brintos / llvm-project-archived public Read only

1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+sse2     | FileCheck %s --check-prefix=X86-SSE23; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+sse4.2   | FileCheck %s --check-prefix=X86-SSE424; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx      | FileCheck %s --check-prefixes=X86-AVX,X86-AVX15; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx2     | FileCheck %s --check-prefixes=X86-AVX,X86-AVX26; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse2   | FileCheck %s --check-prefix=X64-SSE27; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse4.2 | FileCheck %s --check-prefix=X64-SSE428; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx    | FileCheck %s --check-prefixes=X64-AVX,X64-AVX1OR2,X64-AVX19; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx2   | FileCheck %s --check-prefixes=X64-AVX,X64-AVX1OR2,X64-AVX210; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512dq,+avx512vl | FileCheck %s --check-prefixes=X64-AVX,X64-AVX51211 12;13; 128-bit Vectors14;15 16define i64 @test_reduce_v2i64(<2 x i64> %a0) {17; X86-SSE2-LABEL: test_reduce_v2i64:18; X86-SSE2:       ## %bb.0:19; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]20; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0]21; X86-SSE2-NEXT:    movdqa %xmm0, %xmm322; X86-SSE2-NEXT:    pxor %xmm2, %xmm323; X86-SSE2-NEXT:    pxor %xmm1, %xmm224; X86-SSE2-NEXT:    movdqa %xmm2, %xmm425; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm426; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]27; X86-SSE2-NEXT:    pcmpeqd %xmm3, %xmm228; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]29; X86-SSE2-NEXT:    pand %xmm5, %xmm230; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]31; X86-SSE2-NEXT:    por %xmm2, %xmm332; X86-SSE2-NEXT:    pand %xmm3, %xmm033; X86-SSE2-NEXT:    pandn %xmm1, %xmm334; X86-SSE2-NEXT:    por %xmm0, %xmm335; X86-SSE2-NEXT:    movd %xmm3, %eax36; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]37; X86-SSE2-NEXT:    movd %xmm0, %edx38; X86-SSE2-NEXT:    retl39;40; X86-SSE42-LABEL: test_reduce_v2i64:41; X86-SSE42:       ## %bb.0:42; X86-SSE42-NEXT:    movdqa %xmm0, %xmm143; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]44; X86-SSE42-NEXT:    movdqa %xmm2, %xmm045; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm046; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm247; X86-SSE42-NEXT:    movd %xmm2, %eax48; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx49; X86-SSE42-NEXT:    retl50;51; X86-AVX-LABEL: test_reduce_v2i64:52; X86-AVX:       ## %bb.0:53; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]54; X86-AVX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm255; X86-AVX-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm056; X86-AVX-NEXT:    vmovd %xmm0, %eax57; X86-AVX-NEXT:    vpextrd $1, %xmm0, %edx58; X86-AVX-NEXT:    retl59;60; X64-SSE2-LABEL: test_reduce_v2i64:61; X64-SSE2:       ## %bb.0:62; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]63; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]64; X64-SSE2-NEXT:    movdqa %xmm0, %xmm365; X64-SSE2-NEXT:    pxor %xmm2, %xmm366; X64-SSE2-NEXT:    pxor %xmm1, %xmm267; X64-SSE2-NEXT:    movdqa %xmm2, %xmm468; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm469; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]70; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm271; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]72; X64-SSE2-NEXT:    pand %xmm5, %xmm273; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]74; X64-SSE2-NEXT:    por %xmm2, %xmm375; X64-SSE2-NEXT:    pand %xmm3, %xmm076; X64-SSE2-NEXT:    pandn %xmm1, %xmm377; X64-SSE2-NEXT:    por %xmm0, %xmm378; X64-SSE2-NEXT:    movq %xmm3, %rax79; X64-SSE2-NEXT:    retq80;81; X64-SSE42-LABEL: test_reduce_v2i64:82; X64-SSE42:       ## %bb.0:83; X64-SSE42-NEXT:    movdqa %xmm0, %xmm184; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]85; X64-SSE42-NEXT:    movdqa %xmm2, %xmm086; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm087; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm288; X64-SSE42-NEXT:    movq %xmm2, %rax89; X64-SSE42-NEXT:    retq90;91; X64-AVX1OR2-LABEL: test_reduce_v2i64:92; X64-AVX1OR2:       ## %bb.0:93; X64-AVX1OR2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]94; X64-AVX1OR2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm295; X64-AVX1OR2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm096; X64-AVX1OR2-NEXT:    vmovq %xmm0, %rax97; X64-AVX1OR2-NEXT:    retq98;99; X64-AVX512-LABEL: test_reduce_v2i64:100; X64-AVX512:       ## %bb.0:101; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]102; X64-AVX512-NEXT:    vpminsq %xmm1, %xmm0, %xmm0103; X64-AVX512-NEXT:    vmovq %xmm0, %rax104; X64-AVX512-NEXT:    retq105  %1 = shufflevector <2 x i64> %a0, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>106  %2 = icmp slt <2 x i64> %a0, %1107  %3 = select <2 x i1> %2, <2 x i64> %a0, <2 x i64> %1108  %4 = extractelement <2 x i64> %3, i32 0109  ret i64 %4110}111 112define i32 @test_reduce_v4i32(<4 x i32> %a0) {113; X86-SSE2-LABEL: test_reduce_v4i32:114; X86-SSE2:       ## %bb.0:115; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]116; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2117; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm2118; X86-SSE2-NEXT:    pand %xmm2, %xmm0119; X86-SSE2-NEXT:    pandn %xmm1, %xmm2120; X86-SSE2-NEXT:    por %xmm0, %xmm2121; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]122; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1123; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1124; X86-SSE2-NEXT:    pand %xmm1, %xmm2125; X86-SSE2-NEXT:    pandn %xmm0, %xmm1126; X86-SSE2-NEXT:    por %xmm2, %xmm1127; X86-SSE2-NEXT:    movd %xmm1, %eax128; X86-SSE2-NEXT:    retl129;130; X86-SSE42-LABEL: test_reduce_v4i32:131; X86-SSE42:       ## %bb.0:132; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]133; X86-SSE42-NEXT:    pminsd %xmm0, %xmm1134; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]135; X86-SSE42-NEXT:    pminsd %xmm1, %xmm0136; X86-SSE42-NEXT:    movd %xmm0, %eax137; X86-SSE42-NEXT:    retl138;139; X86-AVX-LABEL: test_reduce_v4i32:140; X86-AVX:       ## %bb.0:141; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]142; X86-AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0143; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]144; X86-AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0145; X86-AVX-NEXT:    vmovd %xmm0, %eax146; X86-AVX-NEXT:    retl147;148; X64-SSE2-LABEL: test_reduce_v4i32:149; X64-SSE2:       ## %bb.0:150; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]151; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2152; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm2153; X64-SSE2-NEXT:    pand %xmm2, %xmm0154; X64-SSE2-NEXT:    pandn %xmm1, %xmm2155; X64-SSE2-NEXT:    por %xmm0, %xmm2156; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]157; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1158; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1159; X64-SSE2-NEXT:    pand %xmm1, %xmm2160; X64-SSE2-NEXT:    pandn %xmm0, %xmm1161; X64-SSE2-NEXT:    por %xmm2, %xmm1162; X64-SSE2-NEXT:    movd %xmm1, %eax163; X64-SSE2-NEXT:    retq164;165; X64-SSE42-LABEL: test_reduce_v4i32:166; X64-SSE42:       ## %bb.0:167; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]168; X64-SSE42-NEXT:    pminsd %xmm0, %xmm1169; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]170; X64-SSE42-NEXT:    pminsd %xmm1, %xmm0171; X64-SSE42-NEXT:    movd %xmm0, %eax172; X64-SSE42-NEXT:    retq173;174; X64-AVX-LABEL: test_reduce_v4i32:175; X64-AVX:       ## %bb.0:176; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]177; X64-AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0178; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]179; X64-AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0180; X64-AVX-NEXT:    vmovd %xmm0, %eax181; X64-AVX-NEXT:    retq182  %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>183  %2 = icmp slt <4 x i32> %a0, %1184  %3 = select <4 x i1> %2, <4 x i32> %a0, <4 x i32> %1185  %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>186  %5 = icmp slt <4 x i32> %3, %4187  %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4188  %7 = extractelement <4 x i32> %6, i32 0189  ret i32 %7190}191 192define i16 @test_reduce_v8i16(<8 x i16> %a0) {193; X86-SSE2-LABEL: test_reduce_v8i16:194; X86-SSE2:       ## %bb.0:195; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]196; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1197; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]198; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0199; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1200; X86-SSE2-NEXT:    psrld $16, %xmm1201; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1202; X86-SSE2-NEXT:    movd %xmm1, %eax203; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax204; X86-SSE2-NEXT:    retl205;206; X86-SSE42-LABEL: test_reduce_v8i16:207; X86-SSE42:       ## %bb.0:208; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0209; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0210; X86-SSE42-NEXT:    movd %xmm0, %eax211; X86-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x8000212; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax213; X86-SSE42-NEXT:    retl214;215; X86-AVX-LABEL: test_reduce_v8i16:216; X86-AVX:       ## %bb.0:217; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0218; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0219; X86-AVX-NEXT:    vmovd %xmm0, %eax220; X86-AVX-NEXT:    xorl $32768, %eax ## imm = 0x8000221; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax222; X86-AVX-NEXT:    retl223;224; X64-SSE2-LABEL: test_reduce_v8i16:225; X64-SSE2:       ## %bb.0:226; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]227; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1228; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]229; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0230; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1231; X64-SSE2-NEXT:    psrld $16, %xmm1232; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1233; X64-SSE2-NEXT:    movd %xmm1, %eax234; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax235; X64-SSE2-NEXT:    retq236;237; X64-SSE42-LABEL: test_reduce_v8i16:238; X64-SSE42:       ## %bb.0:239; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0240; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0241; X64-SSE42-NEXT:    movd %xmm0, %eax242; X64-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x8000243; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax244; X64-SSE42-NEXT:    retq245;246; X64-AVX1OR2-LABEL: test_reduce_v8i16:247; X64-AVX1OR2:       ## %bb.0:248; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0249; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0250; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax251; X64-AVX1OR2-NEXT:    xorl $32768, %eax ## imm = 0x8000252; X64-AVX1OR2-NEXT:    ## kill: def $ax killed $ax killed $eax253; X64-AVX1OR2-NEXT:    retq254;255; X64-AVX512-LABEL: test_reduce_v8i16:256; X64-AVX512:       ## %bb.0:257; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0258; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0259; X64-AVX512-NEXT:    vmovd %xmm0, %eax260; X64-AVX512-NEXT:    xorl $32768, %eax ## imm = 0x8000261; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax262; X64-AVX512-NEXT:    retq263  %1  = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>264  %2  = icmp slt <8 x i16> %a0, %1265  %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1266  %4  = shufflevector <8 x i16> %3, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>267  %5  = icmp slt <8 x i16> %3, %4268  %6  = select <8 x i1> %5, <8 x i16> %3, <8 x i16> %4269  %7  = shufflevector <8 x i16> %6, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>270  %8  = icmp slt <8 x i16> %6, %7271  %9  = select <8 x i1> %8, <8 x i16> %6, <8 x i16> %7272  %10 = extractelement <8 x i16> %9, i32 0273  ret i16 %10274}275 276define i8 @test_reduce_v16i8(<16 x i8> %a0) {277; X86-SSE2-LABEL: test_reduce_v16i8:278; X86-SSE2:       ## %bb.0:279; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]280; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2281; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2282; X86-SSE2-NEXT:    pand %xmm2, %xmm0283; X86-SSE2-NEXT:    pandn %xmm1, %xmm2284; X86-SSE2-NEXT:    por %xmm0, %xmm2285; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]286; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0287; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm0288; X86-SSE2-NEXT:    pand %xmm0, %xmm2289; X86-SSE2-NEXT:    pandn %xmm1, %xmm0290; X86-SSE2-NEXT:    por %xmm2, %xmm0291; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2292; X86-SSE2-NEXT:    psrld $16, %xmm2293; X86-SSE2-NEXT:    movdqa %xmm2, %xmm1294; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1295; X86-SSE2-NEXT:    pand %xmm1, %xmm0296; X86-SSE2-NEXT:    pandn %xmm2, %xmm1297; X86-SSE2-NEXT:    por %xmm0, %xmm1298; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0299; X86-SSE2-NEXT:    psrlw $8, %xmm0300; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2301; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2302; X86-SSE2-NEXT:    pand %xmm2, %xmm1303; X86-SSE2-NEXT:    pandn %xmm0, %xmm2304; X86-SSE2-NEXT:    por %xmm1, %xmm2305; X86-SSE2-NEXT:    movd %xmm2, %eax306; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax307; X86-SSE2-NEXT:    retl308;309; X86-SSE42-LABEL: test_reduce_v16i8:310; X86-SSE42:       ## %bb.0:311; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0312; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1313; X86-SSE42-NEXT:    psrlw $8, %xmm1314; X86-SSE42-NEXT:    pminub %xmm0, %xmm1315; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0316; X86-SSE42-NEXT:    movd %xmm0, %eax317; X86-SSE42-NEXT:    addb $-128, %al318; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax319; X86-SSE42-NEXT:    retl320;321; X86-AVX-LABEL: test_reduce_v16i8:322; X86-AVX:       ## %bb.0:323; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0324; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1325; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0326; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0327; X86-AVX-NEXT:    vmovd %xmm0, %eax328; X86-AVX-NEXT:    addb $-128, %al329; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax330; X86-AVX-NEXT:    retl331;332; X64-SSE2-LABEL: test_reduce_v16i8:333; X64-SSE2:       ## %bb.0:334; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]335; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2336; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2337; X64-SSE2-NEXT:    pand %xmm2, %xmm0338; X64-SSE2-NEXT:    pandn %xmm1, %xmm2339; X64-SSE2-NEXT:    por %xmm0, %xmm2340; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]341; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1342; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1343; X64-SSE2-NEXT:    pand %xmm1, %xmm2344; X64-SSE2-NEXT:    pandn %xmm0, %xmm1345; X64-SSE2-NEXT:    por %xmm2, %xmm1346; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0347; X64-SSE2-NEXT:    psrld $16, %xmm0348; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2349; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2350; X64-SSE2-NEXT:    pand %xmm2, %xmm1351; X64-SSE2-NEXT:    pandn %xmm0, %xmm2352; X64-SSE2-NEXT:    por %xmm1, %xmm2353; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0354; X64-SSE2-NEXT:    psrlw $8, %xmm0355; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1356; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1357; X64-SSE2-NEXT:    pand %xmm1, %xmm2358; X64-SSE2-NEXT:    pandn %xmm0, %xmm1359; X64-SSE2-NEXT:    por %xmm2, %xmm1360; X64-SSE2-NEXT:    movd %xmm1, %eax361; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax362; X64-SSE2-NEXT:    retq363;364; X64-SSE42-LABEL: test_reduce_v16i8:365; X64-SSE42:       ## %bb.0:366; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0367; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1368; X64-SSE42-NEXT:    psrlw $8, %xmm1369; X64-SSE42-NEXT:    pminub %xmm0, %xmm1370; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0371; X64-SSE42-NEXT:    movd %xmm0, %eax372; X64-SSE42-NEXT:    addb $-128, %al373; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax374; X64-SSE42-NEXT:    retq375;376; X64-AVX1OR2-LABEL: test_reduce_v16i8:377; X64-AVX1OR2:       ## %bb.0:378; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0379; X64-AVX1OR2-NEXT:    vpsrlw $8, %xmm0, %xmm1380; X64-AVX1OR2-NEXT:    vpminub %xmm1, %xmm0, %xmm0381; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0382; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax383; X64-AVX1OR2-NEXT:    addb $-128, %al384; X64-AVX1OR2-NEXT:    ## kill: def $al killed $al killed $eax385; X64-AVX1OR2-NEXT:    retq386;387; X64-AVX512-LABEL: test_reduce_v16i8:388; X64-AVX512:       ## %bb.0:389; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0390; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1391; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0392; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0393; X64-AVX512-NEXT:    vmovd %xmm0, %eax394; X64-AVX512-NEXT:    addb $-128, %al395; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax396; X64-AVX512-NEXT:    retq397  %1  = shufflevector <16 x i8> %a0, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>398  %2  = icmp slt <16 x i8> %a0, %1399  %3  = select <16 x i1> %2, <16 x i8> %a0, <16 x i8> %1400  %4  = shufflevector <16 x i8> %3, <16 x i8> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>401  %5  = icmp slt <16 x i8> %3, %4402  %6  = select <16 x i1> %5, <16 x i8> %3, <16 x i8> %4403  %7  = shufflevector <16 x i8> %6, <16 x i8> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>404  %8  = icmp slt <16 x i8> %6, %7405  %9  = select <16 x i1> %8, <16 x i8> %6, <16 x i8> %7406  %10 = shufflevector <16 x i8> %9, <16 x i8> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>407  %11 = icmp slt <16 x i8> %9, %10408  %12 = select <16 x i1> %11, <16 x i8> %9, <16 x i8> %10409  %13 = extractelement <16 x i8> %12, i32 0410  ret i8 %13411}412 413;414; 256-bit Vectors415;416 417define i64 @test_reduce_v4i64(<4 x i64> %a0) {418; X86-SSE2-LABEL: test_reduce_v4i64:419; X86-SSE2:       ## %bb.0:420; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0]421; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3422; X86-SSE2-NEXT:    pxor %xmm2, %xmm3423; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4424; X86-SSE2-NEXT:    pxor %xmm2, %xmm4425; X86-SSE2-NEXT:    movdqa %xmm4, %xmm5426; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5427; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]428; X86-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4429; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]430; X86-SSE2-NEXT:    pand %xmm6, %xmm4431; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]432; X86-SSE2-NEXT:    por %xmm4, %xmm3433; X86-SSE2-NEXT:    pand %xmm3, %xmm0434; X86-SSE2-NEXT:    pandn %xmm1, %xmm3435; X86-SSE2-NEXT:    por %xmm0, %xmm3436; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]437; X86-SSE2-NEXT:    movdqa %xmm3, %xmm1438; X86-SSE2-NEXT:    pxor %xmm2, %xmm1439; X86-SSE2-NEXT:    pxor %xmm0, %xmm2440; X86-SSE2-NEXT:    movdqa %xmm2, %xmm4441; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm4442; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]443; X86-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2444; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]445; X86-SSE2-NEXT:    pand %xmm5, %xmm1446; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]447; X86-SSE2-NEXT:    por %xmm1, %xmm2448; X86-SSE2-NEXT:    pand %xmm2, %xmm3449; X86-SSE2-NEXT:    pandn %xmm0, %xmm2450; X86-SSE2-NEXT:    por %xmm3, %xmm2451; X86-SSE2-NEXT:    movd %xmm2, %eax452; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]453; X86-SSE2-NEXT:    movd %xmm0, %edx454; X86-SSE2-NEXT:    retl455;456; X86-SSE42-LABEL: test_reduce_v4i64:457; X86-SSE42:       ## %bb.0:458; X86-SSE42-NEXT:    movdqa %xmm0, %xmm2459; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0460; X86-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0461; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1462; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]463; X86-SSE42-NEXT:    movdqa %xmm2, %xmm0464; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm0465; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2466; X86-SSE42-NEXT:    movd %xmm2, %eax467; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx468; X86-SSE42-NEXT:    retl469;470; X86-AVX1-LABEL: test_reduce_v4i64:471; X86-AVX1:       ## %bb.0:472; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1473; X86-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2474; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0475; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]476; X86-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2477; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0478; X86-AVX1-NEXT:    vmovd %xmm0, %eax479; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx480; X86-AVX1-NEXT:    vzeroupper481; X86-AVX1-NEXT:    retl482;483; X86-AVX2-LABEL: test_reduce_v4i64:484; X86-AVX2:       ## %bb.0:485; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1486; X86-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2487; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0488; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]489; X86-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2490; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0491; X86-AVX2-NEXT:    vmovd %xmm0, %eax492; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx493; X86-AVX2-NEXT:    vzeroupper494; X86-AVX2-NEXT:    retl495;496; X64-SSE2-LABEL: test_reduce_v4i64:497; X64-SSE2:       ## %bb.0:498; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]499; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3500; X64-SSE2-NEXT:    pxor %xmm2, %xmm3501; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4502; X64-SSE2-NEXT:    pxor %xmm2, %xmm4503; X64-SSE2-NEXT:    movdqa %xmm4, %xmm5504; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5505; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]506; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4507; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]508; X64-SSE2-NEXT:    pand %xmm6, %xmm3509; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]510; X64-SSE2-NEXT:    por %xmm3, %xmm4511; X64-SSE2-NEXT:    pand %xmm4, %xmm0512; X64-SSE2-NEXT:    pandn %xmm1, %xmm4513; X64-SSE2-NEXT:    por %xmm0, %xmm4514; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]515; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1516; X64-SSE2-NEXT:    pxor %xmm2, %xmm1517; X64-SSE2-NEXT:    pxor %xmm0, %xmm2518; X64-SSE2-NEXT:    movdqa %xmm2, %xmm3519; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm3520; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]521; X64-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2522; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]523; X64-SSE2-NEXT:    pand %xmm5, %xmm1524; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]525; X64-SSE2-NEXT:    por %xmm1, %xmm2526; X64-SSE2-NEXT:    pand %xmm2, %xmm4527; X64-SSE2-NEXT:    pandn %xmm0, %xmm2528; X64-SSE2-NEXT:    por %xmm4, %xmm2529; X64-SSE2-NEXT:    movq %xmm2, %rax530; X64-SSE2-NEXT:    retq531;532; X64-SSE42-LABEL: test_reduce_v4i64:533; X64-SSE42:       ## %bb.0:534; X64-SSE42-NEXT:    movdqa %xmm0, %xmm2535; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0536; X64-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0537; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1538; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]539; X64-SSE42-NEXT:    movdqa %xmm2, %xmm0540; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm0541; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2542; X64-SSE42-NEXT:    movq %xmm2, %rax543; X64-SSE42-NEXT:    retq544;545; X64-AVX1-LABEL: test_reduce_v4i64:546; X64-AVX1:       ## %bb.0:547; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1548; X64-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2549; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0550; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]551; X64-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2552; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0553; X64-AVX1-NEXT:    vmovq %xmm0, %rax554; X64-AVX1-NEXT:    vzeroupper555; X64-AVX1-NEXT:    retq556;557; X64-AVX2-LABEL: test_reduce_v4i64:558; X64-AVX2:       ## %bb.0:559; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1560; X64-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2561; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0562; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]563; X64-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2564; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0565; X64-AVX2-NEXT:    vmovq %xmm0, %rax566; X64-AVX2-NEXT:    vzeroupper567; X64-AVX2-NEXT:    retq568;569; X64-AVX512-LABEL: test_reduce_v4i64:570; X64-AVX512:       ## %bb.0:571; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1572; X64-AVX512-NEXT:    vpminsq %xmm1, %xmm0, %xmm0573; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]574; X64-AVX512-NEXT:    vpminsq %xmm1, %xmm0, %xmm0575; X64-AVX512-NEXT:    vmovq %xmm0, %rax576; X64-AVX512-NEXT:    vzeroupper577; X64-AVX512-NEXT:    retq578  %1 = shufflevector <4 x i64> %a0, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>579  %2 = icmp slt <4 x i64> %a0, %1580  %3 = select <4 x i1> %2, <4 x i64> %a0, <4 x i64> %1581  %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>582  %5 = icmp slt <4 x i64> %3, %4583  %6 = select <4 x i1> %5, <4 x i64> %3, <4 x i64> %4584  %7 = extractelement <4 x i64> %6, i32 0585  ret i64 %7586}587 588define i32 @test_reduce_v8i32(<8 x i32> %a0) {589; X86-SSE2-LABEL: test_reduce_v8i32:590; X86-SSE2:       ## %bb.0:591; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2592; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm2593; X86-SSE2-NEXT:    pand %xmm2, %xmm0594; X86-SSE2-NEXT:    pandn %xmm1, %xmm2595; X86-SSE2-NEXT:    por %xmm0, %xmm2596; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]597; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1598; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1599; X86-SSE2-NEXT:    pand %xmm1, %xmm2600; X86-SSE2-NEXT:    pandn %xmm0, %xmm1601; X86-SSE2-NEXT:    por %xmm2, %xmm1602; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]603; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2604; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2605; X86-SSE2-NEXT:    pand %xmm2, %xmm1606; X86-SSE2-NEXT:    pandn %xmm0, %xmm2607; X86-SSE2-NEXT:    por %xmm1, %xmm2608; X86-SSE2-NEXT:    movd %xmm2, %eax609; X86-SSE2-NEXT:    retl610;611; X86-SSE42-LABEL: test_reduce_v8i32:612; X86-SSE42:       ## %bb.0:613; X86-SSE42-NEXT:    pminsd %xmm1, %xmm0614; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]615; X86-SSE42-NEXT:    pminsd %xmm0, %xmm1616; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]617; X86-SSE42-NEXT:    pminsd %xmm1, %xmm0618; X86-SSE42-NEXT:    movd %xmm0, %eax619; X86-SSE42-NEXT:    retl620;621; X86-AVX1-LABEL: test_reduce_v8i32:622; X86-AVX1:       ## %bb.0:623; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1624; X86-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0625; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]626; X86-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0627; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]628; X86-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0629; X86-AVX1-NEXT:    vmovd %xmm0, %eax630; X86-AVX1-NEXT:    vzeroupper631; X86-AVX1-NEXT:    retl632;633; X86-AVX2-LABEL: test_reduce_v8i32:634; X86-AVX2:       ## %bb.0:635; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1636; X86-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0637; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]638; X86-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0639; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]640; X86-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0641; X86-AVX2-NEXT:    vmovd %xmm0, %eax642; X86-AVX2-NEXT:    vzeroupper643; X86-AVX2-NEXT:    retl644;645; X64-SSE2-LABEL: test_reduce_v8i32:646; X64-SSE2:       ## %bb.0:647; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2648; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm2649; X64-SSE2-NEXT:    pand %xmm2, %xmm0650; X64-SSE2-NEXT:    pandn %xmm1, %xmm2651; X64-SSE2-NEXT:    por %xmm0, %xmm2652; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]653; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1654; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1655; X64-SSE2-NEXT:    pand %xmm1, %xmm2656; X64-SSE2-NEXT:    pandn %xmm0, %xmm1657; X64-SSE2-NEXT:    por %xmm2, %xmm1658; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]659; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2660; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2661; X64-SSE2-NEXT:    pand %xmm2, %xmm1662; X64-SSE2-NEXT:    pandn %xmm0, %xmm2663; X64-SSE2-NEXT:    por %xmm1, %xmm2664; X64-SSE2-NEXT:    movd %xmm2, %eax665; X64-SSE2-NEXT:    retq666;667; X64-SSE42-LABEL: test_reduce_v8i32:668; X64-SSE42:       ## %bb.0:669; X64-SSE42-NEXT:    pminsd %xmm1, %xmm0670; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]671; X64-SSE42-NEXT:    pminsd %xmm0, %xmm1672; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]673; X64-SSE42-NEXT:    pminsd %xmm1, %xmm0674; X64-SSE42-NEXT:    movd %xmm0, %eax675; X64-SSE42-NEXT:    retq676;677; X64-AVX1-LABEL: test_reduce_v8i32:678; X64-AVX1:       ## %bb.0:679; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1680; X64-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0681; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]682; X64-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0683; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]684; X64-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0685; X64-AVX1-NEXT:    vmovd %xmm0, %eax686; X64-AVX1-NEXT:    vzeroupper687; X64-AVX1-NEXT:    retq688;689; X64-AVX2-LABEL: test_reduce_v8i32:690; X64-AVX2:       ## %bb.0:691; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1692; X64-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0693; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]694; X64-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0695; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]696; X64-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0697; X64-AVX2-NEXT:    vmovd %xmm0, %eax698; X64-AVX2-NEXT:    vzeroupper699; X64-AVX2-NEXT:    retq700;701; X64-AVX512-LABEL: test_reduce_v8i32:702; X64-AVX512:       ## %bb.0:703; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1704; X64-AVX512-NEXT:    vpminsd %xmm1, %xmm0, %xmm0705; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]706; X64-AVX512-NEXT:    vpminsd %xmm1, %xmm0, %xmm0707; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]708; X64-AVX512-NEXT:    vpminsd %xmm1, %xmm0, %xmm0709; X64-AVX512-NEXT:    vmovd %xmm0, %eax710; X64-AVX512-NEXT:    vzeroupper711; X64-AVX512-NEXT:    retq712  %1  = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>713  %2  = icmp slt <8 x i32> %a0, %1714  %3  = select <8 x i1> %2, <8 x i32> %a0, <8 x i32> %1715  %4  = shufflevector <8 x i32> %3, <8 x i32> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>716  %5  = icmp slt <8 x i32> %3, %4717  %6  = select <8 x i1> %5, <8 x i32> %3, <8 x i32> %4718  %7  = shufflevector <8 x i32> %6, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>719  %8  = icmp slt <8 x i32> %6, %7720  %9  = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7721  %10 = extractelement <8 x i32> %9, i32 0722  ret i32 %10723}724 725define i16 @test_reduce_v16i16(<16 x i16> %a0) {726; X86-SSE2-LABEL: test_reduce_v16i16:727; X86-SSE2:       ## %bb.0:728; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0729; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]730; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1731; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]732; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0733; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1734; X86-SSE2-NEXT:    psrld $16, %xmm1735; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1736; X86-SSE2-NEXT:    movd %xmm1, %eax737; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax738; X86-SSE2-NEXT:    retl739;740; X86-SSE42-LABEL: test_reduce_v16i16:741; X86-SSE42:       ## %bb.0:742; X86-SSE42-NEXT:    pminsw %xmm1, %xmm0743; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0744; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0745; X86-SSE42-NEXT:    movd %xmm0, %eax746; X86-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x8000747; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax748; X86-SSE42-NEXT:    retl749;750; X86-AVX1-LABEL: test_reduce_v16i16:751; X86-AVX1:       ## %bb.0:752; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1753; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0754; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0755; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0756; X86-AVX1-NEXT:    vmovd %xmm0, %eax757; X86-AVX1-NEXT:    xorl $32768, %eax ## imm = 0x8000758; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax759; X86-AVX1-NEXT:    vzeroupper760; X86-AVX1-NEXT:    retl761;762; X86-AVX2-LABEL: test_reduce_v16i16:763; X86-AVX2:       ## %bb.0:764; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1765; X86-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0766; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0767; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0768; X86-AVX2-NEXT:    vmovd %xmm0, %eax769; X86-AVX2-NEXT:    xorl $32768, %eax ## imm = 0x8000770; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax771; X86-AVX2-NEXT:    vzeroupper772; X86-AVX2-NEXT:    retl773;774; X64-SSE2-LABEL: test_reduce_v16i16:775; X64-SSE2:       ## %bb.0:776; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0777; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]778; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1779; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]780; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0781; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1782; X64-SSE2-NEXT:    psrld $16, %xmm1783; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1784; X64-SSE2-NEXT:    movd %xmm1, %eax785; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax786; X64-SSE2-NEXT:    retq787;788; X64-SSE42-LABEL: test_reduce_v16i16:789; X64-SSE42:       ## %bb.0:790; X64-SSE42-NEXT:    pminsw %xmm1, %xmm0791; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0792; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0793; X64-SSE42-NEXT:    movd %xmm0, %eax794; X64-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x8000795; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax796; X64-SSE42-NEXT:    retq797;798; X64-AVX1-LABEL: test_reduce_v16i16:799; X64-AVX1:       ## %bb.0:800; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1801; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0802; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0803; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0804; X64-AVX1-NEXT:    vmovd %xmm0, %eax805; X64-AVX1-NEXT:    xorl $32768, %eax ## imm = 0x8000806; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax807; X64-AVX1-NEXT:    vzeroupper808; X64-AVX1-NEXT:    retq809;810; X64-AVX2-LABEL: test_reduce_v16i16:811; X64-AVX2:       ## %bb.0:812; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1813; X64-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0814; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0815; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0816; X64-AVX2-NEXT:    vmovd %xmm0, %eax817; X64-AVX2-NEXT:    xorl $32768, %eax ## imm = 0x8000818; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax819; X64-AVX2-NEXT:    vzeroupper820; X64-AVX2-NEXT:    retq821;822; X64-AVX512-LABEL: test_reduce_v16i16:823; X64-AVX512:       ## %bb.0:824; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1825; X64-AVX512-NEXT:    vpminsw %xmm1, %xmm0, %xmm0826; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0827; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0828; X64-AVX512-NEXT:    vmovd %xmm0, %eax829; X64-AVX512-NEXT:    xorl $32768, %eax ## imm = 0x8000830; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax831; X64-AVX512-NEXT:    vzeroupper832; X64-AVX512-NEXT:    retq833  %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>834  %2  = icmp slt <16 x i16> %a0, %1835  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1836  %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>837  %5  = icmp slt <16 x i16> %3, %4838  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %4839  %7  = shufflevector <16 x i16> %6, <16 x i16> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>840  %8  = icmp slt <16 x i16> %6, %7841  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7842  %10 = shufflevector <16 x i16> %9, <16 x i16> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>843  %11 = icmp slt <16 x i16> %9, %10844  %12 = select <16 x i1> %11, <16 x i16> %9, <16 x i16> %10845  %13 = extractelement <16 x i16> %12, i32 0846  ret i16 %13847}848 849define i8 @test_reduce_v32i8(<32 x i8> %a0) {850; X86-SSE2-LABEL: test_reduce_v32i8:851; X86-SSE2:       ## %bb.0:852; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2853; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2854; X86-SSE2-NEXT:    pand %xmm2, %xmm0855; X86-SSE2-NEXT:    pandn %xmm1, %xmm2856; X86-SSE2-NEXT:    por %xmm0, %xmm2857; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]858; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1859; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1860; X86-SSE2-NEXT:    pand %xmm1, %xmm2861; X86-SSE2-NEXT:    pandn %xmm0, %xmm1862; X86-SSE2-NEXT:    por %xmm2, %xmm1863; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]864; X86-SSE2-NEXT:    movdqa %xmm2, %xmm0865; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm0866; X86-SSE2-NEXT:    pand %xmm0, %xmm1867; X86-SSE2-NEXT:    pandn %xmm2, %xmm0868; X86-SSE2-NEXT:    por %xmm1, %xmm0869; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2870; X86-SSE2-NEXT:    psrld $16, %xmm2871; X86-SSE2-NEXT:    movdqa %xmm2, %xmm1872; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1873; X86-SSE2-NEXT:    pand %xmm1, %xmm0874; X86-SSE2-NEXT:    pandn %xmm2, %xmm1875; X86-SSE2-NEXT:    por %xmm0, %xmm1876; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0877; X86-SSE2-NEXT:    psrlw $8, %xmm0878; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2879; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2880; X86-SSE2-NEXT:    pand %xmm2, %xmm1881; X86-SSE2-NEXT:    pandn %xmm0, %xmm2882; X86-SSE2-NEXT:    por %xmm1, %xmm2883; X86-SSE2-NEXT:    movd %xmm2, %eax884; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax885; X86-SSE2-NEXT:    retl886;887; X86-SSE42-LABEL: test_reduce_v32i8:888; X86-SSE42:       ## %bb.0:889; X86-SSE42-NEXT:    pminsb %xmm1, %xmm0890; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0891; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1892; X86-SSE42-NEXT:    psrlw $8, %xmm1893; X86-SSE42-NEXT:    pminub %xmm0, %xmm1894; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0895; X86-SSE42-NEXT:    movd %xmm0, %eax896; X86-SSE42-NEXT:    addb $-128, %al897; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax898; X86-SSE42-NEXT:    retl899;900; X86-AVX1-LABEL: test_reduce_v32i8:901; X86-AVX1:       ## %bb.0:902; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1903; X86-AVX1-NEXT:    vpminsb %xmm1, %xmm0, %xmm0904; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0905; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1906; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0907; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0908; X86-AVX1-NEXT:    vmovd %xmm0, %eax909; X86-AVX1-NEXT:    addb $-128, %al910; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax911; X86-AVX1-NEXT:    vzeroupper912; X86-AVX1-NEXT:    retl913;914; X86-AVX2-LABEL: test_reduce_v32i8:915; X86-AVX2:       ## %bb.0:916; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1917; X86-AVX2-NEXT:    vpminsb %xmm1, %xmm0, %xmm0918; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0919; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1920; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0921; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0922; X86-AVX2-NEXT:    vmovd %xmm0, %eax923; X86-AVX2-NEXT:    addb $-128, %al924; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax925; X86-AVX2-NEXT:    vzeroupper926; X86-AVX2-NEXT:    retl927;928; X64-SSE2-LABEL: test_reduce_v32i8:929; X64-SSE2:       ## %bb.0:930; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2931; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2932; X64-SSE2-NEXT:    pand %xmm2, %xmm0933; X64-SSE2-NEXT:    pandn %xmm1, %xmm2934; X64-SSE2-NEXT:    por %xmm0, %xmm2935; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]936; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1937; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1938; X64-SSE2-NEXT:    pand %xmm1, %xmm2939; X64-SSE2-NEXT:    pandn %xmm0, %xmm1940; X64-SSE2-NEXT:    por %xmm2, %xmm1941; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]942; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2943; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2944; X64-SSE2-NEXT:    pand %xmm2, %xmm1945; X64-SSE2-NEXT:    pandn %xmm0, %xmm2946; X64-SSE2-NEXT:    por %xmm1, %xmm2947; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0948; X64-SSE2-NEXT:    psrld $16, %xmm0949; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1950; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1951; X64-SSE2-NEXT:    pand %xmm1, %xmm2952; X64-SSE2-NEXT:    pandn %xmm0, %xmm1953; X64-SSE2-NEXT:    por %xmm2, %xmm1954; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0955; X64-SSE2-NEXT:    psrlw $8, %xmm0956; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2957; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2958; X64-SSE2-NEXT:    pand %xmm2, %xmm1959; X64-SSE2-NEXT:    pandn %xmm0, %xmm2960; X64-SSE2-NEXT:    por %xmm1, %xmm2961; X64-SSE2-NEXT:    movd %xmm2, %eax962; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax963; X64-SSE2-NEXT:    retq964;965; X64-SSE42-LABEL: test_reduce_v32i8:966; X64-SSE42:       ## %bb.0:967; X64-SSE42-NEXT:    pminsb %xmm1, %xmm0968; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0969; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1970; X64-SSE42-NEXT:    psrlw $8, %xmm1971; X64-SSE42-NEXT:    pminub %xmm0, %xmm1972; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0973; X64-SSE42-NEXT:    movd %xmm0, %eax974; X64-SSE42-NEXT:    addb $-128, %al975; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax976; X64-SSE42-NEXT:    retq977;978; X64-AVX1-LABEL: test_reduce_v32i8:979; X64-AVX1:       ## %bb.0:980; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1981; X64-AVX1-NEXT:    vpminsb %xmm1, %xmm0, %xmm0982; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0983; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1984; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0985; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0986; X64-AVX1-NEXT:    vmovd %xmm0, %eax987; X64-AVX1-NEXT:    addb $-128, %al988; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax989; X64-AVX1-NEXT:    vzeroupper990; X64-AVX1-NEXT:    retq991;992; X64-AVX2-LABEL: test_reduce_v32i8:993; X64-AVX2:       ## %bb.0:994; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1995; X64-AVX2-NEXT:    vpminsb %xmm1, %xmm0, %xmm0996; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0997; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1998; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0999; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01000; X64-AVX2-NEXT:    vmovd %xmm0, %eax1001; X64-AVX2-NEXT:    addb $-128, %al1002; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1003; X64-AVX2-NEXT:    vzeroupper1004; X64-AVX2-NEXT:    retq1005;1006; X64-AVX512-LABEL: test_reduce_v32i8:1007; X64-AVX512:       ## %bb.0:1008; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11009; X64-AVX512-NEXT:    vpminsb %xmm1, %xmm0, %xmm01010; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01011; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm11012; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01013; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01014; X64-AVX512-NEXT:    vmovd %xmm0, %eax1015; X64-AVX512-NEXT:    addb $-128, %al1016; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax1017; X64-AVX512-NEXT:    vzeroupper1018; X64-AVX512-NEXT:    retq1019  %1  = shufflevector <32 x i8> %a0, <32 x i8> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1020  %2  = icmp slt <32 x i8> %a0, %11021  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %11022  %4  = shufflevector <32 x i8> %3, <32 x i8> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1023  %5  = icmp slt <32 x i8> %3, %41024  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %41025  %7  = shufflevector <32 x i8> %6, <32 x i8> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1026  %8  = icmp slt <32 x i8> %6, %71027  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %71028  %10 = shufflevector <32 x i8> %9, <32 x i8> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1029  %11 = icmp slt <32 x i8> %9, %101030  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %101031  %13 = shufflevector <32 x i8> %12, <32 x i8> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1032  %14 = icmp slt <32 x i8> %12, %131033  %15 = select <32 x i1> %14, <32 x i8> %12, <32 x i8> %131034  %16 = extractelement <32 x i8> %15, i32 01035  ret i8 %161036}1037 1038;1039; 512-bit Vectors1040;1041 1042define i64 @test_reduce_v8i64(<8 x i64> %a0) {1043; X86-SSE2-LABEL: test_reduce_v8i64:1044; X86-SSE2:       ## %bb.0:1045; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,0,2147483648,0]1046; X86-SSE2-NEXT:    movdqa %xmm1, %xmm51047; X86-SSE2-NEXT:    pxor %xmm4, %xmm51048; X86-SSE2-NEXT:    movdqa %xmm3, %xmm61049; X86-SSE2-NEXT:    pxor %xmm4, %xmm61050; X86-SSE2-NEXT:    movdqa %xmm6, %xmm71051; X86-SSE2-NEXT:    pcmpgtd %xmm5, %xmm71052; X86-SSE2-NEXT:    pcmpeqd %xmm5, %xmm61053; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]1054; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1055; X86-SSE2-NEXT:    pand %xmm5, %xmm61056; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1057; X86-SSE2-NEXT:    por %xmm6, %xmm51058; X86-SSE2-NEXT:    pand %xmm5, %xmm11059; X86-SSE2-NEXT:    pandn %xmm3, %xmm51060; X86-SSE2-NEXT:    por %xmm1, %xmm51061; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11062; X86-SSE2-NEXT:    pxor %xmm4, %xmm11063; X86-SSE2-NEXT:    movdqa %xmm2, %xmm31064; X86-SSE2-NEXT:    pxor %xmm4, %xmm31065; X86-SSE2-NEXT:    movdqa %xmm3, %xmm61066; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm61067; X86-SSE2-NEXT:    pcmpeqd %xmm1, %xmm31068; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,0,2,2]1069; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]1070; X86-SSE2-NEXT:    pand %xmm1, %xmm31071; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]1072; X86-SSE2-NEXT:    por %xmm3, %xmm11073; X86-SSE2-NEXT:    pand %xmm1, %xmm01074; X86-SSE2-NEXT:    pandn %xmm2, %xmm11075; X86-SSE2-NEXT:    por %xmm0, %xmm11076; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01077; X86-SSE2-NEXT:    pxor %xmm4, %xmm01078; X86-SSE2-NEXT:    movdqa %xmm5, %xmm21079; X86-SSE2-NEXT:    pxor %xmm4, %xmm21080; X86-SSE2-NEXT:    movdqa %xmm2, %xmm31081; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm31082; X86-SSE2-NEXT:    pcmpeqd %xmm0, %xmm21083; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]1084; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]1085; X86-SSE2-NEXT:    pand %xmm0, %xmm21086; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]1087; X86-SSE2-NEXT:    por %xmm2, %xmm01088; X86-SSE2-NEXT:    pand %xmm0, %xmm11089; X86-SSE2-NEXT:    pandn %xmm5, %xmm01090; X86-SSE2-NEXT:    por %xmm1, %xmm01091; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1092; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21093; X86-SSE2-NEXT:    pxor %xmm4, %xmm21094; X86-SSE2-NEXT:    pxor %xmm1, %xmm41095; X86-SSE2-NEXT:    movdqa %xmm4, %xmm31096; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm31097; X86-SSE2-NEXT:    pcmpeqd %xmm2, %xmm41098; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]1099; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]1100; X86-SSE2-NEXT:    pand %xmm2, %xmm41101; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]1102; X86-SSE2-NEXT:    por %xmm4, %xmm21103; X86-SSE2-NEXT:    pand %xmm2, %xmm01104; X86-SSE2-NEXT:    pandn %xmm1, %xmm21105; X86-SSE2-NEXT:    por %xmm0, %xmm21106; X86-SSE2-NEXT:    movd %xmm2, %eax1107; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1108; X86-SSE2-NEXT:    movd %xmm0, %edx1109; X86-SSE2-NEXT:    retl1110;1111; X86-SSE42-LABEL: test_reduce_v8i64:1112; X86-SSE42:       ## %bb.0:1113; X86-SSE42-NEXT:    movdqa %xmm0, %xmm41114; X86-SSE42-NEXT:    movdqa %xmm2, %xmm01115; X86-SSE42-NEXT:    pcmpgtq %xmm4, %xmm01116; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm21117; X86-SSE42-NEXT:    movdqa %xmm3, %xmm01118; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01119; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm31120; X86-SSE42-NEXT:    movapd %xmm3, %xmm01121; X86-SSE42-NEXT:    pcmpgtq %xmm2, %xmm01122; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm31123; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1124; X86-SSE42-NEXT:    movdqa %xmm1, %xmm01125; X86-SSE42-NEXT:    pcmpgtq %xmm3, %xmm01126; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm11127; X86-SSE42-NEXT:    movd %xmm1, %eax1128; X86-SSE42-NEXT:    pextrd $1, %xmm1, %edx1129; X86-SSE42-NEXT:    retl1130;1131; X86-AVX1-LABEL: test_reduce_v8i64:1132; X86-AVX1:       ## %bb.0:1133; X86-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21134; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm21135; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm01136; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm11137; X86-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm31138; X86-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm01139; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm11140; X86-AVX1-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm01141; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1142; X86-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21143; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01144; X86-AVX1-NEXT:    vmovd %xmm0, %eax1145; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx1146; X86-AVX1-NEXT:    vzeroupper1147; X86-AVX1-NEXT:    retl1148;1149; X86-AVX2-LABEL: test_reduce_v8i64:1150; X86-AVX2:       ## %bb.0:1151; X86-AVX2-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm21152; X86-AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm01153; X86-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm11154; X86-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21155; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01156; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1157; X86-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21158; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01159; X86-AVX2-NEXT:    vmovd %xmm0, %eax1160; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx1161; X86-AVX2-NEXT:    vzeroupper1162; X86-AVX2-NEXT:    retl1163;1164; X64-SSE2-LABEL: test_reduce_v8i64:1165; X64-SSE2:       ## %bb.0:1166; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648]1167; X64-SSE2-NEXT:    movdqa %xmm1, %xmm51168; X64-SSE2-NEXT:    pxor %xmm4, %xmm51169; X64-SSE2-NEXT:    movdqa %xmm3, %xmm61170; X64-SSE2-NEXT:    pxor %xmm4, %xmm61171; X64-SSE2-NEXT:    movdqa %xmm6, %xmm71172; X64-SSE2-NEXT:    pcmpgtd %xmm5, %xmm71173; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]1174; X64-SSE2-NEXT:    pcmpeqd %xmm5, %xmm61175; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1176; X64-SSE2-NEXT:    pand %xmm8, %xmm61177; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1178; X64-SSE2-NEXT:    por %xmm6, %xmm51179; X64-SSE2-NEXT:    pand %xmm5, %xmm11180; X64-SSE2-NEXT:    pandn %xmm3, %xmm51181; X64-SSE2-NEXT:    por %xmm1, %xmm51182; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11183; X64-SSE2-NEXT:    pxor %xmm4, %xmm11184; X64-SSE2-NEXT:    movdqa %xmm2, %xmm31185; X64-SSE2-NEXT:    pxor %xmm4, %xmm31186; X64-SSE2-NEXT:    movdqa %xmm3, %xmm61187; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm61188; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]1189; X64-SSE2-NEXT:    pcmpeqd %xmm1, %xmm31190; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]1191; X64-SSE2-NEXT:    pand %xmm7, %xmm11192; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]1193; X64-SSE2-NEXT:    por %xmm1, %xmm31194; X64-SSE2-NEXT:    pand %xmm3, %xmm01195; X64-SSE2-NEXT:    pandn %xmm2, %xmm31196; X64-SSE2-NEXT:    por %xmm0, %xmm31197; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01198; X64-SSE2-NEXT:    pxor %xmm4, %xmm01199; X64-SSE2-NEXT:    movdqa %xmm5, %xmm11200; X64-SSE2-NEXT:    pxor %xmm4, %xmm11201; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21202; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm21203; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]1204; X64-SSE2-NEXT:    pcmpeqd %xmm0, %xmm11205; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]1206; X64-SSE2-NEXT:    pand %xmm6, %xmm01207; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]1208; X64-SSE2-NEXT:    por %xmm0, %xmm11209; X64-SSE2-NEXT:    pand %xmm1, %xmm31210; X64-SSE2-NEXT:    pandn %xmm5, %xmm11211; X64-SSE2-NEXT:    por %xmm3, %xmm11212; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1213; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21214; X64-SSE2-NEXT:    pxor %xmm4, %xmm21215; X64-SSE2-NEXT:    pxor %xmm0, %xmm41216; X64-SSE2-NEXT:    movdqa %xmm4, %xmm31217; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm31218; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]1219; X64-SSE2-NEXT:    pcmpeqd %xmm2, %xmm41220; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]1221; X64-SSE2-NEXT:    pand %xmm5, %xmm21222; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]1223; X64-SSE2-NEXT:    por %xmm2, %xmm31224; X64-SSE2-NEXT:    pand %xmm3, %xmm11225; X64-SSE2-NEXT:    pandn %xmm0, %xmm31226; X64-SSE2-NEXT:    por %xmm1, %xmm31227; X64-SSE2-NEXT:    movq %xmm3, %rax1228; X64-SSE2-NEXT:    retq1229;1230; X64-SSE42-LABEL: test_reduce_v8i64:1231; X64-SSE42:       ## %bb.0:1232; X64-SSE42-NEXT:    movdqa %xmm0, %xmm41233; X64-SSE42-NEXT:    movdqa %xmm2, %xmm01234; X64-SSE42-NEXT:    pcmpgtq %xmm4, %xmm01235; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm21236; X64-SSE42-NEXT:    movdqa %xmm3, %xmm01237; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01238; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm31239; X64-SSE42-NEXT:    movapd %xmm3, %xmm01240; X64-SSE42-NEXT:    pcmpgtq %xmm2, %xmm01241; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm31242; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1243; X64-SSE42-NEXT:    movdqa %xmm1, %xmm01244; X64-SSE42-NEXT:    pcmpgtq %xmm3, %xmm01245; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm11246; X64-SSE42-NEXT:    movq %xmm1, %rax1247; X64-SSE42-NEXT:    retq1248;1249; X64-AVX1-LABEL: test_reduce_v8i64:1250; X64-AVX1:       ## %bb.0:1251; X64-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21252; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm21253; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm01254; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm11255; X64-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm31256; X64-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm01257; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm11258; X64-AVX1-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm01259; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1260; X64-AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21261; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01262; X64-AVX1-NEXT:    vmovq %xmm0, %rax1263; X64-AVX1-NEXT:    vzeroupper1264; X64-AVX1-NEXT:    retq1265;1266; X64-AVX2-LABEL: test_reduce_v8i64:1267; X64-AVX2:       ## %bb.0:1268; X64-AVX2-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm21269; X64-AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm01270; X64-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm11271; X64-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21272; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01273; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1274; X64-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm21275; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01276; X64-AVX2-NEXT:    vmovq %xmm0, %rax1277; X64-AVX2-NEXT:    vzeroupper1278; X64-AVX2-NEXT:    retq1279;1280; X64-AVX512-LABEL: test_reduce_v8i64:1281; X64-AVX512:       ## %bb.0:1282; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11283; X64-AVX512-NEXT:    vpminsq %zmm1, %zmm0, %zmm01284; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11285; X64-AVX512-NEXT:    vpminsq %xmm1, %xmm0, %xmm01286; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1287; X64-AVX512-NEXT:    vpminsq %xmm1, %xmm0, %xmm01288; X64-AVX512-NEXT:    vmovq %xmm0, %rax1289; X64-AVX512-NEXT:    vzeroupper1290; X64-AVX512-NEXT:    retq1291  %1  = shufflevector <8 x i64> %a0, <8 x i64> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>1292  %2  = icmp slt <8 x i64> %a0, %11293  %3  = select <8 x i1> %2, <8 x i64> %a0, <8 x i64> %11294  %4  = shufflevector <8 x i64> %3, <8 x i64> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1295  %5  = icmp slt <8 x i64> %3, %41296  %6  = select <8 x i1> %5, <8 x i64> %3, <8 x i64> %41297  %7  = shufflevector <8 x i64> %6, <8 x i64> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1298  %8  = icmp slt <8 x i64> %6, %71299  %9  = select <8 x i1> %8, <8 x i64> %6, <8 x i64> %71300  %10 = extractelement <8 x i64> %9, i32 01301  ret i64 %101302}1303 1304define i32 @test_reduce_v16i32(<16 x i32> %a0) {1305; X86-SSE2-LABEL: test_reduce_v16i32:1306; X86-SSE2:       ## %bb.0:1307; X86-SSE2-NEXT:    movdqa %xmm2, %xmm41308; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm41309; X86-SSE2-NEXT:    pand %xmm4, %xmm01310; X86-SSE2-NEXT:    pandn %xmm2, %xmm41311; X86-SSE2-NEXT:    por %xmm0, %xmm41312; X86-SSE2-NEXT:    movdqa %xmm3, %xmm01313; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm01314; X86-SSE2-NEXT:    pand %xmm0, %xmm11315; X86-SSE2-NEXT:    pandn %xmm3, %xmm01316; X86-SSE2-NEXT:    por %xmm1, %xmm01317; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11318; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm11319; X86-SSE2-NEXT:    pand %xmm1, %xmm41320; X86-SSE2-NEXT:    pandn %xmm0, %xmm11321; X86-SSE2-NEXT:    por %xmm4, %xmm11322; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1323; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21324; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm21325; X86-SSE2-NEXT:    pand %xmm2, %xmm11326; X86-SSE2-NEXT:    pandn %xmm0, %xmm21327; X86-SSE2-NEXT:    por %xmm1, %xmm21328; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1329; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11330; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm11331; X86-SSE2-NEXT:    pand %xmm1, %xmm21332; X86-SSE2-NEXT:    pandn %xmm0, %xmm11333; X86-SSE2-NEXT:    por %xmm2, %xmm11334; X86-SSE2-NEXT:    movd %xmm1, %eax1335; X86-SSE2-NEXT:    retl1336;1337; X86-SSE42-LABEL: test_reduce_v16i32:1338; X86-SSE42:       ## %bb.0:1339; X86-SSE42-NEXT:    pminsd %xmm3, %xmm11340; X86-SSE42-NEXT:    pminsd %xmm2, %xmm01341; X86-SSE42-NEXT:    pminsd %xmm1, %xmm01342; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1343; X86-SSE42-NEXT:    pminsd %xmm0, %xmm11344; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1345; X86-SSE42-NEXT:    pminsd %xmm1, %xmm01346; X86-SSE42-NEXT:    movd %xmm0, %eax1347; X86-SSE42-NEXT:    retl1348;1349; X86-AVX1-LABEL: test_reduce_v16i32:1350; X86-AVX1:       ## %bb.0:1351; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21352; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31353; X86-AVX1-NEXT:    vpminsd %xmm2, %xmm3, %xmm21354; X86-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm01355; X86-AVX1-NEXT:    vpminsd %xmm2, %xmm0, %xmm01356; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1357; X86-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm01358; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1359; X86-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm01360; X86-AVX1-NEXT:    vmovd %xmm0, %eax1361; X86-AVX1-NEXT:    vzeroupper1362; X86-AVX1-NEXT:    retl1363;1364; X86-AVX2-LABEL: test_reduce_v16i32:1365; X86-AVX2:       ## %bb.0:1366; X86-AVX2-NEXT:    vpminsd %ymm1, %ymm0, %ymm01367; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11368; X86-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm01369; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1370; X86-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm01371; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1372; X86-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm01373; X86-AVX2-NEXT:    vmovd %xmm0, %eax1374; X86-AVX2-NEXT:    vzeroupper1375; X86-AVX2-NEXT:    retl1376;1377; X64-SSE2-LABEL: test_reduce_v16i32:1378; X64-SSE2:       ## %bb.0:1379; X64-SSE2-NEXT:    movdqa %xmm2, %xmm41380; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm41381; X64-SSE2-NEXT:    pand %xmm4, %xmm01382; X64-SSE2-NEXT:    pandn %xmm2, %xmm41383; X64-SSE2-NEXT:    por %xmm0, %xmm41384; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01385; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm01386; X64-SSE2-NEXT:    pand %xmm0, %xmm11387; X64-SSE2-NEXT:    pandn %xmm3, %xmm01388; X64-SSE2-NEXT:    por %xmm1, %xmm01389; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11390; X64-SSE2-NEXT:    pcmpgtd %xmm4, %xmm11391; X64-SSE2-NEXT:    pand %xmm1, %xmm41392; X64-SSE2-NEXT:    pandn %xmm0, %xmm11393; X64-SSE2-NEXT:    por %xmm4, %xmm11394; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1395; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21396; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm21397; X64-SSE2-NEXT:    pand %xmm2, %xmm11398; X64-SSE2-NEXT:    pandn %xmm0, %xmm21399; X64-SSE2-NEXT:    por %xmm1, %xmm21400; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1401; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11402; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm11403; X64-SSE2-NEXT:    pand %xmm1, %xmm21404; X64-SSE2-NEXT:    pandn %xmm0, %xmm11405; X64-SSE2-NEXT:    por %xmm2, %xmm11406; X64-SSE2-NEXT:    movd %xmm1, %eax1407; X64-SSE2-NEXT:    retq1408;1409; X64-SSE42-LABEL: test_reduce_v16i32:1410; X64-SSE42:       ## %bb.0:1411; X64-SSE42-NEXT:    pminsd %xmm3, %xmm11412; X64-SSE42-NEXT:    pminsd %xmm2, %xmm01413; X64-SSE42-NEXT:    pminsd %xmm1, %xmm01414; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1415; X64-SSE42-NEXT:    pminsd %xmm0, %xmm11416; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1417; X64-SSE42-NEXT:    pminsd %xmm1, %xmm01418; X64-SSE42-NEXT:    movd %xmm0, %eax1419; X64-SSE42-NEXT:    retq1420;1421; X64-AVX1-LABEL: test_reduce_v16i32:1422; X64-AVX1:       ## %bb.0:1423; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21424; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31425; X64-AVX1-NEXT:    vpminsd %xmm2, %xmm3, %xmm21426; X64-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm01427; X64-AVX1-NEXT:    vpminsd %xmm2, %xmm0, %xmm01428; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1429; X64-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm01430; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1431; X64-AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm01432; X64-AVX1-NEXT:    vmovd %xmm0, %eax1433; X64-AVX1-NEXT:    vzeroupper1434; X64-AVX1-NEXT:    retq1435;1436; X64-AVX2-LABEL: test_reduce_v16i32:1437; X64-AVX2:       ## %bb.0:1438; X64-AVX2-NEXT:    vpminsd %ymm1, %ymm0, %ymm01439; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11440; X64-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm01441; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1442; X64-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm01443; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1444; X64-AVX2-NEXT:    vpminsd %xmm1, %xmm0, %xmm01445; X64-AVX2-NEXT:    vmovd %xmm0, %eax1446; X64-AVX2-NEXT:    vzeroupper1447; X64-AVX2-NEXT:    retq1448;1449; X64-AVX512-LABEL: test_reduce_v16i32:1450; X64-AVX512:       ## %bb.0:1451; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11452; X64-AVX512-NEXT:    vpminsd %zmm1, %zmm0, %zmm01453; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11454; X64-AVX512-NEXT:    vpminsd %xmm1, %xmm0, %xmm01455; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1456; X64-AVX512-NEXT:    vpminsd %xmm1, %xmm0, %xmm01457; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1458; X64-AVX512-NEXT:    vpminsd %xmm1, %xmm0, %xmm01459; X64-AVX512-NEXT:    vmovd %xmm0, %eax1460; X64-AVX512-NEXT:    vzeroupper1461; X64-AVX512-NEXT:    retq1462  %1  = shufflevector <16 x i32> %a0, <16 x i32> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1463  %2  = icmp slt <16 x i32> %a0, %11464  %3  = select <16 x i1> %2, <16 x i32> %a0, <16 x i32> %11465  %4  = shufflevector <16 x i32> %3, <16 x i32> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1466  %5  = icmp slt <16 x i32> %3, %41467  %6  = select <16 x i1> %5, <16 x i32> %3, <16 x i32> %41468  %7  = shufflevector <16 x i32> %6, <16 x i32> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1469  %8  = icmp slt <16 x i32> %6, %71470  %9  = select <16 x i1> %8, <16 x i32> %6, <16 x i32> %71471  %10 = shufflevector <16 x i32> %9, <16 x i32> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1472  %11 = icmp slt <16 x i32> %9, %101473  %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %101474  %13 = extractelement <16 x i32> %12, i32 01475  ret i32 %131476}1477 1478define i16 @test_reduce_v32i16(<32 x i16> %a0) {1479; X86-SSE2-LABEL: test_reduce_v32i16:1480; X86-SSE2:       ## %bb.0:1481; X86-SSE2-NEXT:    pminsw %xmm3, %xmm11482; X86-SSE2-NEXT:    pminsw %xmm2, %xmm01483; X86-SSE2-NEXT:    pminsw %xmm1, %xmm01484; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1485; X86-SSE2-NEXT:    pminsw %xmm0, %xmm11486; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1487; X86-SSE2-NEXT:    pminsw %xmm1, %xmm01488; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11489; X86-SSE2-NEXT:    psrld $16, %xmm11490; X86-SSE2-NEXT:    pminsw %xmm0, %xmm11491; X86-SSE2-NEXT:    movd %xmm1, %eax1492; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1493; X86-SSE2-NEXT:    retl1494;1495; X86-SSE42-LABEL: test_reduce_v32i16:1496; X86-SSE42:       ## %bb.0:1497; X86-SSE42-NEXT:    pminsw %xmm3, %xmm11498; X86-SSE42-NEXT:    pminsw %xmm2, %xmm01499; X86-SSE42-NEXT:    pminsw %xmm1, %xmm01500; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01501; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01502; X86-SSE42-NEXT:    movd %xmm0, %eax1503; X86-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x80001504; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1505; X86-SSE42-NEXT:    retl1506;1507; X86-AVX1-LABEL: test_reduce_v32i16:1508; X86-AVX1:       ## %bb.0:1509; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21510; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31511; X86-AVX1-NEXT:    vpminsw %xmm2, %xmm3, %xmm21512; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm01513; X86-AVX1-NEXT:    vpminsw %xmm2, %xmm0, %xmm01514; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01515; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01516; X86-AVX1-NEXT:    vmovd %xmm0, %eax1517; X86-AVX1-NEXT:    xorl $32768, %eax ## imm = 0x80001518; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax1519; X86-AVX1-NEXT:    vzeroupper1520; X86-AVX1-NEXT:    retl1521;1522; X86-AVX2-LABEL: test_reduce_v32i16:1523; X86-AVX2:       ## %bb.0:1524; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm01525; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11526; X86-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm01527; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01528; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01529; X86-AVX2-NEXT:    vmovd %xmm0, %eax1530; X86-AVX2-NEXT:    xorl $32768, %eax ## imm = 0x80001531; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax1532; X86-AVX2-NEXT:    vzeroupper1533; X86-AVX2-NEXT:    retl1534;1535; X64-SSE2-LABEL: test_reduce_v32i16:1536; X64-SSE2:       ## %bb.0:1537; X64-SSE2-NEXT:    pminsw %xmm3, %xmm11538; X64-SSE2-NEXT:    pminsw %xmm2, %xmm01539; X64-SSE2-NEXT:    pminsw %xmm1, %xmm01540; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1541; X64-SSE2-NEXT:    pminsw %xmm0, %xmm11542; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1543; X64-SSE2-NEXT:    pminsw %xmm1, %xmm01544; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11545; X64-SSE2-NEXT:    psrld $16, %xmm11546; X64-SSE2-NEXT:    pminsw %xmm0, %xmm11547; X64-SSE2-NEXT:    movd %xmm1, %eax1548; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1549; X64-SSE2-NEXT:    retq1550;1551; X64-SSE42-LABEL: test_reduce_v32i16:1552; X64-SSE42:       ## %bb.0:1553; X64-SSE42-NEXT:    pminsw %xmm3, %xmm11554; X64-SSE42-NEXT:    pminsw %xmm2, %xmm01555; X64-SSE42-NEXT:    pminsw %xmm1, %xmm01556; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01557; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01558; X64-SSE42-NEXT:    movd %xmm0, %eax1559; X64-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x80001560; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1561; X64-SSE42-NEXT:    retq1562;1563; X64-AVX1-LABEL: test_reduce_v32i16:1564; X64-AVX1:       ## %bb.0:1565; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21566; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31567; X64-AVX1-NEXT:    vpminsw %xmm2, %xmm3, %xmm21568; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm01569; X64-AVX1-NEXT:    vpminsw %xmm2, %xmm0, %xmm01570; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01571; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01572; X64-AVX1-NEXT:    vmovd %xmm0, %eax1573; X64-AVX1-NEXT:    xorl $32768, %eax ## imm = 0x80001574; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax1575; X64-AVX1-NEXT:    vzeroupper1576; X64-AVX1-NEXT:    retq1577;1578; X64-AVX2-LABEL: test_reduce_v32i16:1579; X64-AVX2:       ## %bb.0:1580; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm01581; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11582; X64-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm01583; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01584; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01585; X64-AVX2-NEXT:    vmovd %xmm0, %eax1586; X64-AVX2-NEXT:    xorl $32768, %eax ## imm = 0x80001587; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax1588; X64-AVX2-NEXT:    vzeroupper1589; X64-AVX2-NEXT:    retq1590;1591; X64-AVX512-LABEL: test_reduce_v32i16:1592; X64-AVX512:       ## %bb.0:1593; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11594; X64-AVX512-NEXT:    vpminsw %ymm1, %ymm0, %ymm01595; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11596; X64-AVX512-NEXT:    vpminsw %xmm1, %xmm0, %xmm01597; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01598; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01599; X64-AVX512-NEXT:    vmovd %xmm0, %eax1600; X64-AVX512-NEXT:    xorl $32768, %eax ## imm = 0x80001601; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax1602; X64-AVX512-NEXT:    vzeroupper1603; X64-AVX512-NEXT:    retq1604  %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1605  %2  = icmp slt <32 x i16> %a0, %11606  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %11607  %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1608  %5  = icmp slt <32 x i16> %3, %41609  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %41610  %7  = shufflevector <32 x i16> %6, <32 x i16> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1611  %8  = icmp slt <32 x i16> %6, %71612  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %71613  %10 = shufflevector <32 x i16> %9, <32 x i16> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1614  %11 = icmp slt <32 x i16> %9, %101615  %12 = select <32 x i1> %11, <32 x i16> %9, <32 x i16> %101616  %13 = shufflevector <32 x i16> %12, <32 x i16> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1617  %14 = icmp slt <32 x i16> %12, %131618  %15 = select <32 x i1> %14, <32 x i16> %12, <32 x i16> %131619  %16 = extractelement <32 x i16> %15, i32 01620  ret i16 %161621}1622 1623define i8 @test_reduce_v64i8(<64 x i8> %a0) {1624; X86-SSE2-LABEL: test_reduce_v64i8:1625; X86-SSE2:       ## %bb.0:1626; X86-SSE2-NEXT:    movdqa %xmm2, %xmm41627; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm41628; X86-SSE2-NEXT:    pand %xmm4, %xmm01629; X86-SSE2-NEXT:    pandn %xmm2, %xmm41630; X86-SSE2-NEXT:    por %xmm0, %xmm41631; X86-SSE2-NEXT:    movdqa %xmm3, %xmm01632; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm01633; X86-SSE2-NEXT:    pand %xmm0, %xmm11634; X86-SSE2-NEXT:    pandn %xmm3, %xmm01635; X86-SSE2-NEXT:    por %xmm1, %xmm01636; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11637; X86-SSE2-NEXT:    pcmpgtb %xmm4, %xmm11638; X86-SSE2-NEXT:    pand %xmm1, %xmm41639; X86-SSE2-NEXT:    pandn %xmm0, %xmm11640; X86-SSE2-NEXT:    por %xmm4, %xmm11641; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1642; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21643; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm21644; X86-SSE2-NEXT:    pand %xmm2, %xmm11645; X86-SSE2-NEXT:    pandn %xmm0, %xmm21646; X86-SSE2-NEXT:    por %xmm1, %xmm21647; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]1648; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01649; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm01650; X86-SSE2-NEXT:    pand %xmm0, %xmm21651; X86-SSE2-NEXT:    pandn %xmm1, %xmm01652; X86-SSE2-NEXT:    por %xmm2, %xmm01653; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21654; X86-SSE2-NEXT:    psrld $16, %xmm21655; X86-SSE2-NEXT:    movdqa %xmm2, %xmm11656; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm11657; X86-SSE2-NEXT:    pand %xmm1, %xmm01658; X86-SSE2-NEXT:    pandn %xmm2, %xmm11659; X86-SSE2-NEXT:    por %xmm0, %xmm11660; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01661; X86-SSE2-NEXT:    psrlw $8, %xmm01662; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21663; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm21664; X86-SSE2-NEXT:    pand %xmm2, %xmm11665; X86-SSE2-NEXT:    pandn %xmm0, %xmm21666; X86-SSE2-NEXT:    por %xmm1, %xmm21667; X86-SSE2-NEXT:    movd %xmm2, %eax1668; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1669; X86-SSE2-NEXT:    retl1670;1671; X86-SSE42-LABEL: test_reduce_v64i8:1672; X86-SSE42:       ## %bb.0:1673; X86-SSE42-NEXT:    pminsb %xmm3, %xmm11674; X86-SSE42-NEXT:    pminsb %xmm2, %xmm01675; X86-SSE42-NEXT:    pminsb %xmm1, %xmm01676; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01677; X86-SSE42-NEXT:    movdqa %xmm0, %xmm11678; X86-SSE42-NEXT:    psrlw $8, %xmm11679; X86-SSE42-NEXT:    pminub %xmm0, %xmm11680; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm01681; X86-SSE42-NEXT:    movd %xmm0, %eax1682; X86-SSE42-NEXT:    addb $-128, %al1683; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1684; X86-SSE42-NEXT:    retl1685;1686; X86-AVX1-LABEL: test_reduce_v64i8:1687; X86-AVX1:       ## %bb.0:1688; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21689; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31690; X86-AVX1-NEXT:    vpminsb %xmm2, %xmm3, %xmm21691; X86-AVX1-NEXT:    vpminsb %xmm1, %xmm0, %xmm01692; X86-AVX1-NEXT:    vpminsb %xmm2, %xmm0, %xmm01693; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01694; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11695; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01696; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01697; X86-AVX1-NEXT:    vmovd %xmm0, %eax1698; X86-AVX1-NEXT:    addb $-128, %al1699; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1700; X86-AVX1-NEXT:    vzeroupper1701; X86-AVX1-NEXT:    retl1702;1703; X86-AVX2-LABEL: test_reduce_v64i8:1704; X86-AVX2:       ## %bb.0:1705; X86-AVX2-NEXT:    vpminsb %ymm1, %ymm0, %ymm01706; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11707; X86-AVX2-NEXT:    vpminsb %xmm1, %xmm0, %xmm01708; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01709; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11710; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01711; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01712; X86-AVX2-NEXT:    vmovd %xmm0, %eax1713; X86-AVX2-NEXT:    addb $-128, %al1714; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1715; X86-AVX2-NEXT:    vzeroupper1716; X86-AVX2-NEXT:    retl1717;1718; X64-SSE2-LABEL: test_reduce_v64i8:1719; X64-SSE2:       ## %bb.0:1720; X64-SSE2-NEXT:    movdqa %xmm2, %xmm41721; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm41722; X64-SSE2-NEXT:    pand %xmm4, %xmm01723; X64-SSE2-NEXT:    pandn %xmm2, %xmm41724; X64-SSE2-NEXT:    por %xmm0, %xmm41725; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01726; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm01727; X64-SSE2-NEXT:    pand %xmm0, %xmm11728; X64-SSE2-NEXT:    pandn %xmm3, %xmm01729; X64-SSE2-NEXT:    por %xmm1, %xmm01730; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11731; X64-SSE2-NEXT:    pcmpgtb %xmm4, %xmm11732; X64-SSE2-NEXT:    pand %xmm1, %xmm41733; X64-SSE2-NEXT:    pandn %xmm0, %xmm11734; X64-SSE2-NEXT:    por %xmm4, %xmm11735; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1736; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21737; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm21738; X64-SSE2-NEXT:    pand %xmm2, %xmm11739; X64-SSE2-NEXT:    pandn %xmm0, %xmm21740; X64-SSE2-NEXT:    por %xmm1, %xmm21741; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1742; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11743; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm11744; X64-SSE2-NEXT:    pand %xmm1, %xmm21745; X64-SSE2-NEXT:    pandn %xmm0, %xmm11746; X64-SSE2-NEXT:    por %xmm2, %xmm11747; X64-SSE2-NEXT:    movdqa %xmm1, %xmm01748; X64-SSE2-NEXT:    psrld $16, %xmm01749; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21750; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm21751; X64-SSE2-NEXT:    pand %xmm2, %xmm11752; X64-SSE2-NEXT:    pandn %xmm0, %xmm21753; X64-SSE2-NEXT:    por %xmm1, %xmm21754; X64-SSE2-NEXT:    movdqa %xmm2, %xmm01755; X64-SSE2-NEXT:    psrlw $8, %xmm01756; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11757; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm11758; X64-SSE2-NEXT:    pand %xmm1, %xmm21759; X64-SSE2-NEXT:    pandn %xmm0, %xmm11760; X64-SSE2-NEXT:    por %xmm2, %xmm11761; X64-SSE2-NEXT:    movd %xmm1, %eax1762; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1763; X64-SSE2-NEXT:    retq1764;1765; X64-SSE42-LABEL: test_reduce_v64i8:1766; X64-SSE42:       ## %bb.0:1767; X64-SSE42-NEXT:    pminsb %xmm3, %xmm11768; X64-SSE42-NEXT:    pminsb %xmm2, %xmm01769; X64-SSE42-NEXT:    pminsb %xmm1, %xmm01770; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01771; X64-SSE42-NEXT:    movdqa %xmm0, %xmm11772; X64-SSE42-NEXT:    psrlw $8, %xmm11773; X64-SSE42-NEXT:    pminub %xmm0, %xmm11774; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm01775; X64-SSE42-NEXT:    movd %xmm0, %eax1776; X64-SSE42-NEXT:    addb $-128, %al1777; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1778; X64-SSE42-NEXT:    retq1779;1780; X64-AVX1-LABEL: test_reduce_v64i8:1781; X64-AVX1:       ## %bb.0:1782; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21783; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31784; X64-AVX1-NEXT:    vpminsb %xmm2, %xmm3, %xmm21785; X64-AVX1-NEXT:    vpminsb %xmm1, %xmm0, %xmm01786; X64-AVX1-NEXT:    vpminsb %xmm2, %xmm0, %xmm01787; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01788; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11789; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01790; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01791; X64-AVX1-NEXT:    vmovd %xmm0, %eax1792; X64-AVX1-NEXT:    addb $-128, %al1793; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1794; X64-AVX1-NEXT:    vzeroupper1795; X64-AVX1-NEXT:    retq1796;1797; X64-AVX2-LABEL: test_reduce_v64i8:1798; X64-AVX2:       ## %bb.0:1799; X64-AVX2-NEXT:    vpminsb %ymm1, %ymm0, %ymm01800; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11801; X64-AVX2-NEXT:    vpminsb %xmm1, %xmm0, %xmm01802; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01803; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11804; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01805; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01806; X64-AVX2-NEXT:    vmovd %xmm0, %eax1807; X64-AVX2-NEXT:    addb $-128, %al1808; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1809; X64-AVX2-NEXT:    vzeroupper1810; X64-AVX2-NEXT:    retq1811;1812; X64-AVX512-LABEL: test_reduce_v64i8:1813; X64-AVX512:       ## %bb.0:1814; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11815; X64-AVX512-NEXT:    vpminsb %ymm1, %ymm0, %ymm01816; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11817; X64-AVX512-NEXT:    vpminsb %xmm1, %xmm0, %xmm01818; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01819; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm11820; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01821; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01822; X64-AVX512-NEXT:    vmovd %xmm0, %eax1823; X64-AVX512-NEXT:    addb $-128, %al1824; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax1825; X64-AVX512-NEXT:    vzeroupper1826; X64-AVX512-NEXT:    retq1827  %1  = shufflevector <64 x i8> %a0, <64 x i8> undef, <64 x i32> <i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1828  %2  = icmp slt <64 x i8> %a0, %11829  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %11830  %4  = shufflevector <64 x i8> %3, <64 x i8> undef, <64 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1831  %5  = icmp slt <64 x i8> %3, %41832  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %41833  %7  = shufflevector <64 x i8> %6, <64 x i8> undef, <64 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1834  %8  = icmp slt <64 x i8> %6, %71835  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %71836  %10 = shufflevector <64 x i8> %9, <64 x i8> undef, <64 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1837  %11 = icmp slt <64 x i8> %9, %101838  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %101839  %13 = shufflevector <64 x i8> %12, <64 x i8> undef, <64 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1840  %14 = icmp slt <64 x i8> %12, %131841  %15 = select <64 x i1> %14, <64 x i8> %12, <64 x i8> %131842  %16 = shufflevector <64 x i8> %15, <64 x i8> undef, <64 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1843  %17 = icmp slt <64 x i8> %15, %161844  %18 = select <64 x i1> %17, <64 x i8> %15, <64 x i8> %161845  %19 = extractelement <64 x i8> %18, i32 01846  ret i8 %191847}1848 1849;1850; Partial Vector Reductions1851;1852 1853define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {1854; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:1855; X86-SSE2:       ## %bb.0:1856; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1857; X86-SSE2-NEXT:    pminsw %xmm0, %xmm11858; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1859; X86-SSE2-NEXT:    pminsw %xmm1, %xmm01860; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11861; X86-SSE2-NEXT:    psrld $16, %xmm11862; X86-SSE2-NEXT:    pminsw %xmm0, %xmm11863; X86-SSE2-NEXT:    movd %xmm1, %eax1864; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1865; X86-SSE2-NEXT:    retl1866;1867; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:1868; X86-SSE42:       ## %bb.0:1869; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01870; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01871; X86-SSE42-NEXT:    movd %xmm0, %eax1872; X86-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x80001873; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1874; X86-SSE42-NEXT:    retl1875;1876; X86-AVX-LABEL: test_reduce_v16i16_v8i16:1877; X86-AVX:       ## %bb.0:1878; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01879; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm01880; X86-AVX-NEXT:    vmovd %xmm0, %eax1881; X86-AVX-NEXT:    xorl $32768, %eax ## imm = 0x80001882; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax1883; X86-AVX-NEXT:    vzeroupper1884; X86-AVX-NEXT:    retl1885;1886; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:1887; X64-SSE2:       ## %bb.0:1888; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1889; X64-SSE2-NEXT:    pminsw %xmm0, %xmm11890; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1891; X64-SSE2-NEXT:    pminsw %xmm1, %xmm01892; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11893; X64-SSE2-NEXT:    psrld $16, %xmm11894; X64-SSE2-NEXT:    pminsw %xmm0, %xmm11895; X64-SSE2-NEXT:    movd %xmm1, %eax1896; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1897; X64-SSE2-NEXT:    retq1898;1899; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:1900; X64-SSE42:       ## %bb.0:1901; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01902; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01903; X64-SSE42-NEXT:    movd %xmm0, %eax1904; X64-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x80001905; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1906; X64-SSE42-NEXT:    retq1907;1908; X64-AVX1OR2-LABEL: test_reduce_v16i16_v8i16:1909; X64-AVX1OR2:       ## %bb.0:1910; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01911; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm01912; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax1913; X64-AVX1OR2-NEXT:    xorl $32768, %eax ## imm = 0x80001914; X64-AVX1OR2-NEXT:    ## kill: def $ax killed $ax killed $eax1915; X64-AVX1OR2-NEXT:    vzeroupper1916; X64-AVX1OR2-NEXT:    retq1917;1918; X64-AVX512-LABEL: test_reduce_v16i16_v8i16:1919; X64-AVX512:       ## %bb.0:1920; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01921; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01922; X64-AVX512-NEXT:    vmovd %xmm0, %eax1923; X64-AVX512-NEXT:    xorl $32768, %eax ## imm = 0x80001924; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax1925; X64-AVX512-NEXT:    vzeroupper1926; X64-AVX512-NEXT:    retq1927  %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1928  %2  = icmp slt <16 x i16> %a0, %11929  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %11930  %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1931  %5  = icmp slt <16 x i16> %3, %41932  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %41933  %7  = shufflevector <16 x i16> %6, <16 x i16> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1934  %8  = icmp slt <16 x i16> %6, %71935  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %71936  %10 = extractelement <16 x i16> %9, i32 01937  ret i16 %101938}1939 1940define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {1941; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:1942; X86-SSE2:       ## %bb.0:1943; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1944; X86-SSE2-NEXT:    pminsw %xmm0, %xmm11945; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1946; X86-SSE2-NEXT:    pminsw %xmm1, %xmm01947; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11948; X86-SSE2-NEXT:    psrld $16, %xmm11949; X86-SSE2-NEXT:    pminsw %xmm0, %xmm11950; X86-SSE2-NEXT:    movd %xmm1, %eax1951; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1952; X86-SSE2-NEXT:    retl1953;1954; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:1955; X86-SSE42:       ## %bb.0:1956; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01957; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01958; X86-SSE42-NEXT:    movd %xmm0, %eax1959; X86-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x80001960; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1961; X86-SSE42-NEXT:    retl1962;1963; X86-AVX-LABEL: test_reduce_v32i16_v8i16:1964; X86-AVX:       ## %bb.0:1965; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01966; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm01967; X86-AVX-NEXT:    vmovd %xmm0, %eax1968; X86-AVX-NEXT:    xorl $32768, %eax ## imm = 0x80001969; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax1970; X86-AVX-NEXT:    vzeroupper1971; X86-AVX-NEXT:    retl1972;1973; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:1974; X64-SSE2:       ## %bb.0:1975; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1976; X64-SSE2-NEXT:    pminsw %xmm0, %xmm11977; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1978; X64-SSE2-NEXT:    pminsw %xmm1, %xmm01979; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11980; X64-SSE2-NEXT:    psrld $16, %xmm11981; X64-SSE2-NEXT:    pminsw %xmm0, %xmm11982; X64-SSE2-NEXT:    movd %xmm1, %eax1983; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1984; X64-SSE2-NEXT:    retq1985;1986; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:1987; X64-SSE42:       ## %bb.0:1988; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01989; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01990; X64-SSE42-NEXT:    movd %xmm0, %eax1991; X64-SSE42-NEXT:    xorl $32768, %eax ## imm = 0x80001992; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1993; X64-SSE42-NEXT:    retq1994;1995; X64-AVX1OR2-LABEL: test_reduce_v32i16_v8i16:1996; X64-AVX1OR2:       ## %bb.0:1997; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01998; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm01999; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax2000; X64-AVX1OR2-NEXT:    xorl $32768, %eax ## imm = 0x80002001; X64-AVX1OR2-NEXT:    ## kill: def $ax killed $ax killed $eax2002; X64-AVX1OR2-NEXT:    vzeroupper2003; X64-AVX1OR2-NEXT:    retq2004;2005; X64-AVX512-LABEL: test_reduce_v32i16_v8i16:2006; X64-AVX512:       ## %bb.0:2007; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm02008; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02009; X64-AVX512-NEXT:    vmovd %xmm0, %eax2010; X64-AVX512-NEXT:    xorl $32768, %eax ## imm = 0x80002011; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax2012; X64-AVX512-NEXT:    vzeroupper2013; X64-AVX512-NEXT:    retq2014  %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2015  %2  = icmp slt <32 x i16> %a0, %12016  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %12017  %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2018  %5  = icmp slt <32 x i16> %3, %42019  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %42020  %7  = shufflevector <32 x i16> %6, <32 x i16> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2021  %8  = icmp slt <32 x i16> %6, %72022  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %72023  %10 = extractelement <32 x i16> %9, i32 02024  ret i16 %102025}2026 2027define i8 @test_reduce_v32i8_v16i8(<32 x i8> %a0) {2028; X86-SSE2-LABEL: test_reduce_v32i8_v16i8:2029; X86-SSE2:       ## %bb.0:2030; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2031; X86-SSE2-NEXT:    movdqa %xmm1, %xmm22032; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm22033; X86-SSE2-NEXT:    pand %xmm2, %xmm02034; X86-SSE2-NEXT:    pandn %xmm1, %xmm22035; X86-SSE2-NEXT:    por %xmm0, %xmm22036; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]2037; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02038; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm02039; X86-SSE2-NEXT:    pand %xmm0, %xmm22040; X86-SSE2-NEXT:    pandn %xmm1, %xmm02041; X86-SSE2-NEXT:    por %xmm2, %xmm02042; X86-SSE2-NEXT:    movdqa %xmm0, %xmm22043; X86-SSE2-NEXT:    psrld $16, %xmm22044; X86-SSE2-NEXT:    movdqa %xmm2, %xmm12045; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm12046; X86-SSE2-NEXT:    pand %xmm1, %xmm02047; X86-SSE2-NEXT:    pandn %xmm2, %xmm12048; X86-SSE2-NEXT:    por %xmm0, %xmm12049; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02050; X86-SSE2-NEXT:    psrlw $8, %xmm02051; X86-SSE2-NEXT:    movdqa %xmm0, %xmm22052; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm22053; X86-SSE2-NEXT:    pand %xmm2, %xmm12054; X86-SSE2-NEXT:    pandn %xmm0, %xmm22055; X86-SSE2-NEXT:    por %xmm1, %xmm22056; X86-SSE2-NEXT:    movd %xmm2, %eax2057; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2058; X86-SSE2-NEXT:    retl2059;2060; X86-SSE42-LABEL: test_reduce_v32i8_v16i8:2061; X86-SSE42:       ## %bb.0:2062; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm02063; X86-SSE42-NEXT:    movdqa %xmm0, %xmm12064; X86-SSE42-NEXT:    psrlw $8, %xmm12065; X86-SSE42-NEXT:    pminub %xmm0, %xmm12066; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm02067; X86-SSE42-NEXT:    movd %xmm0, %eax2068; X86-SSE42-NEXT:    addb $-128, %al2069; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2070; X86-SSE42-NEXT:    retl2071;2072; X86-AVX-LABEL: test_reduce_v32i8_v16i8:2073; X86-AVX:       ## %bb.0:2074; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm02075; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12076; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02077; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02078; X86-AVX-NEXT:    vmovd %xmm0, %eax2079; X86-AVX-NEXT:    addb $-128, %al2080; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax2081; X86-AVX-NEXT:    vzeroupper2082; X86-AVX-NEXT:    retl2083;2084; X64-SSE2-LABEL: test_reduce_v32i8_v16i8:2085; X64-SSE2:       ## %bb.0:2086; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2087; X64-SSE2-NEXT:    movdqa %xmm1, %xmm22088; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm22089; X64-SSE2-NEXT:    pand %xmm2, %xmm02090; X64-SSE2-NEXT:    pandn %xmm1, %xmm22091; X64-SSE2-NEXT:    por %xmm0, %xmm22092; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]2093; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12094; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm12095; X64-SSE2-NEXT:    pand %xmm1, %xmm22096; X64-SSE2-NEXT:    pandn %xmm0, %xmm12097; X64-SSE2-NEXT:    por %xmm2, %xmm12098; X64-SSE2-NEXT:    movdqa %xmm1, %xmm02099; X64-SSE2-NEXT:    psrld $16, %xmm02100; X64-SSE2-NEXT:    movdqa %xmm0, %xmm22101; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm22102; X64-SSE2-NEXT:    pand %xmm2, %xmm12103; X64-SSE2-NEXT:    pandn %xmm0, %xmm22104; X64-SSE2-NEXT:    por %xmm1, %xmm22105; X64-SSE2-NEXT:    movdqa %xmm2, %xmm02106; X64-SSE2-NEXT:    psrlw $8, %xmm02107; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12108; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm12109; X64-SSE2-NEXT:    pand %xmm1, %xmm22110; X64-SSE2-NEXT:    pandn %xmm0, %xmm12111; X64-SSE2-NEXT:    por %xmm2, %xmm12112; X64-SSE2-NEXT:    movd %xmm1, %eax2113; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2114; X64-SSE2-NEXT:    retq2115;2116; X64-SSE42-LABEL: test_reduce_v32i8_v16i8:2117; X64-SSE42:       ## %bb.0:2118; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm02119; X64-SSE42-NEXT:    movdqa %xmm0, %xmm12120; X64-SSE42-NEXT:    psrlw $8, %xmm12121; X64-SSE42-NEXT:    pminub %xmm0, %xmm12122; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm02123; X64-SSE42-NEXT:    movd %xmm0, %eax2124; X64-SSE42-NEXT:    addb $-128, %al2125; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2126; X64-SSE42-NEXT:    retq2127;2128; X64-AVX1OR2-LABEL: test_reduce_v32i8_v16i8:2129; X64-AVX1OR2:       ## %bb.0:2130; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm02131; X64-AVX1OR2-NEXT:    vpsrlw $8, %xmm0, %xmm12132; X64-AVX1OR2-NEXT:    vpminub %xmm1, %xmm0, %xmm02133; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm02134; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax2135; X64-AVX1OR2-NEXT:    addb $-128, %al2136; X64-AVX1OR2-NEXT:    ## kill: def $al killed $al killed $eax2137; X64-AVX1OR2-NEXT:    vzeroupper2138; X64-AVX1OR2-NEXT:    retq2139;2140; X64-AVX512-LABEL: test_reduce_v32i8_v16i8:2141; X64-AVX512:       ## %bb.0:2142; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm02143; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm12144; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm02145; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02146; X64-AVX512-NEXT:    vmovd %xmm0, %eax2147; X64-AVX512-NEXT:    addb $-128, %al2148; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax2149; X64-AVX512-NEXT:    vzeroupper2150; X64-AVX512-NEXT:    retq2151  %1  = shufflevector <32 x i8> %a0, <32 x i8> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2152  %2  = icmp slt <32 x i8> %a0, %12153  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %12154  %4  = shufflevector <32 x i8> %3, <32 x i8> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2155  %5  = icmp slt <32 x i8> %3, %42156  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %42157  %7  = shufflevector <32 x i8> %6, <32 x i8> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2158  %8  = icmp slt <32 x i8> %6, %72159  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %72160  %10 = shufflevector <32 x i8> %9, <32 x i8> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2161  %11 = icmp slt <32 x i8> %9, %102162  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %102163  %13 = extractelement <32 x i8> %12, i32 02164  ret i8 %132165}2166 2167define i8 @test_reduce_v64i8_v16i8(<64 x i8> %a0) {2168; X86-SSE2-LABEL: test_reduce_v64i8_v16i8:2169; X86-SSE2:       ## %bb.0:2170; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2171; X86-SSE2-NEXT:    movdqa %xmm1, %xmm22172; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm22173; X86-SSE2-NEXT:    pand %xmm2, %xmm02174; X86-SSE2-NEXT:    pandn %xmm1, %xmm22175; X86-SSE2-NEXT:    por %xmm0, %xmm22176; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]2177; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02178; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm02179; X86-SSE2-NEXT:    pand %xmm0, %xmm22180; X86-SSE2-NEXT:    pandn %xmm1, %xmm02181; X86-SSE2-NEXT:    por %xmm2, %xmm02182; X86-SSE2-NEXT:    movdqa %xmm0, %xmm22183; X86-SSE2-NEXT:    psrld $16, %xmm22184; X86-SSE2-NEXT:    movdqa %xmm2, %xmm12185; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm12186; X86-SSE2-NEXT:    pand %xmm1, %xmm02187; X86-SSE2-NEXT:    pandn %xmm2, %xmm12188; X86-SSE2-NEXT:    por %xmm0, %xmm12189; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02190; X86-SSE2-NEXT:    psrlw $8, %xmm02191; X86-SSE2-NEXT:    movdqa %xmm0, %xmm22192; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm22193; X86-SSE2-NEXT:    pand %xmm2, %xmm12194; X86-SSE2-NEXT:    pandn %xmm0, %xmm22195; X86-SSE2-NEXT:    por %xmm1, %xmm22196; X86-SSE2-NEXT:    movd %xmm2, %eax2197; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2198; X86-SSE2-NEXT:    retl2199;2200; X86-SSE42-LABEL: test_reduce_v64i8_v16i8:2201; X86-SSE42:       ## %bb.0:2202; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm02203; X86-SSE42-NEXT:    movdqa %xmm0, %xmm12204; X86-SSE42-NEXT:    psrlw $8, %xmm12205; X86-SSE42-NEXT:    pminub %xmm0, %xmm12206; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm02207; X86-SSE42-NEXT:    movd %xmm0, %eax2208; X86-SSE42-NEXT:    addb $-128, %al2209; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2210; X86-SSE42-NEXT:    retl2211;2212; X86-AVX-LABEL: test_reduce_v64i8_v16i8:2213; X86-AVX:       ## %bb.0:2214; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm02215; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12216; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02217; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02218; X86-AVX-NEXT:    vmovd %xmm0, %eax2219; X86-AVX-NEXT:    addb $-128, %al2220; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax2221; X86-AVX-NEXT:    vzeroupper2222; X86-AVX-NEXT:    retl2223;2224; X64-SSE2-LABEL: test_reduce_v64i8_v16i8:2225; X64-SSE2:       ## %bb.0:2226; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2227; X64-SSE2-NEXT:    movdqa %xmm1, %xmm22228; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm22229; X64-SSE2-NEXT:    pand %xmm2, %xmm02230; X64-SSE2-NEXT:    pandn %xmm1, %xmm22231; X64-SSE2-NEXT:    por %xmm0, %xmm22232; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]2233; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12234; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm12235; X64-SSE2-NEXT:    pand %xmm1, %xmm22236; X64-SSE2-NEXT:    pandn %xmm0, %xmm12237; X64-SSE2-NEXT:    por %xmm2, %xmm12238; X64-SSE2-NEXT:    movdqa %xmm1, %xmm02239; X64-SSE2-NEXT:    psrld $16, %xmm02240; X64-SSE2-NEXT:    movdqa %xmm0, %xmm22241; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm22242; X64-SSE2-NEXT:    pand %xmm2, %xmm12243; X64-SSE2-NEXT:    pandn %xmm0, %xmm22244; X64-SSE2-NEXT:    por %xmm1, %xmm22245; X64-SSE2-NEXT:    movdqa %xmm2, %xmm02246; X64-SSE2-NEXT:    psrlw $8, %xmm02247; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12248; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm12249; X64-SSE2-NEXT:    pand %xmm1, %xmm22250; X64-SSE2-NEXT:    pandn %xmm0, %xmm12251; X64-SSE2-NEXT:    por %xmm2, %xmm12252; X64-SSE2-NEXT:    movd %xmm1, %eax2253; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2254; X64-SSE2-NEXT:    retq2255;2256; X64-SSE42-LABEL: test_reduce_v64i8_v16i8:2257; X64-SSE42:       ## %bb.0:2258; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm02259; X64-SSE42-NEXT:    movdqa %xmm0, %xmm12260; X64-SSE42-NEXT:    psrlw $8, %xmm12261; X64-SSE42-NEXT:    pminub %xmm0, %xmm12262; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm02263; X64-SSE42-NEXT:    movd %xmm0, %eax2264; X64-SSE42-NEXT:    addb $-128, %al2265; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2266; X64-SSE42-NEXT:    retq2267;2268; X64-AVX1OR2-LABEL: test_reduce_v64i8_v16i8:2269; X64-AVX1OR2:       ## %bb.0:2270; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm02271; X64-AVX1OR2-NEXT:    vpsrlw $8, %xmm0, %xmm12272; X64-AVX1OR2-NEXT:    vpminub %xmm1, %xmm0, %xmm02273; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm02274; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax2275; X64-AVX1OR2-NEXT:    addb $-128, %al2276; X64-AVX1OR2-NEXT:    ## kill: def $al killed $al killed $eax2277; X64-AVX1OR2-NEXT:    vzeroupper2278; X64-AVX1OR2-NEXT:    retq2279;2280; X64-AVX512-LABEL: test_reduce_v64i8_v16i8:2281; X64-AVX512:       ## %bb.0:2282; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm02283; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm12284; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm02285; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02286; X64-AVX512-NEXT:    vmovd %xmm0, %eax2287; X64-AVX512-NEXT:    addb $-128, %al2288; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax2289; X64-AVX512-NEXT:    vzeroupper2290; X64-AVX512-NEXT:    retq2291  %1  = shufflevector <64 x i8> %a0, <64 x i8> undef, <64 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2292  %2  = icmp slt <64 x i8> %a0, %12293  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %12294  %4  = shufflevector <64 x i8> %3, <64 x i8> undef, <64 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2295  %5  = icmp slt <64 x i8> %3, %42296  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %42297  %7  = shufflevector <64 x i8> %6, <64 x i8> undef, <64 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2298  %8  = icmp slt <64 x i8> %6, %72299  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %72300  %10 = shufflevector <64 x i8> %9, <64 x i8> undef, <64 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2301  %11 = icmp slt <64 x i8> %9, %102302  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %102303  %13 = extractelement <64 x i8> %12, i32 02304  ret i8 %132305}2306