2302 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-AVX1OR2,X64-AVX19; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx2 | FileCheck %s --check-prefixes=X64-AVX,X64-AVX1OR2,X64-AVX210; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512dq,+avx512vl | FileCheck %s --check-prefixes=X64-AVX,X64-AVX51211 12;13; 128-bit Vectors14;15 16define i64 @test_reduce_v2i64(<2 x i64> %a0) {17; X86-SSE2-LABEL: test_reduce_v2i64:18; X86-SSE2: ## %bb.0:19; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]20; X86-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0]21; X86-SSE2-NEXT: movdqa %xmm0, %xmm322; X86-SSE2-NEXT: pxor %xmm2, %xmm323; X86-SSE2-NEXT: pxor %xmm1, %xmm224; X86-SSE2-NEXT: movdqa %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: pcmpgtq %xmm2, %xmm045; X86-SSE42-NEXT: blendvpd %xmm0, %xmm1, %xmm246; X86-SSE42-NEXT: movd %xmm2, %eax47; X86-SSE42-NEXT: pextrd $1, %xmm2, %edx48; X86-SSE42-NEXT: retl49;50; X86-AVX-LABEL: test_reduce_v2i64:51; X86-AVX: ## %bb.0:52; X86-AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]53; X86-AVX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm254; X86-AVX-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm055; X86-AVX-NEXT: vmovd %xmm0, %eax56; X86-AVX-NEXT: vpextrd $1, %xmm0, %edx57; X86-AVX-NEXT: retl58;59; X64-SSE2-LABEL: test_reduce_v2i64:60; X64-SSE2: ## %bb.0:61; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]62; X64-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]63; X64-SSE2-NEXT: movdqa %xmm0, %xmm364; X64-SSE2-NEXT: pxor %xmm2, %xmm365; X64-SSE2-NEXT: pxor %xmm1, %xmm266; X64-SSE2-NEXT: movdqa %xmm3, %xmm467; X64-SSE2-NEXT: pcmpgtd %xmm2, %xmm468; X64-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]69; X64-SSE2-NEXT: pcmpeqd %xmm3, %xmm270; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]71; X64-SSE2-NEXT: pand %xmm5, %xmm272; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]73; X64-SSE2-NEXT: por %xmm2, %xmm374; X64-SSE2-NEXT: pand %xmm3, %xmm075; X64-SSE2-NEXT: pandn %xmm1, %xmm376; X64-SSE2-NEXT: por %xmm0, %xmm377; X64-SSE2-NEXT: movq %xmm3, %rax78; X64-SSE2-NEXT: retq79;80; X64-SSE42-LABEL: test_reduce_v2i64:81; X64-SSE42: ## %bb.0:82; X64-SSE42-NEXT: movdqa %xmm0, %xmm183; X64-SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]84; X64-SSE42-NEXT: pcmpgtq %xmm2, %xmm085; X64-SSE42-NEXT: blendvpd %xmm0, %xmm1, %xmm286; X64-SSE42-NEXT: movq %xmm2, %rax87; X64-SSE42-NEXT: retq88;89; X64-AVX1OR2-LABEL: test_reduce_v2i64:90; X64-AVX1OR2: ## %bb.0:91; X64-AVX1OR2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]92; X64-AVX1OR2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm293; X64-AVX1OR2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm094; X64-AVX1OR2-NEXT: vmovq %xmm0, %rax95; X64-AVX1OR2-NEXT: retq96;97; X64-AVX512-LABEL: test_reduce_v2i64:98; X64-AVX512: ## %bb.0:99; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]100; X64-AVX512-NEXT: vpmaxsq %xmm1, %xmm0, %xmm0101; X64-AVX512-NEXT: vmovq %xmm0, %rax102; X64-AVX512-NEXT: retq103 %1 = shufflevector <2 x i64> %a0, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>104 %2 = icmp sgt <2 x i64> %a0, %1105 %3 = select <2 x i1> %2, <2 x i64> %a0, <2 x i64> %1106 %4 = extractelement <2 x i64> %3, i32 0107 ret i64 %4108}109 110define i32 @test_reduce_v4i32(<4 x i32> %a0) {111; X86-SSE2-LABEL: test_reduce_v4i32:112; X86-SSE2: ## %bb.0:113; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]114; X86-SSE2-NEXT: movdqa %xmm0, %xmm2115; X86-SSE2-NEXT: pcmpgtd %xmm1, %xmm2116; X86-SSE2-NEXT: pand %xmm2, %xmm0117; X86-SSE2-NEXT: pandn %xmm1, %xmm2118; X86-SSE2-NEXT: por %xmm0, %xmm2119; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]120; X86-SSE2-NEXT: movdqa %xmm2, %xmm1121; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm1122; X86-SSE2-NEXT: pand %xmm1, %xmm2123; X86-SSE2-NEXT: pandn %xmm0, %xmm1124; X86-SSE2-NEXT: por %xmm2, %xmm1125; X86-SSE2-NEXT: movd %xmm1, %eax126; X86-SSE2-NEXT: retl127;128; X86-SSE42-LABEL: test_reduce_v4i32:129; X86-SSE42: ## %bb.0:130; X86-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]131; X86-SSE42-NEXT: pmaxsd %xmm0, %xmm1132; X86-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]133; X86-SSE42-NEXT: pmaxsd %xmm1, %xmm0134; X86-SSE42-NEXT: movd %xmm0, %eax135; X86-SSE42-NEXT: retl136;137; X86-AVX-LABEL: test_reduce_v4i32:138; X86-AVX: ## %bb.0:139; X86-AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]140; X86-AVX-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0141; X86-AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]142; X86-AVX-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0143; X86-AVX-NEXT: vmovd %xmm0, %eax144; X86-AVX-NEXT: retl145;146; X64-SSE2-LABEL: test_reduce_v4i32:147; X64-SSE2: ## %bb.0:148; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]149; X64-SSE2-NEXT: movdqa %xmm0, %xmm2150; X64-SSE2-NEXT: pcmpgtd %xmm1, %xmm2151; X64-SSE2-NEXT: pand %xmm2, %xmm0152; X64-SSE2-NEXT: pandn %xmm1, %xmm2153; X64-SSE2-NEXT: por %xmm0, %xmm2154; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]155; X64-SSE2-NEXT: movdqa %xmm2, %xmm1156; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm1157; X64-SSE2-NEXT: pand %xmm1, %xmm2158; X64-SSE2-NEXT: pandn %xmm0, %xmm1159; X64-SSE2-NEXT: por %xmm2, %xmm1160; X64-SSE2-NEXT: movd %xmm1, %eax161; X64-SSE2-NEXT: retq162;163; X64-SSE42-LABEL: test_reduce_v4i32:164; X64-SSE42: ## %bb.0:165; X64-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]166; X64-SSE42-NEXT: pmaxsd %xmm0, %xmm1167; X64-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]168; X64-SSE42-NEXT: pmaxsd %xmm1, %xmm0169; X64-SSE42-NEXT: movd %xmm0, %eax170; X64-SSE42-NEXT: retq171;172; X64-AVX-LABEL: test_reduce_v4i32:173; X64-AVX: ## %bb.0:174; X64-AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]175; X64-AVX-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0176; X64-AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]177; X64-AVX-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0178; X64-AVX-NEXT: vmovd %xmm0, %eax179; X64-AVX-NEXT: retq180 %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>181 %2 = icmp sgt <4 x i32> %a0, %1182 %3 = select <4 x i1> %2, <4 x i32> %a0, <4 x i32> %1183 %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>184 %5 = icmp sgt <4 x i32> %3, %4185 %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4186 %7 = extractelement <4 x i32> %6, i32 0187 ret i32 %7188}189 190define i16 @test_reduce_v8i16(<8 x i16> %a0) {191; X86-SSE2-LABEL: test_reduce_v8i16:192; X86-SSE2: ## %bb.0:193; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]194; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm1195; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]196; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm0197; X86-SSE2-NEXT: movdqa %xmm0, %xmm1198; X86-SSE2-NEXT: psrld $16, %xmm1199; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm1200; X86-SSE2-NEXT: movd %xmm1, %eax201; X86-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax202; X86-SSE2-NEXT: retl203;204; X86-SSE42-LABEL: test_reduce_v8i16:205; X86-SSE42: ## %bb.0:206; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0207; X86-SSE42-NEXT: phminposuw %xmm0, %xmm0208; X86-SSE42-NEXT: movd %xmm0, %eax209; X86-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF210; X86-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax211; X86-SSE42-NEXT: retl212;213; X86-AVX-LABEL: test_reduce_v8i16:214; X86-AVX: ## %bb.0:215; X86-AVX-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0216; X86-AVX-NEXT: vphminposuw %xmm0, %xmm0217; X86-AVX-NEXT: vmovd %xmm0, %eax218; X86-AVX-NEXT: xorl $32767, %eax ## imm = 0x7FFF219; X86-AVX-NEXT: ## kill: def $ax killed $ax killed $eax220; X86-AVX-NEXT: retl221;222; X64-SSE2-LABEL: test_reduce_v8i16:223; X64-SSE2: ## %bb.0:224; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]225; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm1226; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]227; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm0228; X64-SSE2-NEXT: movdqa %xmm0, %xmm1229; X64-SSE2-NEXT: psrld $16, %xmm1230; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm1231; X64-SSE2-NEXT: movd %xmm1, %eax232; X64-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax233; X64-SSE2-NEXT: retq234;235; X64-SSE42-LABEL: test_reduce_v8i16:236; X64-SSE42: ## %bb.0:237; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0238; X64-SSE42-NEXT: phminposuw %xmm0, %xmm0239; X64-SSE42-NEXT: movd %xmm0, %eax240; X64-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF241; X64-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax242; X64-SSE42-NEXT: retq243;244; X64-AVX1OR2-LABEL: test_reduce_v8i16:245; X64-AVX1OR2: ## %bb.0:246; X64-AVX1OR2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0247; X64-AVX1OR2-NEXT: vphminposuw %xmm0, %xmm0248; X64-AVX1OR2-NEXT: vmovd %xmm0, %eax249; X64-AVX1OR2-NEXT: xorl $32767, %eax ## imm = 0x7FFF250; X64-AVX1OR2-NEXT: ## kill: def $ax killed $ax killed $eax251; X64-AVX1OR2-NEXT: retq252;253; X64-AVX512-LABEL: test_reduce_v8i16:254; X64-AVX512: ## %bb.0:255; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0256; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm0257; X64-AVX512-NEXT: vmovd %xmm0, %eax258; X64-AVX512-NEXT: xorl $32767, %eax ## imm = 0x7FFF259; X64-AVX512-NEXT: ## kill: def $ax killed $ax killed $eax260; X64-AVX512-NEXT: retq261 %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>262 %2 = icmp sgt <8 x i16> %a0, %1263 %3 = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1264 %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>265 %5 = icmp sgt <8 x i16> %3, %4266 %6 = select <8 x i1> %5, <8 x i16> %3, <8 x i16> %4267 %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>268 %8 = icmp sgt <8 x i16> %6, %7269 %9 = select <8 x i1> %8, <8 x i16> %6, <8 x i16> %7270 %10 = extractelement <8 x i16> %9, i32 0271 ret i16 %10272}273 274define i8 @test_reduce_v16i8(<16 x i8> %a0) {275; X86-SSE2-LABEL: test_reduce_v16i8:276; X86-SSE2: ## %bb.0:277; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]278; X86-SSE2-NEXT: movdqa %xmm0, %xmm2279; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm2280; X86-SSE2-NEXT: pand %xmm2, %xmm0281; X86-SSE2-NEXT: pandn %xmm1, %xmm2282; X86-SSE2-NEXT: por %xmm0, %xmm2283; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]284; X86-SSE2-NEXT: movdqa %xmm2, %xmm0285; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm0286; X86-SSE2-NEXT: pand %xmm0, %xmm2287; X86-SSE2-NEXT: pandn %xmm1, %xmm0288; X86-SSE2-NEXT: por %xmm2, %xmm0289; X86-SSE2-NEXT: movdqa %xmm0, %xmm2290; X86-SSE2-NEXT: psrld $16, %xmm2291; X86-SSE2-NEXT: movdqa %xmm0, %xmm1292; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm1293; X86-SSE2-NEXT: pand %xmm1, %xmm0294; X86-SSE2-NEXT: pandn %xmm2, %xmm1295; X86-SSE2-NEXT: por %xmm0, %xmm1296; X86-SSE2-NEXT: movdqa %xmm1, %xmm0297; X86-SSE2-NEXT: psrlw $8, %xmm0298; X86-SSE2-NEXT: movdqa %xmm1, %xmm2299; X86-SSE2-NEXT: pcmpgtb %xmm0, %xmm2300; X86-SSE2-NEXT: pand %xmm2, %xmm1301; X86-SSE2-NEXT: pandn %xmm0, %xmm2302; X86-SSE2-NEXT: por %xmm1, %xmm2303; X86-SSE2-NEXT: movd %xmm2, %eax304; X86-SSE2-NEXT: ## kill: def $al killed $al killed $eax305; X86-SSE2-NEXT: retl306;307; X86-SSE42-LABEL: test_reduce_v16i8:308; X86-SSE42: ## %bb.0:309; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0310; X86-SSE42-NEXT: movdqa %xmm0, %xmm1311; X86-SSE42-NEXT: psrlw $8, %xmm1312; X86-SSE42-NEXT: pminub %xmm0, %xmm1313; X86-SSE42-NEXT: phminposuw %xmm1, %xmm0314; X86-SSE42-NEXT: movd %xmm0, %eax315; X86-SSE42-NEXT: xorb $127, %al316; X86-SSE42-NEXT: ## kill: def $al killed $al killed $eax317; X86-SSE42-NEXT: retl318;319; X86-AVX-LABEL: test_reduce_v16i8:320; X86-AVX: ## %bb.0:321; X86-AVX-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0322; X86-AVX-NEXT: vpsrlw $8, %xmm0, %xmm1323; X86-AVX-NEXT: vpminub %xmm1, %xmm0, %xmm0324; X86-AVX-NEXT: vphminposuw %xmm0, %xmm0325; X86-AVX-NEXT: vmovd %xmm0, %eax326; X86-AVX-NEXT: xorb $127, %al327; X86-AVX-NEXT: ## kill: def $al killed $al killed $eax328; X86-AVX-NEXT: retl329;330; X64-SSE2-LABEL: test_reduce_v16i8:331; X64-SSE2: ## %bb.0:332; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]333; X64-SSE2-NEXT: movdqa %xmm0, %xmm2334; X64-SSE2-NEXT: pcmpgtb %xmm1, %xmm2335; X64-SSE2-NEXT: pand %xmm2, %xmm0336; X64-SSE2-NEXT: pandn %xmm1, %xmm2337; X64-SSE2-NEXT: por %xmm0, %xmm2338; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]339; X64-SSE2-NEXT: movdqa %xmm2, %xmm1340; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm1341; X64-SSE2-NEXT: pand %xmm1, %xmm2342; X64-SSE2-NEXT: pandn %xmm0, %xmm1343; X64-SSE2-NEXT: por %xmm2, %xmm1344; X64-SSE2-NEXT: movdqa %xmm1, %xmm0345; X64-SSE2-NEXT: psrld $16, %xmm0346; X64-SSE2-NEXT: movdqa %xmm1, %xmm2347; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm2348; X64-SSE2-NEXT: pand %xmm2, %xmm1349; X64-SSE2-NEXT: pandn %xmm0, %xmm2350; X64-SSE2-NEXT: por %xmm1, %xmm2351; X64-SSE2-NEXT: movdqa %xmm2, %xmm0352; X64-SSE2-NEXT: psrlw $8, %xmm0353; X64-SSE2-NEXT: movdqa %xmm2, %xmm1354; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm1355; X64-SSE2-NEXT: pand %xmm1, %xmm2356; X64-SSE2-NEXT: pandn %xmm0, %xmm1357; X64-SSE2-NEXT: por %xmm2, %xmm1358; X64-SSE2-NEXT: movd %xmm1, %eax359; X64-SSE2-NEXT: ## kill: def $al killed $al killed $eax360; X64-SSE2-NEXT: retq361;362; X64-SSE42-LABEL: test_reduce_v16i8:363; X64-SSE42: ## %bb.0:364; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0365; X64-SSE42-NEXT: movdqa %xmm0, %xmm1366; X64-SSE42-NEXT: psrlw $8, %xmm1367; X64-SSE42-NEXT: pminub %xmm0, %xmm1368; X64-SSE42-NEXT: phminposuw %xmm1, %xmm0369; X64-SSE42-NEXT: movd %xmm0, %eax370; X64-SSE42-NEXT: xorb $127, %al371; X64-SSE42-NEXT: ## kill: def $al killed $al killed $eax372; X64-SSE42-NEXT: retq373;374; X64-AVX1OR2-LABEL: test_reduce_v16i8:375; X64-AVX1OR2: ## %bb.0:376; X64-AVX1OR2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0377; X64-AVX1OR2-NEXT: vpsrlw $8, %xmm0, %xmm1378; X64-AVX1OR2-NEXT: vpminub %xmm1, %xmm0, %xmm0379; X64-AVX1OR2-NEXT: vphminposuw %xmm0, %xmm0380; X64-AVX1OR2-NEXT: vmovd %xmm0, %eax381; X64-AVX1OR2-NEXT: xorb $127, %al382; X64-AVX1OR2-NEXT: ## kill: def $al killed $al killed $eax383; X64-AVX1OR2-NEXT: retq384;385; X64-AVX512-LABEL: test_reduce_v16i8:386; X64-AVX512: ## %bb.0:387; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0388; X64-AVX512-NEXT: vpsrlw $8, %xmm0, %xmm1389; X64-AVX512-NEXT: vpminub %xmm1, %xmm0, %xmm0390; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm0391; X64-AVX512-NEXT: vmovd %xmm0, %eax392; X64-AVX512-NEXT: xorb $127, %al393; X64-AVX512-NEXT: ## kill: def $al killed $al killed $eax394; X64-AVX512-NEXT: retq395 %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>396 %2 = icmp sgt <16 x i8> %a0, %1397 %3 = select <16 x i1> %2, <16 x i8> %a0, <16 x i8> %1398 %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>399 %5 = icmp sgt <16 x i8> %3, %4400 %6 = select <16 x i1> %5, <16 x i8> %3, <16 x i8> %4401 %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>402 %8 = icmp sgt <16 x i8> %6, %7403 %9 = select <16 x i1> %8, <16 x i8> %6, <16 x i8> %7404 %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>405 %11 = icmp sgt <16 x i8> %9, %10406 %12 = select <16 x i1> %11, <16 x i8> %9, <16 x i8> %10407 %13 = extractelement <16 x i8> %12, i32 0408 ret i8 %13409}410 411;412; 256-bit Vectors413;414 415define i64 @test_reduce_v4i64(<4 x i64> %a0) {416; X86-SSE2-LABEL: test_reduce_v4i64:417; X86-SSE2: ## %bb.0:418; X86-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0]419; X86-SSE2-NEXT: movdqa %xmm1, %xmm3420; X86-SSE2-NEXT: pxor %xmm2, %xmm3421; X86-SSE2-NEXT: movdqa %xmm0, %xmm4422; X86-SSE2-NEXT: pxor %xmm2, %xmm4423; X86-SSE2-NEXT: movdqa %xmm4, %xmm5424; X86-SSE2-NEXT: pcmpgtd %xmm3, %xmm5425; X86-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]426; X86-SSE2-NEXT: pcmpeqd %xmm3, %xmm4427; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]428; X86-SSE2-NEXT: pand %xmm6, %xmm4429; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]430; X86-SSE2-NEXT: por %xmm4, %xmm3431; X86-SSE2-NEXT: pand %xmm3, %xmm0432; X86-SSE2-NEXT: pandn %xmm1, %xmm3433; X86-SSE2-NEXT: por %xmm0, %xmm3434; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]435; X86-SSE2-NEXT: movdqa %xmm3, %xmm1436; X86-SSE2-NEXT: pxor %xmm2, %xmm1437; X86-SSE2-NEXT: pxor %xmm0, %xmm2438; X86-SSE2-NEXT: movdqa %xmm1, %xmm4439; X86-SSE2-NEXT: pcmpgtd %xmm2, %xmm4440; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]441; X86-SSE2-NEXT: pcmpeqd %xmm1, %xmm2442; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]443; X86-SSE2-NEXT: pand %xmm5, %xmm1444; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]445; X86-SSE2-NEXT: por %xmm1, %xmm2446; X86-SSE2-NEXT: pand %xmm2, %xmm3447; X86-SSE2-NEXT: pandn %xmm0, %xmm2448; X86-SSE2-NEXT: por %xmm3, %xmm2449; X86-SSE2-NEXT: movd %xmm2, %eax450; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]451; X86-SSE2-NEXT: movd %xmm0, %edx452; X86-SSE2-NEXT: retl453;454; X86-SSE42-LABEL: test_reduce_v4i64:455; X86-SSE42: ## %bb.0:456; X86-SSE42-NEXT: movdqa %xmm0, %xmm2457; X86-SSE42-NEXT: pcmpgtq %xmm1, %xmm0458; X86-SSE42-NEXT: blendvpd %xmm0, %xmm2, %xmm1459; X86-SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]460; X86-SSE42-NEXT: movdqa %xmm1, %xmm0461; X86-SSE42-NEXT: pcmpgtq %xmm2, %xmm0462; X86-SSE42-NEXT: blendvpd %xmm0, %xmm1, %xmm2463; X86-SSE42-NEXT: movd %xmm2, %eax464; X86-SSE42-NEXT: pextrd $1, %xmm2, %edx465; X86-SSE42-NEXT: retl466;467; X86-AVX1-LABEL: test_reduce_v4i64:468; X86-AVX1: ## %bb.0:469; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1470; X86-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2471; X86-AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0472; X86-AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]473; X86-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2474; X86-AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0475; X86-AVX1-NEXT: vmovd %xmm0, %eax476; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx477; X86-AVX1-NEXT: vzeroupper478; X86-AVX1-NEXT: retl479;480; X86-AVX2-LABEL: test_reduce_v4i64:481; X86-AVX2: ## %bb.0:482; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1483; X86-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2484; X86-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0485; X86-AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]486; X86-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2487; X86-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0488; X86-AVX2-NEXT: vmovd %xmm0, %eax489; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx490; X86-AVX2-NEXT: vzeroupper491; X86-AVX2-NEXT: retl492;493; X64-SSE2-LABEL: test_reduce_v4i64:494; X64-SSE2: ## %bb.0:495; X64-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]496; X64-SSE2-NEXT: movdqa %xmm1, %xmm3497; X64-SSE2-NEXT: pxor %xmm2, %xmm3498; X64-SSE2-NEXT: movdqa %xmm0, %xmm4499; X64-SSE2-NEXT: pxor %xmm2, %xmm4500; X64-SSE2-NEXT: movdqa %xmm4, %xmm5501; X64-SSE2-NEXT: pcmpgtd %xmm3, %xmm5502; X64-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]503; X64-SSE2-NEXT: pcmpeqd %xmm3, %xmm4504; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]505; X64-SSE2-NEXT: pand %xmm6, %xmm3506; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]507; X64-SSE2-NEXT: por %xmm3, %xmm4508; X64-SSE2-NEXT: pand %xmm4, %xmm0509; X64-SSE2-NEXT: pandn %xmm1, %xmm4510; X64-SSE2-NEXT: por %xmm0, %xmm4511; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]512; X64-SSE2-NEXT: movdqa %xmm4, %xmm1513; X64-SSE2-NEXT: pxor %xmm2, %xmm1514; X64-SSE2-NEXT: pxor %xmm0, %xmm2515; X64-SSE2-NEXT: movdqa %xmm1, %xmm3516; X64-SSE2-NEXT: pcmpgtd %xmm2, %xmm3517; X64-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]518; X64-SSE2-NEXT: pcmpeqd %xmm1, %xmm2519; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]520; X64-SSE2-NEXT: pand %xmm5, %xmm1521; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]522; X64-SSE2-NEXT: por %xmm1, %xmm2523; X64-SSE2-NEXT: pand %xmm2, %xmm4524; X64-SSE2-NEXT: pandn %xmm0, %xmm2525; X64-SSE2-NEXT: por %xmm4, %xmm2526; X64-SSE2-NEXT: movq %xmm2, %rax527; X64-SSE2-NEXT: retq528;529; X64-SSE42-LABEL: test_reduce_v4i64:530; X64-SSE42: ## %bb.0:531; X64-SSE42-NEXT: movdqa %xmm0, %xmm2532; X64-SSE42-NEXT: pcmpgtq %xmm1, %xmm0533; X64-SSE42-NEXT: blendvpd %xmm0, %xmm2, %xmm1534; X64-SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]535; X64-SSE42-NEXT: movdqa %xmm1, %xmm0536; X64-SSE42-NEXT: pcmpgtq %xmm2, %xmm0537; X64-SSE42-NEXT: blendvpd %xmm0, %xmm1, %xmm2538; X64-SSE42-NEXT: movq %xmm2, %rax539; X64-SSE42-NEXT: retq540;541; X64-AVX1-LABEL: test_reduce_v4i64:542; X64-AVX1: ## %bb.0:543; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1544; X64-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2545; X64-AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0546; X64-AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]547; X64-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2548; X64-AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0549; X64-AVX1-NEXT: vmovq %xmm0, %rax550; X64-AVX1-NEXT: vzeroupper551; X64-AVX1-NEXT: retq552;553; X64-AVX2-LABEL: test_reduce_v4i64:554; X64-AVX2: ## %bb.0:555; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1556; X64-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2557; X64-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0558; X64-AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]559; X64-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm2560; X64-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0561; X64-AVX2-NEXT: vmovq %xmm0, %rax562; X64-AVX2-NEXT: vzeroupper563; X64-AVX2-NEXT: retq564;565; X64-AVX512-LABEL: test_reduce_v4i64:566; X64-AVX512: ## %bb.0:567; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1568; X64-AVX512-NEXT: vpmaxsq %xmm1, %xmm0, %xmm0569; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]570; X64-AVX512-NEXT: vpmaxsq %xmm1, %xmm0, %xmm0571; X64-AVX512-NEXT: vmovq %xmm0, %rax572; X64-AVX512-NEXT: vzeroupper573; X64-AVX512-NEXT: retq574 %1 = shufflevector <4 x i64> %a0, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>575 %2 = icmp sgt <4 x i64> %a0, %1576 %3 = select <4 x i1> %2, <4 x i64> %a0, <4 x i64> %1577 %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>578 %5 = icmp sgt <4 x i64> %3, %4579 %6 = select <4 x i1> %5, <4 x i64> %3, <4 x i64> %4580 %7 = extractelement <4 x i64> %6, i32 0581 ret i64 %7582}583 584define i32 @test_reduce_v8i32(<8 x i32> %a0) {585; X86-SSE2-LABEL: test_reduce_v8i32:586; X86-SSE2: ## %bb.0:587; X86-SSE2-NEXT: movdqa %xmm0, %xmm2588; X86-SSE2-NEXT: pcmpgtd %xmm1, %xmm2589; X86-SSE2-NEXT: pand %xmm2, %xmm0590; X86-SSE2-NEXT: pandn %xmm1, %xmm2591; X86-SSE2-NEXT: por %xmm0, %xmm2592; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]593; X86-SSE2-NEXT: movdqa %xmm2, %xmm1594; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm1595; X86-SSE2-NEXT: pand %xmm1, %xmm2596; X86-SSE2-NEXT: pandn %xmm0, %xmm1597; X86-SSE2-NEXT: por %xmm2, %xmm1598; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]599; X86-SSE2-NEXT: movdqa %xmm1, %xmm2600; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm2601; X86-SSE2-NEXT: pand %xmm2, %xmm1602; X86-SSE2-NEXT: pandn %xmm0, %xmm2603; X86-SSE2-NEXT: por %xmm1, %xmm2604; X86-SSE2-NEXT: movd %xmm2, %eax605; X86-SSE2-NEXT: retl606;607; X86-SSE42-LABEL: test_reduce_v8i32:608; X86-SSE42: ## %bb.0:609; X86-SSE42-NEXT: pmaxsd %xmm1, %xmm0610; X86-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]611; X86-SSE42-NEXT: pmaxsd %xmm0, %xmm1612; X86-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]613; X86-SSE42-NEXT: pmaxsd %xmm1, %xmm0614; X86-SSE42-NEXT: movd %xmm0, %eax615; X86-SSE42-NEXT: retl616;617; X86-AVX1-LABEL: test_reduce_v8i32:618; X86-AVX1: ## %bb.0:619; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1620; X86-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0621; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]622; X86-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0623; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]624; X86-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0625; X86-AVX1-NEXT: vmovd %xmm0, %eax626; X86-AVX1-NEXT: vzeroupper627; X86-AVX1-NEXT: retl628;629; X86-AVX2-LABEL: test_reduce_v8i32:630; X86-AVX2: ## %bb.0:631; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1632; X86-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0633; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]634; X86-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0635; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]636; X86-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0637; X86-AVX2-NEXT: vmovd %xmm0, %eax638; X86-AVX2-NEXT: vzeroupper639; X86-AVX2-NEXT: retl640;641; X64-SSE2-LABEL: test_reduce_v8i32:642; X64-SSE2: ## %bb.0:643; X64-SSE2-NEXT: movdqa %xmm0, %xmm2644; X64-SSE2-NEXT: pcmpgtd %xmm1, %xmm2645; X64-SSE2-NEXT: pand %xmm2, %xmm0646; X64-SSE2-NEXT: pandn %xmm1, %xmm2647; X64-SSE2-NEXT: por %xmm0, %xmm2648; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]649; X64-SSE2-NEXT: movdqa %xmm2, %xmm1650; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm1651; X64-SSE2-NEXT: pand %xmm1, %xmm2652; X64-SSE2-NEXT: pandn %xmm0, %xmm1653; X64-SSE2-NEXT: por %xmm2, %xmm1654; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]655; X64-SSE2-NEXT: movdqa %xmm1, %xmm2656; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm2657; X64-SSE2-NEXT: pand %xmm2, %xmm1658; X64-SSE2-NEXT: pandn %xmm0, %xmm2659; X64-SSE2-NEXT: por %xmm1, %xmm2660; X64-SSE2-NEXT: movd %xmm2, %eax661; X64-SSE2-NEXT: retq662;663; X64-SSE42-LABEL: test_reduce_v8i32:664; X64-SSE42: ## %bb.0:665; X64-SSE42-NEXT: pmaxsd %xmm1, %xmm0666; X64-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]667; X64-SSE42-NEXT: pmaxsd %xmm0, %xmm1668; X64-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]669; X64-SSE42-NEXT: pmaxsd %xmm1, %xmm0670; X64-SSE42-NEXT: movd %xmm0, %eax671; X64-SSE42-NEXT: retq672;673; X64-AVX1-LABEL: test_reduce_v8i32:674; X64-AVX1: ## %bb.0:675; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1676; X64-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0677; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]678; X64-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0679; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]680; X64-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0681; X64-AVX1-NEXT: vmovd %xmm0, %eax682; X64-AVX1-NEXT: vzeroupper683; X64-AVX1-NEXT: retq684;685; X64-AVX2-LABEL: test_reduce_v8i32:686; X64-AVX2: ## %bb.0:687; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1688; X64-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0689; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]690; X64-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0691; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]692; X64-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0693; X64-AVX2-NEXT: vmovd %xmm0, %eax694; X64-AVX2-NEXT: vzeroupper695; X64-AVX2-NEXT: retq696;697; X64-AVX512-LABEL: test_reduce_v8i32:698; X64-AVX512: ## %bb.0:699; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1700; X64-AVX512-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0701; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]702; X64-AVX512-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0703; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]704; X64-AVX512-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0705; X64-AVX512-NEXT: vmovd %xmm0, %eax706; X64-AVX512-NEXT: vzeroupper707; X64-AVX512-NEXT: retq708 %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>709 %2 = icmp sgt <8 x i32> %a0, %1710 %3 = select <8 x i1> %2, <8 x i32> %a0, <8 x i32> %1711 %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>712 %5 = icmp sgt <8 x i32> %3, %4713 %6 = select <8 x i1> %5, <8 x i32> %3, <8 x i32> %4714 %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>715 %8 = icmp sgt <8 x i32> %6, %7716 %9 = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7717 %10 = extractelement <8 x i32> %9, i32 0718 ret i32 %10719}720 721define i16 @test_reduce_v16i16(<16 x i16> %a0) {722; X86-SSE2-LABEL: test_reduce_v16i16:723; X86-SSE2: ## %bb.0:724; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm0725; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]726; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm1727; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]728; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm0729; X86-SSE2-NEXT: movdqa %xmm0, %xmm1730; X86-SSE2-NEXT: psrld $16, %xmm1731; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm1732; X86-SSE2-NEXT: movd %xmm1, %eax733; X86-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax734; X86-SSE2-NEXT: retl735;736; X86-SSE42-LABEL: test_reduce_v16i16:737; X86-SSE42: ## %bb.0:738; X86-SSE42-NEXT: pmaxsw %xmm1, %xmm0739; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0740; X86-SSE42-NEXT: phminposuw %xmm0, %xmm0741; X86-SSE42-NEXT: movd %xmm0, %eax742; X86-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF743; X86-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax744; X86-SSE42-NEXT: retl745;746; X86-AVX1-LABEL: test_reduce_v16i16:747; X86-AVX1: ## %bb.0:748; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1749; X86-AVX1-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0750; X86-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0751; X86-AVX1-NEXT: vphminposuw %xmm0, %xmm0752; X86-AVX1-NEXT: vmovd %xmm0, %eax753; X86-AVX1-NEXT: xorl $32767, %eax ## imm = 0x7FFF754; X86-AVX1-NEXT: ## kill: def $ax killed $ax killed $eax755; X86-AVX1-NEXT: vzeroupper756; X86-AVX1-NEXT: retl757;758; X86-AVX2-LABEL: test_reduce_v16i16:759; X86-AVX2: ## %bb.0:760; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1761; X86-AVX2-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0762; X86-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0763; X86-AVX2-NEXT: vphminposuw %xmm0, %xmm0764; X86-AVX2-NEXT: vmovd %xmm0, %eax765; X86-AVX2-NEXT: xorl $32767, %eax ## imm = 0x7FFF766; X86-AVX2-NEXT: ## kill: def $ax killed $ax killed $eax767; X86-AVX2-NEXT: vzeroupper768; X86-AVX2-NEXT: retl769;770; X64-SSE2-LABEL: test_reduce_v16i16:771; X64-SSE2: ## %bb.0:772; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm0773; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]774; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm1775; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]776; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm0777; X64-SSE2-NEXT: movdqa %xmm0, %xmm1778; X64-SSE2-NEXT: psrld $16, %xmm1779; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm1780; X64-SSE2-NEXT: movd %xmm1, %eax781; X64-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax782; X64-SSE2-NEXT: retq783;784; X64-SSE42-LABEL: test_reduce_v16i16:785; X64-SSE42: ## %bb.0:786; X64-SSE42-NEXT: pmaxsw %xmm1, %xmm0787; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0788; X64-SSE42-NEXT: phminposuw %xmm0, %xmm0789; X64-SSE42-NEXT: movd %xmm0, %eax790; X64-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF791; X64-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax792; X64-SSE42-NEXT: retq793;794; X64-AVX1-LABEL: test_reduce_v16i16:795; X64-AVX1: ## %bb.0:796; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1797; X64-AVX1-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0798; X64-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0799; X64-AVX1-NEXT: vphminposuw %xmm0, %xmm0800; X64-AVX1-NEXT: vmovd %xmm0, %eax801; X64-AVX1-NEXT: xorl $32767, %eax ## imm = 0x7FFF802; X64-AVX1-NEXT: ## kill: def $ax killed $ax killed $eax803; X64-AVX1-NEXT: vzeroupper804; X64-AVX1-NEXT: retq805;806; X64-AVX2-LABEL: test_reduce_v16i16:807; X64-AVX2: ## %bb.0:808; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1809; X64-AVX2-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0810; X64-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0811; X64-AVX2-NEXT: vphminposuw %xmm0, %xmm0812; X64-AVX2-NEXT: vmovd %xmm0, %eax813; X64-AVX2-NEXT: xorl $32767, %eax ## imm = 0x7FFF814; X64-AVX2-NEXT: ## kill: def $ax killed $ax killed $eax815; X64-AVX2-NEXT: vzeroupper816; X64-AVX2-NEXT: retq817;818; X64-AVX512-LABEL: test_reduce_v16i16:819; X64-AVX512: ## %bb.0:820; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1821; X64-AVX512-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0822; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0823; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm0824; X64-AVX512-NEXT: vmovd %xmm0, %eax825; X64-AVX512-NEXT: xorl $32767, %eax ## imm = 0x7FFF826; X64-AVX512-NEXT: ## kill: def $ax killed $ax killed $eax827; X64-AVX512-NEXT: vzeroupper828; X64-AVX512-NEXT: retq829 %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>830 %2 = icmp sgt <16 x i16> %a0, %1831 %3 = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1832 %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>833 %5 = icmp sgt <16 x i16> %3, %4834 %6 = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %4835 %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>836 %8 = icmp sgt <16 x i16> %6, %7837 %9 = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7838 %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>839 %11 = icmp sgt <16 x i16> %9, %10840 %12 = select <16 x i1> %11, <16 x i16> %9, <16 x i16> %10841 %13 = extractelement <16 x i16> %12, i32 0842 ret i16 %13843}844 845define i8 @test_reduce_v32i8(<32 x i8> %a0) {846; X86-SSE2-LABEL: test_reduce_v32i8:847; X86-SSE2: ## %bb.0:848; X86-SSE2-NEXT: movdqa %xmm0, %xmm2849; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm2850; X86-SSE2-NEXT: pand %xmm2, %xmm0851; X86-SSE2-NEXT: pandn %xmm1, %xmm2852; X86-SSE2-NEXT: por %xmm0, %xmm2853; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]854; X86-SSE2-NEXT: movdqa %xmm2, %xmm1855; X86-SSE2-NEXT: pcmpgtb %xmm0, %xmm1856; X86-SSE2-NEXT: pand %xmm1, %xmm2857; X86-SSE2-NEXT: pandn %xmm0, %xmm1858; X86-SSE2-NEXT: por %xmm2, %xmm1859; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]860; X86-SSE2-NEXT: movdqa %xmm1, %xmm0861; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm0862; X86-SSE2-NEXT: pand %xmm0, %xmm1863; X86-SSE2-NEXT: pandn %xmm2, %xmm0864; X86-SSE2-NEXT: por %xmm1, %xmm0865; X86-SSE2-NEXT: movdqa %xmm0, %xmm2866; X86-SSE2-NEXT: psrld $16, %xmm2867; X86-SSE2-NEXT: movdqa %xmm0, %xmm1868; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm1869; X86-SSE2-NEXT: pand %xmm1, %xmm0870; X86-SSE2-NEXT: pandn %xmm2, %xmm1871; X86-SSE2-NEXT: por %xmm0, %xmm1872; X86-SSE2-NEXT: movdqa %xmm1, %xmm0873; X86-SSE2-NEXT: psrlw $8, %xmm0874; X86-SSE2-NEXT: movdqa %xmm1, %xmm2875; X86-SSE2-NEXT: pcmpgtb %xmm0, %xmm2876; X86-SSE2-NEXT: pand %xmm2, %xmm1877; X86-SSE2-NEXT: pandn %xmm0, %xmm2878; X86-SSE2-NEXT: por %xmm1, %xmm2879; X86-SSE2-NEXT: movd %xmm2, %eax880; X86-SSE2-NEXT: ## kill: def $al killed $al killed $eax881; X86-SSE2-NEXT: retl882;883; X86-SSE42-LABEL: test_reduce_v32i8:884; X86-SSE42: ## %bb.0:885; X86-SSE42-NEXT: pmaxsb %xmm1, %xmm0886; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0887; X86-SSE42-NEXT: movdqa %xmm0, %xmm1888; X86-SSE42-NEXT: psrlw $8, %xmm1889; X86-SSE42-NEXT: pminub %xmm0, %xmm1890; X86-SSE42-NEXT: phminposuw %xmm1, %xmm0891; X86-SSE42-NEXT: movd %xmm0, %eax892; X86-SSE42-NEXT: xorb $127, %al893; X86-SSE42-NEXT: ## kill: def $al killed $al killed $eax894; X86-SSE42-NEXT: retl895;896; X86-AVX1-LABEL: test_reduce_v32i8:897; X86-AVX1: ## %bb.0:898; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1899; X86-AVX1-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0900; X86-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0901; X86-AVX1-NEXT: vpsrlw $8, %xmm0, %xmm1902; X86-AVX1-NEXT: vpminub %xmm1, %xmm0, %xmm0903; X86-AVX1-NEXT: vphminposuw %xmm0, %xmm0904; X86-AVX1-NEXT: vmovd %xmm0, %eax905; X86-AVX1-NEXT: xorb $127, %al906; X86-AVX1-NEXT: ## kill: def $al killed $al killed $eax907; X86-AVX1-NEXT: vzeroupper908; X86-AVX1-NEXT: retl909;910; X86-AVX2-LABEL: test_reduce_v32i8:911; X86-AVX2: ## %bb.0:912; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1913; X86-AVX2-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0914; X86-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0915; X86-AVX2-NEXT: vpsrlw $8, %xmm0, %xmm1916; X86-AVX2-NEXT: vpminub %xmm1, %xmm0, %xmm0917; X86-AVX2-NEXT: vphminposuw %xmm0, %xmm0918; X86-AVX2-NEXT: vmovd %xmm0, %eax919; X86-AVX2-NEXT: xorb $127, %al920; X86-AVX2-NEXT: ## kill: def $al killed $al killed $eax921; X86-AVX2-NEXT: vzeroupper922; X86-AVX2-NEXT: retl923;924; X64-SSE2-LABEL: test_reduce_v32i8:925; X64-SSE2: ## %bb.0:926; X64-SSE2-NEXT: movdqa %xmm0, %xmm2927; X64-SSE2-NEXT: pcmpgtb %xmm1, %xmm2928; X64-SSE2-NEXT: pand %xmm2, %xmm0929; X64-SSE2-NEXT: pandn %xmm1, %xmm2930; X64-SSE2-NEXT: por %xmm0, %xmm2931; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]932; X64-SSE2-NEXT: movdqa %xmm2, %xmm1933; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm1934; X64-SSE2-NEXT: pand %xmm1, %xmm2935; X64-SSE2-NEXT: pandn %xmm0, %xmm1936; X64-SSE2-NEXT: por %xmm2, %xmm1937; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]938; X64-SSE2-NEXT: movdqa %xmm1, %xmm2939; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm2940; X64-SSE2-NEXT: pand %xmm2, %xmm1941; X64-SSE2-NEXT: pandn %xmm0, %xmm2942; X64-SSE2-NEXT: por %xmm1, %xmm2943; X64-SSE2-NEXT: movdqa %xmm2, %xmm0944; X64-SSE2-NEXT: psrld $16, %xmm0945; X64-SSE2-NEXT: movdqa %xmm2, %xmm1946; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm1947; X64-SSE2-NEXT: pand %xmm1, %xmm2948; X64-SSE2-NEXT: pandn %xmm0, %xmm1949; X64-SSE2-NEXT: por %xmm2, %xmm1950; X64-SSE2-NEXT: movdqa %xmm1, %xmm0951; X64-SSE2-NEXT: psrlw $8, %xmm0952; X64-SSE2-NEXT: movdqa %xmm1, %xmm2953; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm2954; X64-SSE2-NEXT: pand %xmm2, %xmm1955; X64-SSE2-NEXT: pandn %xmm0, %xmm2956; X64-SSE2-NEXT: por %xmm1, %xmm2957; X64-SSE2-NEXT: movd %xmm2, %eax958; X64-SSE2-NEXT: ## kill: def $al killed $al killed $eax959; X64-SSE2-NEXT: retq960;961; X64-SSE42-LABEL: test_reduce_v32i8:962; X64-SSE42: ## %bb.0:963; X64-SSE42-NEXT: pmaxsb %xmm1, %xmm0964; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0965; X64-SSE42-NEXT: movdqa %xmm0, %xmm1966; X64-SSE42-NEXT: psrlw $8, %xmm1967; X64-SSE42-NEXT: pminub %xmm0, %xmm1968; X64-SSE42-NEXT: phminposuw %xmm1, %xmm0969; X64-SSE42-NEXT: movd %xmm0, %eax970; X64-SSE42-NEXT: xorb $127, %al971; X64-SSE42-NEXT: ## kill: def $al killed $al killed $eax972; X64-SSE42-NEXT: retq973;974; X64-AVX1-LABEL: test_reduce_v32i8:975; X64-AVX1: ## %bb.0:976; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1977; X64-AVX1-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0978; X64-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0979; X64-AVX1-NEXT: vpsrlw $8, %xmm0, %xmm1980; X64-AVX1-NEXT: vpminub %xmm1, %xmm0, %xmm0981; X64-AVX1-NEXT: vphminposuw %xmm0, %xmm0982; X64-AVX1-NEXT: vmovd %xmm0, %eax983; X64-AVX1-NEXT: xorb $127, %al984; X64-AVX1-NEXT: ## kill: def $al killed $al killed $eax985; X64-AVX1-NEXT: vzeroupper986; X64-AVX1-NEXT: retq987;988; X64-AVX2-LABEL: test_reduce_v32i8:989; X64-AVX2: ## %bb.0:990; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1991; X64-AVX2-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0992; X64-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0993; X64-AVX2-NEXT: vpsrlw $8, %xmm0, %xmm1994; X64-AVX2-NEXT: vpminub %xmm1, %xmm0, %xmm0995; X64-AVX2-NEXT: vphminposuw %xmm0, %xmm0996; X64-AVX2-NEXT: vmovd %xmm0, %eax997; X64-AVX2-NEXT: xorb $127, %al998; X64-AVX2-NEXT: ## kill: def $al killed $al killed $eax999; X64-AVX2-NEXT: vzeroupper1000; X64-AVX2-NEXT: retq1001;1002; X64-AVX512-LABEL: test_reduce_v32i8:1003; X64-AVX512: ## %bb.0:1004; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm11005; X64-AVX512-NEXT: vpmaxsb %xmm1, %xmm0, %xmm01006; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01007; X64-AVX512-NEXT: vpsrlw $8, %xmm0, %xmm11008; X64-AVX512-NEXT: vpminub %xmm1, %xmm0, %xmm01009; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm01010; X64-AVX512-NEXT: vmovd %xmm0, %eax1011; X64-AVX512-NEXT: xorb $127, %al1012; X64-AVX512-NEXT: ## kill: def $al killed $al killed $eax1013; X64-AVX512-NEXT: vzeroupper1014; X64-AVX512-NEXT: retq1015 %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>1016 %2 = icmp sgt <32 x i8> %a0, %11017 %3 = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %11018 %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>1019 %5 = icmp sgt <32 x i8> %3, %41020 %6 = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %41021 %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>1022 %8 = icmp sgt <32 x i8> %6, %71023 %9 = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %71024 %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>1025 %11 = icmp sgt <32 x i8> %9, %101026 %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %101027 %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>1028 %14 = icmp sgt <32 x i8> %12, %131029 %15 = select <32 x i1> %14, <32 x i8> %12, <32 x i8> %131030 %16 = extractelement <32 x i8> %15, i32 01031 ret i8 %161032}1033 1034;1035; 512-bit Vectors1036;1037 1038define i64 @test_reduce_v8i64(<8 x i64> %a0) {1039; X86-SSE2-LABEL: test_reduce_v8i64:1040; X86-SSE2: ## %bb.0:1041; X86-SSE2-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,0,2147483648,0]1042; X86-SSE2-NEXT: movdqa %xmm2, %xmm51043; X86-SSE2-NEXT: pxor %xmm4, %xmm51044; X86-SSE2-NEXT: movdqa %xmm0, %xmm61045; X86-SSE2-NEXT: pxor %xmm4, %xmm61046; X86-SSE2-NEXT: movdqa %xmm6, %xmm71047; X86-SSE2-NEXT: pcmpgtd %xmm5, %xmm71048; X86-SSE2-NEXT: pcmpeqd %xmm5, %xmm61049; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]1050; X86-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1051; X86-SSE2-NEXT: pand %xmm5, %xmm61052; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1053; X86-SSE2-NEXT: por %xmm6, %xmm51054; X86-SSE2-NEXT: pand %xmm5, %xmm01055; X86-SSE2-NEXT: pandn %xmm2, %xmm51056; X86-SSE2-NEXT: por %xmm0, %xmm51057; X86-SSE2-NEXT: movdqa %xmm3, %xmm01058; X86-SSE2-NEXT: pxor %xmm4, %xmm01059; X86-SSE2-NEXT: movdqa %xmm1, %xmm21060; X86-SSE2-NEXT: pxor %xmm4, %xmm21061; X86-SSE2-NEXT: movdqa %xmm2, %xmm61062; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm61063; X86-SSE2-NEXT: pcmpeqd %xmm0, %xmm21064; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]1065; X86-SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]1066; X86-SSE2-NEXT: pand %xmm0, %xmm71067; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]1068; X86-SSE2-NEXT: por %xmm7, %xmm21069; X86-SSE2-NEXT: pand %xmm2, %xmm11070; X86-SSE2-NEXT: pandn %xmm3, %xmm21071; X86-SSE2-NEXT: por %xmm1, %xmm21072; X86-SSE2-NEXT: movdqa %xmm2, %xmm01073; X86-SSE2-NEXT: pxor %xmm4, %xmm01074; X86-SSE2-NEXT: movdqa %xmm5, %xmm11075; X86-SSE2-NEXT: pxor %xmm4, %xmm11076; X86-SSE2-NEXT: movdqa %xmm1, %xmm31077; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm31078; X86-SSE2-NEXT: pcmpeqd %xmm0, %xmm11079; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]1080; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]1081; X86-SSE2-NEXT: pand %xmm0, %xmm11082; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]1083; X86-SSE2-NEXT: por %xmm1, %xmm01084; X86-SSE2-NEXT: pand %xmm0, %xmm51085; X86-SSE2-NEXT: pandn %xmm2, %xmm01086; X86-SSE2-NEXT: por %xmm5, %xmm01087; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1088; X86-SSE2-NEXT: movdqa %xmm0, %xmm21089; X86-SSE2-NEXT: pxor %xmm4, %xmm21090; X86-SSE2-NEXT: pxor %xmm1, %xmm41091; X86-SSE2-NEXT: movdqa %xmm2, %xmm31092; X86-SSE2-NEXT: pcmpgtd %xmm4, %xmm31093; X86-SSE2-NEXT: pcmpeqd %xmm2, %xmm41094; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]1095; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]1096; X86-SSE2-NEXT: pand %xmm2, %xmm41097; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]1098; X86-SSE2-NEXT: por %xmm4, %xmm21099; X86-SSE2-NEXT: pand %xmm2, %xmm01100; X86-SSE2-NEXT: pandn %xmm1, %xmm21101; X86-SSE2-NEXT: por %xmm0, %xmm21102; X86-SSE2-NEXT: movd %xmm2, %eax1103; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1104; X86-SSE2-NEXT: movd %xmm0, %edx1105; X86-SSE2-NEXT: retl1106;1107; X86-SSE42-LABEL: test_reduce_v8i64:1108; X86-SSE42: ## %bb.0:1109; X86-SSE42-NEXT: movdqa %xmm0, %xmm41110; X86-SSE42-NEXT: movdqa %xmm1, %xmm01111; X86-SSE42-NEXT: pcmpgtq %xmm3, %xmm01112; X86-SSE42-NEXT: blendvpd %xmm0, %xmm1, %xmm31113; X86-SSE42-NEXT: movdqa %xmm4, %xmm01114; X86-SSE42-NEXT: pcmpgtq %xmm2, %xmm01115; X86-SSE42-NEXT: blendvpd %xmm0, %xmm4, %xmm21116; X86-SSE42-NEXT: movapd %xmm2, %xmm01117; X86-SSE42-NEXT: pcmpgtq %xmm3, %xmm01118; X86-SSE42-NEXT: blendvpd %xmm0, %xmm2, %xmm31119; X86-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1120; X86-SSE42-NEXT: movdqa %xmm3, %xmm01121; X86-SSE42-NEXT: pcmpgtq %xmm1, %xmm01122; X86-SSE42-NEXT: blendvpd %xmm0, %xmm3, %xmm11123; X86-SSE42-NEXT: movd %xmm1, %eax1124; X86-SSE42-NEXT: pextrd $1, %xmm1, %edx1125; X86-SSE42-NEXT: retl1126;1127; X86-AVX1-LABEL: test_reduce_v8i64:1128; X86-AVX1: ## %bb.0:1129; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21130; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31131; X86-AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm41132; X86-AVX1-NEXT: vblendvpd %xmm4, %xmm3, %xmm2, %xmm21133; X86-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm31134; X86-AVX1-NEXT: vblendvpd %xmm3, %xmm0, %xmm1, %xmm01135; X86-AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm11136; X86-AVX1-NEXT: vblendvpd %xmm1, %xmm0, %xmm2, %xmm01137; X86-AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1138; X86-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm21139; X86-AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm01140; X86-AVX1-NEXT: vmovd %xmm0, %eax1141; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx1142; X86-AVX1-NEXT: vzeroupper1143; X86-AVX1-NEXT: retl1144;1145; X86-AVX2-LABEL: test_reduce_v8i64:1146; X86-AVX2: ## %bb.0:1147; X86-AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm21148; X86-AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm01149; X86-AVX2-NEXT: vextractf128 $1, %ymm0, %xmm11150; X86-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm21151; X86-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm01152; X86-AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1153; X86-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm21154; X86-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm01155; X86-AVX2-NEXT: vmovd %xmm0, %eax1156; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx1157; X86-AVX2-NEXT: vzeroupper1158; X86-AVX2-NEXT: retl1159;1160; X64-SSE2-LABEL: test_reduce_v8i64:1161; X64-SSE2: ## %bb.0:1162; X64-SSE2-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,2147483648]1163; X64-SSE2-NEXT: movdqa %xmm2, %xmm51164; X64-SSE2-NEXT: pxor %xmm4, %xmm51165; X64-SSE2-NEXT: movdqa %xmm0, %xmm61166; X64-SSE2-NEXT: pxor %xmm4, %xmm61167; X64-SSE2-NEXT: movdqa %xmm6, %xmm71168; X64-SSE2-NEXT: pcmpgtd %xmm5, %xmm71169; X64-SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]1170; X64-SSE2-NEXT: pcmpeqd %xmm5, %xmm61171; X64-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]1172; X64-SSE2-NEXT: pand %xmm8, %xmm61173; X64-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]1174; X64-SSE2-NEXT: por %xmm6, %xmm51175; X64-SSE2-NEXT: pand %xmm5, %xmm01176; X64-SSE2-NEXT: pandn %xmm2, %xmm51177; X64-SSE2-NEXT: por %xmm0, %xmm51178; X64-SSE2-NEXT: movdqa %xmm3, %xmm01179; X64-SSE2-NEXT: pxor %xmm4, %xmm01180; X64-SSE2-NEXT: movdqa %xmm1, %xmm21181; X64-SSE2-NEXT: pxor %xmm4, %xmm21182; X64-SSE2-NEXT: movdqa %xmm2, %xmm61183; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm61184; X64-SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]1185; X64-SSE2-NEXT: pcmpeqd %xmm0, %xmm21186; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]1187; X64-SSE2-NEXT: pand %xmm7, %xmm01188; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]1189; X64-SSE2-NEXT: por %xmm0, %xmm21190; X64-SSE2-NEXT: pand %xmm2, %xmm11191; X64-SSE2-NEXT: pandn %xmm3, %xmm21192; X64-SSE2-NEXT: por %xmm1, %xmm21193; X64-SSE2-NEXT: movdqa %xmm2, %xmm01194; X64-SSE2-NEXT: pxor %xmm4, %xmm01195; X64-SSE2-NEXT: movdqa %xmm5, %xmm11196; X64-SSE2-NEXT: pxor %xmm4, %xmm11197; X64-SSE2-NEXT: movdqa %xmm1, %xmm31198; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm31199; X64-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]1200; X64-SSE2-NEXT: pcmpeqd %xmm0, %xmm11201; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]1202; X64-SSE2-NEXT: pand %xmm6, %xmm01203; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]1204; X64-SSE2-NEXT: por %xmm0, %xmm11205; X64-SSE2-NEXT: pand %xmm1, %xmm51206; X64-SSE2-NEXT: pandn %xmm2, %xmm11207; X64-SSE2-NEXT: por %xmm5, %xmm11208; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]1209; X64-SSE2-NEXT: movdqa %xmm1, %xmm21210; X64-SSE2-NEXT: pxor %xmm4, %xmm21211; X64-SSE2-NEXT: pxor %xmm0, %xmm41212; X64-SSE2-NEXT: movdqa %xmm2, %xmm31213; X64-SSE2-NEXT: pcmpgtd %xmm4, %xmm31214; X64-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]1215; X64-SSE2-NEXT: pcmpeqd %xmm2, %xmm41216; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]1217; X64-SSE2-NEXT: pand %xmm5, %xmm21218; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]1219; X64-SSE2-NEXT: por %xmm2, %xmm31220; X64-SSE2-NEXT: pand %xmm3, %xmm11221; X64-SSE2-NEXT: pandn %xmm0, %xmm31222; X64-SSE2-NEXT: por %xmm1, %xmm31223; X64-SSE2-NEXT: movq %xmm3, %rax1224; X64-SSE2-NEXT: retq1225;1226; X64-SSE42-LABEL: test_reduce_v8i64:1227; X64-SSE42: ## %bb.0:1228; X64-SSE42-NEXT: movdqa %xmm0, %xmm41229; X64-SSE42-NEXT: movdqa %xmm1, %xmm01230; X64-SSE42-NEXT: pcmpgtq %xmm3, %xmm01231; X64-SSE42-NEXT: blendvpd %xmm0, %xmm1, %xmm31232; X64-SSE42-NEXT: movdqa %xmm4, %xmm01233; X64-SSE42-NEXT: pcmpgtq %xmm2, %xmm01234; X64-SSE42-NEXT: blendvpd %xmm0, %xmm4, %xmm21235; X64-SSE42-NEXT: movapd %xmm2, %xmm01236; X64-SSE42-NEXT: pcmpgtq %xmm3, %xmm01237; X64-SSE42-NEXT: blendvpd %xmm0, %xmm2, %xmm31238; X64-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]1239; X64-SSE42-NEXT: movdqa %xmm3, %xmm01240; X64-SSE42-NEXT: pcmpgtq %xmm1, %xmm01241; X64-SSE42-NEXT: blendvpd %xmm0, %xmm3, %xmm11242; X64-SSE42-NEXT: movq %xmm1, %rax1243; X64-SSE42-NEXT: retq1244;1245; X64-AVX1-LABEL: test_reduce_v8i64:1246; X64-AVX1: ## %bb.0:1247; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21248; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31249; X64-AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm41250; X64-AVX1-NEXT: vblendvpd %xmm4, %xmm3, %xmm2, %xmm21251; X64-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm31252; X64-AVX1-NEXT: vblendvpd %xmm3, %xmm0, %xmm1, %xmm01253; X64-AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm11254; X64-AVX1-NEXT: vblendvpd %xmm1, %xmm0, %xmm2, %xmm01255; X64-AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1256; X64-AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm21257; X64-AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm01258; X64-AVX1-NEXT: vmovq %xmm0, %rax1259; X64-AVX1-NEXT: vzeroupper1260; X64-AVX1-NEXT: retq1261;1262; X64-AVX2-LABEL: test_reduce_v8i64:1263; X64-AVX2: ## %bb.0:1264; X64-AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm21265; X64-AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm01266; X64-AVX2-NEXT: vextractf128 $1, %ymm0, %xmm11267; X64-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm21268; X64-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm01269; X64-AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]1270; X64-AVX2-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm21271; X64-AVX2-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm01272; X64-AVX2-NEXT: vmovq %xmm0, %rax1273; X64-AVX2-NEXT: vzeroupper1274; X64-AVX2-NEXT: retq1275;1276; X64-AVX512-LABEL: test_reduce_v8i64:1277; X64-AVX512: ## %bb.0:1278; X64-AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm11279; X64-AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm01280; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm11281; X64-AVX512-NEXT: vpmaxsq %xmm1, %xmm0, %xmm01282; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1283; X64-AVX512-NEXT: vpmaxsq %xmm1, %xmm0, %xmm01284; X64-AVX512-NEXT: vmovq %xmm0, %rax1285; X64-AVX512-NEXT: vzeroupper1286; X64-AVX512-NEXT: retq1287 %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>1288 %2 = icmp sgt <8 x i64> %a0, %11289 %3 = select <8 x i1> %2, <8 x i64> %a0, <8 x i64> %11290 %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>1291 %5 = icmp sgt <8 x i64> %3, %41292 %6 = select <8 x i1> %5, <8 x i64> %3, <8 x i64> %41293 %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>1294 %8 = icmp sgt <8 x i64> %6, %71295 %9 = select <8 x i1> %8, <8 x i64> %6, <8 x i64> %71296 %10 = extractelement <8 x i64> %9, i32 01297 ret i64 %101298}1299 1300define i32 @test_reduce_v16i32(<16 x i32> %a0) {1301; X86-SSE2-LABEL: test_reduce_v16i32:1302; X86-SSE2: ## %bb.0:1303; X86-SSE2-NEXT: movdqa %xmm1, %xmm41304; X86-SSE2-NEXT: pcmpgtd %xmm3, %xmm41305; X86-SSE2-NEXT: pand %xmm4, %xmm11306; X86-SSE2-NEXT: pandn %xmm3, %xmm41307; X86-SSE2-NEXT: por %xmm1, %xmm41308; X86-SSE2-NEXT: movdqa %xmm0, %xmm11309; X86-SSE2-NEXT: pcmpgtd %xmm2, %xmm11310; X86-SSE2-NEXT: pand %xmm1, %xmm01311; X86-SSE2-NEXT: pandn %xmm2, %xmm11312; X86-SSE2-NEXT: por %xmm0, %xmm11313; X86-SSE2-NEXT: movdqa %xmm1, %xmm01314; X86-SSE2-NEXT: pcmpgtd %xmm4, %xmm01315; X86-SSE2-NEXT: pand %xmm0, %xmm11316; X86-SSE2-NEXT: pandn %xmm4, %xmm01317; X86-SSE2-NEXT: por %xmm1, %xmm01318; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1319; X86-SSE2-NEXT: movdqa %xmm0, %xmm21320; X86-SSE2-NEXT: pcmpgtd %xmm1, %xmm21321; X86-SSE2-NEXT: pand %xmm2, %xmm01322; X86-SSE2-NEXT: pandn %xmm1, %xmm21323; X86-SSE2-NEXT: por %xmm0, %xmm21324; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1325; X86-SSE2-NEXT: movdqa %xmm2, %xmm11326; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm11327; X86-SSE2-NEXT: pand %xmm1, %xmm21328; X86-SSE2-NEXT: pandn %xmm0, %xmm11329; X86-SSE2-NEXT: por %xmm2, %xmm11330; X86-SSE2-NEXT: movd %xmm1, %eax1331; X86-SSE2-NEXT: retl1332;1333; X86-SSE42-LABEL: test_reduce_v16i32:1334; X86-SSE42: ## %bb.0:1335; X86-SSE42-NEXT: pmaxsd %xmm3, %xmm11336; X86-SSE42-NEXT: pmaxsd %xmm2, %xmm01337; X86-SSE42-NEXT: pmaxsd %xmm1, %xmm01338; X86-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1339; X86-SSE42-NEXT: pmaxsd %xmm0, %xmm11340; X86-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1341; X86-SSE42-NEXT: pmaxsd %xmm1, %xmm01342; X86-SSE42-NEXT: movd %xmm0, %eax1343; X86-SSE42-NEXT: retl1344;1345; X86-AVX1-LABEL: test_reduce_v16i32:1346; X86-AVX1: ## %bb.0:1347; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21348; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31349; X86-AVX1-NEXT: vpmaxsd %xmm2, %xmm3, %xmm21350; X86-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01351; X86-AVX1-NEXT: vpmaxsd %xmm2, %xmm0, %xmm01352; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1353; X86-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01354; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1355; X86-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01356; X86-AVX1-NEXT: vmovd %xmm0, %eax1357; X86-AVX1-NEXT: vzeroupper1358; X86-AVX1-NEXT: retl1359;1360; X86-AVX2-LABEL: test_reduce_v16i32:1361; X86-AVX2: ## %bb.0:1362; X86-AVX2-NEXT: vpmaxsd %ymm1, %ymm0, %ymm01363; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11364; X86-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01365; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1366; X86-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01367; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1368; X86-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01369; X86-AVX2-NEXT: vmovd %xmm0, %eax1370; X86-AVX2-NEXT: vzeroupper1371; X86-AVX2-NEXT: retl1372;1373; X64-SSE2-LABEL: test_reduce_v16i32:1374; X64-SSE2: ## %bb.0:1375; X64-SSE2-NEXT: movdqa %xmm1, %xmm41376; X64-SSE2-NEXT: pcmpgtd %xmm3, %xmm41377; X64-SSE2-NEXT: pand %xmm4, %xmm11378; X64-SSE2-NEXT: pandn %xmm3, %xmm41379; X64-SSE2-NEXT: por %xmm1, %xmm41380; X64-SSE2-NEXT: movdqa %xmm0, %xmm11381; X64-SSE2-NEXT: pcmpgtd %xmm2, %xmm11382; X64-SSE2-NEXT: pand %xmm1, %xmm01383; X64-SSE2-NEXT: pandn %xmm2, %xmm11384; X64-SSE2-NEXT: por %xmm0, %xmm11385; X64-SSE2-NEXT: movdqa %xmm1, %xmm01386; X64-SSE2-NEXT: pcmpgtd %xmm4, %xmm01387; X64-SSE2-NEXT: pand %xmm0, %xmm11388; X64-SSE2-NEXT: pandn %xmm4, %xmm01389; X64-SSE2-NEXT: por %xmm1, %xmm01390; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1391; X64-SSE2-NEXT: movdqa %xmm0, %xmm21392; X64-SSE2-NEXT: pcmpgtd %xmm1, %xmm21393; X64-SSE2-NEXT: pand %xmm2, %xmm01394; X64-SSE2-NEXT: pandn %xmm1, %xmm21395; X64-SSE2-NEXT: por %xmm0, %xmm21396; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1397; X64-SSE2-NEXT: movdqa %xmm2, %xmm11398; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm11399; X64-SSE2-NEXT: pand %xmm1, %xmm21400; X64-SSE2-NEXT: pandn %xmm0, %xmm11401; X64-SSE2-NEXT: por %xmm2, %xmm11402; X64-SSE2-NEXT: movd %xmm1, %eax1403; X64-SSE2-NEXT: retq1404;1405; X64-SSE42-LABEL: test_reduce_v16i32:1406; X64-SSE42: ## %bb.0:1407; X64-SSE42-NEXT: pmaxsd %xmm3, %xmm11408; X64-SSE42-NEXT: pmaxsd %xmm2, %xmm01409; X64-SSE42-NEXT: pmaxsd %xmm1, %xmm01410; X64-SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1411; X64-SSE42-NEXT: pmaxsd %xmm0, %xmm11412; X64-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1413; X64-SSE42-NEXT: pmaxsd %xmm1, %xmm01414; X64-SSE42-NEXT: movd %xmm0, %eax1415; X64-SSE42-NEXT: retq1416;1417; X64-AVX1-LABEL: test_reduce_v16i32:1418; X64-AVX1: ## %bb.0:1419; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21420; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31421; X64-AVX1-NEXT: vpmaxsd %xmm2, %xmm3, %xmm21422; X64-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01423; X64-AVX1-NEXT: vpmaxsd %xmm2, %xmm0, %xmm01424; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1425; X64-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01426; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1427; X64-AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01428; X64-AVX1-NEXT: vmovd %xmm0, %eax1429; X64-AVX1-NEXT: vzeroupper1430; X64-AVX1-NEXT: retq1431;1432; X64-AVX2-LABEL: test_reduce_v16i32:1433; X64-AVX2: ## %bb.0:1434; X64-AVX2-NEXT: vpmaxsd %ymm1, %ymm0, %ymm01435; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11436; X64-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01437; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1438; X64-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01439; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1440; X64-AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01441; X64-AVX2-NEXT: vmovd %xmm0, %eax1442; X64-AVX2-NEXT: vzeroupper1443; X64-AVX2-NEXT: retq1444;1445; X64-AVX512-LABEL: test_reduce_v16i32:1446; X64-AVX512: ## %bb.0:1447; X64-AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm11448; X64-AVX512-NEXT: vpmaxsd %zmm1, %zmm0, %zmm01449; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm11450; X64-AVX512-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01451; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1452; X64-AVX512-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01453; X64-AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]1454; X64-AVX512-NEXT: vpmaxsd %xmm1, %xmm0, %xmm01455; X64-AVX512-NEXT: vmovd %xmm0, %eax1456; X64-AVX512-NEXT: vzeroupper1457; X64-AVX512-NEXT: retq1458 %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>1459 %2 = icmp sgt <16 x i32> %a0, %11460 %3 = select <16 x i1> %2, <16 x i32> %a0, <16 x i32> %11461 %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>1462 %5 = icmp sgt <16 x i32> %3, %41463 %6 = select <16 x i1> %5, <16 x i32> %3, <16 x i32> %41464 %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>1465 %8 = icmp sgt <16 x i32> %6, %71466 %9 = select <16 x i1> %8, <16 x i32> %6, <16 x i32> %71467 %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>1468 %11 = icmp sgt <16 x i32> %9, %101469 %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %101470 %13 = extractelement <16 x i32> %12, i32 01471 ret i32 %131472}1473 1474define i16 @test_reduce_v32i16(<32 x i16> %a0) {1475; X86-SSE2-LABEL: test_reduce_v32i16:1476; X86-SSE2: ## %bb.0:1477; X86-SSE2-NEXT: pmaxsw %xmm3, %xmm11478; X86-SSE2-NEXT: pmaxsw %xmm2, %xmm01479; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm01480; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1481; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm11482; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1483; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm01484; X86-SSE2-NEXT: movdqa %xmm0, %xmm11485; X86-SSE2-NEXT: psrld $16, %xmm11486; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm11487; X86-SSE2-NEXT: movd %xmm1, %eax1488; X86-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax1489; X86-SSE2-NEXT: retl1490;1491; X86-SSE42-LABEL: test_reduce_v32i16:1492; X86-SSE42: ## %bb.0:1493; X86-SSE42-NEXT: pmaxsw %xmm3, %xmm11494; X86-SSE42-NEXT: pmaxsw %xmm2, %xmm01495; X86-SSE42-NEXT: pmaxsw %xmm1, %xmm01496; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01497; X86-SSE42-NEXT: phminposuw %xmm0, %xmm01498; X86-SSE42-NEXT: movd %xmm0, %eax1499; X86-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF1500; X86-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax1501; X86-SSE42-NEXT: retl1502;1503; X86-AVX1-LABEL: test_reduce_v32i16:1504; X86-AVX1: ## %bb.0:1505; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21506; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31507; X86-AVX1-NEXT: vpmaxsw %xmm2, %xmm3, %xmm21508; X86-AVX1-NEXT: vpmaxsw %xmm1, %xmm0, %xmm01509; X86-AVX1-NEXT: vpmaxsw %xmm2, %xmm0, %xmm01510; X86-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01511; X86-AVX1-NEXT: vphminposuw %xmm0, %xmm01512; X86-AVX1-NEXT: vmovd %xmm0, %eax1513; X86-AVX1-NEXT: xorl $32767, %eax ## imm = 0x7FFF1514; X86-AVX1-NEXT: ## kill: def $ax killed $ax killed $eax1515; X86-AVX1-NEXT: vzeroupper1516; X86-AVX1-NEXT: retl1517;1518; X86-AVX2-LABEL: test_reduce_v32i16:1519; X86-AVX2: ## %bb.0:1520; X86-AVX2-NEXT: vpmaxsw %ymm1, %ymm0, %ymm01521; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11522; X86-AVX2-NEXT: vpmaxsw %xmm1, %xmm0, %xmm01523; X86-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01524; X86-AVX2-NEXT: vphminposuw %xmm0, %xmm01525; X86-AVX2-NEXT: vmovd %xmm0, %eax1526; X86-AVX2-NEXT: xorl $32767, %eax ## imm = 0x7FFF1527; X86-AVX2-NEXT: ## kill: def $ax killed $ax killed $eax1528; X86-AVX2-NEXT: vzeroupper1529; X86-AVX2-NEXT: retl1530;1531; X64-SSE2-LABEL: test_reduce_v32i16:1532; X64-SSE2: ## %bb.0:1533; X64-SSE2-NEXT: pmaxsw %xmm3, %xmm11534; X64-SSE2-NEXT: pmaxsw %xmm2, %xmm01535; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm01536; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1537; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm11538; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1539; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm01540; X64-SSE2-NEXT: movdqa %xmm0, %xmm11541; X64-SSE2-NEXT: psrld $16, %xmm11542; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm11543; X64-SSE2-NEXT: movd %xmm1, %eax1544; X64-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax1545; X64-SSE2-NEXT: retq1546;1547; X64-SSE42-LABEL: test_reduce_v32i16:1548; X64-SSE42: ## %bb.0:1549; X64-SSE42-NEXT: pmaxsw %xmm3, %xmm11550; X64-SSE42-NEXT: pmaxsw %xmm2, %xmm01551; X64-SSE42-NEXT: pmaxsw %xmm1, %xmm01552; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01553; X64-SSE42-NEXT: phminposuw %xmm0, %xmm01554; X64-SSE42-NEXT: movd %xmm0, %eax1555; X64-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF1556; X64-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax1557; X64-SSE42-NEXT: retq1558;1559; X64-AVX1-LABEL: test_reduce_v32i16:1560; X64-AVX1: ## %bb.0:1561; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21562; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31563; X64-AVX1-NEXT: vpmaxsw %xmm2, %xmm3, %xmm21564; X64-AVX1-NEXT: vpmaxsw %xmm1, %xmm0, %xmm01565; X64-AVX1-NEXT: vpmaxsw %xmm2, %xmm0, %xmm01566; X64-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01567; X64-AVX1-NEXT: vphminposuw %xmm0, %xmm01568; X64-AVX1-NEXT: vmovd %xmm0, %eax1569; X64-AVX1-NEXT: xorl $32767, %eax ## imm = 0x7FFF1570; X64-AVX1-NEXT: ## kill: def $ax killed $ax killed $eax1571; X64-AVX1-NEXT: vzeroupper1572; X64-AVX1-NEXT: retq1573;1574; X64-AVX2-LABEL: test_reduce_v32i16:1575; X64-AVX2: ## %bb.0:1576; X64-AVX2-NEXT: vpmaxsw %ymm1, %ymm0, %ymm01577; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11578; X64-AVX2-NEXT: vpmaxsw %xmm1, %xmm0, %xmm01579; X64-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01580; X64-AVX2-NEXT: vphminposuw %xmm0, %xmm01581; X64-AVX2-NEXT: vmovd %xmm0, %eax1582; X64-AVX2-NEXT: xorl $32767, %eax ## imm = 0x7FFF1583; X64-AVX2-NEXT: ## kill: def $ax killed $ax killed $eax1584; X64-AVX2-NEXT: vzeroupper1585; X64-AVX2-NEXT: retq1586;1587; X64-AVX512-LABEL: test_reduce_v32i16:1588; X64-AVX512: ## %bb.0:1589; X64-AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm11590; X64-AVX512-NEXT: vpmaxsw %ymm1, %ymm0, %ymm01591; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm11592; X64-AVX512-NEXT: vpmaxsw %xmm1, %xmm0, %xmm01593; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01594; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm01595; X64-AVX512-NEXT: vmovd %xmm0, %eax1596; X64-AVX512-NEXT: xorl $32767, %eax ## imm = 0x7FFF1597; X64-AVX512-NEXT: ## kill: def $ax killed $ax killed $eax1598; X64-AVX512-NEXT: vzeroupper1599; X64-AVX512-NEXT: retq1600 %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>1601 %2 = icmp sgt <32 x i16> %a0, %11602 %3 = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %11603 %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>1604 %5 = icmp sgt <32 x i16> %3, %41605 %6 = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %41606 %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>1607 %8 = icmp sgt <32 x i16> %6, %71608 %9 = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %71609 %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>1610 %11 = icmp sgt <32 x i16> %9, %101611 %12 = select <32 x i1> %11, <32 x i16> %9, <32 x i16> %101612 %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>1613 %14 = icmp sgt <32 x i16> %12, %131614 %15 = select <32 x i1> %14, <32 x i16> %12, <32 x i16> %131615 %16 = extractelement <32 x i16> %15, i32 01616 ret i16 %161617}1618 1619define i8 @test_reduce_v64i8(<64 x i8> %a0) {1620; X86-SSE2-LABEL: test_reduce_v64i8:1621; X86-SSE2: ## %bb.0:1622; X86-SSE2-NEXT: movdqa %xmm1, %xmm41623; X86-SSE2-NEXT: pcmpgtb %xmm3, %xmm41624; X86-SSE2-NEXT: pand %xmm4, %xmm11625; X86-SSE2-NEXT: pandn %xmm3, %xmm41626; X86-SSE2-NEXT: por %xmm1, %xmm41627; X86-SSE2-NEXT: movdqa %xmm0, %xmm11628; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm11629; X86-SSE2-NEXT: pand %xmm1, %xmm01630; X86-SSE2-NEXT: pandn %xmm2, %xmm11631; X86-SSE2-NEXT: por %xmm0, %xmm11632; X86-SSE2-NEXT: movdqa %xmm1, %xmm01633; X86-SSE2-NEXT: pcmpgtb %xmm4, %xmm01634; X86-SSE2-NEXT: pand %xmm0, %xmm11635; X86-SSE2-NEXT: pandn %xmm4, %xmm01636; X86-SSE2-NEXT: por %xmm1, %xmm01637; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1638; X86-SSE2-NEXT: movdqa %xmm0, %xmm21639; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm21640; X86-SSE2-NEXT: pand %xmm2, %xmm01641; X86-SSE2-NEXT: pandn %xmm1, %xmm21642; X86-SSE2-NEXT: por %xmm0, %xmm21643; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]1644; X86-SSE2-NEXT: movdqa %xmm2, %xmm01645; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm01646; X86-SSE2-NEXT: pand %xmm0, %xmm21647; X86-SSE2-NEXT: pandn %xmm1, %xmm01648; X86-SSE2-NEXT: por %xmm2, %xmm01649; X86-SSE2-NEXT: movdqa %xmm0, %xmm21650; X86-SSE2-NEXT: psrld $16, %xmm21651; X86-SSE2-NEXT: movdqa %xmm0, %xmm11652; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm11653; X86-SSE2-NEXT: pand %xmm1, %xmm01654; X86-SSE2-NEXT: pandn %xmm2, %xmm11655; X86-SSE2-NEXT: por %xmm0, %xmm11656; X86-SSE2-NEXT: movdqa %xmm1, %xmm01657; X86-SSE2-NEXT: psrlw $8, %xmm01658; X86-SSE2-NEXT: movdqa %xmm1, %xmm21659; X86-SSE2-NEXT: pcmpgtb %xmm0, %xmm21660; X86-SSE2-NEXT: pand %xmm2, %xmm11661; X86-SSE2-NEXT: pandn %xmm0, %xmm21662; X86-SSE2-NEXT: por %xmm1, %xmm21663; X86-SSE2-NEXT: movd %xmm2, %eax1664; X86-SSE2-NEXT: ## kill: def $al killed $al killed $eax1665; X86-SSE2-NEXT: retl1666;1667; X86-SSE42-LABEL: test_reduce_v64i8:1668; X86-SSE42: ## %bb.0:1669; X86-SSE42-NEXT: pmaxsb %xmm3, %xmm11670; X86-SSE42-NEXT: pmaxsb %xmm2, %xmm01671; X86-SSE42-NEXT: pmaxsb %xmm1, %xmm01672; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01673; X86-SSE42-NEXT: movdqa %xmm0, %xmm11674; X86-SSE42-NEXT: psrlw $8, %xmm11675; X86-SSE42-NEXT: pminub %xmm0, %xmm11676; X86-SSE42-NEXT: phminposuw %xmm1, %xmm01677; X86-SSE42-NEXT: movd %xmm0, %eax1678; X86-SSE42-NEXT: xorb $127, %al1679; X86-SSE42-NEXT: ## kill: def $al killed $al killed $eax1680; X86-SSE42-NEXT: retl1681;1682; X86-AVX1-LABEL: test_reduce_v64i8:1683; X86-AVX1: ## %bb.0:1684; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21685; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31686; X86-AVX1-NEXT: vpmaxsb %xmm2, %xmm3, %xmm21687; X86-AVX1-NEXT: vpmaxsb %xmm1, %xmm0, %xmm01688; X86-AVX1-NEXT: vpmaxsb %xmm2, %xmm0, %xmm01689; X86-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01690; X86-AVX1-NEXT: vpsrlw $8, %xmm0, %xmm11691; X86-AVX1-NEXT: vpminub %xmm1, %xmm0, %xmm01692; X86-AVX1-NEXT: vphminposuw %xmm0, %xmm01693; X86-AVX1-NEXT: vmovd %xmm0, %eax1694; X86-AVX1-NEXT: xorb $127, %al1695; X86-AVX1-NEXT: ## kill: def $al killed $al killed $eax1696; X86-AVX1-NEXT: vzeroupper1697; X86-AVX1-NEXT: retl1698;1699; X86-AVX2-LABEL: test_reduce_v64i8:1700; X86-AVX2: ## %bb.0:1701; X86-AVX2-NEXT: vpmaxsb %ymm1, %ymm0, %ymm01702; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11703; X86-AVX2-NEXT: vpmaxsb %xmm1, %xmm0, %xmm01704; X86-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01705; X86-AVX2-NEXT: vpsrlw $8, %xmm0, %xmm11706; X86-AVX2-NEXT: vpminub %xmm1, %xmm0, %xmm01707; X86-AVX2-NEXT: vphminposuw %xmm0, %xmm01708; X86-AVX2-NEXT: vmovd %xmm0, %eax1709; X86-AVX2-NEXT: xorb $127, %al1710; X86-AVX2-NEXT: ## kill: def $al killed $al killed $eax1711; X86-AVX2-NEXT: vzeroupper1712; X86-AVX2-NEXT: retl1713;1714; X64-SSE2-LABEL: test_reduce_v64i8:1715; X64-SSE2: ## %bb.0:1716; X64-SSE2-NEXT: movdqa %xmm1, %xmm41717; X64-SSE2-NEXT: pcmpgtb %xmm3, %xmm41718; X64-SSE2-NEXT: pand %xmm4, %xmm11719; X64-SSE2-NEXT: pandn %xmm3, %xmm41720; X64-SSE2-NEXT: por %xmm1, %xmm41721; X64-SSE2-NEXT: movdqa %xmm0, %xmm11722; X64-SSE2-NEXT: pcmpgtb %xmm2, %xmm11723; X64-SSE2-NEXT: pand %xmm1, %xmm01724; X64-SSE2-NEXT: pandn %xmm2, %xmm11725; X64-SSE2-NEXT: por %xmm0, %xmm11726; X64-SSE2-NEXT: movdqa %xmm1, %xmm01727; X64-SSE2-NEXT: pcmpgtb %xmm4, %xmm01728; X64-SSE2-NEXT: pand %xmm0, %xmm11729; X64-SSE2-NEXT: pandn %xmm4, %xmm01730; X64-SSE2-NEXT: por %xmm1, %xmm01731; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1732; X64-SSE2-NEXT: movdqa %xmm0, %xmm21733; X64-SSE2-NEXT: pcmpgtb %xmm1, %xmm21734; X64-SSE2-NEXT: pand %xmm2, %xmm01735; X64-SSE2-NEXT: pandn %xmm1, %xmm21736; X64-SSE2-NEXT: por %xmm0, %xmm21737; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]1738; X64-SSE2-NEXT: movdqa %xmm2, %xmm11739; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm11740; X64-SSE2-NEXT: pand %xmm1, %xmm21741; X64-SSE2-NEXT: pandn %xmm0, %xmm11742; X64-SSE2-NEXT: por %xmm2, %xmm11743; X64-SSE2-NEXT: movdqa %xmm1, %xmm01744; X64-SSE2-NEXT: psrld $16, %xmm01745; X64-SSE2-NEXT: movdqa %xmm1, %xmm21746; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm21747; X64-SSE2-NEXT: pand %xmm2, %xmm11748; X64-SSE2-NEXT: pandn %xmm0, %xmm21749; X64-SSE2-NEXT: por %xmm1, %xmm21750; X64-SSE2-NEXT: movdqa %xmm2, %xmm01751; X64-SSE2-NEXT: psrlw $8, %xmm01752; X64-SSE2-NEXT: movdqa %xmm2, %xmm11753; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm11754; X64-SSE2-NEXT: pand %xmm1, %xmm21755; X64-SSE2-NEXT: pandn %xmm0, %xmm11756; X64-SSE2-NEXT: por %xmm2, %xmm11757; X64-SSE2-NEXT: movd %xmm1, %eax1758; X64-SSE2-NEXT: ## kill: def $al killed $al killed $eax1759; X64-SSE2-NEXT: retq1760;1761; X64-SSE42-LABEL: test_reduce_v64i8:1762; X64-SSE42: ## %bb.0:1763; X64-SSE42-NEXT: pmaxsb %xmm3, %xmm11764; X64-SSE42-NEXT: pmaxsb %xmm2, %xmm01765; X64-SSE42-NEXT: pmaxsb %xmm1, %xmm01766; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01767; X64-SSE42-NEXT: movdqa %xmm0, %xmm11768; X64-SSE42-NEXT: psrlw $8, %xmm11769; X64-SSE42-NEXT: pminub %xmm0, %xmm11770; X64-SSE42-NEXT: phminposuw %xmm1, %xmm01771; X64-SSE42-NEXT: movd %xmm0, %eax1772; X64-SSE42-NEXT: xorb $127, %al1773; X64-SSE42-NEXT: ## kill: def $al killed $al killed $eax1774; X64-SSE42-NEXT: retq1775;1776; X64-AVX1-LABEL: test_reduce_v64i8:1777; X64-AVX1: ## %bb.0:1778; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21779; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31780; X64-AVX1-NEXT: vpmaxsb %xmm2, %xmm3, %xmm21781; X64-AVX1-NEXT: vpmaxsb %xmm1, %xmm0, %xmm01782; X64-AVX1-NEXT: vpmaxsb %xmm2, %xmm0, %xmm01783; X64-AVX1-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01784; X64-AVX1-NEXT: vpsrlw $8, %xmm0, %xmm11785; X64-AVX1-NEXT: vpminub %xmm1, %xmm0, %xmm01786; X64-AVX1-NEXT: vphminposuw %xmm0, %xmm01787; X64-AVX1-NEXT: vmovd %xmm0, %eax1788; X64-AVX1-NEXT: xorb $127, %al1789; X64-AVX1-NEXT: ## kill: def $al killed $al killed $eax1790; X64-AVX1-NEXT: vzeroupper1791; X64-AVX1-NEXT: retq1792;1793; X64-AVX2-LABEL: test_reduce_v64i8:1794; X64-AVX2: ## %bb.0:1795; X64-AVX2-NEXT: vpmaxsb %ymm1, %ymm0, %ymm01796; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11797; X64-AVX2-NEXT: vpmaxsb %xmm1, %xmm0, %xmm01798; X64-AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01799; X64-AVX2-NEXT: vpsrlw $8, %xmm0, %xmm11800; X64-AVX2-NEXT: vpminub %xmm1, %xmm0, %xmm01801; X64-AVX2-NEXT: vphminposuw %xmm0, %xmm01802; X64-AVX2-NEXT: vmovd %xmm0, %eax1803; X64-AVX2-NEXT: xorb $127, %al1804; X64-AVX2-NEXT: ## kill: def $al killed $al killed $eax1805; X64-AVX2-NEXT: vzeroupper1806; X64-AVX2-NEXT: retq1807;1808; X64-AVX512-LABEL: test_reduce_v64i8:1809; X64-AVX512: ## %bb.0:1810; X64-AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm11811; X64-AVX512-NEXT: vpmaxsb %ymm1, %ymm0, %ymm01812; X64-AVX512-NEXT: vextracti128 $1, %ymm0, %xmm11813; X64-AVX512-NEXT: vpmaxsb %xmm1, %xmm0, %xmm01814; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01815; X64-AVX512-NEXT: vpsrlw $8, %xmm0, %xmm11816; X64-AVX512-NEXT: vpminub %xmm1, %xmm0, %xmm01817; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm01818; X64-AVX512-NEXT: vmovd %xmm0, %eax1819; X64-AVX512-NEXT: xorb $127, %al1820; X64-AVX512-NEXT: ## kill: def $al killed $al killed $eax1821; X64-AVX512-NEXT: vzeroupper1822; X64-AVX512-NEXT: retq1823 %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>1824 %2 = icmp sgt <64 x i8> %a0, %11825 %3 = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %11826 %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>1827 %5 = icmp sgt <64 x i8> %3, %41828 %6 = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %41829 %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>1830 %8 = icmp sgt <64 x i8> %6, %71831 %9 = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %71832 %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>1833 %11 = icmp sgt <64 x i8> %9, %101834 %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %101835 %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>1836 %14 = icmp sgt <64 x i8> %12, %131837 %15 = select <64 x i1> %14, <64 x i8> %12, <64 x i8> %131838 %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>1839 %17 = icmp sgt <64 x i8> %15, %161840 %18 = select <64 x i1> %17, <64 x i8> %15, <64 x i8> %161841 %19 = extractelement <64 x i8> %18, i32 01842 ret i8 %191843}1844 1845;1846; Partial Vector Reductions1847;1848 1849define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {1850; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:1851; X86-SSE2: ## %bb.0:1852; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1853; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm11854; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1855; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm01856; X86-SSE2-NEXT: movdqa %xmm0, %xmm11857; X86-SSE2-NEXT: psrld $16, %xmm11858; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm11859; X86-SSE2-NEXT: movd %xmm1, %eax1860; X86-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax1861; X86-SSE2-NEXT: retl1862;1863; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:1864; X86-SSE42: ## %bb.0:1865; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01866; X86-SSE42-NEXT: phminposuw %xmm0, %xmm01867; X86-SSE42-NEXT: movd %xmm0, %eax1868; X86-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF1869; X86-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax1870; X86-SSE42-NEXT: retl1871;1872; X86-AVX-LABEL: test_reduce_v16i16_v8i16:1873; X86-AVX: ## %bb.0:1874; X86-AVX-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01875; X86-AVX-NEXT: vphminposuw %xmm0, %xmm01876; X86-AVX-NEXT: vmovd %xmm0, %eax1877; X86-AVX-NEXT: xorl $32767, %eax ## imm = 0x7FFF1878; X86-AVX-NEXT: ## kill: def $ax killed $ax killed $eax1879; X86-AVX-NEXT: vzeroupper1880; X86-AVX-NEXT: retl1881;1882; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:1883; X64-SSE2: ## %bb.0:1884; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1885; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm11886; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1887; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm01888; X64-SSE2-NEXT: movdqa %xmm0, %xmm11889; X64-SSE2-NEXT: psrld $16, %xmm11890; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm11891; X64-SSE2-NEXT: movd %xmm1, %eax1892; X64-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax1893; X64-SSE2-NEXT: retq1894;1895; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:1896; X64-SSE42: ## %bb.0:1897; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01898; X64-SSE42-NEXT: phminposuw %xmm0, %xmm01899; X64-SSE42-NEXT: movd %xmm0, %eax1900; X64-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF1901; X64-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax1902; X64-SSE42-NEXT: retq1903;1904; X64-AVX1OR2-LABEL: test_reduce_v16i16_v8i16:1905; X64-AVX1OR2: ## %bb.0:1906; X64-AVX1OR2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01907; X64-AVX1OR2-NEXT: vphminposuw %xmm0, %xmm01908; X64-AVX1OR2-NEXT: vmovd %xmm0, %eax1909; X64-AVX1OR2-NEXT: xorl $32767, %eax ## imm = 0x7FFF1910; X64-AVX1OR2-NEXT: ## kill: def $ax killed $ax killed $eax1911; X64-AVX1OR2-NEXT: vzeroupper1912; X64-AVX1OR2-NEXT: retq1913;1914; X64-AVX512-LABEL: test_reduce_v16i16_v8i16:1915; X64-AVX512: ## %bb.0:1916; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm01917; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm01918; X64-AVX512-NEXT: vmovd %xmm0, %eax1919; X64-AVX512-NEXT: xorl $32767, %eax ## imm = 0x7FFF1920; X64-AVX512-NEXT: ## kill: def $ax killed $ax killed $eax1921; X64-AVX512-NEXT: vzeroupper1922; X64-AVX512-NEXT: retq1923 %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>1924 %2 = icmp sgt <16 x i16> %a0, %11925 %3 = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %11926 %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>1927 %5 = icmp sgt <16 x i16> %3, %41928 %6 = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %41929 %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>1930 %8 = icmp sgt <16 x i16> %6, %71931 %9 = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %71932 %10 = extractelement <16 x i16> %9, i32 01933 ret i16 %101934}1935 1936define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {1937; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:1938; X86-SSE2: ## %bb.0:1939; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1940; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm11941; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1942; X86-SSE2-NEXT: pmaxsw %xmm1, %xmm01943; X86-SSE2-NEXT: movdqa %xmm0, %xmm11944; X86-SSE2-NEXT: psrld $16, %xmm11945; X86-SSE2-NEXT: pmaxsw %xmm0, %xmm11946; X86-SSE2-NEXT: movd %xmm1, %eax1947; X86-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax1948; X86-SSE2-NEXT: retl1949;1950; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:1951; X86-SSE42: ## %bb.0:1952; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm01953; X86-SSE42-NEXT: phminposuw %xmm0, %xmm01954; X86-SSE42-NEXT: movd %xmm0, %eax1955; X86-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF1956; X86-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax1957; X86-SSE42-NEXT: retl1958;1959; X86-AVX-LABEL: test_reduce_v32i16_v8i16:1960; X86-AVX: ## %bb.0:1961; X86-AVX-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm01962; X86-AVX-NEXT: vphminposuw %xmm0, %xmm01963; X86-AVX-NEXT: vmovd %xmm0, %eax1964; X86-AVX-NEXT: xorl $32767, %eax ## imm = 0x7FFF1965; X86-AVX-NEXT: ## kill: def $ax killed $ax killed $eax1966; X86-AVX-NEXT: vzeroupper1967; X86-AVX-NEXT: retl1968;1969; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:1970; X64-SSE2: ## %bb.0:1971; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]1972; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm11973; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]1974; X64-SSE2-NEXT: pmaxsw %xmm1, %xmm01975; X64-SSE2-NEXT: movdqa %xmm0, %xmm11976; X64-SSE2-NEXT: psrld $16, %xmm11977; X64-SSE2-NEXT: pmaxsw %xmm0, %xmm11978; X64-SSE2-NEXT: movd %xmm1, %eax1979; X64-SSE2-NEXT: ## kill: def $ax killed $ax killed $eax1980; X64-SSE2-NEXT: retq1981;1982; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:1983; X64-SSE42: ## %bb.0:1984; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01985; X64-SSE42-NEXT: phminposuw %xmm0, %xmm01986; X64-SSE42-NEXT: movd %xmm0, %eax1987; X64-SSE42-NEXT: xorl $32767, %eax ## imm = 0x7FFF1988; X64-SSE42-NEXT: ## kill: def $ax killed $ax killed $eax1989; X64-SSE42-NEXT: retq1990;1991; X64-AVX1OR2-LABEL: test_reduce_v32i16_v8i16:1992; X64-AVX1OR2: ## %bb.0:1993; X64-AVX1OR2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm01994; X64-AVX1OR2-NEXT: vphminposuw %xmm0, %xmm01995; X64-AVX1OR2-NEXT: vmovd %xmm0, %eax1996; X64-AVX1OR2-NEXT: xorl $32767, %eax ## imm = 0x7FFF1997; X64-AVX1OR2-NEXT: ## kill: def $ax killed $ax killed $eax1998; X64-AVX1OR2-NEXT: vzeroupper1999; X64-AVX1OR2-NEXT: retq2000;2001; X64-AVX512-LABEL: test_reduce_v32i16_v8i16:2002; X64-AVX512: ## %bb.0:2003; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm02004; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm02005; X64-AVX512-NEXT: vmovd %xmm0, %eax2006; X64-AVX512-NEXT: xorl $32767, %eax ## imm = 0x7FFF2007; X64-AVX512-NEXT: ## kill: def $ax killed $ax killed $eax2008; X64-AVX512-NEXT: vzeroupper2009; X64-AVX512-NEXT: retq2010 %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>2011 %2 = icmp sgt <32 x i16> %a0, %12012 %3 = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %12013 %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>2014 %5 = icmp sgt <32 x i16> %3, %42015 %6 = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %42016 %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>2017 %8 = icmp sgt <32 x i16> %6, %72018 %9 = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %72019 %10 = extractelement <32 x i16> %9, i32 02020 ret i16 %102021}2022 2023define i8 @test_reduce_v32i8_v16i8(<32 x i8> %a0) {2024; X86-SSE2-LABEL: test_reduce_v32i8_v16i8:2025; X86-SSE2: ## %bb.0:2026; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2027; X86-SSE2-NEXT: movdqa %xmm0, %xmm22028; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm22029; X86-SSE2-NEXT: pand %xmm2, %xmm02030; X86-SSE2-NEXT: pandn %xmm1, %xmm22031; X86-SSE2-NEXT: por %xmm0, %xmm22032; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]2033; X86-SSE2-NEXT: movdqa %xmm2, %xmm02034; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm02035; X86-SSE2-NEXT: pand %xmm0, %xmm22036; X86-SSE2-NEXT: pandn %xmm1, %xmm02037; X86-SSE2-NEXT: por %xmm2, %xmm02038; X86-SSE2-NEXT: movdqa %xmm0, %xmm22039; X86-SSE2-NEXT: psrld $16, %xmm22040; X86-SSE2-NEXT: movdqa %xmm0, %xmm12041; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm12042; X86-SSE2-NEXT: pand %xmm1, %xmm02043; X86-SSE2-NEXT: pandn %xmm2, %xmm12044; X86-SSE2-NEXT: por %xmm0, %xmm12045; X86-SSE2-NEXT: movdqa %xmm1, %xmm02046; X86-SSE2-NEXT: psrlw $8, %xmm02047; X86-SSE2-NEXT: movdqa %xmm1, %xmm22048; X86-SSE2-NEXT: pcmpgtb %xmm0, %xmm22049; X86-SSE2-NEXT: pand %xmm2, %xmm12050; X86-SSE2-NEXT: pandn %xmm0, %xmm22051; X86-SSE2-NEXT: por %xmm1, %xmm22052; X86-SSE2-NEXT: movd %xmm2, %eax2053; X86-SSE2-NEXT: ## kill: def $al killed $al killed $eax2054; X86-SSE2-NEXT: retl2055;2056; X86-SSE42-LABEL: test_reduce_v32i8_v16i8:2057; X86-SSE42: ## %bb.0:2058; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm02059; X86-SSE42-NEXT: movdqa %xmm0, %xmm12060; X86-SSE42-NEXT: psrlw $8, %xmm12061; X86-SSE42-NEXT: pminub %xmm0, %xmm12062; X86-SSE42-NEXT: phminposuw %xmm1, %xmm02063; X86-SSE42-NEXT: movd %xmm0, %eax2064; X86-SSE42-NEXT: xorb $127, %al2065; X86-SSE42-NEXT: ## kill: def $al killed $al killed $eax2066; X86-SSE42-NEXT: retl2067;2068; X86-AVX-LABEL: test_reduce_v32i8_v16i8:2069; X86-AVX: ## %bb.0:2070; X86-AVX-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm02071; X86-AVX-NEXT: vpsrlw $8, %xmm0, %xmm12072; X86-AVX-NEXT: vpminub %xmm1, %xmm0, %xmm02073; X86-AVX-NEXT: vphminposuw %xmm0, %xmm02074; X86-AVX-NEXT: vmovd %xmm0, %eax2075; X86-AVX-NEXT: xorb $127, %al2076; X86-AVX-NEXT: ## kill: def $al killed $al killed $eax2077; X86-AVX-NEXT: vzeroupper2078; X86-AVX-NEXT: retl2079;2080; X64-SSE2-LABEL: test_reduce_v32i8_v16i8:2081; X64-SSE2: ## %bb.0:2082; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2083; X64-SSE2-NEXT: movdqa %xmm0, %xmm22084; X64-SSE2-NEXT: pcmpgtb %xmm1, %xmm22085; X64-SSE2-NEXT: pand %xmm2, %xmm02086; X64-SSE2-NEXT: pandn %xmm1, %xmm22087; X64-SSE2-NEXT: por %xmm0, %xmm22088; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]2089; X64-SSE2-NEXT: movdqa %xmm2, %xmm12090; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm12091; X64-SSE2-NEXT: pand %xmm1, %xmm22092; X64-SSE2-NEXT: pandn %xmm0, %xmm12093; X64-SSE2-NEXT: por %xmm2, %xmm12094; X64-SSE2-NEXT: movdqa %xmm1, %xmm02095; X64-SSE2-NEXT: psrld $16, %xmm02096; X64-SSE2-NEXT: movdqa %xmm1, %xmm22097; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm22098; X64-SSE2-NEXT: pand %xmm2, %xmm12099; X64-SSE2-NEXT: pandn %xmm0, %xmm22100; X64-SSE2-NEXT: por %xmm1, %xmm22101; X64-SSE2-NEXT: movdqa %xmm2, %xmm02102; X64-SSE2-NEXT: psrlw $8, %xmm02103; X64-SSE2-NEXT: movdqa %xmm2, %xmm12104; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm12105; X64-SSE2-NEXT: pand %xmm1, %xmm22106; X64-SSE2-NEXT: pandn %xmm0, %xmm12107; X64-SSE2-NEXT: por %xmm2, %xmm12108; X64-SSE2-NEXT: movd %xmm1, %eax2109; X64-SSE2-NEXT: ## kill: def $al killed $al killed $eax2110; X64-SSE2-NEXT: retq2111;2112; X64-SSE42-LABEL: test_reduce_v32i8_v16i8:2113; X64-SSE42: ## %bb.0:2114; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm02115; X64-SSE42-NEXT: movdqa %xmm0, %xmm12116; X64-SSE42-NEXT: psrlw $8, %xmm12117; X64-SSE42-NEXT: pminub %xmm0, %xmm12118; X64-SSE42-NEXT: phminposuw %xmm1, %xmm02119; X64-SSE42-NEXT: movd %xmm0, %eax2120; X64-SSE42-NEXT: xorb $127, %al2121; X64-SSE42-NEXT: ## kill: def $al killed $al killed $eax2122; X64-SSE42-NEXT: retq2123;2124; X64-AVX1OR2-LABEL: test_reduce_v32i8_v16i8:2125; X64-AVX1OR2: ## %bb.0:2126; X64-AVX1OR2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm02127; X64-AVX1OR2-NEXT: vpsrlw $8, %xmm0, %xmm12128; X64-AVX1OR2-NEXT: vpminub %xmm1, %xmm0, %xmm02129; X64-AVX1OR2-NEXT: vphminposuw %xmm0, %xmm02130; X64-AVX1OR2-NEXT: vmovd %xmm0, %eax2131; X64-AVX1OR2-NEXT: xorb $127, %al2132; X64-AVX1OR2-NEXT: ## kill: def $al killed $al killed $eax2133; X64-AVX1OR2-NEXT: vzeroupper2134; X64-AVX1OR2-NEXT: retq2135;2136; X64-AVX512-LABEL: test_reduce_v32i8_v16i8:2137; X64-AVX512: ## %bb.0:2138; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm02139; X64-AVX512-NEXT: vpsrlw $8, %xmm0, %xmm12140; X64-AVX512-NEXT: vpminub %xmm1, %xmm0, %xmm02141; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm02142; X64-AVX512-NEXT: vmovd %xmm0, %eax2143; X64-AVX512-NEXT: xorb $127, %al2144; X64-AVX512-NEXT: ## kill: def $al killed $al killed $eax2145; X64-AVX512-NEXT: vzeroupper2146; X64-AVX512-NEXT: retq2147 %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>2148 %2 = icmp sgt <32 x i8> %a0, %12149 %3 = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %12150 %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>2151 %5 = icmp sgt <32 x i8> %3, %42152 %6 = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %42153 %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>2154 %8 = icmp sgt <32 x i8> %6, %72155 %9 = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %72156 %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>2157 %11 = icmp sgt <32 x i8> %9, %102158 %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %102159 %13 = extractelement <32 x i8> %12, i32 02160 ret i8 %132161}2162 2163define i8 @test_reduce_v64i8_v16i8(<64 x i8> %a0) {2164; X86-SSE2-LABEL: test_reduce_v64i8_v16i8:2165; X86-SSE2: ## %bb.0:2166; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2167; X86-SSE2-NEXT: movdqa %xmm0, %xmm22168; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm22169; X86-SSE2-NEXT: pand %xmm2, %xmm02170; X86-SSE2-NEXT: pandn %xmm1, %xmm22171; X86-SSE2-NEXT: por %xmm0, %xmm22172; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]2173; X86-SSE2-NEXT: movdqa %xmm2, %xmm02174; X86-SSE2-NEXT: pcmpgtb %xmm1, %xmm02175; X86-SSE2-NEXT: pand %xmm0, %xmm22176; X86-SSE2-NEXT: pandn %xmm1, %xmm02177; X86-SSE2-NEXT: por %xmm2, %xmm02178; X86-SSE2-NEXT: movdqa %xmm0, %xmm22179; X86-SSE2-NEXT: psrld $16, %xmm22180; X86-SSE2-NEXT: movdqa %xmm0, %xmm12181; X86-SSE2-NEXT: pcmpgtb %xmm2, %xmm12182; X86-SSE2-NEXT: pand %xmm1, %xmm02183; X86-SSE2-NEXT: pandn %xmm2, %xmm12184; X86-SSE2-NEXT: por %xmm0, %xmm12185; X86-SSE2-NEXT: movdqa %xmm1, %xmm02186; X86-SSE2-NEXT: psrlw $8, %xmm02187; X86-SSE2-NEXT: movdqa %xmm1, %xmm22188; X86-SSE2-NEXT: pcmpgtb %xmm0, %xmm22189; X86-SSE2-NEXT: pand %xmm2, %xmm12190; X86-SSE2-NEXT: pandn %xmm0, %xmm22191; X86-SSE2-NEXT: por %xmm1, %xmm22192; X86-SSE2-NEXT: movd %xmm2, %eax2193; X86-SSE2-NEXT: ## kill: def $al killed $al killed $eax2194; X86-SSE2-NEXT: retl2195;2196; X86-SSE42-LABEL: test_reduce_v64i8_v16i8:2197; X86-SSE42: ## %bb.0:2198; X86-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm02199; X86-SSE42-NEXT: movdqa %xmm0, %xmm12200; X86-SSE42-NEXT: psrlw $8, %xmm12201; X86-SSE42-NEXT: pminub %xmm0, %xmm12202; X86-SSE42-NEXT: phminposuw %xmm1, %xmm02203; X86-SSE42-NEXT: movd %xmm0, %eax2204; X86-SSE42-NEXT: xorb $127, %al2205; X86-SSE42-NEXT: ## kill: def $al killed $al killed $eax2206; X86-SSE42-NEXT: retl2207;2208; X86-AVX-LABEL: test_reduce_v64i8_v16i8:2209; X86-AVX: ## %bb.0:2210; X86-AVX-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm02211; X86-AVX-NEXT: vpsrlw $8, %xmm0, %xmm12212; X86-AVX-NEXT: vpminub %xmm1, %xmm0, %xmm02213; X86-AVX-NEXT: vphminposuw %xmm0, %xmm02214; X86-AVX-NEXT: vmovd %xmm0, %eax2215; X86-AVX-NEXT: xorb $127, %al2216; X86-AVX-NEXT: ## kill: def $al killed $al killed $eax2217; X86-AVX-NEXT: vzeroupper2218; X86-AVX-NEXT: retl2219;2220; X64-SSE2-LABEL: test_reduce_v64i8_v16i8:2221; X64-SSE2: ## %bb.0:2222; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]2223; X64-SSE2-NEXT: movdqa %xmm0, %xmm22224; X64-SSE2-NEXT: pcmpgtb %xmm1, %xmm22225; X64-SSE2-NEXT: pand %xmm2, %xmm02226; X64-SSE2-NEXT: pandn %xmm1, %xmm22227; X64-SSE2-NEXT: por %xmm0, %xmm22228; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]2229; X64-SSE2-NEXT: movdqa %xmm2, %xmm12230; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm12231; X64-SSE2-NEXT: pand %xmm1, %xmm22232; X64-SSE2-NEXT: pandn %xmm0, %xmm12233; X64-SSE2-NEXT: por %xmm2, %xmm12234; X64-SSE2-NEXT: movdqa %xmm1, %xmm02235; X64-SSE2-NEXT: psrld $16, %xmm02236; X64-SSE2-NEXT: movdqa %xmm1, %xmm22237; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm22238; X64-SSE2-NEXT: pand %xmm2, %xmm12239; X64-SSE2-NEXT: pandn %xmm0, %xmm22240; X64-SSE2-NEXT: por %xmm1, %xmm22241; X64-SSE2-NEXT: movdqa %xmm2, %xmm02242; X64-SSE2-NEXT: psrlw $8, %xmm02243; X64-SSE2-NEXT: movdqa %xmm2, %xmm12244; X64-SSE2-NEXT: pcmpgtb %xmm0, %xmm12245; X64-SSE2-NEXT: pand %xmm1, %xmm22246; X64-SSE2-NEXT: pandn %xmm0, %xmm12247; X64-SSE2-NEXT: por %xmm2, %xmm12248; X64-SSE2-NEXT: movd %xmm1, %eax2249; X64-SSE2-NEXT: ## kill: def $al killed $al killed $eax2250; X64-SSE2-NEXT: retq2251;2252; X64-SSE42-LABEL: test_reduce_v64i8_v16i8:2253; X64-SSE42: ## %bb.0:2254; X64-SSE42-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm02255; X64-SSE42-NEXT: movdqa %xmm0, %xmm12256; X64-SSE42-NEXT: psrlw $8, %xmm12257; X64-SSE42-NEXT: pminub %xmm0, %xmm12258; X64-SSE42-NEXT: phminposuw %xmm1, %xmm02259; X64-SSE42-NEXT: movd %xmm0, %eax2260; X64-SSE42-NEXT: xorb $127, %al2261; X64-SSE42-NEXT: ## kill: def $al killed $al killed $eax2262; X64-SSE42-NEXT: retq2263;2264; X64-AVX1OR2-LABEL: test_reduce_v64i8_v16i8:2265; X64-AVX1OR2: ## %bb.0:2266; X64-AVX1OR2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm02267; X64-AVX1OR2-NEXT: vpsrlw $8, %xmm0, %xmm12268; X64-AVX1OR2-NEXT: vpminub %xmm1, %xmm0, %xmm02269; X64-AVX1OR2-NEXT: vphminposuw %xmm0, %xmm02270; X64-AVX1OR2-NEXT: vmovd %xmm0, %eax2271; X64-AVX1OR2-NEXT: xorb $127, %al2272; X64-AVX1OR2-NEXT: ## kill: def $al killed $al killed $eax2273; X64-AVX1OR2-NEXT: vzeroupper2274; X64-AVX1OR2-NEXT: retq2275;2276; X64-AVX512-LABEL: test_reduce_v64i8_v16i8:2277; X64-AVX512: ## %bb.0:2278; X64-AVX512-NEXT: vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm02279; X64-AVX512-NEXT: vpsrlw $8, %xmm0, %xmm12280; X64-AVX512-NEXT: vpminub %xmm1, %xmm0, %xmm02281; X64-AVX512-NEXT: vphminposuw %xmm0, %xmm02282; X64-AVX512-NEXT: vmovd %xmm0, %eax2283; X64-AVX512-NEXT: xorb $127, %al2284; X64-AVX512-NEXT: ## kill: def $al killed $al killed $eax2285; X64-AVX512-NEXT: vzeroupper2286; X64-AVX512-NEXT: retq2287 %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>2288 %2 = icmp sgt <64 x i8> %a0, %12289 %3 = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %12290 %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>2291 %5 = icmp sgt <64 x i8> %3, %42292 %6 = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %42293 %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>2294 %8 = icmp sgt <64 x i8> %6, %72295 %9 = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %72296 %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>2297 %11 = icmp sgt <64 x i8> %9, %102298 %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %102299 %13 = extractelement <64 x i8> %12, i32 02300 ret i8 %132301}2302