; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13841; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm713842; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm6, %zmm713843; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13844; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm713845; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm27, %zmm6, %zmm713846; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13847; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm713848; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm713849; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13850; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm713851; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm20, %zmm6, %zmm713852; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13853; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm613854; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13855; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,3,9,15,21,27,0,0,0,3,9,15,21,27]13856; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]13857; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm813858; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm613859; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm813860; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13861; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [1,7,13,0,0,0,21,27,1,7,13,0,0,0,21,27]13862; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]13863; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm013864; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm013865; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13866; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm013867; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm013868; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13869; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm713870; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13871; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,20,26,0,6,12,0,0,0,20,26,0,6,12]13872; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13873; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm713874; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm713875; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13876; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, %zmm713877; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm713878; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13879; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm713880; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm713881; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13882; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm013883; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13884; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,0,21,27,1,7,13,0,0,0,21,27,1,7,13]13885; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13886; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm713887; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm0, %zmm713888; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13889; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, %zmm713890; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm0, %zmm713891; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13892; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm713893; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm4, %zmm0, %zmm713894; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13895; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm013896; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13897; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,0,0,6,12,18,24,30,0,0,0,6,12,18,24,30]13898; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]13899; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm013900; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm27, %zmm7, %zmm013901; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13902; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,0,1,7,13,19,25,31,0,0,1,7,13,19,25,31]13903; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13904; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm27, %zmm0, %zmm313905; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13906; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm313907; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm7, %zmm313908; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13909; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm0, %zmm413910; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13911; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm313912; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm20, %zmm7, %zmm313913; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13914; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm713915; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13916; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm21, %zmm0, %zmm113917; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13918; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm113919; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm8, %zmm113920; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13921; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm20, %zmm0, %zmm613922; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13923; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm31 = [18,24,30,0,0,0,6,12,18,24,30,0,0,0,6,12]13924; AVX512DQ-BW-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3]13925; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013926; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm31, %zmm013927; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13928; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [19,25,31,0,0,1,7,13,19,25,31,0,0,1,7,13]13929; AVX512DQ-BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]13930; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013931; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm25, %zmm013932; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13933; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [20,26,0,0,0,2,8,14,20,26,0,0,0,2,8,14]13934; AVX512DQ-BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]13935; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013936; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm3, %zmm013937; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13938; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [21,27,0,0,0,3,9,15,21,27,0,0,0,3,9,15]13939; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]13940; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm0, %zmm213941; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13942; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm2613943; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm113944; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213945; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm26, %zmm5, %zmm213946; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13947; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213948; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm26, %zmm8, %zmm213949; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13950; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm213951; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm213952; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13953; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm213954; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm213955; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13956; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm213957; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm213958; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13959; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2613960; AVX512DQ-BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm2213961; AVX512DQ-BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm113962; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213963; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm22, %zmm5, %zmm213964; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13965; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm1913966; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm213967; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm513968; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13969; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm413970; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm22, %zmm8, %zmm413971; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13972; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm19, %zmm2, %zmm813973; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13974; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, %zmm2813975; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm31, %zmm2813976; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm3113977; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, %zmm2913978; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm25, %zmm2913979; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm2513980; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, %zmm413981; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm3, %zmm413982; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13983; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm19, %zmm313984; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13985; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm0, %zmm1913986; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm0, %zmm2213987; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm213988; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm013989; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,6,12,18,24,30,0,0]13990; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2013991; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm2013992; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,7,13,19,25,31,0,0]13993; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2113994; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm4, %zmm2113995; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [2,8,14,20,26,0,0,0]13996; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113997; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm12, %zmm113998; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13999; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,9,15,21,27,0,0,0]14000; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114001; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm14, %zmm114002; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill14003; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [20,26,0,6,12,0,0,0]14004; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm2314005; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm2314006; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm24 = [21,27,1,7,13,0,0,0]14007; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm214008; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm014009; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm614010; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1514011; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm3, %zmm1514012; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1614013; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm1614014; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2714015; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm12, %zmm2714016; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3014017; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm14, %zmm3014018; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1714019; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm1714020; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm614021; AVX512DQ-BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm1014022; AVX512DQ-BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm714023; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm814024; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm3, %zmm814025; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm114026; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm014027; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm314028; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm914029; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm4, %zmm914030; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm414031; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm1314032; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm12, %zmm1314033; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1214034; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm1814035; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm10, %zmm14, %zmm1814036; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm1414037; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm1114038; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm1114039; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm514040; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm24, %zmm114041; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm24, %zmm1014042; AVX512DQ-BW-FCP-NEXT: movb $56, %al14043; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k214044; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14045; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k2}14046; AVX512DQ-BW-FCP-NEXT: movw $-2048, %ax # imm = 0xF80014047; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k114048; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14049; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}14050; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14051; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k2}14052; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14053; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}14054; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14055; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k2}14056; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14057; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}14058; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14059; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k2}14060; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14061; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}14062; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14063; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm9 {%k2}14064; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14065; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}14066; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14067; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm21 {%k2}14068; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14069; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm21 {%k1}14070; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14071; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k2}14072; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14073; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}14074; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14075; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm4 {%k2}14076; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14077; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}14078; AVX512DQ-BW-FCP-NEXT: movw $31, %ax14079; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k214080; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm13, %zmm28 {%k2}14081; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14082; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm28 {%k1}14083; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload14084; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14085; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}14086; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14087; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}14088; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm1314089; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload14090; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm27, %zmm7 {%k2}14091; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14092; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}14093; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2414094; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm31 {%k2}14095; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14096; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm31 {%k1}14097; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm18, %zmm29 {%k2}14098; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14099; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1}14100; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload14101; AVX512DQ-BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload14102; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}14103; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14104; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}14105; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload14106; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm30, %zmm12 {%k2}14107; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14108; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm12 {%k1}14109; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm14, %zmm25 {%k2}14110; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14111; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm25 {%k1}14112; AVX512DQ-BW-FCP-NEXT: movw $992, %ax # imm = 0x3E014113; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k114114; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14115; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm11 {%k1}14116; AVX512DQ-BW-FCP-NEXT: movb $-32, %al14117; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k214118; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14119; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k2}14120; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14121; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}14122; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14123; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm23 {%k2}14124; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14125; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm17 {%k1}14126; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14127; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm17 {%k2}14128; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14129; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}14130; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14131; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5 {%k2}14132; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14133; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}14134; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14135; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k2}14136; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm26, %zmm6 {%k1}14137; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14138; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k2}14139; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm1 {%k1}14140; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14141; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1 {%k2}14142; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm22, %zmm10 {%k1}14143; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14144; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm10 {%k2}14145; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 192(%rsi)14146; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 128(%rsi)14147; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 64(%rsi)14148; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, (%rsi)14149; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 192(%rdx)14150; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, (%rdx)14151; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 64(%rdx)14152; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 128(%rdx)14153; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, 192(%rcx)14154; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, (%rcx)14155; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 64(%rcx)14156; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, 128(%rcx)14157; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, 192(%r8)14158; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, (%r8)14159; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 64(%r8)14160; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 128(%r8)14161; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 192(%r9)14162; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, (%r9)14163; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, 64(%r9)14164; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, 128(%r9)14165; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax14166; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)14167; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)14168; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, (%rax)14169; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)14170; AVX512DQ-BW-FCP-NEXT: addq $2632, %rsp # imm = 0xA4814171; AVX512DQ-BW-FCP-NEXT: vzeroupper14172; AVX512DQ-BW-FCP-NEXT: retq14173 %wide.vec = load <384 x i32>, ptr %in.vec, align 6414174 %strided.vec0 = shufflevector <384 x i32> %wide.vec, <384 x i32> poison, <64 x i32> <i32 0, i32 6, i32 12, i32 18, i32 24, i32 30, i32 36, i32 42, i32 48, i32 54, i32 60, i32 66, i32 72, i32 78, i32 84, i32 90, i32 96, i32 102, i32 108, i32 114, i32 120, i32 126, i32 132, i32 138, i32 144, i32 150, i32 156, i32 162, i32 168, i32 174, i32 180, i32 186, i32 192, i32 198, i32 204, i32 210, i32 216, i32 222, i32 228, i32 234, i32 240, i32 246, i32 252, i32 258, i32 264, i32 270, i32 276, i32 282, i32 288, i32 294, i32 300, i32 306, i32 312, i32 318, i32 324, i32 330, i32 336, i32 342, i32 348, i32 354, i32 360, i32 366, i32 372, i32 378>14175 %strided.vec1 = shufflevector <384 x i32> %wide.vec, <384 x i32> poison, <64 x i32> <i32 1, i32 7, i32 13, i32 19, i32 25, i32 31, i32 37, i32 43, i32 49, i32 55, i32 61, i32 67, i32 73, i32 79, i32 85, i32 91, i32 97, i32 103, i32 109, i32 115, i32 121, i32 127, i32 133, i32 139, i32 145, i32 151, i32 157, i32 163, i32 169, i32 175, i32 181, i32 187, i32 193, i32 199, i32 205, i32 211, i32 217, i32 223, i32 229, i32 235, i32 241, i32 247, i32 253, i32 259, i32 265, i32 271, i32 277, i32 283, i32 289, i32 295, i32 301, i32 307, i32 313, i32 319, i32 325, i32 331, i32 337, i32 343, i32 349, i32 355, i32 361, i32 367, i32 373, i32 379>14176 %strided.vec2 = shufflevector <384 x i32> %wide.vec, <384 x i32> poison, <64 x i32> <i32 2, i32 8, i32 14, i32 20, i32 26, i32 32, i32 38, i32 44, i32 50, i32 56, i32 62, i32 68, i32 74, i32 80, i32 86, i32 92, i32 98, i32 104, i32 110, i32 116, i32 122, i32 128, i32 134, i32 140, i32 146, i32 152, i32 158, i32 164, i32 170, i32 176, i32 182, i32 188, i32 194, i32 200, i32 206, i32 212, i32 218, i32 224, i32 230, i32 236, i32 242, i32 248, i32 254, i32 260, i32 266, i32 272, i32 278, i32 284, i32 290, i32 296, i32 302, i32 308, i32 314, i32 320, i32 326, i32 332, i32 338, i32 344, i32 350, i32 356, i32 362, i32 368, i32 374, i32 380>14177 %strided.vec3 = shufflevector <384 x i32> %wide.vec, <384 x i32> poison, <64 x i32> <i32 3, i32 9, i32 15, i32 21, i32 27, i32 33, i32 39, i32 45, i32 51, i32 57, i32 63, i32 69, i32 75, i32 81, i32 87, i32 93, i32 99, i32 105, i32 111, i32 117, i32 123, i32 129, i32 135, i32 141, i32 147, i32 153, i32 159, i32 165, i32 171, i32 177, i32 183, i32 189, i32 195, i32 201, i32 207, i32 213, i32 219, i32 225, i32 231, i32 237, i32 243, i32 249, i32 255, i32 261, i32 267, i32 273, i32 279, i32 285, i32 291, i32 297, i32 303, i32 309, i32 315, i32 321, i32 327, i32 333, i32 339, i32 345, i32 351, i32 357, i32 363, i32 369, i32 375, i32 381>14178 %strided.vec4 = shufflevector <384 x i32> %wide.vec, <384 x i32> poison, <64 x i32> <i32 4, i32 10, i32 16, i32 22, i32 28, i32 34, i32 40, i32 46, i32 52, i32 58, i32 64, i32 70, i32 76, i32 82, i32 88, i32 94, i32 100, i32 106, i32 112, i32 118, i32 124, i32 130, i32 136, i32 142, i32 148, i32 154, i32 160, i32 166, i32 172, i32 178, i32 184, i32 190, i32 196, i32 202, i32 208, i32 214, i32 220, i32 226, i32 232, i32 238, i32 244, i32 250, i32 256, i32 262, i32 268, i32 274, i32 280, i32 286, i32 292, i32 298, i32 304, i32 310, i32 316, i32 322, i32 328, i32 334, i32 340, i32 346, i32 352, i32 358, i32 364, i32 370, i32 376, i32 382>14179 %strided.vec5 = shufflevector <384 x i32> %wide.vec, <384 x i32> poison, <64 x i32> <i32 5, i32 11, i32 17, i32 23, i32 29, i32 35, i32 41, i32 47, i32 53, i32 59, i32 65, i32 71, i32 77, i32 83, i32 89, i32 95, i32 101, i32 107, i32 113, i32 119, i32 125, i32 131, i32 137, i32 143, i32 149, i32 155, i32 161, i32 167, i32 173, i32 179, i32 185, i32 191, i32 197, i32 203, i32 209, i32 215, i32 221, i32 227, i32 233, i32 239, i32 245, i32 251, i32 257, i32 263, i32 269, i32 275, i32 281, i32 287, i32 293, i32 299, i32 305, i32 311, i32 317, i32 323, i32 329, i32 335, i32 341, i32 347, i32 353, i32 359, i32 365, i32 371, i32 377, i32 383>14180 store <64 x i32> %strided.vec0, ptr %out.vec0, align 6414181 store <64 x i32> %strided.vec1, ptr %out.vec1, align 6414182 store <64 x i32> %strided.vec2, ptr %out.vec2, align 6414183 store <64 x i32> %strided.vec3, ptr %out.vec3, align 6414184 store <64 x i32> %strided.vec4, ptr %out.vec4, align 6414185 store <64 x i32> %strided.vec5, ptr %out.vec5, align 6414186 ret void14187}14188