brintos

brintos / llvm-project-archived public Read only

0
0
Text · 111.6 KiB · 9946267 Raw
2453 lines · plain
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 %xmm3, %xmm425; X86-SSE2-NEXT:    pcmpgtd %xmm2, %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 {{.*#+}} xmm3 = [0,2147483648,0,2147483648]45; X86-SSE42-NEXT:    pxor %xmm3, %xmm046; X86-SSE42-NEXT:    pxor %xmm2, %xmm347; X86-SSE42-NEXT:    pcmpgtq %xmm3, %xmm048; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm249; X86-SSE42-NEXT:    movd %xmm2, %eax50; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx51; X86-SSE42-NEXT:    retl52;53; X86-AVX1-LABEL: test_reduce_v2i64:54; X86-AVX1:       ## %bb.0:55; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]56; X86-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [0,2147483648,0,2147483648]57; X86-AVX1-NEXT:    ## xmm2 = mem[0,0]58; X86-AVX1-NEXT:    vxorps %xmm2, %xmm0, %xmm359; X86-AVX1-NEXT:    vxorps %xmm2, %xmm1, %xmm260; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm261; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm062; X86-AVX1-NEXT:    vmovd %xmm0, %eax63; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx64; X86-AVX1-NEXT:    retl65;66; X86-AVX2-LABEL: test_reduce_v2i64:67; X86-AVX2:       ## %bb.0:68; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]69; X86-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [0,2147483648,0,2147483648]70; X86-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm371; X86-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm272; X86-AVX2-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm273; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm074; X86-AVX2-NEXT:    vmovd %xmm0, %eax75; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx76; X86-AVX2-NEXT:    retl77;78; X64-SSE2-LABEL: test_reduce_v2i64:79; X64-SSE2:       ## %bb.0:80; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]81; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]82; X64-SSE2-NEXT:    movdqa %xmm0, %xmm383; X64-SSE2-NEXT:    pxor %xmm2, %xmm384; X64-SSE2-NEXT:    pxor %xmm1, %xmm285; X64-SSE2-NEXT:    movdqa %xmm3, %xmm486; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm487; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]88; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm289; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]90; X64-SSE2-NEXT:    pand %xmm5, %xmm291; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]92; X64-SSE2-NEXT:    por %xmm2, %xmm393; X64-SSE2-NEXT:    pand %xmm3, %xmm094; X64-SSE2-NEXT:    pandn %xmm1, %xmm395; X64-SSE2-NEXT:    por %xmm0, %xmm396; X64-SSE2-NEXT:    movq %xmm3, %rax97; X64-SSE2-NEXT:    retq98;99; X64-SSE42-LABEL: test_reduce_v2i64:100; X64-SSE42:       ## %bb.0:101; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1102; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]103; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]104; X64-SSE42-NEXT:    pxor %xmm3, %xmm0105; X64-SSE42-NEXT:    pxor %xmm2, %xmm3106; X64-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0107; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2108; X64-SSE42-NEXT:    movq %xmm2, %rax109; X64-SSE42-NEXT:    retq110;111; X64-AVX1-LABEL: test_reduce_v2i64:112; X64-AVX1:       ## %bb.0:113; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]114; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]115; X64-AVX1-NEXT:    ## xmm2 = mem[0,0]116; X64-AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm3117; X64-AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm2118; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2119; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0120; X64-AVX1-NEXT:    vmovq %xmm0, %rax121; X64-AVX1-NEXT:    retq122;123; X64-AVX2-LABEL: test_reduce_v2i64:124; X64-AVX2:       ## %bb.0:125; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]126; X64-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]127; X64-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm3128; X64-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm2129; X64-AVX2-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2130; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0131; X64-AVX2-NEXT:    vmovq %xmm0, %rax132; X64-AVX2-NEXT:    retq133;134; X64-AVX512-LABEL: test_reduce_v2i64:135; X64-AVX512:       ## %bb.0:136; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]137; X64-AVX512-NEXT:    vpmaxuq %xmm1, %xmm0, %xmm0138; X64-AVX512-NEXT:    vmovq %xmm0, %rax139; X64-AVX512-NEXT:    retq140  %1 = shufflevector <2 x i64> %a0, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>141  %2 = icmp ugt <2 x i64> %a0, %1142  %3 = select <2 x i1> %2, <2 x i64> %a0, <2 x i64> %1143  %4 = extractelement <2 x i64> %3, i32 0144  ret i64 %4145}146 147define i32 @test_reduce_v4i32(<4 x i32> %a0) {148; X86-SSE2-LABEL: test_reduce_v4i32:149; X86-SSE2:       ## %bb.0:150; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]151; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]152; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1153; X86-SSE2-NEXT:    pxor %xmm2, %xmm1154; X86-SSE2-NEXT:    movdqa %xmm3, %xmm4155; X86-SSE2-NEXT:    pxor %xmm2, %xmm4156; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm1157; X86-SSE2-NEXT:    pand %xmm1, %xmm0158; X86-SSE2-NEXT:    pandn %xmm3, %xmm1159; X86-SSE2-NEXT:    por %xmm0, %xmm1160; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]161; X86-SSE2-NEXT:    movdqa %xmm1, %xmm3162; X86-SSE2-NEXT:    pxor %xmm2, %xmm3163; X86-SSE2-NEXT:    pxor %xmm0, %xmm2164; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm3165; X86-SSE2-NEXT:    pand %xmm3, %xmm1166; X86-SSE2-NEXT:    pandn %xmm0, %xmm3167; X86-SSE2-NEXT:    por %xmm1, %xmm3168; X86-SSE2-NEXT:    movd %xmm3, %eax169; X86-SSE2-NEXT:    retl170;171; X86-SSE42-LABEL: test_reduce_v4i32:172; X86-SSE42:       ## %bb.0:173; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]174; X86-SSE42-NEXT:    pmaxud %xmm0, %xmm1175; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]176; X86-SSE42-NEXT:    pmaxud %xmm1, %xmm0177; X86-SSE42-NEXT:    movd %xmm0, %eax178; X86-SSE42-NEXT:    retl179;180; X86-AVX-LABEL: test_reduce_v4i32:181; X86-AVX:       ## %bb.0:182; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]183; X86-AVX-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0184; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]185; X86-AVX-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0186; X86-AVX-NEXT:    vmovd %xmm0, %eax187; X86-AVX-NEXT:    retl188;189; X64-SSE2-LABEL: test_reduce_v4i32:190; X64-SSE2:       ## %bb.0:191; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]192; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]193; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3194; X64-SSE2-NEXT:    pxor %xmm2, %xmm3195; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4196; X64-SSE2-NEXT:    pxor %xmm2, %xmm4197; X64-SSE2-NEXT:    pcmpgtd %xmm4, %xmm3198; X64-SSE2-NEXT:    pand %xmm3, %xmm0199; X64-SSE2-NEXT:    pandn %xmm1, %xmm3200; X64-SSE2-NEXT:    por %xmm0, %xmm3201; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]202; X64-SSE2-NEXT:    movdqa %xmm3, %xmm1203; X64-SSE2-NEXT:    pxor %xmm2, %xmm1204; X64-SSE2-NEXT:    pxor %xmm0, %xmm2205; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1206; X64-SSE2-NEXT:    pand %xmm1, %xmm3207; X64-SSE2-NEXT:    pandn %xmm0, %xmm1208; X64-SSE2-NEXT:    por %xmm3, %xmm1209; X64-SSE2-NEXT:    movd %xmm1, %eax210; X64-SSE2-NEXT:    retq211;212; X64-SSE42-LABEL: test_reduce_v4i32:213; X64-SSE42:       ## %bb.0:214; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]215; X64-SSE42-NEXT:    pmaxud %xmm0, %xmm1216; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]217; X64-SSE42-NEXT:    pmaxud %xmm1, %xmm0218; X64-SSE42-NEXT:    movd %xmm0, %eax219; X64-SSE42-NEXT:    retq220;221; X64-AVX-LABEL: test_reduce_v4i32:222; X64-AVX:       ## %bb.0:223; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]224; X64-AVX-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0225; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]226; X64-AVX-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0227; X64-AVX-NEXT:    vmovd %xmm0, %eax228; X64-AVX-NEXT:    retq229  %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>230  %2 = icmp ugt <4 x i32> %a0, %1231  %3 = select <4 x i1> %2, <4 x i32> %a0, <4 x i32> %1232  %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>233  %5 = icmp ugt <4 x i32> %3, %4234  %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4235  %7 = extractelement <4 x i32> %6, i32 0236  ret i32 %7237}238 239define i16 @test_reduce_v8i16(<8 x i16> %a0) {240; X86-SSE2-LABEL: test_reduce_v8i16:241; X86-SSE2:       ## %bb.0:242; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]243; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1244; X86-SSE2-NEXT:    paddw %xmm0, %xmm1245; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]246; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0247; X86-SSE2-NEXT:    paddw %xmm1, %xmm0248; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1249; X86-SSE2-NEXT:    psrld $16, %xmm1250; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1251; X86-SSE2-NEXT:    paddw %xmm0, %xmm1252; X86-SSE2-NEXT:    movd %xmm1, %eax253; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax254; X86-SSE2-NEXT:    retl255;256; X86-SSE42-LABEL: test_reduce_v8i16:257; X86-SSE42:       ## %bb.0:258; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1259; X86-SSE42-NEXT:    pxor %xmm0, %xmm1260; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0261; X86-SSE42-NEXT:    movd %xmm0, %eax262; X86-SSE42-NEXT:    notl %eax263; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax264; X86-SSE42-NEXT:    retl265;266; X86-AVX-LABEL: test_reduce_v8i16:267; X86-AVX:       ## %bb.0:268; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1269; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0270; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0271; X86-AVX-NEXT:    vmovd %xmm0, %eax272; X86-AVX-NEXT:    notl %eax273; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax274; X86-AVX-NEXT:    retl275;276; X64-SSE2-LABEL: test_reduce_v8i16:277; X64-SSE2:       ## %bb.0:278; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]279; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1280; X64-SSE2-NEXT:    paddw %xmm0, %xmm1281; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]282; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0283; X64-SSE2-NEXT:    paddw %xmm1, %xmm0284; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1285; X64-SSE2-NEXT:    psrld $16, %xmm1286; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1287; X64-SSE2-NEXT:    paddw %xmm0, %xmm1288; X64-SSE2-NEXT:    movd %xmm1, %eax289; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax290; X64-SSE2-NEXT:    retq291;292; X64-SSE42-LABEL: test_reduce_v8i16:293; X64-SSE42:       ## %bb.0:294; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1295; X64-SSE42-NEXT:    pxor %xmm0, %xmm1296; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0297; X64-SSE42-NEXT:    movd %xmm0, %eax298; X64-SSE42-NEXT:    notl %eax299; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax300; X64-SSE42-NEXT:    retq301;302; X64-AVX1-LABEL: test_reduce_v8i16:303; X64-AVX1:       ## %bb.0:304; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1305; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0306; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0307; X64-AVX1-NEXT:    vmovd %xmm0, %eax308; X64-AVX1-NEXT:    notl %eax309; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax310; X64-AVX1-NEXT:    retq311;312; X64-AVX2-LABEL: test_reduce_v8i16:313; X64-AVX2:       ## %bb.0:314; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1315; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0316; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0317; X64-AVX2-NEXT:    vmovd %xmm0, %eax318; X64-AVX2-NEXT:    notl %eax319; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax320; X64-AVX2-NEXT:    retq321;322; X64-AVX512-LABEL: test_reduce_v8i16:323; X64-AVX512:       ## %bb.0:324; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0325; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0326; X64-AVX512-NEXT:    vmovd %xmm0, %eax327; X64-AVX512-NEXT:    notl %eax328; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax329; X64-AVX512-NEXT:    retq330  %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>331  %2  = icmp ugt <8 x i16> %a0, %1332  %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1333  %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>334  %5  = icmp ugt <8 x i16> %3, %4335  %6  = select <8 x i1> %5, <8 x i16> %3, <8 x i16> %4336  %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>337  %8  = icmp ugt <8 x i16> %6, %7338  %9  = select <8 x i1> %8, <8 x i16> %6, <8 x i16> %7339  %10 = extractelement <8 x i16> %9, i32 0340  ret i16 %10341}342 343define i8 @test_reduce_v16i8(<16 x i8> %a0) {344; X86-SSE2-LABEL: test_reduce_v16i8:345; X86-SSE2:       ## %bb.0:346; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]347; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm1348; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]349; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm0350; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1351; X86-SSE2-NEXT:    psrld $16, %xmm1352; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm1353; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0354; X86-SSE2-NEXT:    psrlw $8, %xmm0355; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm0356; X86-SSE2-NEXT:    movd %xmm0, %eax357; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax358; X86-SSE2-NEXT:    retl359;360; X86-SSE42-LABEL: test_reduce_v16i8:361; X86-SSE42:       ## %bb.0:362; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1363; X86-SSE42-NEXT:    pxor %xmm0, %xmm1364; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0365; X86-SSE42-NEXT:    psrlw $8, %xmm0366; X86-SSE42-NEXT:    pminub %xmm1, %xmm0367; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0368; X86-SSE42-NEXT:    movd %xmm0, %eax369; X86-SSE42-NEXT:    notb %al370; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax371; X86-SSE42-NEXT:    retl372;373; X86-AVX-LABEL: test_reduce_v16i8:374; X86-AVX:       ## %bb.0:375; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1376; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0377; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1378; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0379; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0380; X86-AVX-NEXT:    vmovd %xmm0, %eax381; X86-AVX-NEXT:    notb %al382; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax383; X86-AVX-NEXT:    retl384;385; X64-SSE2-LABEL: test_reduce_v16i8:386; X64-SSE2:       ## %bb.0:387; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]388; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm1389; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]390; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm0391; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1392; X64-SSE2-NEXT:    psrld $16, %xmm1393; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm1394; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0395; X64-SSE2-NEXT:    psrlw $8, %xmm0396; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm0397; X64-SSE2-NEXT:    movd %xmm0, %eax398; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax399; X64-SSE2-NEXT:    retq400;401; X64-SSE42-LABEL: test_reduce_v16i8:402; X64-SSE42:       ## %bb.0:403; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1404; X64-SSE42-NEXT:    pxor %xmm0, %xmm1405; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0406; X64-SSE42-NEXT:    psrlw $8, %xmm0407; X64-SSE42-NEXT:    pminub %xmm1, %xmm0408; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0409; X64-SSE42-NEXT:    movd %xmm0, %eax410; X64-SSE42-NEXT:    notb %al411; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax412; X64-SSE42-NEXT:    retq413;414; X64-AVX1-LABEL: test_reduce_v16i8:415; X64-AVX1:       ## %bb.0:416; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1417; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0418; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1419; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0420; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0421; X64-AVX1-NEXT:    vmovd %xmm0, %eax422; X64-AVX1-NEXT:    notb %al423; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax424; X64-AVX1-NEXT:    retq425;426; X64-AVX2-LABEL: test_reduce_v16i8:427; X64-AVX2:       ## %bb.0:428; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1429; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0430; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1431; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0432; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0433; X64-AVX2-NEXT:    vmovd %xmm0, %eax434; X64-AVX2-NEXT:    notb %al435; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax436; X64-AVX2-NEXT:    retq437;438; X64-AVX512-LABEL: test_reduce_v16i8:439; X64-AVX512:       ## %bb.0:440; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0441; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1442; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0443; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0444; X64-AVX512-NEXT:    vmovd %xmm0, %eax445; X64-AVX512-NEXT:    notb %al446; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax447; X64-AVX512-NEXT:    retq448  %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>449  %2  = icmp ugt <16 x i8> %a0, %1450  %3  = select <16 x i1> %2, <16 x i8> %a0, <16 x i8> %1451  %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>452  %5  = icmp ugt <16 x i8> %3, %4453  %6  = select <16 x i1> %5, <16 x i8> %3, <16 x i8> %4454  %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>455  %8  = icmp ugt <16 x i8> %6, %7456  %9  = select <16 x i1> %8, <16 x i8> %6, <16 x i8> %7457  %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>458  %11 = icmp ugt <16 x i8> %9, %10459  %12 = select <16 x i1> %11, <16 x i8> %9, <16 x i8> %10460  %13 = extractelement <16 x i8> %12, i32 0461  ret i8 %13462}463 464;465; 256-bit Vectors466;467 468define i64 @test_reduce_v4i64(<4 x i64> %a0) {469; X86-SSE2-LABEL: test_reduce_v4i64:470; X86-SSE2:       ## %bb.0:471; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]472; X86-SSE2-NEXT:    movdqa %xmm1, %xmm3473; X86-SSE2-NEXT:    pxor %xmm2, %xmm3474; X86-SSE2-NEXT:    movdqa %xmm0, %xmm4475; X86-SSE2-NEXT:    pxor %xmm2, %xmm4476; X86-SSE2-NEXT:    movdqa %xmm4, %xmm5477; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5478; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]479; X86-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4480; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]481; X86-SSE2-NEXT:    pand %xmm6, %xmm4482; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]483; X86-SSE2-NEXT:    por %xmm4, %xmm3484; X86-SSE2-NEXT:    pand %xmm3, %xmm0485; X86-SSE2-NEXT:    pandn %xmm1, %xmm3486; X86-SSE2-NEXT:    por %xmm0, %xmm3487; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]488; X86-SSE2-NEXT:    movdqa %xmm3, %xmm1489; X86-SSE2-NEXT:    pxor %xmm2, %xmm1490; X86-SSE2-NEXT:    pxor %xmm0, %xmm2491; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4492; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm4493; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]494; X86-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2495; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]496; X86-SSE2-NEXT:    pand %xmm5, %xmm1497; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]498; X86-SSE2-NEXT:    por %xmm1, %xmm2499; X86-SSE2-NEXT:    pand %xmm2, %xmm3500; X86-SSE2-NEXT:    pandn %xmm0, %xmm2501; X86-SSE2-NEXT:    por %xmm3, %xmm2502; X86-SSE2-NEXT:    movd %xmm2, %eax503; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]504; X86-SSE2-NEXT:    movd %xmm0, %edx505; X86-SSE2-NEXT:    retl506;507; X86-SSE42-LABEL: test_reduce_v4i64:508; X86-SSE42:       ## %bb.0:509; X86-SSE42-NEXT:    movdqa %xmm0, %xmm2510; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm3 = [0,2147483648,0,2147483648]511; X86-SSE42-NEXT:    movdqa %xmm1, %xmm4512; X86-SSE42-NEXT:    pxor %xmm3, %xmm4513; X86-SSE42-NEXT:    pxor %xmm3, %xmm0514; X86-SSE42-NEXT:    pcmpgtq %xmm4, %xmm0515; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1516; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]517; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0518; X86-SSE42-NEXT:    pxor %xmm3, %xmm0519; X86-SSE42-NEXT:    pxor %xmm2, %xmm3520; X86-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0521; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2522; X86-SSE42-NEXT:    movd %xmm2, %eax523; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx524; X86-SSE42-NEXT:    retl525;526; X86-AVX1-LABEL: test_reduce_v4i64:527; X86-AVX1:       ## %bb.0:528; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1529; X86-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [0,2147483648,0,2147483648]530; X86-AVX1-NEXT:    ## xmm2 = mem[0,0]531; X86-AVX1-NEXT:    vxorps %xmm2, %xmm1, %xmm3532; X86-AVX1-NEXT:    vxorps %xmm2, %xmm0, %xmm4533; X86-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3534; X86-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0535; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]536; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm3537; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm1, %xmm2538; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2539; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0540; X86-AVX1-NEXT:    vmovd %xmm0, %eax541; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx542; X86-AVX1-NEXT:    vzeroupper543; X86-AVX1-NEXT:    retl544;545; X86-AVX2-LABEL: test_reduce_v4i64:546; X86-AVX2:       ## %bb.0:547; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1548; X86-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [0,2147483648,0,2147483648]549; X86-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm3550; X86-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm4551; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3552; X86-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0553; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]554; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm3555; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm2556; X86-AVX2-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2557; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0558; X86-AVX2-NEXT:    vmovd %xmm0, %eax559; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx560; X86-AVX2-NEXT:    vzeroupper561; X86-AVX2-NEXT:    retl562;563; X64-SSE2-LABEL: test_reduce_v4i64:564; X64-SSE2:       ## %bb.0:565; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]566; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3567; X64-SSE2-NEXT:    pxor %xmm2, %xmm3568; X64-SSE2-NEXT:    movdqa %xmm0, %xmm4569; X64-SSE2-NEXT:    pxor %xmm2, %xmm4570; X64-SSE2-NEXT:    movdqa %xmm4, %xmm5571; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5572; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]573; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4574; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]575; X64-SSE2-NEXT:    pand %xmm6, %xmm3576; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]577; X64-SSE2-NEXT:    por %xmm3, %xmm4578; X64-SSE2-NEXT:    pand %xmm4, %xmm0579; X64-SSE2-NEXT:    pandn %xmm1, %xmm4580; X64-SSE2-NEXT:    por %xmm0, %xmm4581; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]582; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1583; X64-SSE2-NEXT:    pxor %xmm2, %xmm1584; X64-SSE2-NEXT:    pxor %xmm0, %xmm2585; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3586; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm3587; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]588; X64-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2589; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]590; X64-SSE2-NEXT:    pand %xmm5, %xmm1591; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]592; X64-SSE2-NEXT:    por %xmm1, %xmm2593; X64-SSE2-NEXT:    pand %xmm2, %xmm4594; X64-SSE2-NEXT:    pandn %xmm0, %xmm2595; X64-SSE2-NEXT:    por %xmm4, %xmm2596; X64-SSE2-NEXT:    movq %xmm2, %rax597; X64-SSE2-NEXT:    retq598;599; X64-SSE42-LABEL: test_reduce_v4i64:600; X64-SSE42:       ## %bb.0:601; X64-SSE42-NEXT:    movdqa %xmm0, %xmm2602; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]603; X64-SSE42-NEXT:    movdqa %xmm1, %xmm4604; X64-SSE42-NEXT:    pxor %xmm3, %xmm4605; X64-SSE42-NEXT:    pxor %xmm3, %xmm0606; X64-SSE42-NEXT:    pcmpgtq %xmm4, %xmm0607; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1608; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]609; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0610; X64-SSE42-NEXT:    pxor %xmm3, %xmm0611; X64-SSE42-NEXT:    pxor %xmm2, %xmm3612; X64-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0613; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2614; X64-SSE42-NEXT:    movq %xmm2, %rax615; X64-SSE42-NEXT:    retq616;617; X64-AVX1-LABEL: test_reduce_v4i64:618; X64-AVX1:       ## %bb.0:619; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1620; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]621; X64-AVX1-NEXT:    ## xmm2 = mem[0,0]622; X64-AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm3623; X64-AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm4624; X64-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3625; X64-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0626; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]627; X64-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm3628; X64-AVX1-NEXT:    vxorpd %xmm2, %xmm1, %xmm2629; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2630; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0631; X64-AVX1-NEXT:    vmovq %xmm0, %rax632; X64-AVX1-NEXT:    vzeroupper633; X64-AVX1-NEXT:    retq634;635; X64-AVX2-LABEL: test_reduce_v4i64:636; X64-AVX2:       ## %bb.0:637; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1638; X64-AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]639; X64-AVX2-NEXT:    vpxor %xmm2, %xmm1, %xmm3640; X64-AVX2-NEXT:    vpxor %xmm2, %xmm0, %xmm4641; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3642; X64-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0643; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]644; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm3645; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm2646; X64-AVX2-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2647; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0648; X64-AVX2-NEXT:    vmovq %xmm0, %rax649; X64-AVX2-NEXT:    vzeroupper650; X64-AVX2-NEXT:    retq651;652; X64-AVX512-LABEL: test_reduce_v4i64:653; X64-AVX512:       ## %bb.0:654; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1655; X64-AVX512-NEXT:    vpmaxuq %xmm1, %xmm0, %xmm0656; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]657; X64-AVX512-NEXT:    vpmaxuq %xmm1, %xmm0, %xmm0658; X64-AVX512-NEXT:    vmovq %xmm0, %rax659; X64-AVX512-NEXT:    vzeroupper660; X64-AVX512-NEXT:    retq661  %1 = shufflevector <4 x i64> %a0, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>662  %2 = icmp ugt <4 x i64> %a0, %1663  %3 = select <4 x i1> %2, <4 x i64> %a0, <4 x i64> %1664  %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>665  %5 = icmp ugt <4 x i64> %3, %4666  %6 = select <4 x i1> %5, <4 x i64> %3, <4 x i64> %4667  %7 = extractelement <4 x i64> %6, i32 0668  ret i64 %7669}670 671define i32 @test_reduce_v8i32(<8 x i32> %a0) {672; X86-SSE2-LABEL: test_reduce_v8i32:673; X86-SSE2:       ## %bb.0:674; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]675; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4676; X86-SSE2-NEXT:    pxor %xmm2, %xmm4677; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3678; X86-SSE2-NEXT:    pxor %xmm2, %xmm3679; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm3680; X86-SSE2-NEXT:    pand %xmm3, %xmm0681; X86-SSE2-NEXT:    pandn %xmm1, %xmm3682; X86-SSE2-NEXT:    por %xmm0, %xmm3683; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]684; X86-SSE2-NEXT:    movdqa %xmm3, %xmm0685; X86-SSE2-NEXT:    pxor %xmm2, %xmm0686; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4687; X86-SSE2-NEXT:    pxor %xmm2, %xmm4688; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm0689; X86-SSE2-NEXT:    pand %xmm0, %xmm3690; X86-SSE2-NEXT:    pandn %xmm1, %xmm0691; X86-SSE2-NEXT:    por %xmm3, %xmm0692; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]693; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3694; X86-SSE2-NEXT:    pxor %xmm2, %xmm3695; X86-SSE2-NEXT:    pxor %xmm1, %xmm2696; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm3697; X86-SSE2-NEXT:    pand %xmm3, %xmm0698; X86-SSE2-NEXT:    pandn %xmm1, %xmm3699; X86-SSE2-NEXT:    por %xmm0, %xmm3700; X86-SSE2-NEXT:    movd %xmm3, %eax701; X86-SSE2-NEXT:    retl702;703; X86-SSE42-LABEL: test_reduce_v8i32:704; X86-SSE42:       ## %bb.0:705; X86-SSE42-NEXT:    pmaxud %xmm1, %xmm0706; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]707; X86-SSE42-NEXT:    pmaxud %xmm0, %xmm1708; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]709; X86-SSE42-NEXT:    pmaxud %xmm1, %xmm0710; X86-SSE42-NEXT:    movd %xmm0, %eax711; X86-SSE42-NEXT:    retl712;713; X86-AVX1-LABEL: test_reduce_v8i32:714; X86-AVX1:       ## %bb.0:715; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1716; X86-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0717; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]718; X86-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0719; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]720; X86-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0721; X86-AVX1-NEXT:    vmovd %xmm0, %eax722; X86-AVX1-NEXT:    vzeroupper723; X86-AVX1-NEXT:    retl724;725; X86-AVX2-LABEL: test_reduce_v8i32:726; X86-AVX2:       ## %bb.0:727; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1728; X86-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0729; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]730; X86-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0731; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]732; X86-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0733; X86-AVX2-NEXT:    vmovd %xmm0, %eax734; X86-AVX2-NEXT:    vzeroupper735; X86-AVX2-NEXT:    retl736;737; X64-SSE2-LABEL: test_reduce_v8i32:738; X64-SSE2:       ## %bb.0:739; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]740; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3741; X64-SSE2-NEXT:    pxor %xmm2, %xmm3742; X64-SSE2-NEXT:    movdqa %xmm0, %xmm4743; X64-SSE2-NEXT:    pxor %xmm2, %xmm4744; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm4745; X64-SSE2-NEXT:    pand %xmm4, %xmm0746; X64-SSE2-NEXT:    pandn %xmm1, %xmm4747; X64-SSE2-NEXT:    por %xmm0, %xmm4748; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]749; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1750; X64-SSE2-NEXT:    pxor %xmm2, %xmm1751; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3752; X64-SSE2-NEXT:    pxor %xmm2, %xmm3753; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm1754; X64-SSE2-NEXT:    pand %xmm1, %xmm4755; X64-SSE2-NEXT:    pandn %xmm0, %xmm1756; X64-SSE2-NEXT:    por %xmm4, %xmm1757; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]758; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3759; X64-SSE2-NEXT:    pxor %xmm2, %xmm3760; X64-SSE2-NEXT:    pxor %xmm0, %xmm2761; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm3762; X64-SSE2-NEXT:    pand %xmm3, %xmm1763; X64-SSE2-NEXT:    pandn %xmm0, %xmm3764; X64-SSE2-NEXT:    por %xmm1, %xmm3765; X64-SSE2-NEXT:    movd %xmm3, %eax766; X64-SSE2-NEXT:    retq767;768; X64-SSE42-LABEL: test_reduce_v8i32:769; X64-SSE42:       ## %bb.0:770; X64-SSE42-NEXT:    pmaxud %xmm1, %xmm0771; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]772; X64-SSE42-NEXT:    pmaxud %xmm0, %xmm1773; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]774; X64-SSE42-NEXT:    pmaxud %xmm1, %xmm0775; X64-SSE42-NEXT:    movd %xmm0, %eax776; X64-SSE42-NEXT:    retq777;778; X64-AVX1-LABEL: test_reduce_v8i32:779; X64-AVX1:       ## %bb.0:780; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1781; X64-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0782; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]783; X64-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0784; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]785; X64-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0786; X64-AVX1-NEXT:    vmovd %xmm0, %eax787; X64-AVX1-NEXT:    vzeroupper788; X64-AVX1-NEXT:    retq789;790; X64-AVX2-LABEL: test_reduce_v8i32:791; X64-AVX2:       ## %bb.0:792; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1793; X64-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0794; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]795; X64-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0796; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]797; X64-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0798; X64-AVX2-NEXT:    vmovd %xmm0, %eax799; X64-AVX2-NEXT:    vzeroupper800; X64-AVX2-NEXT:    retq801;802; X64-AVX512-LABEL: test_reduce_v8i32:803; X64-AVX512:       ## %bb.0:804; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1805; X64-AVX512-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0806; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]807; X64-AVX512-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0808; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]809; X64-AVX512-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0810; X64-AVX512-NEXT:    vmovd %xmm0, %eax811; X64-AVX512-NEXT:    vzeroupper812; X64-AVX512-NEXT:    retq813  %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>814  %2  = icmp ugt <8 x i32> %a0, %1815  %3  = select <8 x i1> %2, <8 x i32> %a0, <8 x i32> %1816  %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>817  %5  = icmp ugt <8 x i32> %3, %4818  %6  = select <8 x i1> %5, <8 x i32> %3, <8 x i32> %4819  %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>820  %8  = icmp ugt <8 x i32> %6, %7821  %9  = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7822  %10 = extractelement <8 x i32> %9, i32 0823  ret i32 %10824}825 826define i16 @test_reduce_v16i16(<16 x i16> %a0) {827; X86-SSE2-LABEL: test_reduce_v16i16:828; X86-SSE2:       ## %bb.0:829; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1830; X86-SSE2-NEXT:    paddw %xmm0, %xmm1831; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]832; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0833; X86-SSE2-NEXT:    paddw %xmm1, %xmm0834; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]835; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1836; X86-SSE2-NEXT:    paddw %xmm0, %xmm1837; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0838; X86-SSE2-NEXT:    psrld $16, %xmm0839; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0840; X86-SSE2-NEXT:    paddw %xmm1, %xmm0841; X86-SSE2-NEXT:    movd %xmm0, %eax842; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax843; X86-SSE2-NEXT:    retl844;845; X86-SSE42-LABEL: test_reduce_v16i16:846; X86-SSE42:       ## %bb.0:847; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0848; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1849; X86-SSE42-NEXT:    pxor %xmm0, %xmm1850; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0851; X86-SSE42-NEXT:    movd %xmm0, %eax852; X86-SSE42-NEXT:    notl %eax853; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax854; X86-SSE42-NEXT:    retl855;856; X86-AVX1-LABEL: test_reduce_v16i16:857; X86-AVX1:       ## %bb.0:858; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1859; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0860; X86-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1861; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0862; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0863; X86-AVX1-NEXT:    vmovd %xmm0, %eax864; X86-AVX1-NEXT:    notl %eax865; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax866; X86-AVX1-NEXT:    vzeroupper867; X86-AVX1-NEXT:    retl868;869; X86-AVX2-LABEL: test_reduce_v16i16:870; X86-AVX2:       ## %bb.0:871; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1872; X86-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0873; X86-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1874; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0875; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0876; X86-AVX2-NEXT:    vmovd %xmm0, %eax877; X86-AVX2-NEXT:    notl %eax878; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax879; X86-AVX2-NEXT:    vzeroupper880; X86-AVX2-NEXT:    retl881;882; X64-SSE2-LABEL: test_reduce_v16i16:883; X64-SSE2:       ## %bb.0:884; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1885; X64-SSE2-NEXT:    paddw %xmm0, %xmm1886; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]887; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0888; X64-SSE2-NEXT:    paddw %xmm1, %xmm0889; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]890; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1891; X64-SSE2-NEXT:    paddw %xmm0, %xmm1892; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0893; X64-SSE2-NEXT:    psrld $16, %xmm0894; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0895; X64-SSE2-NEXT:    paddw %xmm1, %xmm0896; X64-SSE2-NEXT:    movd %xmm0, %eax897; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax898; X64-SSE2-NEXT:    retq899;900; X64-SSE42-LABEL: test_reduce_v16i16:901; X64-SSE42:       ## %bb.0:902; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0903; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1904; X64-SSE42-NEXT:    pxor %xmm0, %xmm1905; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0906; X64-SSE42-NEXT:    movd %xmm0, %eax907; X64-SSE42-NEXT:    notl %eax908; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax909; X64-SSE42-NEXT:    retq910;911; X64-AVX1-LABEL: test_reduce_v16i16:912; X64-AVX1:       ## %bb.0:913; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1914; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0915; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1916; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0917; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0918; X64-AVX1-NEXT:    vmovd %xmm0, %eax919; X64-AVX1-NEXT:    notl %eax920; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax921; X64-AVX1-NEXT:    vzeroupper922; X64-AVX1-NEXT:    retq923;924; X64-AVX2-LABEL: test_reduce_v16i16:925; X64-AVX2:       ## %bb.0:926; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1927; X64-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0928; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1929; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0930; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0931; X64-AVX2-NEXT:    vmovd %xmm0, %eax932; X64-AVX2-NEXT:    notl %eax933; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax934; X64-AVX2-NEXT:    vzeroupper935; X64-AVX2-NEXT:    retq936;937; X64-AVX512-LABEL: test_reduce_v16i16:938; X64-AVX512:       ## %bb.0:939; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1940; X64-AVX512-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0941; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0942; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0943; X64-AVX512-NEXT:    vmovd %xmm0, %eax944; X64-AVX512-NEXT:    notl %eax945; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax946; X64-AVX512-NEXT:    vzeroupper947; X64-AVX512-NEXT:    retq948  %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>949  %2  = icmp ugt <16 x i16> %a0, %1950  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1951  %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>952  %5  = icmp ugt <16 x i16> %3, %4953  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %4954  %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>955  %8  = icmp ugt <16 x i16> %6, %7956  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7957  %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>958  %11 = icmp ugt <16 x i16> %9, %10959  %12 = select <16 x i1> %11, <16 x i16> %9, <16 x i16> %10960  %13 = extractelement <16 x i16> %12, i32 0961  ret i16 %13962}963 964define i8 @test_reduce_v32i8(<32 x i8> %a0) {965; X86-SSE2-LABEL: test_reduce_v32i8:966; X86-SSE2:       ## %bb.0:967; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm0968; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]969; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm1970; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]971; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm0972; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1973; X86-SSE2-NEXT:    psrld $16, %xmm1974; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm1975; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0976; X86-SSE2-NEXT:    psrlw $8, %xmm0977; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm0978; X86-SSE2-NEXT:    movd %xmm0, %eax979; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax980; X86-SSE2-NEXT:    retl981;982; X86-SSE42-LABEL: test_reduce_v32i8:983; X86-SSE42:       ## %bb.0:984; X86-SSE42-NEXT:    pmaxub %xmm1, %xmm0985; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1986; X86-SSE42-NEXT:    pxor %xmm0, %xmm1987; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0988; X86-SSE42-NEXT:    psrlw $8, %xmm0989; X86-SSE42-NEXT:    pminub %xmm1, %xmm0990; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0991; X86-SSE42-NEXT:    movd %xmm0, %eax992; X86-SSE42-NEXT:    notb %al993; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax994; X86-SSE42-NEXT:    retl995;996; X86-AVX1-LABEL: test_reduce_v32i8:997; X86-AVX1:       ## %bb.0:998; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1999; X86-AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01000; X86-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11001; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm01002; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11003; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01004; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01005; X86-AVX1-NEXT:    vmovd %xmm0, %eax1006; X86-AVX1-NEXT:    notb %al1007; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1008; X86-AVX1-NEXT:    vzeroupper1009; X86-AVX1-NEXT:    retl1010;1011; X86-AVX2-LABEL: test_reduce_v32i8:1012; X86-AVX2:       ## %bb.0:1013; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11014; X86-AVX2-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01015; X86-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11016; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm01017; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11018; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01019; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01020; X86-AVX2-NEXT:    vmovd %xmm0, %eax1021; X86-AVX2-NEXT:    notb %al1022; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1023; X86-AVX2-NEXT:    vzeroupper1024; X86-AVX2-NEXT:    retl1025;1026; X64-SSE2-LABEL: test_reduce_v32i8:1027; X64-SSE2:       ## %bb.0:1028; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm01029; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1030; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm11031; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1032; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm01033; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11034; X64-SSE2-NEXT:    psrld $16, %xmm11035; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm11036; X64-SSE2-NEXT:    movdqa %xmm1, %xmm01037; X64-SSE2-NEXT:    psrlw $8, %xmm01038; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm01039; X64-SSE2-NEXT:    movd %xmm0, %eax1040; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1041; X64-SSE2-NEXT:    retq1042;1043; X64-SSE42-LABEL: test_reduce_v32i8:1044; X64-SSE42:       ## %bb.0:1045; X64-SSE42-NEXT:    pmaxub %xmm1, %xmm01046; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm11047; X64-SSE42-NEXT:    pxor %xmm0, %xmm11048; X64-SSE42-NEXT:    movdqa %xmm1, %xmm01049; X64-SSE42-NEXT:    psrlw $8, %xmm01050; X64-SSE42-NEXT:    pminub %xmm1, %xmm01051; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01052; X64-SSE42-NEXT:    movd %xmm0, %eax1053; X64-SSE42-NEXT:    notb %al1054; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1055; X64-SSE42-NEXT:    retq1056;1057; X64-AVX1-LABEL: test_reduce_v32i8:1058; X64-AVX1:       ## %bb.0:1059; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm11060; X64-AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01061; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11062; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm01063; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11064; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01065; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01066; X64-AVX1-NEXT:    vmovd %xmm0, %eax1067; X64-AVX1-NEXT:    notb %al1068; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1069; X64-AVX1-NEXT:    vzeroupper1070; X64-AVX1-NEXT:    retq1071;1072; X64-AVX2-LABEL: test_reduce_v32i8:1073; X64-AVX2:       ## %bb.0:1074; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11075; X64-AVX2-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01076; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11077; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm01078; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11079; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01080; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01081; X64-AVX2-NEXT:    vmovd %xmm0, %eax1082; X64-AVX2-NEXT:    notb %al1083; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1084; X64-AVX2-NEXT:    vzeroupper1085; X64-AVX2-NEXT:    retq1086;1087; X64-AVX512-LABEL: test_reduce_v32i8:1088; X64-AVX512:       ## %bb.0:1089; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11090; X64-AVX512-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01091; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm01092; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm11093; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01094; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01095; X64-AVX512-NEXT:    vmovd %xmm0, %eax1096; X64-AVX512-NEXT:    notb %al1097; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax1098; X64-AVX512-NEXT:    vzeroupper1099; X64-AVX512-NEXT:    retq1100  %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>1101  %2  = icmp ugt <32 x i8> %a0, %11102  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %11103  %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>1104  %5  = icmp ugt <32 x i8> %3, %41105  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %41106  %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>1107  %8  = icmp ugt <32 x i8> %6, %71108  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %71109  %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>1110  %11 = icmp ugt <32 x i8> %9, %101111  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %101112  %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>1113  %14 = icmp ugt <32 x i8> %12, %131114  %15 = select <32 x i1> %14, <32 x i8> %12, <32 x i8> %131115  %16 = extractelement <32 x i8> %15, i32 01116  ret i8 %161117}1118 1119;1120; 512-bit Vectors1121;1122 1123define i64 @test_reduce_v8i64(<8 x i64> %a0) {1124; X86-SSE2-LABEL: test_reduce_v8i64:1125; X86-SSE2:       ## %bb.0:1126; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1127; X86-SSE2-NEXT:    movdqa %xmm2, %xmm51128; X86-SSE2-NEXT:    pxor %xmm4, %xmm51129; X86-SSE2-NEXT:    movdqa %xmm0, %xmm61130; X86-SSE2-NEXT:    pxor %xmm4, %xmm61131; X86-SSE2-NEXT:    movdqa %xmm6, %xmm71132; X86-SSE2-NEXT:    pcmpgtd %xmm5, %xmm71133; X86-SSE2-NEXT:    pcmpeqd %xmm5, %xmm61134; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]1135; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1136; X86-SSE2-NEXT:    pand %xmm5, %xmm61137; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1138; X86-SSE2-NEXT:    por %xmm6, %xmm51139; X86-SSE2-NEXT:    pand %xmm5, %xmm01140; X86-SSE2-NEXT:    pandn %xmm2, %xmm51141; X86-SSE2-NEXT:    por %xmm0, %xmm51142; X86-SSE2-NEXT:    movdqa %xmm3, %xmm01143; X86-SSE2-NEXT:    pxor %xmm4, %xmm01144; X86-SSE2-NEXT:    movdqa %xmm1, %xmm21145; X86-SSE2-NEXT:    pxor %xmm4, %xmm21146; X86-SSE2-NEXT:    movdqa %xmm2, %xmm61147; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm61148; X86-SSE2-NEXT:    pcmpeqd %xmm0, %xmm21149; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]1150; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]1151; X86-SSE2-NEXT:    pand %xmm0, %xmm71152; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]1153; X86-SSE2-NEXT:    por %xmm7, %xmm21154; X86-SSE2-NEXT:    pand %xmm2, %xmm11155; X86-SSE2-NEXT:    pandn %xmm3, %xmm21156; X86-SSE2-NEXT:    por %xmm1, %xmm21157; X86-SSE2-NEXT:    movdqa %xmm2, %xmm01158; X86-SSE2-NEXT:    pxor %xmm4, %xmm01159; X86-SSE2-NEXT:    movdqa %xmm5, %xmm11160; X86-SSE2-NEXT:    pxor %xmm4, %xmm11161; X86-SSE2-NEXT:    movdqa %xmm1, %xmm31162; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm31163; X86-SSE2-NEXT:    pcmpeqd %xmm0, %xmm11164; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]1165; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]1166; X86-SSE2-NEXT:    pand %xmm0, %xmm11167; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]1168; X86-SSE2-NEXT:    por %xmm1, %xmm01169; X86-SSE2-NEXT:    pand %xmm0, %xmm51170; X86-SSE2-NEXT:    pandn %xmm2, %xmm01171; X86-SSE2-NEXT:    por %xmm5, %xmm01172; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1173; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21174; X86-SSE2-NEXT:    pxor %xmm4, %xmm21175; X86-SSE2-NEXT:    pxor %xmm1, %xmm41176; X86-SSE2-NEXT:    movdqa %xmm2, %xmm31177; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm31178; X86-SSE2-NEXT:    pcmpeqd %xmm2, %xmm41179; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]1180; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]1181; X86-SSE2-NEXT:    pand %xmm2, %xmm41182; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]1183; X86-SSE2-NEXT:    por %xmm4, %xmm21184; X86-SSE2-NEXT:    pand %xmm2, %xmm01185; X86-SSE2-NEXT:    pandn %xmm1, %xmm21186; X86-SSE2-NEXT:    por %xmm0, %xmm21187; X86-SSE2-NEXT:    movd %xmm2, %eax1188; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1189; X86-SSE2-NEXT:    movd %xmm0, %edx1190; X86-SSE2-NEXT:    retl1191;1192; X86-SSE42-LABEL: test_reduce_v8i64:1193; X86-SSE42:       ## %bb.0:1194; X86-SSE42-NEXT:    movdqa %xmm0, %xmm41195; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm5 = [0,2147483648,0,2147483648]1196; X86-SSE42-NEXT:    movdqa %xmm2, %xmm61197; X86-SSE42-NEXT:    pxor %xmm5, %xmm61198; X86-SSE42-NEXT:    pxor %xmm5, %xmm01199; X86-SSE42-NEXT:    pcmpgtq %xmm6, %xmm01200; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm21201; X86-SSE42-NEXT:    movdqa %xmm3, %xmm41202; X86-SSE42-NEXT:    pxor %xmm5, %xmm41203; X86-SSE42-NEXT:    movdqa %xmm1, %xmm01204; X86-SSE42-NEXT:    pxor %xmm5, %xmm01205; X86-SSE42-NEXT:    pcmpgtq %xmm4, %xmm01206; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm31207; X86-SSE42-NEXT:    movapd %xmm3, %xmm11208; X86-SSE42-NEXT:    xorpd %xmm5, %xmm11209; X86-SSE42-NEXT:    movapd %xmm2, %xmm01210; X86-SSE42-NEXT:    xorpd %xmm5, %xmm01211; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01212; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm31213; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1214; X86-SSE42-NEXT:    movdqa %xmm3, %xmm01215; X86-SSE42-NEXT:    pxor %xmm5, %xmm01216; X86-SSE42-NEXT:    pxor %xmm1, %xmm51217; X86-SSE42-NEXT:    pcmpgtq %xmm5, %xmm01218; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm11219; X86-SSE42-NEXT:    movd %xmm1, %eax1220; X86-SSE42-NEXT:    pextrd $1, %xmm1, %edx1221; X86-SSE42-NEXT:    retl1222;1223; X86-AVX1-LABEL: test_reduce_v8i64:1224; X86-AVX1:       ## %bb.0:1225; X86-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [0,2147483648,0,2147483648]1226; X86-AVX1-NEXT:    ## xmm2 = mem[0,0]1227; X86-AVX1-NEXT:    vxorps %xmm2, %xmm1, %xmm31228; X86-AVX1-NEXT:    vxorps %xmm2, %xmm0, %xmm41229; X86-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm31230; X86-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm31231; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm11232; X86-AVX1-NEXT:    vxorps %xmm2, %xmm1, %xmm41233; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm01234; X86-AVX1-NEXT:    vxorps %xmm2, %xmm0, %xmm51235; X86-AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm41236; X86-AVX1-NEXT:    vblendvpd %xmm4, %xmm0, %xmm1, %xmm01237; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm11238; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm3, %xmm41239; X86-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm11240; X86-AVX1-NEXT:    vblendvpd %xmm1, %xmm3, %xmm0, %xmm01241; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1242; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm31243; X86-AVX1-NEXT:    vxorpd %xmm2, %xmm1, %xmm21244; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm21245; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01246; X86-AVX1-NEXT:    vmovd %xmm0, %eax1247; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx1248; X86-AVX1-NEXT:    vzeroupper1249; X86-AVX1-NEXT:    retl1250;1251; X86-AVX2-LABEL: test_reduce_v8i64:1252; X86-AVX2:       ## %bb.0:1253; X86-AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [0,2147483648,0,2147483648,0,2147483648,0,2147483648]1254; X86-AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm31255; X86-AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm41256; X86-AVX2-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm31257; X86-AVX2-NEXT:    vblendvpd %ymm3, %ymm0, %ymm1, %ymm01258; X86-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm11259; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm31260; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm41261; X86-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm31262; X86-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm01263; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1264; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm31265; X86-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm21266; X86-AVX2-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm21267; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01268; X86-AVX2-NEXT:    vmovd %xmm0, %eax1269; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx1270; X86-AVX2-NEXT:    vzeroupper1271; X86-AVX2-NEXT:    retl1272;1273; X64-SSE2-LABEL: test_reduce_v8i64:1274; X64-SSE2:       ## %bb.0:1275; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]1276; X64-SSE2-NEXT:    movdqa %xmm2, %xmm51277; X64-SSE2-NEXT:    pxor %xmm4, %xmm51278; X64-SSE2-NEXT:    movdqa %xmm0, %xmm61279; X64-SSE2-NEXT:    pxor %xmm4, %xmm61280; X64-SSE2-NEXT:    movdqa %xmm6, %xmm71281; X64-SSE2-NEXT:    pcmpgtd %xmm5, %xmm71282; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]1283; X64-SSE2-NEXT:    pcmpeqd %xmm5, %xmm61284; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1285; X64-SSE2-NEXT:    pand %xmm8, %xmm61286; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1287; X64-SSE2-NEXT:    por %xmm6, %xmm51288; X64-SSE2-NEXT:    pand %xmm5, %xmm01289; X64-SSE2-NEXT:    pandn %xmm2, %xmm51290; X64-SSE2-NEXT:    por %xmm0, %xmm51291; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01292; X64-SSE2-NEXT:    pxor %xmm4, %xmm01293; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21294; X64-SSE2-NEXT:    pxor %xmm4, %xmm21295; X64-SSE2-NEXT:    movdqa %xmm2, %xmm61296; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm61297; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]1298; X64-SSE2-NEXT:    pcmpeqd %xmm0, %xmm21299; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]1300; X64-SSE2-NEXT:    pand %xmm7, %xmm01301; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]1302; X64-SSE2-NEXT:    por %xmm0, %xmm21303; X64-SSE2-NEXT:    pand %xmm2, %xmm11304; X64-SSE2-NEXT:    pandn %xmm3, %xmm21305; X64-SSE2-NEXT:    por %xmm1, %xmm21306; X64-SSE2-NEXT:    movdqa %xmm2, %xmm01307; X64-SSE2-NEXT:    pxor %xmm4, %xmm01308; X64-SSE2-NEXT:    movdqa %xmm5, %xmm11309; X64-SSE2-NEXT:    pxor %xmm4, %xmm11310; X64-SSE2-NEXT:    movdqa %xmm1, %xmm31311; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm31312; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]1313; X64-SSE2-NEXT:    pcmpeqd %xmm0, %xmm11314; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]1315; X64-SSE2-NEXT:    pand %xmm6, %xmm01316; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]1317; X64-SSE2-NEXT:    por %xmm0, %xmm11318; X64-SSE2-NEXT:    pand %xmm1, %xmm51319; X64-SSE2-NEXT:    pandn %xmm2, %xmm11320; X64-SSE2-NEXT:    por %xmm5, %xmm11321; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1322; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21323; X64-SSE2-NEXT:    pxor %xmm4, %xmm21324; X64-SSE2-NEXT:    pxor %xmm0, %xmm41325; X64-SSE2-NEXT:    movdqa %xmm2, %xmm31326; X64-SSE2-NEXT:    pcmpgtd %xmm4, %xmm31327; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]1328; X64-SSE2-NEXT:    pcmpeqd %xmm2, %xmm41329; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]1330; X64-SSE2-NEXT:    pand %xmm5, %xmm21331; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]1332; X64-SSE2-NEXT:    por %xmm2, %xmm31333; X64-SSE2-NEXT:    pand %xmm3, %xmm11334; X64-SSE2-NEXT:    pandn %xmm0, %xmm31335; X64-SSE2-NEXT:    por %xmm1, %xmm31336; X64-SSE2-NEXT:    movq %xmm3, %rax1337; X64-SSE2-NEXT:    retq1338;1339; X64-SSE42-LABEL: test_reduce_v8i64:1340; X64-SSE42:       ## %bb.0:1341; X64-SSE42-NEXT:    movdqa %xmm0, %xmm41342; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm5 = [9223372036854775808,9223372036854775808]1343; X64-SSE42-NEXT:    movdqa %xmm2, %xmm61344; X64-SSE42-NEXT:    pxor %xmm5, %xmm61345; X64-SSE42-NEXT:    pxor %xmm5, %xmm01346; X64-SSE42-NEXT:    pcmpgtq %xmm6, %xmm01347; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm21348; X64-SSE42-NEXT:    movdqa %xmm3, %xmm41349; X64-SSE42-NEXT:    pxor %xmm5, %xmm41350; X64-SSE42-NEXT:    movdqa %xmm1, %xmm01351; X64-SSE42-NEXT:    pxor %xmm5, %xmm01352; X64-SSE42-NEXT:    pcmpgtq %xmm4, %xmm01353; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm31354; X64-SSE42-NEXT:    movapd %xmm3, %xmm11355; X64-SSE42-NEXT:    xorpd %xmm5, %xmm11356; X64-SSE42-NEXT:    movapd %xmm2, %xmm01357; X64-SSE42-NEXT:    xorpd %xmm5, %xmm01358; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm01359; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm31360; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1361; X64-SSE42-NEXT:    movdqa %xmm3, %xmm01362; X64-SSE42-NEXT:    pxor %xmm5, %xmm01363; X64-SSE42-NEXT:    pxor %xmm1, %xmm51364; X64-SSE42-NEXT:    pcmpgtq %xmm5, %xmm01365; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm11366; X64-SSE42-NEXT:    movq %xmm1, %rax1367; X64-SSE42-NEXT:    retq1368;1369; X64-AVX1-LABEL: test_reduce_v8i64:1370; X64-AVX1:       ## %bb.0:1371; X64-AVX1-NEXT:    vmovddup {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]1372; X64-AVX1-NEXT:    ## xmm2 = mem[0,0]1373; X64-AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm31374; X64-AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm41375; X64-AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm31376; X64-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm31377; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm11378; X64-AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm41379; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm01380; X64-AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm51381; X64-AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm41382; X64-AVX1-NEXT:    vblendvpd %xmm4, %xmm0, %xmm1, %xmm01383; X64-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm11384; X64-AVX1-NEXT:    vxorpd %xmm2, %xmm3, %xmm41385; X64-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm11386; X64-AVX1-NEXT:    vblendvpd %xmm1, %xmm3, %xmm0, %xmm01387; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1388; X64-AVX1-NEXT:    vxorpd %xmm2, %xmm0, %xmm31389; X64-AVX1-NEXT:    vxorpd %xmm2, %xmm1, %xmm21390; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm21391; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01392; X64-AVX1-NEXT:    vmovq %xmm0, %rax1393; X64-AVX1-NEXT:    vzeroupper1394; X64-AVX1-NEXT:    retq1395;1396; X64-AVX2-LABEL: test_reduce_v8i64:1397; X64-AVX2:       ## %bb.0:1398; X64-AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]1399; X64-AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm31400; X64-AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm41401; X64-AVX2-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm31402; X64-AVX2-NEXT:    vblendvpd %ymm3, %ymm0, %ymm1, %ymm01403; X64-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm11404; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm31405; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm41406; X64-AVX2-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm31407; X64-AVX2-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm01408; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1409; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm0, %xmm31410; X64-AVX2-NEXT:    vxorpd %xmm2, %xmm1, %xmm21411; X64-AVX2-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm21412; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm01413; X64-AVX2-NEXT:    vmovq %xmm0, %rax1414; X64-AVX2-NEXT:    vzeroupper1415; X64-AVX2-NEXT:    retq1416;1417; X64-AVX512-LABEL: test_reduce_v8i64:1418; X64-AVX512:       ## %bb.0:1419; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11420; X64-AVX512-NEXT:    vpmaxuq %zmm1, %zmm0, %zmm01421; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11422; X64-AVX512-NEXT:    vpmaxuq %xmm1, %xmm0, %xmm01423; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1424; X64-AVX512-NEXT:    vpmaxuq %xmm1, %xmm0, %xmm01425; X64-AVX512-NEXT:    vmovq %xmm0, %rax1426; X64-AVX512-NEXT:    vzeroupper1427; X64-AVX512-NEXT:    retq1428  %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>1429  %2  = icmp ugt <8 x i64> %a0, %11430  %3  = select <8 x i1> %2, <8 x i64> %a0, <8 x i64> %11431  %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>1432  %5  = icmp ugt <8 x i64> %3, %41433  %6  = select <8 x i1> %5, <8 x i64> %3, <8 x i64> %41434  %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>1435  %8  = icmp ugt <8 x i64> %6, %71436  %9  = select <8 x i1> %8, <8 x i64> %6, <8 x i64> %71437  %10 = extractelement <8 x i64> %9, i32 01438  ret i64 %101439}1440 1441define i32 @test_reduce_v16i32(<16 x i32> %a0) {1442; X86-SSE2-LABEL: test_reduce_v16i32:1443; X86-SSE2:       ## %bb.0:1444; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1445; X86-SSE2-NEXT:    movdqa %xmm2, %xmm61446; X86-SSE2-NEXT:    pxor %xmm4, %xmm61447; X86-SSE2-NEXT:    movdqa %xmm0, %xmm51448; X86-SSE2-NEXT:    pxor %xmm4, %xmm51449; X86-SSE2-NEXT:    pcmpgtd %xmm6, %xmm51450; X86-SSE2-NEXT:    pand %xmm5, %xmm01451; X86-SSE2-NEXT:    pandn %xmm2, %xmm51452; X86-SSE2-NEXT:    por %xmm0, %xmm51453; X86-SSE2-NEXT:    movdqa %xmm3, %xmm21454; X86-SSE2-NEXT:    pxor %xmm4, %xmm21455; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01456; X86-SSE2-NEXT:    pxor %xmm4, %xmm01457; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm01458; X86-SSE2-NEXT:    pand %xmm0, %xmm11459; X86-SSE2-NEXT:    pandn %xmm3, %xmm01460; X86-SSE2-NEXT:    por %xmm1, %xmm01461; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21462; X86-SSE2-NEXT:    pxor %xmm4, %xmm21463; X86-SSE2-NEXT:    movdqa %xmm5, %xmm11464; X86-SSE2-NEXT:    pxor %xmm4, %xmm11465; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm11466; X86-SSE2-NEXT:    pand %xmm1, %xmm51467; X86-SSE2-NEXT:    pandn %xmm0, %xmm11468; X86-SSE2-NEXT:    por %xmm5, %xmm11469; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]1470; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01471; X86-SSE2-NEXT:    pxor %xmm4, %xmm01472; X86-SSE2-NEXT:    movdqa %xmm2, %xmm31473; X86-SSE2-NEXT:    pxor %xmm4, %xmm31474; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm01475; X86-SSE2-NEXT:    pand %xmm0, %xmm11476; X86-SSE2-NEXT:    pandn %xmm2, %xmm01477; X86-SSE2-NEXT:    por %xmm1, %xmm01478; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1479; X86-SSE2-NEXT:    movdqa %xmm0, %xmm21480; X86-SSE2-NEXT:    pxor %xmm4, %xmm21481; X86-SSE2-NEXT:    pxor %xmm1, %xmm41482; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm21483; X86-SSE2-NEXT:    pand %xmm2, %xmm01484; X86-SSE2-NEXT:    pandn %xmm1, %xmm21485; X86-SSE2-NEXT:    por %xmm0, %xmm21486; X86-SSE2-NEXT:    movd %xmm2, %eax1487; X86-SSE2-NEXT:    retl1488;1489; X86-SSE42-LABEL: test_reduce_v16i32:1490; X86-SSE42:       ## %bb.0:1491; X86-SSE42-NEXT:    pmaxud %xmm3, %xmm11492; X86-SSE42-NEXT:    pmaxud %xmm2, %xmm01493; X86-SSE42-NEXT:    pmaxud %xmm1, %xmm01494; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1495; X86-SSE42-NEXT:    pmaxud %xmm0, %xmm11496; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1497; X86-SSE42-NEXT:    pmaxud %xmm1, %xmm01498; X86-SSE42-NEXT:    movd %xmm0, %eax1499; X86-SSE42-NEXT:    retl1500;1501; X86-AVX1-LABEL: test_reduce_v16i32:1502; X86-AVX1:       ## %bb.0:1503; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21504; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31505; X86-AVX1-NEXT:    vpmaxud %xmm2, %xmm3, %xmm21506; X86-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01507; X86-AVX1-NEXT:    vpmaxud %xmm2, %xmm0, %xmm01508; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1509; X86-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01510; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1511; X86-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01512; X86-AVX1-NEXT:    vmovd %xmm0, %eax1513; X86-AVX1-NEXT:    vzeroupper1514; X86-AVX1-NEXT:    retl1515;1516; X86-AVX2-LABEL: test_reduce_v16i32:1517; X86-AVX2:       ## %bb.0:1518; X86-AVX2-NEXT:    vpmaxud %ymm1, %ymm0, %ymm01519; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11520; X86-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01521; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1522; X86-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01523; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1524; X86-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01525; X86-AVX2-NEXT:    vmovd %xmm0, %eax1526; X86-AVX2-NEXT:    vzeroupper1527; X86-AVX2-NEXT:    retl1528;1529; X64-SSE2-LABEL: test_reduce_v16i32:1530; X64-SSE2:       ## %bb.0:1531; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1532; X64-SSE2-NEXT:    movdqa %xmm2, %xmm61533; X64-SSE2-NEXT:    pxor %xmm4, %xmm61534; X64-SSE2-NEXT:    movdqa %xmm0, %xmm51535; X64-SSE2-NEXT:    pxor %xmm4, %xmm51536; X64-SSE2-NEXT:    pcmpgtd %xmm6, %xmm51537; X64-SSE2-NEXT:    pand %xmm5, %xmm01538; X64-SSE2-NEXT:    pandn %xmm2, %xmm51539; X64-SSE2-NEXT:    por %xmm0, %xmm51540; X64-SSE2-NEXT:    movdqa %xmm3, %xmm01541; X64-SSE2-NEXT:    pxor %xmm4, %xmm01542; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21543; X64-SSE2-NEXT:    pxor %xmm4, %xmm21544; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm21545; X64-SSE2-NEXT:    pand %xmm2, %xmm11546; X64-SSE2-NEXT:    pandn %xmm3, %xmm21547; X64-SSE2-NEXT:    por %xmm1, %xmm21548; X64-SSE2-NEXT:    movdqa %xmm2, %xmm01549; X64-SSE2-NEXT:    pxor %xmm4, %xmm01550; X64-SSE2-NEXT:    movdqa %xmm5, %xmm11551; X64-SSE2-NEXT:    pxor %xmm4, %xmm11552; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm11553; X64-SSE2-NEXT:    pand %xmm1, %xmm51554; X64-SSE2-NEXT:    pandn %xmm2, %xmm11555; X64-SSE2-NEXT:    por %xmm5, %xmm11556; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1557; X64-SSE2-NEXT:    movdqa %xmm1, %xmm21558; X64-SSE2-NEXT:    pxor %xmm4, %xmm21559; X64-SSE2-NEXT:    movdqa %xmm0, %xmm31560; X64-SSE2-NEXT:    pxor %xmm4, %xmm31561; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm21562; X64-SSE2-NEXT:    pand %xmm2, %xmm11563; X64-SSE2-NEXT:    pandn %xmm0, %xmm21564; X64-SSE2-NEXT:    por %xmm1, %xmm21565; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1566; X64-SSE2-NEXT:    movdqa %xmm2, %xmm11567; X64-SSE2-NEXT:    pxor %xmm4, %xmm11568; X64-SSE2-NEXT:    pxor %xmm0, %xmm41569; X64-SSE2-NEXT:    pcmpgtd %xmm4, %xmm11570; X64-SSE2-NEXT:    pand %xmm1, %xmm21571; X64-SSE2-NEXT:    pandn %xmm0, %xmm11572; X64-SSE2-NEXT:    por %xmm2, %xmm11573; X64-SSE2-NEXT:    movd %xmm1, %eax1574; X64-SSE2-NEXT:    retq1575;1576; X64-SSE42-LABEL: test_reduce_v16i32:1577; X64-SSE42:       ## %bb.0:1578; X64-SSE42-NEXT:    pmaxud %xmm3, %xmm11579; X64-SSE42-NEXT:    pmaxud %xmm2, %xmm01580; X64-SSE42-NEXT:    pmaxud %xmm1, %xmm01581; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1582; X64-SSE42-NEXT:    pmaxud %xmm0, %xmm11583; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1584; X64-SSE42-NEXT:    pmaxud %xmm1, %xmm01585; X64-SSE42-NEXT:    movd %xmm0, %eax1586; X64-SSE42-NEXT:    retq1587;1588; X64-AVX1-LABEL: test_reduce_v16i32:1589; X64-AVX1:       ## %bb.0:1590; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21591; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31592; X64-AVX1-NEXT:    vpmaxud %xmm2, %xmm3, %xmm21593; X64-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01594; X64-AVX1-NEXT:    vpmaxud %xmm2, %xmm0, %xmm01595; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1596; X64-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01597; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1598; X64-AVX1-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01599; X64-AVX1-NEXT:    vmovd %xmm0, %eax1600; X64-AVX1-NEXT:    vzeroupper1601; X64-AVX1-NEXT:    retq1602;1603; X64-AVX2-LABEL: test_reduce_v16i32:1604; X64-AVX2:       ## %bb.0:1605; X64-AVX2-NEXT:    vpmaxud %ymm1, %ymm0, %ymm01606; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11607; X64-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01608; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1609; X64-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01610; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1611; X64-AVX2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01612; X64-AVX2-NEXT:    vmovd %xmm0, %eax1613; X64-AVX2-NEXT:    vzeroupper1614; X64-AVX2-NEXT:    retq1615;1616; X64-AVX512-LABEL: test_reduce_v16i32:1617; X64-AVX512:       ## %bb.0:1618; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11619; X64-AVX512-NEXT:    vpmaxud %zmm1, %zmm0, %zmm01620; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11621; X64-AVX512-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01622; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1623; X64-AVX512-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01624; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1625; X64-AVX512-NEXT:    vpmaxud %xmm1, %xmm0, %xmm01626; X64-AVX512-NEXT:    vmovd %xmm0, %eax1627; X64-AVX512-NEXT:    vzeroupper1628; X64-AVX512-NEXT:    retq1629  %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>1630  %2  = icmp ugt <16 x i32> %a0, %11631  %3  = select <16 x i1> %2, <16 x i32> %a0, <16 x i32> %11632  %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>1633  %5  = icmp ugt <16 x i32> %3, %41634  %6  = select <16 x i1> %5, <16 x i32> %3, <16 x i32> %41635  %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>1636  %8  = icmp ugt <16 x i32> %6, %71637  %9  = select <16 x i1> %8, <16 x i32> %6, <16 x i32> %71638  %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>1639  %11 = icmp ugt <16 x i32> %9, %101640  %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %101641  %13 = extractelement <16 x i32> %12, i32 01642  ret i32 %131643}1644 1645define i16 @test_reduce_v32i16(<32 x i16> %a0) {1646; X86-SSE2-LABEL: test_reduce_v32i16:1647; X86-SSE2:       ## %bb.0:1648; X86-SSE2-NEXT:    psubusw %xmm0, %xmm21649; X86-SSE2-NEXT:    paddw %xmm0, %xmm21650; X86-SSE2-NEXT:    psubusw %xmm1, %xmm31651; X86-SSE2-NEXT:    paddw %xmm1, %xmm31652; X86-SSE2-NEXT:    psubusw %xmm2, %xmm31653; X86-SSE2-NEXT:    paddw %xmm2, %xmm31654; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]1655; X86-SSE2-NEXT:    psubusw %xmm3, %xmm01656; X86-SSE2-NEXT:    paddw %xmm3, %xmm01657; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1658; X86-SSE2-NEXT:    psubusw %xmm0, %xmm11659; X86-SSE2-NEXT:    paddw %xmm0, %xmm11660; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01661; X86-SSE2-NEXT:    psrld $16, %xmm01662; X86-SSE2-NEXT:    psubusw %xmm1, %xmm01663; X86-SSE2-NEXT:    paddw %xmm1, %xmm01664; X86-SSE2-NEXT:    movd %xmm0, %eax1665; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1666; X86-SSE2-NEXT:    retl1667;1668; X86-SSE42-LABEL: test_reduce_v32i16:1669; X86-SSE42:       ## %bb.0:1670; X86-SSE42-NEXT:    pmaxuw %xmm3, %xmm11671; X86-SSE42-NEXT:    pmaxuw %xmm2, %xmm01672; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm01673; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm11674; X86-SSE42-NEXT:    pxor %xmm0, %xmm11675; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm01676; X86-SSE42-NEXT:    movd %xmm0, %eax1677; X86-SSE42-NEXT:    notl %eax1678; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1679; X86-SSE42-NEXT:    retl1680;1681; X86-AVX1-LABEL: test_reduce_v32i16:1682; X86-AVX1:       ## %bb.0:1683; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21684; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31685; X86-AVX1-NEXT:    vpmaxuw %xmm2, %xmm3, %xmm21686; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm01687; X86-AVX1-NEXT:    vpmaxuw %xmm2, %xmm0, %xmm01688; X86-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11689; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm01690; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01691; X86-AVX1-NEXT:    vmovd %xmm0, %eax1692; X86-AVX1-NEXT:    notl %eax1693; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax1694; X86-AVX1-NEXT:    vzeroupper1695; X86-AVX1-NEXT:    retl1696;1697; X86-AVX2-LABEL: test_reduce_v32i16:1698; X86-AVX2:       ## %bb.0:1699; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm01700; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11701; X86-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm01702; X86-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11703; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm01704; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01705; X86-AVX2-NEXT:    vmovd %xmm0, %eax1706; X86-AVX2-NEXT:    notl %eax1707; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax1708; X86-AVX2-NEXT:    vzeroupper1709; X86-AVX2-NEXT:    retl1710;1711; X64-SSE2-LABEL: test_reduce_v32i16:1712; X64-SSE2:       ## %bb.0:1713; X64-SSE2-NEXT:    psubusw %xmm0, %xmm21714; X64-SSE2-NEXT:    paddw %xmm0, %xmm21715; X64-SSE2-NEXT:    psubusw %xmm1, %xmm31716; X64-SSE2-NEXT:    paddw %xmm1, %xmm31717; X64-SSE2-NEXT:    psubusw %xmm2, %xmm31718; X64-SSE2-NEXT:    paddw %xmm2, %xmm31719; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]1720; X64-SSE2-NEXT:    psubusw %xmm3, %xmm01721; X64-SSE2-NEXT:    paddw %xmm3, %xmm01722; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1723; X64-SSE2-NEXT:    psubusw %xmm0, %xmm11724; X64-SSE2-NEXT:    paddw %xmm0, %xmm11725; X64-SSE2-NEXT:    movdqa %xmm1, %xmm01726; X64-SSE2-NEXT:    psrld $16, %xmm01727; X64-SSE2-NEXT:    psubusw %xmm1, %xmm01728; X64-SSE2-NEXT:    paddw %xmm1, %xmm01729; X64-SSE2-NEXT:    movd %xmm0, %eax1730; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax1731; X64-SSE2-NEXT:    retq1732;1733; X64-SSE42-LABEL: test_reduce_v32i16:1734; X64-SSE42:       ## %bb.0:1735; X64-SSE42-NEXT:    pmaxuw %xmm3, %xmm11736; X64-SSE42-NEXT:    pmaxuw %xmm2, %xmm01737; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm01738; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm11739; X64-SSE42-NEXT:    pxor %xmm0, %xmm11740; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm01741; X64-SSE42-NEXT:    movd %xmm0, %eax1742; X64-SSE42-NEXT:    notl %eax1743; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax1744; X64-SSE42-NEXT:    retq1745;1746; X64-AVX1-LABEL: test_reduce_v32i16:1747; X64-AVX1:       ## %bb.0:1748; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21749; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31750; X64-AVX1-NEXT:    vpmaxuw %xmm2, %xmm3, %xmm21751; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm01752; X64-AVX1-NEXT:    vpmaxuw %xmm2, %xmm0, %xmm01753; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11754; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm01755; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01756; X64-AVX1-NEXT:    vmovd %xmm0, %eax1757; X64-AVX1-NEXT:    notl %eax1758; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax1759; X64-AVX1-NEXT:    vzeroupper1760; X64-AVX1-NEXT:    retq1761;1762; X64-AVX2-LABEL: test_reduce_v32i16:1763; X64-AVX2:       ## %bb.0:1764; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm01765; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11766; X64-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm01767; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11768; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm01769; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01770; X64-AVX2-NEXT:    vmovd %xmm0, %eax1771; X64-AVX2-NEXT:    notl %eax1772; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax1773; X64-AVX2-NEXT:    vzeroupper1774; X64-AVX2-NEXT:    retq1775;1776; X64-AVX512-LABEL: test_reduce_v32i16:1777; X64-AVX512:       ## %bb.0:1778; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11779; X64-AVX512-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm01780; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11781; X64-AVX512-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm01782; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm01783; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01784; X64-AVX512-NEXT:    vmovd %xmm0, %eax1785; X64-AVX512-NEXT:    notl %eax1786; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax1787; X64-AVX512-NEXT:    vzeroupper1788; X64-AVX512-NEXT:    retq1789  %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>1790  %2  = icmp ugt <32 x i16> %a0, %11791  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %11792  %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>1793  %5  = icmp ugt <32 x i16> %3, %41794  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %41795  %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>1796  %8  = icmp ugt <32 x i16> %6, %71797  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %71798  %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>1799  %11 = icmp ugt <32 x i16> %9, %101800  %12 = select <32 x i1> %11, <32 x i16> %9, <32 x i16> %101801  %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>1802  %14 = icmp ugt <32 x i16> %12, %131803  %15 = select <32 x i1> %14, <32 x i16> %12, <32 x i16> %131804  %16 = extractelement <32 x i16> %15, i32 01805  ret i16 %161806}1807 1808define i8 @test_reduce_v64i8(<64 x i8> %a0) {1809; X86-SSE2-LABEL: test_reduce_v64i8:1810; X86-SSE2:       ## %bb.0:1811; X86-SSE2-NEXT:    pmaxub %xmm3, %xmm11812; X86-SSE2-NEXT:    pmaxub %xmm2, %xmm01813; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm01814; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1815; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm11816; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1817; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm01818; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11819; X86-SSE2-NEXT:    psrld $16, %xmm11820; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm11821; X86-SSE2-NEXT:    movdqa %xmm1, %xmm01822; X86-SSE2-NEXT:    psrlw $8, %xmm01823; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm01824; X86-SSE2-NEXT:    movd %xmm0, %eax1825; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1826; X86-SSE2-NEXT:    retl1827;1828; X86-SSE42-LABEL: test_reduce_v64i8:1829; X86-SSE42:       ## %bb.0:1830; X86-SSE42-NEXT:    pmaxub %xmm3, %xmm11831; X86-SSE42-NEXT:    pmaxub %xmm2, %xmm01832; X86-SSE42-NEXT:    pmaxub %xmm1, %xmm01833; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm11834; X86-SSE42-NEXT:    pxor %xmm0, %xmm11835; X86-SSE42-NEXT:    movdqa %xmm1, %xmm01836; X86-SSE42-NEXT:    psrlw $8, %xmm01837; X86-SSE42-NEXT:    pminub %xmm1, %xmm01838; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm01839; X86-SSE42-NEXT:    movd %xmm0, %eax1840; X86-SSE42-NEXT:    notb %al1841; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1842; X86-SSE42-NEXT:    retl1843;1844; X86-AVX1-LABEL: test_reduce_v64i8:1845; X86-AVX1:       ## %bb.0:1846; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21847; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31848; X86-AVX1-NEXT:    vpmaxub %xmm2, %xmm3, %xmm21849; X86-AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01850; X86-AVX1-NEXT:    vpmaxub %xmm2, %xmm0, %xmm01851; X86-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11852; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm01853; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11854; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01855; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm01856; X86-AVX1-NEXT:    vmovd %xmm0, %eax1857; X86-AVX1-NEXT:    notb %al1858; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1859; X86-AVX1-NEXT:    vzeroupper1860; X86-AVX1-NEXT:    retl1861;1862; X86-AVX2-LABEL: test_reduce_v64i8:1863; X86-AVX2:       ## %bb.0:1864; X86-AVX2-NEXT:    vpmaxub %ymm1, %ymm0, %ymm01865; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11866; X86-AVX2-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01867; X86-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11868; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm01869; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11870; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01871; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm01872; X86-AVX2-NEXT:    vmovd %xmm0, %eax1873; X86-AVX2-NEXT:    notb %al1874; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1875; X86-AVX2-NEXT:    vzeroupper1876; X86-AVX2-NEXT:    retl1877;1878; X64-SSE2-LABEL: test_reduce_v64i8:1879; X64-SSE2:       ## %bb.0:1880; X64-SSE2-NEXT:    pmaxub %xmm3, %xmm11881; X64-SSE2-NEXT:    pmaxub %xmm2, %xmm01882; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm01883; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1884; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm11885; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1886; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm01887; X64-SSE2-NEXT:    movdqa %xmm0, %xmm11888; X64-SSE2-NEXT:    psrld $16, %xmm11889; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm11890; X64-SSE2-NEXT:    movdqa %xmm1, %xmm01891; X64-SSE2-NEXT:    psrlw $8, %xmm01892; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm01893; X64-SSE2-NEXT:    movd %xmm0, %eax1894; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax1895; X64-SSE2-NEXT:    retq1896;1897; X64-SSE42-LABEL: test_reduce_v64i8:1898; X64-SSE42:       ## %bb.0:1899; X64-SSE42-NEXT:    pmaxub %xmm3, %xmm11900; X64-SSE42-NEXT:    pmaxub %xmm2, %xmm01901; X64-SSE42-NEXT:    pmaxub %xmm1, %xmm01902; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm11903; X64-SSE42-NEXT:    pxor %xmm0, %xmm11904; X64-SSE42-NEXT:    movdqa %xmm1, %xmm01905; X64-SSE42-NEXT:    psrlw $8, %xmm01906; X64-SSE42-NEXT:    pminub %xmm1, %xmm01907; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm01908; X64-SSE42-NEXT:    movd %xmm0, %eax1909; X64-SSE42-NEXT:    notb %al1910; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax1911; X64-SSE42-NEXT:    retq1912;1913; X64-AVX1-LABEL: test_reduce_v64i8:1914; X64-AVX1:       ## %bb.0:1915; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm21916; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm31917; X64-AVX1-NEXT:    vpmaxub %xmm2, %xmm3, %xmm21918; X64-AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01919; X64-AVX1-NEXT:    vpmaxub %xmm2, %xmm0, %xmm01920; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11921; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm01922; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm11923; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm01924; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm01925; X64-AVX1-NEXT:    vmovd %xmm0, %eax1926; X64-AVX1-NEXT:    notb %al1927; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax1928; X64-AVX1-NEXT:    vzeroupper1929; X64-AVX1-NEXT:    retq1930;1931; X64-AVX2-LABEL: test_reduce_v64i8:1932; X64-AVX2:       ## %bb.0:1933; X64-AVX2-NEXT:    vpmaxub %ymm1, %ymm0, %ymm01934; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm11935; X64-AVX2-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01936; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm11937; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm01938; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm11939; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm01940; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm01941; X64-AVX2-NEXT:    vmovd %xmm0, %eax1942; X64-AVX2-NEXT:    notb %al1943; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax1944; X64-AVX2-NEXT:    vzeroupper1945; X64-AVX2-NEXT:    retq1946;1947; X64-AVX512-LABEL: test_reduce_v64i8:1948; X64-AVX512:       ## %bb.0:1949; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm11950; X64-AVX512-NEXT:    vpmaxub %ymm1, %ymm0, %ymm01951; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm11952; X64-AVX512-NEXT:    vpmaxub %xmm1, %xmm0, %xmm01953; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm01954; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm11955; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm01956; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm01957; X64-AVX512-NEXT:    vmovd %xmm0, %eax1958; X64-AVX512-NEXT:    notb %al1959; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax1960; X64-AVX512-NEXT:    vzeroupper1961; X64-AVX512-NEXT:    retq1962  %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>1963  %2  = icmp ugt <64 x i8> %a0, %11964  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %11965  %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>1966  %5  = icmp ugt <64 x i8> %3, %41967  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %41968  %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>1969  %8  = icmp ugt <64 x i8> %6, %71970  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %71971  %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>1972  %11 = icmp ugt <64 x i8> %9, %101973  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %101974  %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>1975  %14 = icmp ugt <64 x i8> %12, %131976  %15 = select <64 x i1> %14, <64 x i8> %12, <64 x i8> %131977  %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>1978  %17 = icmp ugt <64 x i8> %15, %161979  %18 = select <64 x i1> %17, <64 x i8> %15, <64 x i8> %161980  %19 = extractelement <64 x i8> %18, i32 01981  ret i8 %191982}1983 1984;1985; Partial Vector Reductions1986;1987 1988define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {1989; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:1990; X86-SSE2:       ## %bb.0:1991; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1992; X86-SSE2-NEXT:    psubusw %xmm0, %xmm11993; X86-SSE2-NEXT:    paddw %xmm0, %xmm11994; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1995; X86-SSE2-NEXT:    psubusw %xmm1, %xmm01996; X86-SSE2-NEXT:    paddw %xmm1, %xmm01997; X86-SSE2-NEXT:    movdqa %xmm0, %xmm11998; X86-SSE2-NEXT:    psrld $16, %xmm11999; X86-SSE2-NEXT:    psubusw %xmm0, %xmm12000; X86-SSE2-NEXT:    paddw %xmm0, %xmm12001; X86-SSE2-NEXT:    movd %xmm1, %eax2002; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax2003; X86-SSE2-NEXT:    retl2004;2005; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:2006; X86-SSE42:       ## %bb.0:2007; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12008; X86-SSE42-NEXT:    pxor %xmm0, %xmm12009; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm02010; X86-SSE42-NEXT:    movd %xmm0, %eax2011; X86-SSE42-NEXT:    notl %eax2012; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax2013; X86-SSE42-NEXT:    retl2014;2015; X86-AVX-LABEL: test_reduce_v16i16_v8i16:2016; X86-AVX:       ## %bb.0:2017; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12018; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm02019; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02020; X86-AVX-NEXT:    vmovd %xmm0, %eax2021; X86-AVX-NEXT:    notl %eax2022; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax2023; X86-AVX-NEXT:    vzeroupper2024; X86-AVX-NEXT:    retl2025;2026; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:2027; X64-SSE2:       ## %bb.0:2028; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2029; X64-SSE2-NEXT:    psubusw %xmm0, %xmm12030; X64-SSE2-NEXT:    paddw %xmm0, %xmm12031; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2032; X64-SSE2-NEXT:    psubusw %xmm1, %xmm02033; X64-SSE2-NEXT:    paddw %xmm1, %xmm02034; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12035; X64-SSE2-NEXT:    psrld $16, %xmm12036; X64-SSE2-NEXT:    psubusw %xmm0, %xmm12037; X64-SSE2-NEXT:    paddw %xmm0, %xmm12038; X64-SSE2-NEXT:    movd %xmm1, %eax2039; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax2040; X64-SSE2-NEXT:    retq2041;2042; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:2043; X64-SSE42:       ## %bb.0:2044; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12045; X64-SSE42-NEXT:    pxor %xmm0, %xmm12046; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm02047; X64-SSE42-NEXT:    movd %xmm0, %eax2048; X64-SSE42-NEXT:    notl %eax2049; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax2050; X64-SSE42-NEXT:    retq2051;2052; X64-AVX1-LABEL: test_reduce_v16i16_v8i16:2053; X64-AVX1:       ## %bb.0:2054; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12055; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm02056; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm02057; X64-AVX1-NEXT:    vmovd %xmm0, %eax2058; X64-AVX1-NEXT:    notl %eax2059; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax2060; X64-AVX1-NEXT:    vzeroupper2061; X64-AVX1-NEXT:    retq2062;2063; X64-AVX2-LABEL: test_reduce_v16i16_v8i16:2064; X64-AVX2:       ## %bb.0:2065; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12066; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm02067; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm02068; X64-AVX2-NEXT:    vmovd %xmm0, %eax2069; X64-AVX2-NEXT:    notl %eax2070; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax2071; X64-AVX2-NEXT:    vzeroupper2072; X64-AVX2-NEXT:    retq2073;2074; X64-AVX512-LABEL: test_reduce_v16i16_v8i16:2075; X64-AVX512:       ## %bb.0:2076; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm02077; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02078; X64-AVX512-NEXT:    vmovd %xmm0, %eax2079; X64-AVX512-NEXT:    notl %eax2080; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax2081; X64-AVX512-NEXT:    vzeroupper2082; X64-AVX512-NEXT:    retq2083  %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>2084  %2  = icmp ugt <16 x i16> %a0, %12085  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %12086  %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>2087  %5  = icmp ugt <16 x i16> %3, %42088  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %42089  %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>2090  %8  = icmp ugt <16 x i16> %6, %72091  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %72092  %10 = extractelement <16 x i16> %9, i32 02093  ret i16 %102094}2095 2096define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {2097; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:2098; X86-SSE2:       ## %bb.0:2099; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2100; X86-SSE2-NEXT:    psubusw %xmm0, %xmm12101; X86-SSE2-NEXT:    paddw %xmm0, %xmm12102; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2103; X86-SSE2-NEXT:    psubusw %xmm1, %xmm02104; X86-SSE2-NEXT:    paddw %xmm1, %xmm02105; X86-SSE2-NEXT:    movdqa %xmm0, %xmm12106; X86-SSE2-NEXT:    psrld $16, %xmm12107; X86-SSE2-NEXT:    psubusw %xmm0, %xmm12108; X86-SSE2-NEXT:    paddw %xmm0, %xmm12109; X86-SSE2-NEXT:    movd %xmm1, %eax2110; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax2111; X86-SSE2-NEXT:    retl2112;2113; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:2114; X86-SSE42:       ## %bb.0:2115; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12116; X86-SSE42-NEXT:    pxor %xmm0, %xmm12117; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm02118; X86-SSE42-NEXT:    movd %xmm0, %eax2119; X86-SSE42-NEXT:    notl %eax2120; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax2121; X86-SSE42-NEXT:    retl2122;2123; X86-AVX-LABEL: test_reduce_v32i16_v8i16:2124; X86-AVX:       ## %bb.0:2125; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12126; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm02127; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02128; X86-AVX-NEXT:    vmovd %xmm0, %eax2129; X86-AVX-NEXT:    notl %eax2130; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax2131; X86-AVX-NEXT:    vzeroupper2132; X86-AVX-NEXT:    retl2133;2134; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:2135; X64-SSE2:       ## %bb.0:2136; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2137; X64-SSE2-NEXT:    psubusw %xmm0, %xmm12138; X64-SSE2-NEXT:    paddw %xmm0, %xmm12139; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2140; X64-SSE2-NEXT:    psubusw %xmm1, %xmm02141; X64-SSE2-NEXT:    paddw %xmm1, %xmm02142; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12143; X64-SSE2-NEXT:    psrld $16, %xmm12144; X64-SSE2-NEXT:    psubusw %xmm0, %xmm12145; X64-SSE2-NEXT:    paddw %xmm0, %xmm12146; X64-SSE2-NEXT:    movd %xmm1, %eax2147; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax2148; X64-SSE2-NEXT:    retq2149;2150; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:2151; X64-SSE42:       ## %bb.0:2152; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12153; X64-SSE42-NEXT:    pxor %xmm0, %xmm12154; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm02155; X64-SSE42-NEXT:    movd %xmm0, %eax2156; X64-SSE42-NEXT:    notl %eax2157; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax2158; X64-SSE42-NEXT:    retq2159;2160; X64-AVX1-LABEL: test_reduce_v32i16_v8i16:2161; X64-AVX1:       ## %bb.0:2162; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12163; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm02164; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm02165; X64-AVX1-NEXT:    vmovd %xmm0, %eax2166; X64-AVX1-NEXT:    notl %eax2167; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax2168; X64-AVX1-NEXT:    vzeroupper2169; X64-AVX1-NEXT:    retq2170;2171; X64-AVX2-LABEL: test_reduce_v32i16_v8i16:2172; X64-AVX2:       ## %bb.0:2173; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12174; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm02175; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm02176; X64-AVX2-NEXT:    vmovd %xmm0, %eax2177; X64-AVX2-NEXT:    notl %eax2178; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax2179; X64-AVX2-NEXT:    vzeroupper2180; X64-AVX2-NEXT:    retq2181;2182; X64-AVX512-LABEL: test_reduce_v32i16_v8i16:2183; X64-AVX512:       ## %bb.0:2184; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm02185; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02186; X64-AVX512-NEXT:    vmovd %xmm0, %eax2187; X64-AVX512-NEXT:    notl %eax2188; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax2189; X64-AVX512-NEXT:    vzeroupper2190; X64-AVX512-NEXT:    retq2191  %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>2192  %2  = icmp ugt <32 x i16> %a0, %12193  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %12194  %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>2195  %5  = icmp ugt <32 x i16> %3, %42196  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %42197  %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>2198  %8  = icmp ugt <32 x i16> %6, %72199  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %72200  %10 = extractelement <32 x i16> %9, i32 02201  ret i16 %102202}2203 2204define i8 @test_reduce_v32i8_v16i8(<32 x i8> %a0) {2205; X86-SSE2-LABEL: test_reduce_v32i8_v16i8:2206; X86-SSE2:       ## %bb.0:2207; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2208; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm12209; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2210; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm02211; X86-SSE2-NEXT:    movdqa %xmm0, %xmm12212; X86-SSE2-NEXT:    psrld $16, %xmm12213; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm12214; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02215; X86-SSE2-NEXT:    psrlw $8, %xmm02216; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm02217; X86-SSE2-NEXT:    movd %xmm0, %eax2218; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2219; X86-SSE2-NEXT:    retl2220;2221; X86-SSE42-LABEL: test_reduce_v32i8_v16i8:2222; X86-SSE42:       ## %bb.0:2223; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12224; X86-SSE42-NEXT:    pxor %xmm0, %xmm12225; X86-SSE42-NEXT:    movdqa %xmm1, %xmm02226; X86-SSE42-NEXT:    psrlw $8, %xmm02227; X86-SSE42-NEXT:    pminub %xmm1, %xmm02228; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm02229; X86-SSE42-NEXT:    movd %xmm0, %eax2230; X86-SSE42-NEXT:    notb %al2231; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2232; X86-SSE42-NEXT:    retl2233;2234; X86-AVX-LABEL: test_reduce_v32i8_v16i8:2235; X86-AVX:       ## %bb.0:2236; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12237; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm02238; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12239; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02240; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02241; X86-AVX-NEXT:    vmovd %xmm0, %eax2242; X86-AVX-NEXT:    notb %al2243; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax2244; X86-AVX-NEXT:    vzeroupper2245; X86-AVX-NEXT:    retl2246;2247; X64-SSE2-LABEL: test_reduce_v32i8_v16i8:2248; X64-SSE2:       ## %bb.0:2249; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2250; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm12251; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2252; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm02253; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12254; X64-SSE2-NEXT:    psrld $16, %xmm12255; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm12256; X64-SSE2-NEXT:    movdqa %xmm1, %xmm02257; X64-SSE2-NEXT:    psrlw $8, %xmm02258; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm02259; X64-SSE2-NEXT:    movd %xmm0, %eax2260; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2261; X64-SSE2-NEXT:    retq2262;2263; X64-SSE42-LABEL: test_reduce_v32i8_v16i8:2264; X64-SSE42:       ## %bb.0:2265; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12266; X64-SSE42-NEXT:    pxor %xmm0, %xmm12267; X64-SSE42-NEXT:    movdqa %xmm1, %xmm02268; X64-SSE42-NEXT:    psrlw $8, %xmm02269; X64-SSE42-NEXT:    pminub %xmm1, %xmm02270; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm02271; X64-SSE42-NEXT:    movd %xmm0, %eax2272; X64-SSE42-NEXT:    notb %al2273; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2274; X64-SSE42-NEXT:    retq2275;2276; X64-AVX1-LABEL: test_reduce_v32i8_v16i8:2277; X64-AVX1:       ## %bb.0:2278; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12279; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm02280; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm12281; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm02282; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm02283; X64-AVX1-NEXT:    vmovd %xmm0, %eax2284; X64-AVX1-NEXT:    notb %al2285; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax2286; X64-AVX1-NEXT:    vzeroupper2287; X64-AVX1-NEXT:    retq2288;2289; X64-AVX2-LABEL: test_reduce_v32i8_v16i8:2290; X64-AVX2:       ## %bb.0:2291; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12292; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm02293; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm12294; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm02295; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm02296; X64-AVX2-NEXT:    vmovd %xmm0, %eax2297; X64-AVX2-NEXT:    notb %al2298; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax2299; X64-AVX2-NEXT:    vzeroupper2300; X64-AVX2-NEXT:    retq2301;2302; X64-AVX512-LABEL: test_reduce_v32i8_v16i8:2303; X64-AVX512:       ## %bb.0:2304; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm02305; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm12306; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm02307; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02308; X64-AVX512-NEXT:    vmovd %xmm0, %eax2309; X64-AVX512-NEXT:    notb %al2310; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax2311; X64-AVX512-NEXT:    vzeroupper2312; X64-AVX512-NEXT:    retq2313  %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>2314  %2  = icmp ugt <32 x i8> %a0, %12315  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %12316  %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>2317  %5  = icmp ugt <32 x i8> %3, %42318  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %42319  %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>2320  %8  = icmp ugt <32 x i8> %6, %72321  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %72322  %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>2323  %11 = icmp ugt <32 x i8> %9, %102324  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %102325  %13 = extractelement <32 x i8> %12, i32 02326  ret i8 %132327}2328 2329define i8 @test_reduce_v64i8_v16i8(<64 x i8> %a0) {2330; X86-SSE2-LABEL: test_reduce_v64i8_v16i8:2331; X86-SSE2:       ## %bb.0:2332; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2333; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm12334; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2335; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm02336; X86-SSE2-NEXT:    movdqa %xmm0, %xmm12337; X86-SSE2-NEXT:    psrld $16, %xmm12338; X86-SSE2-NEXT:    pmaxub %xmm0, %xmm12339; X86-SSE2-NEXT:    movdqa %xmm1, %xmm02340; X86-SSE2-NEXT:    psrlw $8, %xmm02341; X86-SSE2-NEXT:    pmaxub %xmm1, %xmm02342; X86-SSE2-NEXT:    movd %xmm0, %eax2343; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2344; X86-SSE2-NEXT:    retl2345;2346; X86-SSE42-LABEL: test_reduce_v64i8_v16i8:2347; X86-SSE42:       ## %bb.0:2348; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12349; X86-SSE42-NEXT:    pxor %xmm0, %xmm12350; X86-SSE42-NEXT:    movdqa %xmm1, %xmm02351; X86-SSE42-NEXT:    psrlw $8, %xmm02352; X86-SSE42-NEXT:    pminub %xmm1, %xmm02353; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm02354; X86-SSE42-NEXT:    movd %xmm0, %eax2355; X86-SSE42-NEXT:    notb %al2356; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2357; X86-SSE42-NEXT:    retl2358;2359; X86-AVX-LABEL: test_reduce_v64i8_v16i8:2360; X86-AVX:       ## %bb.0:2361; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12362; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm02363; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm12364; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm02365; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm02366; X86-AVX-NEXT:    vmovd %xmm0, %eax2367; X86-AVX-NEXT:    notb %al2368; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax2369; X86-AVX-NEXT:    vzeroupper2370; X86-AVX-NEXT:    retl2371;2372; X64-SSE2-LABEL: test_reduce_v64i8_v16i8:2373; X64-SSE2:       ## %bb.0:2374; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2375; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm12376; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]2377; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm02378; X64-SSE2-NEXT:    movdqa %xmm0, %xmm12379; X64-SSE2-NEXT:    psrld $16, %xmm12380; X64-SSE2-NEXT:    pmaxub %xmm0, %xmm12381; X64-SSE2-NEXT:    movdqa %xmm1, %xmm02382; X64-SSE2-NEXT:    psrlw $8, %xmm02383; X64-SSE2-NEXT:    pmaxub %xmm1, %xmm02384; X64-SSE2-NEXT:    movd %xmm0, %eax2385; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax2386; X64-SSE2-NEXT:    retq2387;2388; X64-SSE42-LABEL: test_reduce_v64i8_v16i8:2389; X64-SSE42:       ## %bb.0:2390; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm12391; X64-SSE42-NEXT:    pxor %xmm0, %xmm12392; X64-SSE42-NEXT:    movdqa %xmm1, %xmm02393; X64-SSE42-NEXT:    psrlw $8, %xmm02394; X64-SSE42-NEXT:    pminub %xmm1, %xmm02395; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm02396; X64-SSE42-NEXT:    movd %xmm0, %eax2397; X64-SSE42-NEXT:    notb %al2398; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax2399; X64-SSE42-NEXT:    retq2400;2401; X64-AVX1-LABEL: test_reduce_v64i8_v16i8:2402; X64-AVX1:       ## %bb.0:2403; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12404; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm02405; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm12406; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm02407; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm02408; X64-AVX1-NEXT:    vmovd %xmm0, %eax2409; X64-AVX1-NEXT:    notb %al2410; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax2411; X64-AVX1-NEXT:    vzeroupper2412; X64-AVX1-NEXT:    retq2413;2414; X64-AVX2-LABEL: test_reduce_v64i8_v16i8:2415; X64-AVX2:       ## %bb.0:2416; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm12417; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm02418; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm12419; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm02420; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm02421; X64-AVX2-NEXT:    vmovd %xmm0, %eax2422; X64-AVX2-NEXT:    notb %al2423; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax2424; X64-AVX2-NEXT:    vzeroupper2425; X64-AVX2-NEXT:    retq2426;2427; X64-AVX512-LABEL: test_reduce_v64i8_v16i8:2428; X64-AVX512:       ## %bb.0:2429; X64-AVX512-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm02430; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm12431; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm02432; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm02433; X64-AVX512-NEXT:    vmovd %xmm0, %eax2434; X64-AVX512-NEXT:    notb %al2435; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax2436; X64-AVX512-NEXT:    vzeroupper2437; X64-AVX512-NEXT:    retq2438  %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>2439  %2  = icmp ugt <64 x i8> %a0, %12440  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %12441  %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>2442  %5  = icmp ugt <64 x i8> %3, %42443  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %42444  %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>2445  %8  = icmp ugt <64 x i8> %6, %72446  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %72447  %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>2448  %11 = icmp ugt <64 x i8> %9, %102449  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %102450  %13 = extractelement <64 x i8> %12, i32 02451  ret i8 %132452}2453