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-AVX19; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx2   | FileCheck %s --check-prefixes=X64-AVX,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 = [2147483648,2147483648,2147483648,2147483648]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 {{.*#+}} xmm0 = [0,2147483648,0,2147483648]45; X86-SSE42-NEXT:    movdqa %xmm1, %xmm346; X86-SSE42-NEXT:    pxor %xmm0, %xmm347; X86-SSE42-NEXT:    pxor %xmm2, %xmm048; X86-SSE42-NEXT:    pcmpgtq %xmm3, %xmm049; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm250; X86-SSE42-NEXT:    movd %xmm2, %eax51; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx52; X86-SSE42-NEXT:    retl53;54; X86-AVX1-LABEL: test_reduce_v2i64:55; X86-AVX1:       ## %bb.0:56; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]57; X86-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [0,2147483648,0,2147483648]58; X86-AVX1-NEXT:    ## xmm2 = mem[0,0]59; X86-AVX1-NEXT:    vxorps %xmm2, %xmm0, %xmm360; X86-AVX1-NEXT:    vxorps %xmm2, %xmm1, %xmm261; X86-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm262; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm063; X86-AVX1-NEXT:    vmovd %xmm0, %eax64; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx65; X86-AVX1-NEXT:    retl66;67; X86-AVX2-LABEL: test_reduce_v2i64:68; X86-AVX2:       ## %bb.0:69; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]70; X86-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [0,2147483648,0,2147483648]71; X86-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm372; X86-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm273; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm274; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm075; X86-AVX2-NEXT:    vmovd %xmm0, %eax76; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx77; X86-AVX2-NEXT:    retl78;79; X64-SSE2-LABEL: test_reduce_v2i64:80; X64-SSE2:       ## %bb.0:81; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]82; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]83; X64-SSE2-NEXT:    movdqa %xmm0, %xmm384; X64-SSE2-NEXT:    pxor %xmm2, %xmm385; X64-SSE2-NEXT:    pxor %xmm1, %xmm286; X64-SSE2-NEXT:    movdqa %xmm2, %xmm487; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm488; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]89; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm290; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]91; X64-SSE2-NEXT:    pand %xmm5, %xmm292; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]93; X64-SSE2-NEXT:    por %xmm2, %xmm394; X64-SSE2-NEXT:    pand %xmm3, %xmm095; X64-SSE2-NEXT:    pandn %xmm1, %xmm396; X64-SSE2-NEXT:    por %xmm0, %xmm397; X64-SSE2-NEXT:    movq %xmm3, %rax98; X64-SSE2-NEXT:    retq99;100; X64-SSE42-LABEL: test_reduce_v2i64:101; X64-SSE42:       ## %bb.0:102; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1103; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]104; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm0 = [9223372036854775808,9223372036854775808]105; X64-SSE42-NEXT:    movdqa %xmm1, %xmm3106; X64-SSE42-NEXT:    pxor %xmm0, %xmm3107; X64-SSE42-NEXT:    pxor %xmm2, %xmm0108; X64-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0109; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2110; X64-SSE42-NEXT:    movq %xmm2, %rax111; X64-SSE42-NEXT:    retq112;113; X64-AVX1-LABEL: test_reduce_v2i64:114; X64-AVX1:       ## %bb.0:115; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]116; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]117; X64-AVX1-NEXT:    ## xmm2 = mem[0,0]118; X64-AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm3119; X64-AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm2120; X64-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm2121; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0122; X64-AVX1-NEXT:    vmovq %xmm0, %rax123; X64-AVX1-NEXT:    retq124;125; X64-AVX2-LABEL: test_reduce_v2i64:126; X64-AVX2:       ## %bb.0:127; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]128; X64-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]129; X64-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm3130; X64-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm2131; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm2132; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0133; X64-AVX2-NEXT:    vmovq %xmm0, %rax134; X64-AVX2-NEXT:    retq135;136; X64-AVX512-LABEL: test_reduce_v2i64:137; X64-AVX512:       ## %bb.0:138; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]139; X64-AVX512-NEXT:    vpminuq %xmm1, %xmm0, %xmm0140; X64-AVX512-NEXT:    vmovq %xmm0, %rax141; X64-AVX512-NEXT:    retq142  %1 = shufflevector <2 x i64> %a0, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>143  %2 = icmp ult <2 x i64> %a0, %1144  %3 = select <2 x i1> %2, <2 x i64> %a0, <2 x i64> %1145  %4 = extractelement <2 x i64> %3, i32 0146  ret i64 %4147}148 149define i32 @test_reduce_v4i32(<4 x i32> %a0) {150; X86-SSE2-LABEL: test_reduce_v4i32:151; X86-SSE2:       ## %bb.0:152; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]153; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]154; X86-SSE2-NEXT:    movdqa %xmm0, %xmm4155; X86-SSE2-NEXT:    pxor %xmm1, %xmm4156; X86-SSE2-NEXT:    movdqa %xmm3, %xmm2157; X86-SSE2-NEXT:    pxor %xmm1, %xmm2158; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm2159; X86-SSE2-NEXT:    pand %xmm2, %xmm0160; X86-SSE2-NEXT:    pandn %xmm3, %xmm2161; X86-SSE2-NEXT:    por %xmm0, %xmm2162; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]163; X86-SSE2-NEXT:    movdqa %xmm2, %xmm3164; X86-SSE2-NEXT:    pxor %xmm1, %xmm3165; X86-SSE2-NEXT:    pxor %xmm0, %xmm1166; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm1167; X86-SSE2-NEXT:    pand %xmm1, %xmm2168; X86-SSE2-NEXT:    pandn %xmm0, %xmm1169; X86-SSE2-NEXT:    por %xmm2, %xmm1170; X86-SSE2-NEXT:    movd %xmm1, %eax171; X86-SSE2-NEXT:    retl172;173; X86-SSE42-LABEL: test_reduce_v4i32:174; X86-SSE42:       ## %bb.0:175; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]176; X86-SSE42-NEXT:    pminud %xmm0, %xmm1177; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]178; X86-SSE42-NEXT:    pminud %xmm1, %xmm0179; X86-SSE42-NEXT:    movd %xmm0, %eax180; X86-SSE42-NEXT:    retl181;182; X86-AVX-LABEL: test_reduce_v4i32:183; X86-AVX:       ## %bb.0:184; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]185; X86-AVX-NEXT:    vpminud %xmm1, %xmm0, %xmm0186; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]187; X86-AVX-NEXT:    vpminud %xmm1, %xmm0, %xmm0188; X86-AVX-NEXT:    vmovd %xmm0, %eax189; X86-AVX-NEXT:    retl190;191; X64-SSE2-LABEL: test_reduce_v4i32:192; X64-SSE2:       ## %bb.0:193; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]194; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]195; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3196; X64-SSE2-NEXT:    pxor %xmm2, %xmm3197; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4198; X64-SSE2-NEXT:    pxor %xmm2, %xmm4199; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm4200; X64-SSE2-NEXT:    pand %xmm4, %xmm0201; X64-SSE2-NEXT:    pandn %xmm1, %xmm4202; X64-SSE2-NEXT:    por %xmm0, %xmm4203; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,1,1]204; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1205; X64-SSE2-NEXT:    pxor %xmm2, %xmm1206; X64-SSE2-NEXT:    pxor %xmm0, %xmm2207; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2208; X64-SSE2-NEXT:    pand %xmm2, %xmm4209; X64-SSE2-NEXT:    pandn %xmm0, %xmm2210; X64-SSE2-NEXT:    por %xmm4, %xmm2211; X64-SSE2-NEXT:    movd %xmm2, %eax212; X64-SSE2-NEXT:    retq213;214; X64-SSE42-LABEL: test_reduce_v4i32:215; X64-SSE42:       ## %bb.0:216; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]217; X64-SSE42-NEXT:    pminud %xmm0, %xmm1218; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]219; X64-SSE42-NEXT:    pminud %xmm1, %xmm0220; X64-SSE42-NEXT:    movd %xmm0, %eax221; X64-SSE42-NEXT:    retq222;223; X64-AVX-LABEL: test_reduce_v4i32:224; X64-AVX:       ## %bb.0:225; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]226; X64-AVX-NEXT:    vpminud %xmm1, %xmm0, %xmm0227; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]228; X64-AVX-NEXT:    vpminud %xmm1, %xmm0, %xmm0229; X64-AVX-NEXT:    vmovd %xmm0, %eax230; X64-AVX-NEXT:    retq231  %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>232  %2 = icmp ult <4 x i32> %a0, %1233  %3 = select <4 x i1> %2, <4 x i32> %a0, <4 x i32> %1234  %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>235  %5 = icmp ult <4 x i32> %3, %4236  %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4237  %7 = extractelement <4 x i32> %6, i32 0238  ret i32 %7239}240 241define i16 @test_reduce_v8i16(<8 x i16> %a0) {242; X86-SSE2-LABEL: test_reduce_v8i16:243; X86-SSE2:       ## %bb.0:244; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]245; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2246; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2247; X86-SSE2-NEXT:    psubw %xmm2, %xmm0248; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]249; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2250; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2251; X86-SSE2-NEXT:    psubw %xmm2, %xmm0252; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1253; X86-SSE2-NEXT:    psrld $16, %xmm1254; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2255; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2256; X86-SSE2-NEXT:    psubw %xmm2, %xmm0257; X86-SSE2-NEXT:    movd %xmm0, %eax258; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax259; X86-SSE2-NEXT:    retl260;261; X86-SSE42-LABEL: test_reduce_v8i16:262; X86-SSE42:       ## %bb.0:263; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0264; X86-SSE42-NEXT:    movd %xmm0, %eax265; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax266; X86-SSE42-NEXT:    retl267;268; X86-AVX-LABEL: test_reduce_v8i16:269; X86-AVX:       ## %bb.0:270; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0271; X86-AVX-NEXT:    vmovd %xmm0, %eax272; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax273; X86-AVX-NEXT:    retl274;275; X64-SSE2-LABEL: test_reduce_v8i16:276; X64-SSE2:       ## %bb.0:277; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]278; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2279; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2280; X64-SSE2-NEXT:    psubw %xmm2, %xmm0281; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]282; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2283; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2284; X64-SSE2-NEXT:    psubw %xmm2, %xmm0285; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1286; X64-SSE2-NEXT:    psrld $16, %xmm1287; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2288; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2289; X64-SSE2-NEXT:    psubw %xmm2, %xmm0290; X64-SSE2-NEXT:    movd %xmm0, %eax291; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax292; X64-SSE2-NEXT:    retq293;294; X64-SSE42-LABEL: test_reduce_v8i16:295; X64-SSE42:       ## %bb.0:296; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0297; X64-SSE42-NEXT:    movd %xmm0, %eax298; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax299; X64-SSE42-NEXT:    retq300;301; X64-AVX-LABEL: test_reduce_v8i16:302; X64-AVX:       ## %bb.0:303; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm0304; X64-AVX-NEXT:    vmovd %xmm0, %eax305; X64-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax306; X64-AVX-NEXT:    retq307  %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>308  %2  = icmp ult <8 x i16> %a0, %1309  %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1310  %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>311  %5  = icmp ult <8 x i16> %3, %4312  %6  = select <8 x i1> %5, <8 x i16> %3, <8 x i16> %4313  %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>314  %8  = icmp ult <8 x i16> %6, %7315  %9  = select <8 x i1> %8, <8 x i16> %6, <8 x i16> %7316  %10 = extractelement <8 x i16> %9, i32 0317  ret i16 %10318}319 320define i8 @test_reduce_v16i8(<16 x i8> %a0) {321; X86-SSE2-LABEL: test_reduce_v16i8:322; X86-SSE2:       ## %bb.0:323; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]324; X86-SSE2-NEXT:    pminub %xmm0, %xmm1325; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]326; X86-SSE2-NEXT:    pminub %xmm1, %xmm0327; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1328; X86-SSE2-NEXT:    psrld $16, %xmm1329; X86-SSE2-NEXT:    pminub %xmm0, %xmm1330; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0331; X86-SSE2-NEXT:    psrlw $8, %xmm0332; X86-SSE2-NEXT:    pminub %xmm1, %xmm0333; X86-SSE2-NEXT:    movd %xmm0, %eax334; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax335; X86-SSE2-NEXT:    retl336;337; X86-SSE42-LABEL: test_reduce_v16i8:338; X86-SSE42:       ## %bb.0:339; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1340; X86-SSE42-NEXT:    psrlw $8, %xmm1341; X86-SSE42-NEXT:    pminub %xmm0, %xmm1342; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0343; X86-SSE42-NEXT:    movd %xmm0, %eax344; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax345; X86-SSE42-NEXT:    retl346;347; X86-AVX-LABEL: test_reduce_v16i8:348; X86-AVX:       ## %bb.0:349; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1350; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0351; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0352; X86-AVX-NEXT:    vmovd %xmm0, %eax353; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax354; X86-AVX-NEXT:    retl355;356; X64-SSE2-LABEL: test_reduce_v16i8:357; X64-SSE2:       ## %bb.0:358; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]359; X64-SSE2-NEXT:    pminub %xmm0, %xmm1360; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]361; X64-SSE2-NEXT:    pminub %xmm1, %xmm0362; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1363; X64-SSE2-NEXT:    psrld $16, %xmm1364; X64-SSE2-NEXT:    pminub %xmm0, %xmm1365; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0366; X64-SSE2-NEXT:    psrlw $8, %xmm0367; X64-SSE2-NEXT:    pminub %xmm1, %xmm0368; X64-SSE2-NEXT:    movd %xmm0, %eax369; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax370; X64-SSE2-NEXT:    retq371;372; X64-SSE42-LABEL: test_reduce_v16i8:373; X64-SSE42:       ## %bb.0:374; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1375; X64-SSE42-NEXT:    psrlw $8, %xmm1376; X64-SSE42-NEXT:    pminub %xmm0, %xmm1377; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0378; X64-SSE42-NEXT:    movd %xmm0, %eax379; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax380; X64-SSE42-NEXT:    retq381;382; X64-AVX-LABEL: test_reduce_v16i8:383; X64-AVX:       ## %bb.0:384; X64-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1385; X64-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0386; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm0387; X64-AVX-NEXT:    vmovd %xmm0, %eax388; X64-AVX-NEXT:    ## kill: def $al killed $al killed $eax389; X64-AVX-NEXT:    retq390  %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>391  %2  = icmp ult <16 x i8> %a0, %1392  %3  = select <16 x i1> %2, <16 x i8> %a0, <16 x i8> %1393  %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>394  %5  = icmp ult <16 x i8> %3, %4395  %6  = select <16 x i1> %5, <16 x i8> %3, <16 x i8> %4396  %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>397  %8  = icmp ult <16 x i8> %6, %7398  %9  = select <16 x i1> %8, <16 x i8> %6, <16 x i8> %7399  %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>400  %11 = icmp ult <16 x i8> %9, %10401  %12 = select <16 x i1> %11, <16 x i8> %9, <16 x i8> %10402  %13 = extractelement <16 x i8> %12, i32 0403  ret i8 %13404}405 406;407; 256-bit Vectors408;409 410define i64 @test_reduce_v4i64(<4 x i64> %a0) {411; X86-SSE2-LABEL: test_reduce_v4i64:412; X86-SSE2:       ## %bb.0:413; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]414; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3415; X86-SSE2-NEXT:    pxor %xmm2, %xmm3416; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4417; X86-SSE2-NEXT:    pxor %xmm2, %xmm4418; X86-SSE2-NEXT:    movdqa %xmm4, %xmm5419; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5420; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]421; X86-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4422; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]423; X86-SSE2-NEXT:    pand %xmm6, %xmm4424; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]425; X86-SSE2-NEXT:    por %xmm4, %xmm3426; X86-SSE2-NEXT:    pand %xmm3, %xmm0427; X86-SSE2-NEXT:    pandn %xmm1, %xmm3428; X86-SSE2-NEXT:    por %xmm0, %xmm3429; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]430; X86-SSE2-NEXT:    movdqa %xmm3, %xmm1431; X86-SSE2-NEXT:    pxor %xmm2, %xmm1432; X86-SSE2-NEXT:    pxor %xmm0, %xmm2433; X86-SSE2-NEXT:    movdqa %xmm2, %xmm4434; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm4435; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]436; X86-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2437; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]438; X86-SSE2-NEXT:    pand %xmm5, %xmm1439; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]440; X86-SSE2-NEXT:    por %xmm1, %xmm2441; X86-SSE2-NEXT:    pand %xmm2, %xmm3442; X86-SSE2-NEXT:    pandn %xmm0, %xmm2443; X86-SSE2-NEXT:    por %xmm3, %xmm2444; X86-SSE2-NEXT:    movd %xmm2, %eax445; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]446; X86-SSE2-NEXT:    movd %xmm0, %edx447; X86-SSE2-NEXT:    retl448;449; X86-SSE42-LABEL: test_reduce_v4i64:450; X86-SSE42:       ## %bb.0:451; X86-SSE42-NEXT:    movdqa %xmm0, %xmm3452; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm2 = [0,2147483648,0,2147483648]453; X86-SSE42-NEXT:    movdqa %xmm0, %xmm4454; X86-SSE42-NEXT:    pxor %xmm2, %xmm4455; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0456; X86-SSE42-NEXT:    pxor %xmm2, %xmm0457; X86-SSE42-NEXT:    pcmpgtq %xmm4, %xmm0458; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm1459; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]460; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0461; X86-SSE42-NEXT:    pxor %xmm2, %xmm0462; X86-SSE42-NEXT:    pxor %xmm3, %xmm2463; X86-SSE42-NEXT:    pcmpgtq %xmm0, %xmm2464; X86-SSE42-NEXT:    movdqa %xmm2, %xmm0465; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm3466; X86-SSE42-NEXT:    movd %xmm3, %eax467; X86-SSE42-NEXT:    pextrd $1, %xmm3, %edx468; X86-SSE42-NEXT:    retl469;470; X86-AVX1-LABEL: test_reduce_v4i64:471; X86-AVX1:       ## %bb.0:472; X86-AVX1-NEXT:    vmovddup {{.*#+}} xmm1 = [0,2147483648,0,2147483648]473; X86-AVX1-NEXT:    ## xmm1 = mem[0,0]474; X86-AVX1-NEXT:    vxorps %xmm1, %xmm0, %xmm2475; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3476; X86-AVX1-NEXT:    vxorps %xmm1, %xmm3, %xmm4477; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm4, %xmm2478; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm3, %xmm0479; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm0[2,3,2,3]480; X86-AVX1-NEXT:    vxorpd %xmm1, %xmm0, %xmm3481; X86-AVX1-NEXT:    vxorpd %xmm1, %xmm2, %xmm1482; X86-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm1, %xmm1483; X86-AVX1-NEXT:    vblendvpd %xmm1, %xmm0, %xmm2, %xmm0484; X86-AVX1-NEXT:    vmovd %xmm0, %eax485; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx486; X86-AVX1-NEXT:    vzeroupper487; X86-AVX1-NEXT:    retl488;489; X86-AVX2-LABEL: test_reduce_v4i64:490; X86-AVX2:       ## %bb.0:491; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1492; X86-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [0,2147483648,0,2147483648]493; X86-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm3494; X86-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm4495; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3496; X86-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0497; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]498; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm3499; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm2500; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm2501; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0502; X86-AVX2-NEXT:    vmovd %xmm0, %eax503; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx504; X86-AVX2-NEXT:    vzeroupper505; X86-AVX2-NEXT:    retl506;507; X64-SSE2-LABEL: test_reduce_v4i64:508; X64-SSE2:       ## %bb.0:509; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]510; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3511; X64-SSE2-NEXT:    pxor %xmm2, %xmm3512; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4513; X64-SSE2-NEXT:    pxor %xmm2, %xmm4514; X64-SSE2-NEXT:    movdqa %xmm4, %xmm5515; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5516; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]517; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4518; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]519; X64-SSE2-NEXT:    pand %xmm6, %xmm3520; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]521; X64-SSE2-NEXT:    por %xmm3, %xmm4522; X64-SSE2-NEXT:    pand %xmm4, %xmm0523; X64-SSE2-NEXT:    pandn %xmm1, %xmm4524; X64-SSE2-NEXT:    por %xmm0, %xmm4525; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]526; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1527; X64-SSE2-NEXT:    pxor %xmm2, %xmm1528; X64-SSE2-NEXT:    pxor %xmm0, %xmm2529; X64-SSE2-NEXT:    movdqa %xmm2, %xmm3530; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm3531; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]532; X64-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2533; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]534; X64-SSE2-NEXT:    pand %xmm5, %xmm1535; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]536; X64-SSE2-NEXT:    por %xmm1, %xmm2537; X64-SSE2-NEXT:    pand %xmm2, %xmm4538; X64-SSE2-NEXT:    pandn %xmm0, %xmm2539; X64-SSE2-NEXT:    por %xmm4, %xmm2540; X64-SSE2-NEXT:    movq %xmm2, %rax541; X64-SSE2-NEXT:    retq542;543; X64-SSE42-LABEL: test_reduce_v4i64:544; X64-SSE42:       ## %bb.0:545; X64-SSE42-NEXT:    movdqa %xmm0, %xmm2546; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]547; X64-SSE42-NEXT:    movdqa %xmm0, %xmm4548; X64-SSE42-NEXT:    pxor %xmm3, %xmm4549; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0550; X64-SSE42-NEXT:    pxor %xmm3, %xmm0551; X64-SSE42-NEXT:    pcmpgtq %xmm4, %xmm0552; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1553; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]554; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0555; X64-SSE42-NEXT:    pxor %xmm3, %xmm0556; X64-SSE42-NEXT:    pxor %xmm2, %xmm3557; X64-SSE42-NEXT:    pcmpgtq %xmm0, %xmm3558; X64-SSE42-NEXT:    movdqa %xmm3, %xmm0559; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2560; X64-SSE42-NEXT:    movq %xmm2, %rax561; X64-SSE42-NEXT:    retq562;563; X64-AVX1-LABEL: test_reduce_v4i64:564; X64-AVX1:       ## %bb.0:565; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]566; X64-AVX1-NEXT:    ## xmm1 = mem[0,0]567; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm2568; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3569; X64-AVX1-NEXT:    vpxor %xmm1, %xmm3, %xmm4570; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm4, %xmm2571; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm3, %xmm0572; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm0[2,3,2,3]573; X64-AVX1-NEXT:    vxorpd %xmm1, %xmm0, %xmm3574; X64-AVX1-NEXT:    vxorpd %xmm1, %xmm2, %xmm1575; X64-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm1, %xmm1576; X64-AVX1-NEXT:    vblendvpd %xmm1, %xmm0, %xmm2, %xmm0577; X64-AVX1-NEXT:    vmovq %xmm0, %rax578; X64-AVX1-NEXT:    vzeroupper579; X64-AVX1-NEXT:    retq580;581; X64-AVX2-LABEL: test_reduce_v4i64:582; X64-AVX2:       ## %bb.0:583; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1584; X64-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]585; X64-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm3586; X64-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm4587; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3588; X64-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0589; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]590; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm3591; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm2592; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm2593; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0594; X64-AVX2-NEXT:    vmovq %xmm0, %rax595; X64-AVX2-NEXT:    vzeroupper596; X64-AVX2-NEXT:    retq597;598; X64-AVX512-LABEL: test_reduce_v4i64:599; X64-AVX512:       ## %bb.0:600; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1601; X64-AVX512-NEXT:    vpminuq %xmm1, %xmm0, %xmm0602; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]603; X64-AVX512-NEXT:    vpminuq %xmm1, %xmm0, %xmm0604; X64-AVX512-NEXT:    vmovq %xmm0, %rax605; X64-AVX512-NEXT:    vzeroupper606; X64-AVX512-NEXT:    retq607  %1 = shufflevector <4 x i64> %a0, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>608  %2 = icmp ult <4 x i64> %a0, %1609  %3 = select <4 x i1> %2, <4 x i64> %a0, <4 x i64> %1610  %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>611  %5 = icmp ult <4 x i64> %3, %4612  %6 = select <4 x i1> %5, <4 x i64> %3, <4 x i64> %4613  %7 = extractelement <4 x i64> %6, i32 0614  ret i64 %7615}616 617define i32 @test_reduce_v8i32(<8 x i32> %a0) {618; X86-SSE2-LABEL: test_reduce_v8i32:619; X86-SSE2:       ## %bb.0:620; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]621; X86-SSE2-NEXT:    movdqa %xmm0, %xmm4622; X86-SSE2-NEXT:    pxor %xmm2, %xmm4623; X86-SSE2-NEXT:    movdqa %xmm1, %xmm3624; X86-SSE2-NEXT:    pxor %xmm2, %xmm3625; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm3626; X86-SSE2-NEXT:    pand %xmm3, %xmm0627; X86-SSE2-NEXT:    pandn %xmm1, %xmm3628; X86-SSE2-NEXT:    por %xmm0, %xmm3629; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]630; X86-SSE2-NEXT:    movdqa %xmm3, %xmm4631; X86-SSE2-NEXT:    pxor %xmm2, %xmm4632; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0633; X86-SSE2-NEXT:    pxor %xmm2, %xmm0634; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm0635; X86-SSE2-NEXT:    pand %xmm0, %xmm3636; X86-SSE2-NEXT:    pandn %xmm1, %xmm0637; X86-SSE2-NEXT:    por %xmm3, %xmm0638; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]639; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3640; X86-SSE2-NEXT:    pxor %xmm2, %xmm3641; X86-SSE2-NEXT:    pxor %xmm1, %xmm2642; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm2643; X86-SSE2-NEXT:    pand %xmm2, %xmm0644; X86-SSE2-NEXT:    pandn %xmm1, %xmm2645; X86-SSE2-NEXT:    por %xmm0, %xmm2646; X86-SSE2-NEXT:    movd %xmm2, %eax647; X86-SSE2-NEXT:    retl648;649; X86-SSE42-LABEL: test_reduce_v8i32:650; X86-SSE42:       ## %bb.0:651; X86-SSE42-NEXT:    pminud %xmm1, %xmm0652; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]653; X86-SSE42-NEXT:    pminud %xmm0, %xmm1654; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]655; X86-SSE42-NEXT:    pminud %xmm1, %xmm0656; X86-SSE42-NEXT:    movd %xmm0, %eax657; X86-SSE42-NEXT:    retl658;659; X86-AVX1-LABEL: test_reduce_v8i32:660; X86-AVX1:       ## %bb.0:661; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1662; X86-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm0663; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]664; X86-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm0665; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]666; X86-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm0667; X86-AVX1-NEXT:    vmovd %xmm0, %eax668; X86-AVX1-NEXT:    vzeroupper669; X86-AVX1-NEXT:    retl670;671; X86-AVX2-LABEL: test_reduce_v8i32:672; X86-AVX2:       ## %bb.0:673; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1674; X86-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm0675; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]676; X86-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm0677; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]678; X86-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm0679; X86-AVX2-NEXT:    vmovd %xmm0, %eax680; X86-AVX2-NEXT:    vzeroupper681; X86-AVX2-NEXT:    retl682;683; X64-SSE2-LABEL: test_reduce_v8i32:684; X64-SSE2:       ## %bb.0:685; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]686; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3687; X64-SSE2-NEXT:    pxor %xmm2, %xmm3688; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4689; X64-SSE2-NEXT:    pxor %xmm2, %xmm4690; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm4691; X64-SSE2-NEXT:    pand %xmm4, %xmm0692; X64-SSE2-NEXT:    pandn %xmm1, %xmm4693; X64-SSE2-NEXT:    por %xmm0, %xmm4694; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]695; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1696; X64-SSE2-NEXT:    pxor %xmm2, %xmm1697; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3698; X64-SSE2-NEXT:    pxor %xmm2, %xmm3699; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm3700; X64-SSE2-NEXT:    pand %xmm3, %xmm4701; X64-SSE2-NEXT:    pandn %xmm0, %xmm3702; X64-SSE2-NEXT:    por %xmm4, %xmm3703; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]704; X64-SSE2-NEXT:    movdqa %xmm3, %xmm1705; X64-SSE2-NEXT:    pxor %xmm2, %xmm1706; X64-SSE2-NEXT:    pxor %xmm0, %xmm2707; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2708; X64-SSE2-NEXT:    pand %xmm2, %xmm3709; X64-SSE2-NEXT:    pandn %xmm0, %xmm2710; X64-SSE2-NEXT:    por %xmm3, %xmm2711; X64-SSE2-NEXT:    movd %xmm2, %eax712; X64-SSE2-NEXT:    retq713;714; X64-SSE42-LABEL: test_reduce_v8i32:715; X64-SSE42:       ## %bb.0:716; X64-SSE42-NEXT:    pminud %xmm1, %xmm0717; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]718; X64-SSE42-NEXT:    pminud %xmm0, %xmm1719; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]720; X64-SSE42-NEXT:    pminud %xmm1, %xmm0721; X64-SSE42-NEXT:    movd %xmm0, %eax722; X64-SSE42-NEXT:    retq723;724; X64-AVX1-LABEL: test_reduce_v8i32:725; X64-AVX1:       ## %bb.0:726; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1727; X64-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm0728; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]729; X64-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm0730; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]731; X64-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm0732; X64-AVX1-NEXT:    vmovd %xmm0, %eax733; X64-AVX1-NEXT:    vzeroupper734; X64-AVX1-NEXT:    retq735;736; X64-AVX2-LABEL: test_reduce_v8i32:737; X64-AVX2:       ## %bb.0:738; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1739; X64-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm0740; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]741; X64-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm0742; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]743; X64-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm0744; X64-AVX2-NEXT:    vmovd %xmm0, %eax745; X64-AVX2-NEXT:    vzeroupper746; X64-AVX2-NEXT:    retq747;748; X64-AVX512-LABEL: test_reduce_v8i32:749; X64-AVX512:       ## %bb.0:750; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1751; X64-AVX512-NEXT:    vpminud %xmm1, %xmm0, %xmm0752; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]753; X64-AVX512-NEXT:    vpminud %xmm1, %xmm0, %xmm0754; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]755; X64-AVX512-NEXT:    vpminud %xmm1, %xmm0, %xmm0756; X64-AVX512-NEXT:    vmovd %xmm0, %eax757; X64-AVX512-NEXT:    vzeroupper758; X64-AVX512-NEXT:    retq759  %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>760  %2  = icmp ult <8 x i32> %a0, %1761  %3  = select <8 x i1> %2, <8 x i32> %a0, <8 x i32> %1762  %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>763  %5  = icmp ult <8 x i32> %3, %4764  %6  = select <8 x i1> %5, <8 x i32> %3, <8 x i32> %4765  %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>766  %8  = icmp ult <8 x i32> %6, %7767  %9  = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7768  %10 = extractelement <8 x i32> %9, i32 0769  ret i32 %10770}771 772define i16 @test_reduce_v16i16(<16 x i16> %a0) {773; X86-SSE2-LABEL: test_reduce_v16i16:774; X86-SSE2:       ## %bb.0:775; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2776; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2777; X86-SSE2-NEXT:    psubw %xmm2, %xmm0778; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]779; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2780; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2781; X86-SSE2-NEXT:    psubw %xmm2, %xmm0782; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]783; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2784; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2785; X86-SSE2-NEXT:    psubw %xmm2, %xmm0786; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1787; X86-SSE2-NEXT:    psrld $16, %xmm1788; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2789; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2790; X86-SSE2-NEXT:    psubw %xmm2, %xmm0791; X86-SSE2-NEXT:    movd %xmm0, %eax792; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax793; X86-SSE2-NEXT:    retl794;795; X86-SSE42-LABEL: test_reduce_v16i16:796; X86-SSE42:       ## %bb.0:797; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0798; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0799; X86-SSE42-NEXT:    movd %xmm0, %eax800; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax801; X86-SSE42-NEXT:    retl802;803; X86-AVX1-LABEL: test_reduce_v16i16:804; X86-AVX1:       ## %bb.0:805; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1806; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0807; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0808; X86-AVX1-NEXT:    vmovd %xmm0, %eax809; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax810; X86-AVX1-NEXT:    vzeroupper811; X86-AVX1-NEXT:    retl812;813; X86-AVX2-LABEL: test_reduce_v16i16:814; X86-AVX2:       ## %bb.0:815; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1816; X86-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0817; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0818; X86-AVX2-NEXT:    vmovd %xmm0, %eax819; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax820; X86-AVX2-NEXT:    vzeroupper821; X86-AVX2-NEXT:    retl822;823; X64-SSE2-LABEL: test_reduce_v16i16:824; X64-SSE2:       ## %bb.0:825; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2826; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2827; X64-SSE2-NEXT:    psubw %xmm2, %xmm0828; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]829; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2830; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2831; X64-SSE2-NEXT:    psubw %xmm2, %xmm0832; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]833; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2834; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2835; X64-SSE2-NEXT:    psubw %xmm2, %xmm0836; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1837; X64-SSE2-NEXT:    psrld $16, %xmm1838; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2839; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2840; X64-SSE2-NEXT:    psubw %xmm2, %xmm0841; X64-SSE2-NEXT:    movd %xmm0, %eax842; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax843; X64-SSE2-NEXT:    retq844;845; X64-SSE42-LABEL: test_reduce_v16i16:846; X64-SSE42:       ## %bb.0:847; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0848; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0849; X64-SSE42-NEXT:    movd %xmm0, %eax850; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax851; X64-SSE42-NEXT:    retq852;853; X64-AVX1-LABEL: test_reduce_v16i16:854; X64-AVX1:       ## %bb.0:855; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1856; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0857; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0858; X64-AVX1-NEXT:    vmovd %xmm0, %eax859; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax860; X64-AVX1-NEXT:    vzeroupper861; X64-AVX1-NEXT:    retq862;863; X64-AVX2-LABEL: test_reduce_v16i16:864; X64-AVX2:       ## %bb.0:865; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1866; X64-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0867; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0868; X64-AVX2-NEXT:    vmovd %xmm0, %eax869; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax870; X64-AVX2-NEXT:    vzeroupper871; X64-AVX2-NEXT:    retq872;873; X64-AVX512-LABEL: test_reduce_v16i16:874; X64-AVX512:       ## %bb.0:875; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1876; X64-AVX512-NEXT:    vpminuw %xmm1, %xmm0, %xmm0877; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0878; X64-AVX512-NEXT:    vmovd %xmm0, %eax879; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax880; X64-AVX512-NEXT:    vzeroupper881; X64-AVX512-NEXT:    retq882  %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>883  %2  = icmp ult <16 x i16> %a0, %1884  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1885  %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>886  %5  = icmp ult <16 x i16> %3, %4887  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %4888  %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>889  %8  = icmp ult <16 x i16> %6, %7890  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7891  %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>892  %11 = icmp ult <16 x i16> %9, %10893  %12 = select <16 x i1> %11, <16 x i16> %9, <16 x i16> %10894  %13 = extractelement <16 x i16> %12, i32 0895  ret i16 %13896}897 898define i8 @test_reduce_v32i8(<32 x i8> %a0) {899; X86-SSE2-LABEL: test_reduce_v32i8:900; X86-SSE2:       ## %bb.0:901; X86-SSE2-NEXT:    pminub %xmm1, %xmm0902; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]903; X86-SSE2-NEXT:    pminub %xmm0, %xmm1904; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]905; X86-SSE2-NEXT:    pminub %xmm1, %xmm0906; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1907; X86-SSE2-NEXT:    psrld $16, %xmm1908; X86-SSE2-NEXT:    pminub %xmm0, %xmm1909; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0910; X86-SSE2-NEXT:    psrlw $8, %xmm0911; X86-SSE2-NEXT:    pminub %xmm1, %xmm0912; X86-SSE2-NEXT:    movd %xmm0, %eax913; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax914; X86-SSE2-NEXT:    retl915;916; X86-SSE42-LABEL: test_reduce_v32i8:917; X86-SSE42:       ## %bb.0:918; X86-SSE42-NEXT:    pminub %xmm1, %xmm0919; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1920; X86-SSE42-NEXT:    psrlw $8, %xmm1921; X86-SSE42-NEXT:    pminub %xmm0, %xmm1922; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0923; X86-SSE42-NEXT:    movd %xmm0, %eax924; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax925; X86-SSE42-NEXT:    retl926;927; X86-AVX1-LABEL: test_reduce_v32i8:928; X86-AVX1:       ## %bb.0:929; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1930; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0931; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1932; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0933; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0934; X86-AVX1-NEXT:    vmovd %xmm0, %eax935; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax936; X86-AVX1-NEXT:    vzeroupper937; X86-AVX1-NEXT:    retl938;939; X86-AVX2-LABEL: test_reduce_v32i8:940; X86-AVX2:       ## %bb.0:941; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1942; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0943; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1944; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0945; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0946; X86-AVX2-NEXT:    vmovd %xmm0, %eax947; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax948; X86-AVX2-NEXT:    vzeroupper949; X86-AVX2-NEXT:    retl950;951; X64-SSE2-LABEL: test_reduce_v32i8:952; X64-SSE2:       ## %bb.0:953; X64-SSE2-NEXT:    pminub %xmm1, %xmm0954; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]955; X64-SSE2-NEXT:    pminub %xmm0, %xmm1956; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]957; X64-SSE2-NEXT:    pminub %xmm1, %xmm0958; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1959; X64-SSE2-NEXT:    psrld $16, %xmm1960; X64-SSE2-NEXT:    pminub %xmm0, %xmm1961; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0962; X64-SSE2-NEXT:    psrlw $8, %xmm0963; X64-SSE2-NEXT:    pminub %xmm1, %xmm0964; X64-SSE2-NEXT:    movd %xmm0, %eax965; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax966; X64-SSE2-NEXT:    retq967;968; X64-SSE42-LABEL: test_reduce_v32i8:969; X64-SSE42:       ## %bb.0:970; X64-SSE42-NEXT:    pminub %xmm1, %xmm0971; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1972; X64-SSE42-NEXT:    psrlw $8, %xmm1973; X64-SSE42-NEXT:    pminub %xmm0, %xmm1974; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0975; X64-SSE42-NEXT:    movd %xmm0, %eax976; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax977; X64-SSE42-NEXT:    retq978;979; X64-AVX1-LABEL: test_reduce_v32i8:980; X64-AVX1:       ## %bb.0:981; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1982; X64-AVX1-NEXT:    vpminub %xmm1, %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:    ## kill: def $al killed $al killed $eax988; X64-AVX1-NEXT:    vzeroupper989; X64-AVX1-NEXT:    retq990;991; X64-AVX2-LABEL: test_reduce_v32i8:992; X64-AVX2:       ## %bb.0:993; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1994; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0995; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1996; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0997; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0998; X64-AVX2-NEXT:    vmovd %xmm0, %eax999; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1000; X64-AVX2-NEXT:    vzeroupper1001; X64-AVX2-NEXT:    retq1002;1003; X64-AVX512-LABEL: test_reduce_v32i8:1004; X64-AVX512:       ## %bb.0:1005; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11006; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01007; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm11008; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01009; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01010; X64-AVX512-NEXT:    vmovd %xmm0, %eax1011; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax1012; X64-AVX512-NEXT:    vzeroupper1013; X64-AVX512-NEXT:    retq1014  %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>1015  %2  = icmp ult <32 x i8> %a0, %11016  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %11017  %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>1018  %5  = icmp ult <32 x i8> %3, %41019  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %41020  %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>1021  %8  = icmp ult <32 x i8> %6, %71022  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %71023  %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>1024  %11 = icmp ult <32 x i8> %9, %101025  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %101026  %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>1027  %14 = icmp ult <32 x i8> %12, %131028  %15 = select <32 x i1> %14, <32 x i8> %12, <32 x i8> %131029  %16 = extractelement <32 x i8> %15, i32 01030  ret i8 %161031}1032 1033;1034; 512-bit Vectors1035;1036 1037define i64 @test_reduce_v8i64(<8 x i64> %a0) {1038; X86-SSE2-LABEL: test_reduce_v8i64:1039; X86-SSE2:       ## %bb.0:1040; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1041; X86-SSE2-NEXT:    movdqa %xmm1, %xmm51042; X86-SSE2-NEXT:    pxor %xmm4, %xmm51043; X86-SSE2-NEXT:    movdqa %xmm3, %xmm61044; X86-SSE2-NEXT:    pxor %xmm4, %xmm61045; X86-SSE2-NEXT:    movdqa %xmm6, %xmm71046; X86-SSE2-NEXT:    pcmpgtd %xmm5, %xmm71047; X86-SSE2-NEXT:    pcmpeqd %xmm5, %xmm61048; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]1049; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1050; X86-SSE2-NEXT:    pand %xmm5, %xmm61051; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1052; X86-SSE2-NEXT:    por %xmm6, %xmm51053; X86-SSE2-NEXT:    pand %xmm5, %xmm11054; X86-SSE2-NEXT:    pandn %xmm3, %xmm51055; X86-SSE2-NEXT:    por %xmm1, %xmm51056; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11057; X86-SSE2-NEXT:    pxor %xmm4, %xmm11058; X86-SSE2-NEXT:    movdqa %xmm2, %xmm31059; X86-SSE2-NEXT:    pxor %xmm4, %xmm31060; X86-SSE2-NEXT:    movdqa %xmm3, %xmm61061; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm61062; X86-SSE2-NEXT:    pcmpeqd %xmm1, %xmm31063; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,0,2,2]1064; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]1065; X86-SSE2-NEXT:    pand %xmm1, %xmm31066; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]1067; X86-SSE2-NEXT:    por %xmm3, %xmm11068; X86-SSE2-NEXT:    pand %xmm1, %xmm01069; X86-SSE2-NEXT:    pandn %xmm2, %xmm11070; X86-SSE2-NEXT:    por %xmm0, %xmm11071; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01072; X86-SSE2-NEXT:    pxor %xmm4, %xmm01073; X86-SSE2-NEXT:    movdqa %xmm5, %xmm21074; X86-SSE2-NEXT:    pxor %xmm4, %xmm21075; X86-SSE2-NEXT:    movdqa %xmm2, %xmm31076; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm31077; X86-SSE2-NEXT:    pcmpeqd %xmm0, %xmm21078; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]1079; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]1080; X86-SSE2-NEXT:    pand %xmm0, %xmm21081; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]1082; X86-SSE2-NEXT:    por %xmm2, %xmm01083; X86-SSE2-NEXT:    pand %xmm0, %xmm11084; X86-SSE2-NEXT:    pandn %xmm5, %xmm01085; X86-SSE2-NEXT:    por %xmm1, %xmm01086; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1087; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21088; X86-SSE2-NEXT:    pxor %xmm4, %xmm21089; X86-SSE2-NEXT:    pxor %xmm1, %xmm41090; X86-SSE2-NEXT:    movdqa %xmm4, %xmm31091; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm31092; X86-SSE2-NEXT:    pcmpeqd %xmm2, %xmm41093; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]1094; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]1095; X86-SSE2-NEXT:    pand %xmm2, %xmm41096; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]1097; X86-SSE2-NEXT:    por %xmm4, %xmm21098; X86-SSE2-NEXT:    pand %xmm2, %xmm01099; X86-SSE2-NEXT:    pandn %xmm1, %xmm21100; X86-SSE2-NEXT:    por %xmm0, %xmm21101; X86-SSE2-NEXT:    movd %xmm2, %eax1102; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1103; X86-SSE2-NEXT:    movd %xmm0, %edx1104; X86-SSE2-NEXT:    retl1105;1106; X86-SSE42-LABEL: test_reduce_v8i64:1107; X86-SSE42:       ## %bb.0:1108; X86-SSE42-NEXT:    movdqa %xmm0, %xmm41109; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm5 = [0,2147483648,0,2147483648]1110; X86-SSE42-NEXT:    movdqa %xmm1, %xmm61111; X86-SSE42-NEXT:    pxor %xmm5, %xmm61112; X86-SSE42-NEXT:    movdqa %xmm3, %xmm01113; X86-SSE42-NEXT:    pxor %xmm5, %xmm01114; X86-SSE42-NEXT:    pcmpgtq %xmm6, %xmm01115; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm31116; X86-SSE42-NEXT:    movdqa %xmm4, %xmm11117; X86-SSE42-NEXT:    pxor %xmm5, %xmm11118; X86-SSE42-NEXT:    movdqa %xmm2, %xmm01119; X86-SSE42-NEXT:    pxor %xmm5, %xmm01120; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01121; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm21122; X86-SSE42-NEXT:    movapd %xmm2, %xmm11123; X86-SSE42-NEXT:    xorpd %xmm5, %xmm11124; X86-SSE42-NEXT:    movapd %xmm3, %xmm01125; X86-SSE42-NEXT:    xorpd %xmm5, %xmm01126; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01127; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm31128; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1129; X86-SSE42-NEXT:    movdqa %xmm3, %xmm01130; X86-SSE42-NEXT:    pxor %xmm5, %xmm01131; X86-SSE42-NEXT:    pxor %xmm1, %xmm51132; X86-SSE42-NEXT:    pcmpgtq %xmm0, %xmm51133; X86-SSE42-NEXT:    movdqa %xmm5, %xmm01134; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm11135; X86-SSE42-NEXT:    movd %xmm1, %eax1136; X86-SSE42-NEXT:    pextrd $1, %xmm1, %edx1137; X86-SSE42-NEXT:    retl1138;1139; X86-AVX1-LABEL: test_reduce_v8i64:1140; X86-AVX1:       ## %bb.0:1141; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31142; X86-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [0,2147483648,0,2147483648]1143; X86-AVX1-NEXT:    ## xmm2 = mem[0,0]1144; X86-AVX1-NEXT:    vxorps %xmm2, %xmm3, %xmm41145; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm51146; X86-AVX1-NEXT:    vxorps %xmm2, %xmm5, %xmm61147; X86-AVX1-NEXT:    vpcmpgtq %xmm4, %xmm6, %xmm41148; X86-AVX1-NEXT:    vblendvpd %xmm4, %xmm3, %xmm5, %xmm31149; X86-AVX1-NEXT:    vxorps %xmm2, %xmm0, %xmm41150; X86-AVX1-NEXT:    vxorps %xmm2, %xmm1, %xmm51151; X86-AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm41152; X86-AVX1-NEXT:    vblendvpd %xmm4, %xmm0, %xmm1, %xmm01153; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm11154; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm3, %xmm41155; X86-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm11156; X86-AVX1-NEXT:    vblendvpd %xmm1, %xmm0, %xmm3, %xmm01157; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1158; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm31159; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm1, %xmm21160; X86-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm21161; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01162; X86-AVX1-NEXT:    vmovd %xmm0, %eax1163; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx1164; X86-AVX1-NEXT:    vzeroupper1165; X86-AVX1-NEXT:    retl1166;1167; X86-AVX2-LABEL: test_reduce_v8i64:1168; X86-AVX2:       ## %bb.0:1169; X86-AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [0,2147483648,0,2147483648,0,2147483648,0,2147483648]1170; X86-AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm31171; X86-AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm41172; X86-AVX2-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm31173; X86-AVX2-NEXT:    vblendvpd %ymm3, %ymm0, %ymm1, %ymm01174; X86-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm11175; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm31176; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm41177; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm31178; X86-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm01179; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1180; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm31181; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm21182; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm21183; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01184; X86-AVX2-NEXT:    vmovd %xmm0, %eax1185; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx1186; X86-AVX2-NEXT:    vzeroupper1187; X86-AVX2-NEXT:    retl1188;1189; X64-SSE2-LABEL: test_reduce_v8i64:1190; X64-SSE2:       ## %bb.0:1191; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]1192; X64-SSE2-NEXT:    movdqa %xmm1, %xmm51193; X64-SSE2-NEXT:    pxor %xmm4, %xmm51194; X64-SSE2-NEXT:    movdqa %xmm3, %xmm61195; X64-SSE2-NEXT:    pxor %xmm4, %xmm61196; X64-SSE2-NEXT:    movdqa %xmm6, %xmm71197; X64-SSE2-NEXT:    pcmpgtd %xmm5, %xmm71198; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]1199; X64-SSE2-NEXT:    pcmpeqd %xmm5, %xmm61200; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1201; X64-SSE2-NEXT:    pand %xmm8, %xmm61202; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1203; X64-SSE2-NEXT:    por %xmm6, %xmm51204; X64-SSE2-NEXT:    pand %xmm5, %xmm11205; X64-SSE2-NEXT:    pandn %xmm3, %xmm51206; X64-SSE2-NEXT:    por %xmm1, %xmm51207; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11208; X64-SSE2-NEXT:    pxor %xmm4, %xmm11209; X64-SSE2-NEXT:    movdqa %xmm2, %xmm31210; X64-SSE2-NEXT:    pxor %xmm4, %xmm31211; X64-SSE2-NEXT:    movdqa %xmm3, %xmm61212; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm61213; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]1214; X64-SSE2-NEXT:    pcmpeqd %xmm1, %xmm31215; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]1216; X64-SSE2-NEXT:    pand %xmm7, %xmm11217; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]1218; X64-SSE2-NEXT:    por %xmm1, %xmm31219; X64-SSE2-NEXT:    pand %xmm3, %xmm01220; X64-SSE2-NEXT:    pandn %xmm2, %xmm31221; X64-SSE2-NEXT:    por %xmm0, %xmm31222; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01223; X64-SSE2-NEXT:    pxor %xmm4, %xmm01224; X64-SSE2-NEXT:    movdqa %xmm5, %xmm11225; X64-SSE2-NEXT:    pxor %xmm4, %xmm11226; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21227; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm21228; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]1229; X64-SSE2-NEXT:    pcmpeqd %xmm0, %xmm11230; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]1231; X64-SSE2-NEXT:    pand %xmm6, %xmm01232; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]1233; X64-SSE2-NEXT:    por %xmm0, %xmm11234; X64-SSE2-NEXT:    pand %xmm1, %xmm31235; X64-SSE2-NEXT:    pandn %xmm5, %xmm11236; X64-SSE2-NEXT:    por %xmm3, %xmm11237; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1238; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21239; X64-SSE2-NEXT:    pxor %xmm4, %xmm21240; X64-SSE2-NEXT:    pxor %xmm0, %xmm41241; X64-SSE2-NEXT:    movdqa %xmm4, %xmm31242; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm31243; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]1244; X64-SSE2-NEXT:    pcmpeqd %xmm2, %xmm41245; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]1246; X64-SSE2-NEXT:    pand %xmm5, %xmm21247; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]1248; X64-SSE2-NEXT:    por %xmm2, %xmm31249; X64-SSE2-NEXT:    pand %xmm3, %xmm11250; X64-SSE2-NEXT:    pandn %xmm0, %xmm31251; X64-SSE2-NEXT:    por %xmm1, %xmm31252; X64-SSE2-NEXT:    movq %xmm3, %rax1253; X64-SSE2-NEXT:    retq1254;1255; X64-SSE42-LABEL: test_reduce_v8i64:1256; X64-SSE42:       ## %bb.0:1257; X64-SSE42-NEXT:    movdqa %xmm0, %xmm51258; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm4 = [9223372036854775808,9223372036854775808]1259; X64-SSE42-NEXT:    movdqa %xmm1, %xmm61260; X64-SSE42-NEXT:    pxor %xmm4, %xmm61261; X64-SSE42-NEXT:    movdqa %xmm3, %xmm01262; X64-SSE42-NEXT:    pxor %xmm4, %xmm01263; X64-SSE42-NEXT:    pcmpgtq %xmm6, %xmm01264; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm31265; X64-SSE42-NEXT:    movdqa %xmm5, %xmm11266; X64-SSE42-NEXT:    pxor %xmm4, %xmm11267; X64-SSE42-NEXT:    movdqa %xmm2, %xmm01268; X64-SSE42-NEXT:    pxor %xmm4, %xmm01269; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01270; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm5, %xmm21271; X64-SSE42-NEXT:    movapd %xmm2, %xmm11272; X64-SSE42-NEXT:    xorpd %xmm4, %xmm11273; X64-SSE42-NEXT:    movapd %xmm3, %xmm01274; X64-SSE42-NEXT:    xorpd %xmm4, %xmm01275; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01276; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm31277; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1278; X64-SSE42-NEXT:    movdqa %xmm3, %xmm01279; X64-SSE42-NEXT:    pxor %xmm4, %xmm01280; X64-SSE42-NEXT:    pxor %xmm1, %xmm41281; X64-SSE42-NEXT:    pcmpgtq %xmm0, %xmm41282; X64-SSE42-NEXT:    movdqa %xmm4, %xmm01283; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm11284; X64-SSE42-NEXT:    movq %xmm1, %rax1285; X64-SSE42-NEXT:    retq1286;1287; X64-AVX1-LABEL: test_reduce_v8i64:1288; X64-AVX1:       ## %bb.0:1289; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm21290; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]1291; X64-AVX1-NEXT:    ## xmm3 = mem[0,0]1292; X64-AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm41293; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm51294; X64-AVX1-NEXT:    vpxor %xmm3, %xmm5, %xmm61295; X64-AVX1-NEXT:    vpcmpgtq %xmm4, %xmm6, %xmm41296; X64-AVX1-NEXT:    vblendvpd %xmm4, %xmm2, %xmm5, %xmm21297; X64-AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm41298; X64-AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm51299; X64-AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm41300; X64-AVX1-NEXT:    vblendvpd %xmm4, %xmm0, %xmm1, %xmm01301; X64-AVX1-NEXT:    vxorpd %xmm3, %xmm0, %xmm11302; X64-AVX1-NEXT:    vxorpd %xmm3, %xmm2, %xmm41303; X64-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm11304; X64-AVX1-NEXT:    vblendvpd %xmm1, %xmm0, %xmm2, %xmm01305; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1306; X64-AVX1-NEXT:    vxorpd %xmm3, %xmm0, %xmm21307; X64-AVX1-NEXT:    vxorpd %xmm3, %xmm1, %xmm31308; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm21309; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01310; X64-AVX1-NEXT:    vmovq %xmm0, %rax1311; X64-AVX1-NEXT:    vzeroupper1312; X64-AVX1-NEXT:    retq1313;1314; X64-AVX2-LABEL: test_reduce_v8i64:1315; X64-AVX2:       ## %bb.0:1316; X64-AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]1317; X64-AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm31318; X64-AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm41319; X64-AVX2-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm31320; X64-AVX2-NEXT:    vblendvpd %ymm3, %ymm0, %ymm1, %ymm01321; X64-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm11322; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm31323; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm41324; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm31325; X64-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm01326; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1327; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm31328; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm21329; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm21330; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01331; X64-AVX2-NEXT:    vmovq %xmm0, %rax1332; X64-AVX2-NEXT:    vzeroupper1333; X64-AVX2-NEXT:    retq1334;1335; X64-AVX512-LABEL: test_reduce_v8i64:1336; X64-AVX512:       ## %bb.0:1337; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11338; X64-AVX512-NEXT:    vpminuq %zmm1, %zmm0, %zmm01339; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11340; X64-AVX512-NEXT:    vpminuq %xmm1, %xmm0, %xmm01341; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1342; X64-AVX512-NEXT:    vpminuq %xmm1, %xmm0, %xmm01343; X64-AVX512-NEXT:    vmovq %xmm0, %rax1344; X64-AVX512-NEXT:    vzeroupper1345; X64-AVX512-NEXT:    retq1346  %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>1347  %2  = icmp ult <8 x i64> %a0, %11348  %3  = select <8 x i1> %2, <8 x i64> %a0, <8 x i64> %11349  %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>1350  %5  = icmp ult <8 x i64> %3, %41351  %6  = select <8 x i1> %5, <8 x i64> %3, <8 x i64> %41352  %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>1353  %8  = icmp ult <8 x i64> %6, %71354  %9  = select <8 x i1> %8, <8 x i64> %6, <8 x i64> %71355  %10 = extractelement <8 x i64> %9, i32 01356  ret i64 %101357}1358 1359define i32 @test_reduce_v16i32(<16 x i32> %a0) {1360; X86-SSE2-LABEL: test_reduce_v16i32:1361; X86-SSE2:       ## %bb.0:1362; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1363; X86-SSE2-NEXT:    movdqa %xmm1, %xmm61364; X86-SSE2-NEXT:    pxor %xmm4, %xmm61365; X86-SSE2-NEXT:    movdqa %xmm3, %xmm51366; X86-SSE2-NEXT:    pxor %xmm4, %xmm51367; X86-SSE2-NEXT:    pcmpgtd %xmm6, %xmm51368; X86-SSE2-NEXT:    pand %xmm5, %xmm11369; X86-SSE2-NEXT:    pandn %xmm3, %xmm51370; X86-SSE2-NEXT:    por %xmm1, %xmm51371; X86-SSE2-NEXT:    movdqa %xmm0, %xmm31372; X86-SSE2-NEXT:    pxor %xmm4, %xmm31373; X86-SSE2-NEXT:    movdqa %xmm2, %xmm11374; X86-SSE2-NEXT:    pxor %xmm4, %xmm11375; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm11376; X86-SSE2-NEXT:    pand %xmm1, %xmm01377; X86-SSE2-NEXT:    pandn %xmm2, %xmm11378; X86-SSE2-NEXT:    por %xmm0, %xmm11379; X86-SSE2-NEXT:    movdqa %xmm1, %xmm21380; X86-SSE2-NEXT:    pxor %xmm4, %xmm21381; X86-SSE2-NEXT:    movdqa %xmm5, %xmm01382; X86-SSE2-NEXT:    pxor %xmm4, %xmm01383; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm01384; X86-SSE2-NEXT:    pand %xmm0, %xmm11385; X86-SSE2-NEXT:    pandn %xmm5, %xmm01386; X86-SSE2-NEXT:    por %xmm1, %xmm01387; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]1388; X86-SSE2-NEXT:    movdqa %xmm0, %xmm31389; X86-SSE2-NEXT:    pxor %xmm4, %xmm31390; X86-SSE2-NEXT:    movdqa %xmm2, %xmm11391; X86-SSE2-NEXT:    pxor %xmm4, %xmm11392; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm11393; X86-SSE2-NEXT:    pand %xmm1, %xmm01394; X86-SSE2-NEXT:    pandn %xmm2, %xmm11395; X86-SSE2-NEXT:    por %xmm0, %xmm11396; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1397; X86-SSE2-NEXT:    movdqa %xmm1, %xmm21398; X86-SSE2-NEXT:    pxor %xmm4, %xmm21399; X86-SSE2-NEXT:    pxor %xmm0, %xmm41400; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm41401; X86-SSE2-NEXT:    pand %xmm4, %xmm11402; X86-SSE2-NEXT:    pandn %xmm0, %xmm41403; X86-SSE2-NEXT:    por %xmm1, %xmm41404; X86-SSE2-NEXT:    movd %xmm4, %eax1405; X86-SSE2-NEXT:    retl1406;1407; X86-SSE42-LABEL: test_reduce_v16i32:1408; X86-SSE42:       ## %bb.0:1409; X86-SSE42-NEXT:    pminud %xmm3, %xmm11410; X86-SSE42-NEXT:    pminud %xmm2, %xmm01411; X86-SSE42-NEXT:    pminud %xmm1, %xmm01412; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1413; X86-SSE42-NEXT:    pminud %xmm0, %xmm11414; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1415; X86-SSE42-NEXT:    pminud %xmm1, %xmm01416; X86-SSE42-NEXT:    movd %xmm0, %eax1417; X86-SSE42-NEXT:    retl1418;1419; X86-AVX1-LABEL: test_reduce_v16i32:1420; X86-AVX1:       ## %bb.0:1421; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21422; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31423; X86-AVX1-NEXT:    vpminud %xmm2, %xmm3, %xmm21424; X86-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm01425; X86-AVX1-NEXT:    vpminud %xmm2, %xmm0, %xmm01426; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1427; X86-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm01428; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1429; X86-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm01430; X86-AVX1-NEXT:    vmovd %xmm0, %eax1431; X86-AVX1-NEXT:    vzeroupper1432; X86-AVX1-NEXT:    retl1433;1434; X86-AVX2-LABEL: test_reduce_v16i32:1435; X86-AVX2:       ## %bb.0:1436; X86-AVX2-NEXT:    vpminud %ymm1, %ymm0, %ymm01437; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11438; X86-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm01439; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1440; X86-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm01441; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1442; X86-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm01443; X86-AVX2-NEXT:    vmovd %xmm0, %eax1444; X86-AVX2-NEXT:    vzeroupper1445; X86-AVX2-NEXT:    retl1446;1447; X64-SSE2-LABEL: test_reduce_v16i32:1448; X64-SSE2:       ## %bb.0:1449; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1450; X64-SSE2-NEXT:    movdqa %xmm1, %xmm51451; X64-SSE2-NEXT:    pxor %xmm4, %xmm51452; X64-SSE2-NEXT:    movdqa %xmm3, %xmm61453; X64-SSE2-NEXT:    pxor %xmm4, %xmm61454; X64-SSE2-NEXT:    pcmpgtd %xmm5, %xmm61455; X64-SSE2-NEXT:    pand %xmm6, %xmm11456; X64-SSE2-NEXT:    pandn %xmm3, %xmm61457; X64-SSE2-NEXT:    por %xmm1, %xmm61458; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11459; X64-SSE2-NEXT:    pxor %xmm4, %xmm11460; X64-SSE2-NEXT:    movdqa %xmm2, %xmm31461; X64-SSE2-NEXT:    pxor %xmm4, %xmm31462; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm31463; X64-SSE2-NEXT:    pand %xmm3, %xmm01464; X64-SSE2-NEXT:    pandn %xmm2, %xmm31465; X64-SSE2-NEXT:    por %xmm0, %xmm31466; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01467; X64-SSE2-NEXT:    pxor %xmm4, %xmm01468; X64-SSE2-NEXT:    movdqa %xmm6, %xmm11469; X64-SSE2-NEXT:    pxor %xmm4, %xmm11470; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm11471; X64-SSE2-NEXT:    pand %xmm1, %xmm31472; X64-SSE2-NEXT:    pandn %xmm6, %xmm11473; X64-SSE2-NEXT:    por %xmm3, %xmm11474; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1475; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21476; X64-SSE2-NEXT:    pxor %xmm4, %xmm21477; X64-SSE2-NEXT:    movdqa %xmm0, %xmm31478; X64-SSE2-NEXT:    pxor %xmm4, %xmm31479; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm31480; X64-SSE2-NEXT:    pand %xmm3, %xmm11481; X64-SSE2-NEXT:    pandn %xmm0, %xmm31482; X64-SSE2-NEXT:    por %xmm1, %xmm31483; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]1484; X64-SSE2-NEXT:    movdqa %xmm3, %xmm11485; X64-SSE2-NEXT:    pxor %xmm4, %xmm11486; X64-SSE2-NEXT:    pxor %xmm0, %xmm41487; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm41488; X64-SSE2-NEXT:    pand %xmm4, %xmm31489; X64-SSE2-NEXT:    pandn %xmm0, %xmm41490; X64-SSE2-NEXT:    por %xmm3, %xmm41491; X64-SSE2-NEXT:    movd %xmm4, %eax1492; X64-SSE2-NEXT:    retq1493;1494; X64-SSE42-LABEL: test_reduce_v16i32:1495; X64-SSE42:       ## %bb.0:1496; X64-SSE42-NEXT:    pminud %xmm3, %xmm11497; X64-SSE42-NEXT:    pminud %xmm2, %xmm01498; X64-SSE42-NEXT:    pminud %xmm1, %xmm01499; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1500; X64-SSE42-NEXT:    pminud %xmm0, %xmm11501; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1502; X64-SSE42-NEXT:    pminud %xmm1, %xmm01503; X64-SSE42-NEXT:    movd %xmm0, %eax1504; X64-SSE42-NEXT:    retq1505;1506; X64-AVX1-LABEL: test_reduce_v16i32:1507; X64-AVX1:       ## %bb.0:1508; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21509; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31510; X64-AVX1-NEXT:    vpminud %xmm2, %xmm3, %xmm21511; X64-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm01512; X64-AVX1-NEXT:    vpminud %xmm2, %xmm0, %xmm01513; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1514; X64-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm01515; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1516; X64-AVX1-NEXT:    vpminud %xmm1, %xmm0, %xmm01517; X64-AVX1-NEXT:    vmovd %xmm0, %eax1518; X64-AVX1-NEXT:    vzeroupper1519; X64-AVX1-NEXT:    retq1520;1521; X64-AVX2-LABEL: test_reduce_v16i32:1522; X64-AVX2:       ## %bb.0:1523; X64-AVX2-NEXT:    vpminud %ymm1, %ymm0, %ymm01524; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11525; X64-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm01526; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1527; X64-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm01528; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1529; X64-AVX2-NEXT:    vpminud %xmm1, %xmm0, %xmm01530; X64-AVX2-NEXT:    vmovd %xmm0, %eax1531; X64-AVX2-NEXT:    vzeroupper1532; X64-AVX2-NEXT:    retq1533;1534; X64-AVX512-LABEL: test_reduce_v16i32:1535; X64-AVX512:       ## %bb.0:1536; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11537; X64-AVX512-NEXT:    vpminud %zmm1, %zmm0, %zmm01538; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11539; X64-AVX512-NEXT:    vpminud %xmm1, %xmm0, %xmm01540; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1541; X64-AVX512-NEXT:    vpminud %xmm1, %xmm0, %xmm01542; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1543; X64-AVX512-NEXT:    vpminud %xmm1, %xmm0, %xmm01544; X64-AVX512-NEXT:    vmovd %xmm0, %eax1545; X64-AVX512-NEXT:    vzeroupper1546; X64-AVX512-NEXT:    retq1547  %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>1548  %2  = icmp ult <16 x i32> %a0, %11549  %3  = select <16 x i1> %2, <16 x i32> %a0, <16 x i32> %11550  %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>1551  %5  = icmp ult <16 x i32> %3, %41552  %6  = select <16 x i1> %5, <16 x i32> %3, <16 x i32> %41553  %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>1554  %8  = icmp ult <16 x i32> %6, %71555  %9  = select <16 x i1> %8, <16 x i32> %6, <16 x i32> %71556  %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>1557  %11 = icmp ult <16 x i32> %9, %101558  %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %101559  %13 = extractelement <16 x i32> %12, i32 01560  ret i32 %131561}1562 1563define i16 @test_reduce_v32i16(<32 x i16> %a0) {1564; X86-SSE2-LABEL: test_reduce_v32i16:1565; X86-SSE2:       ## %bb.0:1566; X86-SSE2-NEXT:    movdqa %xmm1, %xmm41567; X86-SSE2-NEXT:    psubusw %xmm3, %xmm41568; X86-SSE2-NEXT:    psubw %xmm4, %xmm11569; X86-SSE2-NEXT:    movdqa %xmm0, %xmm31570; X86-SSE2-NEXT:    psubusw %xmm2, %xmm31571; X86-SSE2-NEXT:    psubw %xmm3, %xmm01572; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21573; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21574; X86-SSE2-NEXT:    psubw %xmm2, %xmm01575; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1576; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21577; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21578; X86-SSE2-NEXT:    psubw %xmm2, %xmm01579; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1580; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21581; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21582; X86-SSE2-NEXT:    psubw %xmm2, %xmm01583; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11584; X86-SSE2-NEXT:    psrld $16, %xmm11585; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21586; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21587; X86-SSE2-NEXT:    psubw %xmm2, %xmm01588; X86-SSE2-NEXT:    movd %xmm0, %eax1589; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1590; X86-SSE2-NEXT:    retl1591;1592; X86-SSE42-LABEL: test_reduce_v32i16:1593; X86-SSE42:       ## %bb.0:1594; X86-SSE42-NEXT:    pminuw %xmm3, %xmm11595; X86-SSE42-NEXT:    pminuw %xmm2, %xmm01596; X86-SSE42-NEXT:    pminuw %xmm1, %xmm01597; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01598; X86-SSE42-NEXT:    movd %xmm0, %eax1599; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1600; X86-SSE42-NEXT:    retl1601;1602; X86-AVX1-LABEL: test_reduce_v32i16:1603; X86-AVX1:       ## %bb.0:1604; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21605; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31606; X86-AVX1-NEXT:    vpminuw %xmm2, %xmm3, %xmm21607; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm01608; X86-AVX1-NEXT:    vpminuw %xmm2, %xmm0, %xmm01609; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01610; X86-AVX1-NEXT:    vmovd %xmm0, %eax1611; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax1612; X86-AVX1-NEXT:    vzeroupper1613; X86-AVX1-NEXT:    retl1614;1615; X86-AVX2-LABEL: test_reduce_v32i16:1616; X86-AVX2:       ## %bb.0:1617; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm01618; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11619; X86-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm01620; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01621; X86-AVX2-NEXT:    vmovd %xmm0, %eax1622; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax1623; X86-AVX2-NEXT:    vzeroupper1624; X86-AVX2-NEXT:    retl1625;1626; X64-SSE2-LABEL: test_reduce_v32i16:1627; X64-SSE2:       ## %bb.0:1628; X64-SSE2-NEXT:    movdqa %xmm1, %xmm41629; X64-SSE2-NEXT:    psubusw %xmm3, %xmm41630; X64-SSE2-NEXT:    psubw %xmm4, %xmm11631; X64-SSE2-NEXT:    movdqa %xmm0, %xmm31632; X64-SSE2-NEXT:    psubusw %xmm2, %xmm31633; X64-SSE2-NEXT:    psubw %xmm3, %xmm01634; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21635; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21636; X64-SSE2-NEXT:    psubw %xmm2, %xmm01637; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1638; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21639; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21640; X64-SSE2-NEXT:    psubw %xmm2, %xmm01641; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1642; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21643; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21644; X64-SSE2-NEXT:    psubw %xmm2, %xmm01645; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11646; X64-SSE2-NEXT:    psrld $16, %xmm11647; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21648; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21649; X64-SSE2-NEXT:    psubw %xmm2, %xmm01650; X64-SSE2-NEXT:    movd %xmm0, %eax1651; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1652; X64-SSE2-NEXT:    retq1653;1654; X64-SSE42-LABEL: test_reduce_v32i16:1655; X64-SSE42:       ## %bb.0:1656; X64-SSE42-NEXT:    pminuw %xmm3, %xmm11657; X64-SSE42-NEXT:    pminuw %xmm2, %xmm01658; X64-SSE42-NEXT:    pminuw %xmm1, %xmm01659; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01660; X64-SSE42-NEXT:    movd %xmm0, %eax1661; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1662; X64-SSE42-NEXT:    retq1663;1664; X64-AVX1-LABEL: test_reduce_v32i16:1665; X64-AVX1:       ## %bb.0:1666; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21667; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31668; X64-AVX1-NEXT:    vpminuw %xmm2, %xmm3, %xmm21669; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm01670; X64-AVX1-NEXT:    vpminuw %xmm2, %xmm0, %xmm01671; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01672; X64-AVX1-NEXT:    vmovd %xmm0, %eax1673; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax1674; X64-AVX1-NEXT:    vzeroupper1675; X64-AVX1-NEXT:    retq1676;1677; X64-AVX2-LABEL: test_reduce_v32i16:1678; X64-AVX2:       ## %bb.0:1679; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm01680; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11681; X64-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm01682; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01683; X64-AVX2-NEXT:    vmovd %xmm0, %eax1684; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax1685; X64-AVX2-NEXT:    vzeroupper1686; X64-AVX2-NEXT:    retq1687;1688; X64-AVX512-LABEL: test_reduce_v32i16:1689; X64-AVX512:       ## %bb.0:1690; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11691; X64-AVX512-NEXT:    vpminuw %ymm1, %ymm0, %ymm01692; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11693; X64-AVX512-NEXT:    vpminuw %xmm1, %xmm0, %xmm01694; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01695; X64-AVX512-NEXT:    vmovd %xmm0, %eax1696; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax1697; X64-AVX512-NEXT:    vzeroupper1698; X64-AVX512-NEXT:    retq1699  %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>1700  %2  = icmp ult <32 x i16> %a0, %11701  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %11702  %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>1703  %5  = icmp ult <32 x i16> %3, %41704  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %41705  %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>1706  %8  = icmp ult <32 x i16> %6, %71707  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %71708  %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>1709  %11 = icmp ult <32 x i16> %9, %101710  %12 = select <32 x i1> %11, <32 x i16> %9, <32 x i16> %101711  %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>1712  %14 = icmp ult <32 x i16> %12, %131713  %15 = select <32 x i1> %14, <32 x i16> %12, <32 x i16> %131714  %16 = extractelement <32 x i16> %15, i32 01715  ret i16 %161716}1717 1718define i8 @test_reduce_v64i8(<64 x i8> %a0) {1719; X86-SSE2-LABEL: test_reduce_v64i8:1720; X86-SSE2:       ## %bb.0:1721; X86-SSE2-NEXT:    pminub %xmm3, %xmm11722; X86-SSE2-NEXT:    pminub %xmm2, %xmm01723; X86-SSE2-NEXT:    pminub %xmm1, %xmm01724; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1725; X86-SSE2-NEXT:    pminub %xmm0, %xmm11726; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1727; X86-SSE2-NEXT:    pminub %xmm1, %xmm01728; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11729; X86-SSE2-NEXT:    psrld $16, %xmm11730; X86-SSE2-NEXT:    pminub %xmm0, %xmm11731; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01732; X86-SSE2-NEXT:    psrlw $8, %xmm01733; X86-SSE2-NEXT:    pminub %xmm1, %xmm01734; X86-SSE2-NEXT:    movd %xmm0, %eax1735; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1736; X86-SSE2-NEXT:    retl1737;1738; X86-SSE42-LABEL: test_reduce_v64i8:1739; X86-SSE42:       ## %bb.0:1740; X86-SSE42-NEXT:    pminub %xmm3, %xmm11741; X86-SSE42-NEXT:    pminub %xmm2, %xmm01742; X86-SSE42-NEXT:    pminub %xmm1, %xmm01743; X86-SSE42-NEXT:    movdqa %xmm0, %xmm11744; X86-SSE42-NEXT:    psrlw $8, %xmm11745; X86-SSE42-NEXT:    pminub %xmm0, %xmm11746; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm01747; X86-SSE42-NEXT:    movd %xmm0, %eax1748; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1749; X86-SSE42-NEXT:    retl1750;1751; X86-AVX1-LABEL: test_reduce_v64i8:1752; X86-AVX1:       ## %bb.0:1753; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21754; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31755; X86-AVX1-NEXT:    vpminub %xmm2, %xmm3, %xmm21756; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01757; X86-AVX1-NEXT:    vpminub %xmm2, %xmm0, %xmm01758; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11759; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01760; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01761; X86-AVX1-NEXT:    vmovd %xmm0, %eax1762; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1763; X86-AVX1-NEXT:    vzeroupper1764; X86-AVX1-NEXT:    retl1765;1766; X86-AVX2-LABEL: test_reduce_v64i8:1767; X86-AVX2:       ## %bb.0:1768; X86-AVX2-NEXT:    vpminub %ymm1, %ymm0, %ymm01769; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11770; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01771; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11772; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01773; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01774; X86-AVX2-NEXT:    vmovd %xmm0, %eax1775; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1776; X86-AVX2-NEXT:    vzeroupper1777; X86-AVX2-NEXT:    retl1778;1779; X64-SSE2-LABEL: test_reduce_v64i8:1780; X64-SSE2:       ## %bb.0:1781; X64-SSE2-NEXT:    pminub %xmm3, %xmm11782; X64-SSE2-NEXT:    pminub %xmm2, %xmm01783; X64-SSE2-NEXT:    pminub %xmm1, %xmm01784; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1785; X64-SSE2-NEXT:    pminub %xmm0, %xmm11786; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1787; X64-SSE2-NEXT:    pminub %xmm1, %xmm01788; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11789; X64-SSE2-NEXT:    psrld $16, %xmm11790; X64-SSE2-NEXT:    pminub %xmm0, %xmm11791; X64-SSE2-NEXT:    movdqa %xmm1, %xmm01792; X64-SSE2-NEXT:    psrlw $8, %xmm01793; X64-SSE2-NEXT:    pminub %xmm1, %xmm01794; X64-SSE2-NEXT:    movd %xmm0, %eax1795; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1796; X64-SSE2-NEXT:    retq1797;1798; X64-SSE42-LABEL: test_reduce_v64i8:1799; X64-SSE42:       ## %bb.0:1800; X64-SSE42-NEXT:    pminub %xmm3, %xmm11801; X64-SSE42-NEXT:    pminub %xmm2, %xmm01802; X64-SSE42-NEXT:    pminub %xmm1, %xmm01803; X64-SSE42-NEXT:    movdqa %xmm0, %xmm11804; X64-SSE42-NEXT:    psrlw $8, %xmm11805; X64-SSE42-NEXT:    pminub %xmm0, %xmm11806; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm01807; X64-SSE42-NEXT:    movd %xmm0, %eax1808; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1809; X64-SSE42-NEXT:    retq1810;1811; X64-AVX1-LABEL: test_reduce_v64i8:1812; X64-AVX1:       ## %bb.0:1813; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21814; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31815; X64-AVX1-NEXT:    vpminub %xmm2, %xmm3, %xmm21816; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01817; X64-AVX1-NEXT:    vpminub %xmm2, %xmm0, %xmm01818; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11819; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01820; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01821; X64-AVX1-NEXT:    vmovd %xmm0, %eax1822; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1823; X64-AVX1-NEXT:    vzeroupper1824; X64-AVX1-NEXT:    retq1825;1826; X64-AVX2-LABEL: test_reduce_v64i8:1827; X64-AVX2:       ## %bb.0:1828; X64-AVX2-NEXT:    vpminub %ymm1, %ymm0, %ymm01829; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11830; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01831; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11832; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01833; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01834; X64-AVX2-NEXT:    vmovd %xmm0, %eax1835; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1836; X64-AVX2-NEXT:    vzeroupper1837; X64-AVX2-NEXT:    retq1838;1839; X64-AVX512-LABEL: test_reduce_v64i8:1840; X64-AVX512:       ## %bb.0:1841; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11842; X64-AVX512-NEXT:    vpminub %ymm1, %ymm0, %ymm01843; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11844; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01845; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm11846; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01847; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01848; X64-AVX512-NEXT:    vmovd %xmm0, %eax1849; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax1850; X64-AVX512-NEXT:    vzeroupper1851; X64-AVX512-NEXT:    retq1852  %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>1853  %2  = icmp ult <64 x i8> %a0, %11854  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %11855  %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>1856  %5  = icmp ult <64 x i8> %3, %41857  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %41858  %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>1859  %8  = icmp ult <64 x i8> %6, %71860  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %71861  %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>1862  %11 = icmp ult <64 x i8> %9, %101863  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %101864  %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>1865  %14 = icmp ult <64 x i8> %12, %131866  %15 = select <64 x i1> %14, <64 x i8> %12, <64 x i8> %131867  %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>1868  %17 = icmp ult <64 x i8> %15, %161869  %18 = select <64 x i1> %17, <64 x i8> %15, <64 x i8> %161870  %19 = extractelement <64 x i8> %18, i32 01871  ret i8 %191872}1873 1874;1875; Partial Vector Reductions1876;1877 1878define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {1879; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:1880; X86-SSE2:       ## %bb.0:1881; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1882; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21883; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21884; X86-SSE2-NEXT:    psubw %xmm2, %xmm01885; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1886; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21887; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21888; X86-SSE2-NEXT:    psubw %xmm2, %xmm01889; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11890; X86-SSE2-NEXT:    psrld $16, %xmm11891; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21892; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21893; X86-SSE2-NEXT:    psubw %xmm2, %xmm01894; X86-SSE2-NEXT:    movd %xmm0, %eax1895; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1896; X86-SSE2-NEXT:    retl1897;1898; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:1899; X86-SSE42:       ## %bb.0:1900; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01901; X86-SSE42-NEXT:    movd %xmm0, %eax1902; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1903; X86-SSE42-NEXT:    retl1904;1905; X86-AVX-LABEL: test_reduce_v16i16_v8i16:1906; X86-AVX:       ## %bb.0:1907; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm01908; X86-AVX-NEXT:    vmovd %xmm0, %eax1909; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax1910; X86-AVX-NEXT:    vzeroupper1911; X86-AVX-NEXT:    retl1912;1913; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:1914; X64-SSE2:       ## %bb.0:1915; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1916; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21917; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21918; X64-SSE2-NEXT:    psubw %xmm2, %xmm01919; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1920; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21921; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21922; X64-SSE2-NEXT:    psubw %xmm2, %xmm01923; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11924; X64-SSE2-NEXT:    psrld $16, %xmm11925; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21926; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21927; X64-SSE2-NEXT:    psubw %xmm2, %xmm01928; X64-SSE2-NEXT:    movd %xmm0, %eax1929; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1930; X64-SSE2-NEXT:    retq1931;1932; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:1933; X64-SSE42:       ## %bb.0:1934; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01935; X64-SSE42-NEXT:    movd %xmm0, %eax1936; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1937; X64-SSE42-NEXT:    retq1938;1939; X64-AVX-LABEL: test_reduce_v16i16_v8i16:1940; X64-AVX:       ## %bb.0:1941; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm01942; X64-AVX-NEXT:    vmovd %xmm0, %eax1943; X64-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax1944; X64-AVX-NEXT:    vzeroupper1945; X64-AVX-NEXT:    retq1946  %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>1947  %2  = icmp ult <16 x i16> %a0, %11948  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %11949  %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>1950  %5  = icmp ult <16 x i16> %3, %41951  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %41952  %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>1953  %8  = icmp ult <16 x i16> %6, %71954  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %71955  %10 = extractelement <16 x i16> %9, i32 01956  ret i16 %101957}1958 1959define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {1960; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:1961; X86-SSE2:       ## %bb.0:1962; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1963; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21964; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21965; X86-SSE2-NEXT:    psubw %xmm2, %xmm01966; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1967; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21968; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21969; X86-SSE2-NEXT:    psubw %xmm2, %xmm01970; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11971; X86-SSE2-NEXT:    psrld $16, %xmm11972; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21973; X86-SSE2-NEXT:    psubusw %xmm1, %xmm21974; X86-SSE2-NEXT:    psubw %xmm2, %xmm01975; X86-SSE2-NEXT:    movd %xmm0, %eax1976; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1977; X86-SSE2-NEXT:    retl1978;1979; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:1980; X86-SSE42:       ## %bb.0:1981; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01982; X86-SSE42-NEXT:    movd %xmm0, %eax1983; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1984; X86-SSE42-NEXT:    retl1985;1986; X86-AVX-LABEL: test_reduce_v32i16_v8i16:1987; X86-AVX:       ## %bb.0:1988; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm01989; X86-AVX-NEXT:    vmovd %xmm0, %eax1990; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax1991; X86-AVX-NEXT:    vzeroupper1992; X86-AVX-NEXT:    retl1993;1994; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:1995; X64-SSE2:       ## %bb.0:1996; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1997; X64-SSE2-NEXT:    movdqa %xmm0, %xmm21998; X64-SSE2-NEXT:    psubusw %xmm1, %xmm21999; X64-SSE2-NEXT:    psubw %xmm2, %xmm02000; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]2001; X64-SSE2-NEXT:    movdqa %xmm0, %xmm22002; X64-SSE2-NEXT:    psubusw %xmm1, %xmm22003; X64-SSE2-NEXT:    psubw %xmm2, %xmm02004; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12005; X64-SSE2-NEXT:    psrld $16, %xmm12006; X64-SSE2-NEXT:    movdqa %xmm0, %xmm22007; X64-SSE2-NEXT:    psubusw %xmm1, %xmm22008; X64-SSE2-NEXT:    psubw %xmm2, %xmm02009; X64-SSE2-NEXT:    movd %xmm0, %eax2010; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax2011; X64-SSE2-NEXT:    retq2012;2013; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:2014; X64-SSE42:       ## %bb.0:2015; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm02016; X64-SSE42-NEXT:    movd %xmm0, %eax2017; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax2018; X64-SSE42-NEXT:    retq2019;2020; X64-AVX-LABEL: test_reduce_v32i16_v8i16:2021; X64-AVX:       ## %bb.0:2022; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm02023; X64-AVX-NEXT:    vmovd %xmm0, %eax2024; X64-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax2025; X64-AVX-NEXT:    vzeroupper2026; X64-AVX-NEXT:    retq2027  %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>2028  %2  = icmp ult <32 x i16> %a0, %12029  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %12030  %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>2031  %5  = icmp ult <32 x i16> %3, %42032  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %42033  %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>2034  %8  = icmp ult <32 x i16> %6, %72035  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %72036  %10 = extractelement <32 x i16> %9, i32 02037  ret i16 %102038}2039 2040define i8 @test_reduce_v32i8_v16i8(<32 x i8> %a0) {2041; X86-SSE2-LABEL: test_reduce_v32i8_v16i8:2042; X86-SSE2:       ## %bb.0:2043; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2044; X86-SSE2-NEXT:    pminub %xmm0, %xmm12045; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2046; X86-SSE2-NEXT:    pminub %xmm1, %xmm02047; X86-SSE2-NEXT:    movdqa %xmm0, %xmm12048; X86-SSE2-NEXT:    psrld $16, %xmm12049; X86-SSE2-NEXT:    pminub %xmm0, %xmm12050; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02051; X86-SSE2-NEXT:    psrlw $8, %xmm02052; X86-SSE2-NEXT:    pminub %xmm1, %xmm02053; X86-SSE2-NEXT:    movd %xmm0, %eax2054; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2055; X86-SSE2-NEXT:    retl2056;2057; X86-SSE42-LABEL: test_reduce_v32i8_v16i8:2058; X86-SSE42:       ## %bb.0:2059; X86-SSE42-NEXT:    movdqa %xmm0, %xmm12060; X86-SSE42-NEXT:    psrlw $8, %xmm12061; X86-SSE42-NEXT:    pminub %xmm0, %xmm12062; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm02063; X86-SSE42-NEXT:    movd %xmm0, %eax2064; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2065; X86-SSE42-NEXT:    retl2066;2067; X86-AVX-LABEL: test_reduce_v32i8_v16i8:2068; X86-AVX:       ## %bb.0:2069; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12070; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02071; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02072; X86-AVX-NEXT:    vmovd %xmm0, %eax2073; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax2074; X86-AVX-NEXT:    vzeroupper2075; X86-AVX-NEXT:    retl2076;2077; X64-SSE2-LABEL: test_reduce_v32i8_v16i8:2078; X64-SSE2:       ## %bb.0:2079; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2080; X64-SSE2-NEXT:    pminub %xmm0, %xmm12081; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2082; X64-SSE2-NEXT:    pminub %xmm1, %xmm02083; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12084; X64-SSE2-NEXT:    psrld $16, %xmm12085; X64-SSE2-NEXT:    pminub %xmm0, %xmm12086; X64-SSE2-NEXT:    movdqa %xmm1, %xmm02087; X64-SSE2-NEXT:    psrlw $8, %xmm02088; X64-SSE2-NEXT:    pminub %xmm1, %xmm02089; X64-SSE2-NEXT:    movd %xmm0, %eax2090; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2091; X64-SSE2-NEXT:    retq2092;2093; X64-SSE42-LABEL: test_reduce_v32i8_v16i8:2094; X64-SSE42:       ## %bb.0:2095; X64-SSE42-NEXT:    movdqa %xmm0, %xmm12096; X64-SSE42-NEXT:    psrlw $8, %xmm12097; X64-SSE42-NEXT:    pminub %xmm0, %xmm12098; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm02099; X64-SSE42-NEXT:    movd %xmm0, %eax2100; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2101; X64-SSE42-NEXT:    retq2102;2103; X64-AVX-LABEL: test_reduce_v32i8_v16i8:2104; X64-AVX:       ## %bb.0:2105; X64-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12106; X64-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02107; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm02108; X64-AVX-NEXT:    vmovd %xmm0, %eax2109; X64-AVX-NEXT:    ## kill: def $al killed $al killed $eax2110; X64-AVX-NEXT:    vzeroupper2111; X64-AVX-NEXT:    retq2112  %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>2113  %2  = icmp ult <32 x i8> %a0, %12114  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %12115  %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>2116  %5  = icmp ult <32 x i8> %3, %42117  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %42118  %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>2119  %8  = icmp ult <32 x i8> %6, %72120  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %72121  %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>2122  %11 = icmp ult <32 x i8> %9, %102123  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %102124  %13 = extractelement <32 x i8> %12, i32 02125  ret i8 %132126}2127 2128define i8 @test_reduce_v64i8_v16i8(<64 x i8> %a0) {2129; X86-SSE2-LABEL: test_reduce_v64i8_v16i8:2130; X86-SSE2:       ## %bb.0:2131; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2132; X86-SSE2-NEXT:    pminub %xmm0, %xmm12133; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2134; X86-SSE2-NEXT:    pminub %xmm1, %xmm02135; X86-SSE2-NEXT:    movdqa %xmm0, %xmm12136; X86-SSE2-NEXT:    psrld $16, %xmm12137; X86-SSE2-NEXT:    pminub %xmm0, %xmm12138; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02139; X86-SSE2-NEXT:    psrlw $8, %xmm02140; X86-SSE2-NEXT:    pminub %xmm1, %xmm02141; X86-SSE2-NEXT:    movd %xmm0, %eax2142; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2143; X86-SSE2-NEXT:    retl2144;2145; X86-SSE42-LABEL: test_reduce_v64i8_v16i8:2146; X86-SSE42:       ## %bb.0:2147; X86-SSE42-NEXT:    movdqa %xmm0, %xmm12148; X86-SSE42-NEXT:    psrlw $8, %xmm12149; X86-SSE42-NEXT:    pminub %xmm0, %xmm12150; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm02151; X86-SSE42-NEXT:    movd %xmm0, %eax2152; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2153; X86-SSE42-NEXT:    retl2154;2155; X86-AVX-LABEL: test_reduce_v64i8_v16i8:2156; X86-AVX:       ## %bb.0:2157; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12158; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02159; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02160; X86-AVX-NEXT:    vmovd %xmm0, %eax2161; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax2162; X86-AVX-NEXT:    vzeroupper2163; X86-AVX-NEXT:    retl2164;2165; X64-SSE2-LABEL: test_reduce_v64i8_v16i8:2166; X64-SSE2:       ## %bb.0:2167; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2168; X64-SSE2-NEXT:    pminub %xmm0, %xmm12169; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2170; X64-SSE2-NEXT:    pminub %xmm1, %xmm02171; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12172; X64-SSE2-NEXT:    psrld $16, %xmm12173; X64-SSE2-NEXT:    pminub %xmm0, %xmm12174; X64-SSE2-NEXT:    movdqa %xmm1, %xmm02175; X64-SSE2-NEXT:    psrlw $8, %xmm02176; X64-SSE2-NEXT:    pminub %xmm1, %xmm02177; X64-SSE2-NEXT:    movd %xmm0, %eax2178; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2179; X64-SSE2-NEXT:    retq2180;2181; X64-SSE42-LABEL: test_reduce_v64i8_v16i8:2182; X64-SSE42:       ## %bb.0:2183; X64-SSE42-NEXT:    movdqa %xmm0, %xmm12184; X64-SSE42-NEXT:    psrlw $8, %xmm12185; X64-SSE42-NEXT:    pminub %xmm0, %xmm12186; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm02187; X64-SSE42-NEXT:    movd %xmm0, %eax2188; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2189; X64-SSE42-NEXT:    retq2190;2191; X64-AVX-LABEL: test_reduce_v64i8_v16i8:2192; X64-AVX:       ## %bb.0:2193; X64-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12194; X64-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02195; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm02196; X64-AVX-NEXT:    vmovd %xmm0, %eax2197; X64-AVX-NEXT:    ## kill: def $al killed $al killed $eax2198; X64-AVX-NEXT:    vzeroupper2199; X64-AVX-NEXT:    retq2200  %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>2201  %2  = icmp ult <64 x i8> %a0, %12202  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %12203  %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>2204  %5  = icmp ult <64 x i8> %3, %42205  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %42206  %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>2207  %8  = icmp ult <64 x i8> %6, %72208  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %72209  %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>2210  %11 = icmp ult <64 x i8> %9, %102211  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %102212  %13 = extractelement <64 x i8> %12, i32 02213  ret i8 %132214}2215