606 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=SSE3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX12,AVX14; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX12,AVX25; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512F6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=AVX512BW7 8define i8 @v8i64(<8 x i64> %a, <8 x i64> %b, <8 x i64> %c, <8 x i64> %d) {9; SSE-LABEL: v8i64:10; SSE: # %bb.0:11; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm812; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm913; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm1014; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm1115; SSE-NEXT: pcmpgtq %xmm7, %xmm316; SSE-NEXT: pcmpgtq %xmm6, %xmm217; SSE-NEXT: packssdw %xmm3, %xmm218; SSE-NEXT: pcmpgtq %xmm5, %xmm119; SSE-NEXT: pcmpgtq %xmm4, %xmm020; SSE-NEXT: packssdw %xmm1, %xmm021; SSE-NEXT: packssdw %xmm2, %xmm022; SSE-NEXT: pcmpgtq {{[0-9]+}}(%rsp), %xmm1123; SSE-NEXT: pcmpgtq {{[0-9]+}}(%rsp), %xmm1024; SSE-NEXT: packssdw %xmm11, %xmm1025; SSE-NEXT: pcmpgtq {{[0-9]+}}(%rsp), %xmm926; SSE-NEXT: pcmpgtq {{[0-9]+}}(%rsp), %xmm827; SSE-NEXT: packssdw %xmm9, %xmm828; SSE-NEXT: packssdw %xmm10, %xmm829; SSE-NEXT: pand %xmm0, %xmm830; SSE-NEXT: packsswb %xmm8, %xmm831; SSE-NEXT: pmovmskb %xmm8, %eax32; SSE-NEXT: # kill: def $al killed $al killed $eax33; SSE-NEXT: retq34;35; AVX1-LABEL: v8i64:36; AVX1: # %bb.0:37; AVX1-NEXT: vextractf128 $1, %ymm7, %xmm838; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm939; AVX1-NEXT: vpcmpgtq %xmm8, %xmm9, %xmm840; AVX1-NEXT: vpcmpgtq %xmm7, %xmm5, %xmm541; AVX1-NEXT: vpackssdw %xmm8, %xmm5, %xmm542; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm743; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm844; AVX1-NEXT: vpcmpgtq %xmm7, %xmm8, %xmm745; AVX1-NEXT: vpcmpgtq %xmm6, %xmm4, %xmm446; AVX1-NEXT: vpackssdw %xmm7, %xmm4, %xmm447; AVX1-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm448; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm549; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm650; AVX1-NEXT: vpcmpgtq %xmm5, %xmm6, %xmm551; AVX1-NEXT: vpcmpgtq %xmm3, %xmm1, %xmm152; AVX1-NEXT: vpackssdw %xmm5, %xmm1, %xmm153; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm354; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm555; AVX1-NEXT: vpcmpgtq %xmm3, %xmm5, %xmm356; AVX1-NEXT: vpcmpgtq %xmm2, %xmm0, %xmm057; AVX1-NEXT: vpackssdw %xmm3, %xmm0, %xmm058; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm059; AVX1-NEXT: vandps %ymm4, %ymm0, %ymm060; AVX1-NEXT: vmovmskps %ymm0, %eax61; AVX1-NEXT: # kill: def $al killed $al killed $eax62; AVX1-NEXT: vzeroupper63; AVX1-NEXT: retq64;65; AVX2-LABEL: v8i64:66; AVX2: # %bb.0:67; AVX2-NEXT: vpcmpgtq %ymm7, %ymm5, %ymm568; AVX2-NEXT: vpcmpgtq %ymm3, %ymm1, %ymm169; AVX2-NEXT: vpand %ymm5, %ymm1, %ymm170; AVX2-NEXT: vpcmpgtq %ymm6, %ymm4, %ymm371; AVX2-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm072; AVX2-NEXT: vpand %ymm3, %ymm0, %ymm073; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm074; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]75; AVX2-NEXT: vmovmskps %ymm0, %eax76; AVX2-NEXT: # kill: def $al killed $al killed $eax77; AVX2-NEXT: vzeroupper78; AVX2-NEXT: retq79;80; AVX512F-LABEL: v8i64:81; AVX512F: # %bb.0:82; AVX512F-NEXT: vpcmpgtq %zmm1, %zmm0, %k183; AVX512F-NEXT: vpcmpgtq %zmm3, %zmm2, %k0 {%k1}84; AVX512F-NEXT: kmovw %k0, %eax85; AVX512F-NEXT: # kill: def $al killed $al killed $eax86; AVX512F-NEXT: vzeroupper87; AVX512F-NEXT: retq88;89; AVX512BW-LABEL: v8i64:90; AVX512BW: # %bb.0:91; AVX512BW-NEXT: vpcmpgtq %zmm1, %zmm0, %k192; AVX512BW-NEXT: vpcmpgtq %zmm3, %zmm2, %k0 {%k1}93; AVX512BW-NEXT: kmovd %k0, %eax94; AVX512BW-NEXT: # kill: def $al killed $al killed $eax95; AVX512BW-NEXT: vzeroupper96; AVX512BW-NEXT: retq97 %x0 = icmp sgt <8 x i64> %a, %b98 %x1 = icmp sgt <8 x i64> %c, %d99 %y = and <8 x i1> %x0, %x1100 %res = bitcast <8 x i1> %y to i8101 ret i8 %res102}103 104define i8 @v8f64(<8 x double> %a, <8 x double> %b, <8 x double> %c, <8 x double> %d) {105; SSE-LABEL: v8f64:106; SSE: # %bb.0:107; SSE-NEXT: movapd {{[0-9]+}}(%rsp), %xmm8108; SSE-NEXT: movapd {{[0-9]+}}(%rsp), %xmm9109; SSE-NEXT: movapd {{[0-9]+}}(%rsp), %xmm10110; SSE-NEXT: movapd {{[0-9]+}}(%rsp), %xmm11111; SSE-NEXT: cmpltpd %xmm3, %xmm7112; SSE-NEXT: cmpltpd %xmm2, %xmm6113; SSE-NEXT: packssdw %xmm7, %xmm6114; SSE-NEXT: cmpltpd %xmm1, %xmm5115; SSE-NEXT: cmpltpd %xmm0, %xmm4116; SSE-NEXT: packssdw %xmm5, %xmm4117; SSE-NEXT: packssdw %xmm6, %xmm4118; SSE-NEXT: cmpltpd {{[0-9]+}}(%rsp), %xmm11119; SSE-NEXT: cmpltpd {{[0-9]+}}(%rsp), %xmm10120; SSE-NEXT: packssdw %xmm11, %xmm10121; SSE-NEXT: cmpltpd {{[0-9]+}}(%rsp), %xmm9122; SSE-NEXT: cmpltpd {{[0-9]+}}(%rsp), %xmm8123; SSE-NEXT: packssdw %xmm9, %xmm8124; SSE-NEXT: packssdw %xmm10, %xmm8125; SSE-NEXT: pand %xmm4, %xmm8126; SSE-NEXT: packsswb %xmm8, %xmm8127; SSE-NEXT: pmovmskb %xmm8, %eax128; SSE-NEXT: # kill: def $al killed $al killed $eax129; SSE-NEXT: retq130;131; AVX1-LABEL: v8f64:132; AVX1: # %bb.0:133; AVX1-NEXT: vcmpltpd %ymm5, %ymm7, %ymm5134; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm7135; AVX1-NEXT: vpackssdw %xmm7, %xmm5, %xmm5136; AVX1-NEXT: vcmpltpd %ymm4, %ymm6, %ymm4137; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm6138; AVX1-NEXT: vpackssdw %xmm6, %xmm4, %xmm4139; AVX1-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4140; AVX1-NEXT: vcmpltpd %ymm1, %ymm3, %ymm1141; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3142; AVX1-NEXT: vpackssdw %xmm3, %xmm1, %xmm1143; AVX1-NEXT: vcmpltpd %ymm0, %ymm2, %ymm0144; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2145; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0146; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0147; AVX1-NEXT: vandps %ymm4, %ymm0, %ymm0148; AVX1-NEXT: vmovmskps %ymm0, %eax149; AVX1-NEXT: # kill: def $al killed $al killed $eax150; AVX1-NEXT: vzeroupper151; AVX1-NEXT: retq152;153; AVX2-LABEL: v8f64:154; AVX2: # %bb.0:155; AVX2-NEXT: vcmpltpd %ymm5, %ymm7, %ymm5156; AVX2-NEXT: vcmpltpd %ymm1, %ymm3, %ymm1157; AVX2-NEXT: vandpd %ymm5, %ymm1, %ymm1158; AVX2-NEXT: vcmpltpd %ymm4, %ymm6, %ymm3159; AVX2-NEXT: vcmpltpd %ymm0, %ymm2, %ymm0160; AVX2-NEXT: vandpd %ymm3, %ymm0, %ymm0161; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0162; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]163; AVX2-NEXT: vmovmskps %ymm0, %eax164; AVX2-NEXT: # kill: def $al killed $al killed $eax165; AVX2-NEXT: vzeroupper166; AVX2-NEXT: retq167;168; AVX512F-LABEL: v8f64:169; AVX512F: # %bb.0:170; AVX512F-NEXT: vcmpltpd %zmm0, %zmm1, %k1171; AVX512F-NEXT: vcmpltpd %zmm2, %zmm3, %k0 {%k1}172; AVX512F-NEXT: kmovw %k0, %eax173; AVX512F-NEXT: # kill: def $al killed $al killed $eax174; AVX512F-NEXT: vzeroupper175; AVX512F-NEXT: retq176;177; AVX512BW-LABEL: v8f64:178; AVX512BW: # %bb.0:179; AVX512BW-NEXT: vcmpltpd %zmm0, %zmm1, %k1180; AVX512BW-NEXT: vcmpltpd %zmm2, %zmm3, %k0 {%k1}181; AVX512BW-NEXT: kmovd %k0, %eax182; AVX512BW-NEXT: # kill: def $al killed $al killed $eax183; AVX512BW-NEXT: vzeroupper184; AVX512BW-NEXT: retq185 %x0 = fcmp ogt <8 x double> %a, %b186 %x1 = fcmp ogt <8 x double> %c, %d187 %y = and <8 x i1> %x0, %x1188 %res = bitcast <8 x i1> %y to i8189 ret i8 %res190}191 192define i32 @v32i16(<32 x i16> %a, <32 x i16> %b, <32 x i16> %c, <32 x i16> %d) {193; SSE-LABEL: v32i16:194; SSE: # %bb.0:195; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm8196; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm9197; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm10198; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm11199; SSE-NEXT: pcmpgtw %xmm4, %xmm0200; SSE-NEXT: pcmpgtw %xmm5, %xmm1201; SSE-NEXT: pcmpgtw %xmm6, %xmm2202; SSE-NEXT: pcmpgtw %xmm7, %xmm3203; SSE-NEXT: pcmpgtw {{[0-9]+}}(%rsp), %xmm11204; SSE-NEXT: pand %xmm0, %xmm11205; SSE-NEXT: pcmpgtw {{[0-9]+}}(%rsp), %xmm10206; SSE-NEXT: pand %xmm1, %xmm10207; SSE-NEXT: packsswb %xmm10, %xmm11208; SSE-NEXT: pcmpgtw {{[0-9]+}}(%rsp), %xmm9209; SSE-NEXT: pand %xmm2, %xmm9210; SSE-NEXT: pcmpgtw {{[0-9]+}}(%rsp), %xmm8211; SSE-NEXT: pand %xmm3, %xmm8212; SSE-NEXT: packsswb %xmm8, %xmm9213; SSE-NEXT: pmovmskb %xmm11, %ecx214; SSE-NEXT: pmovmskb %xmm9, %eax215; SSE-NEXT: shll $16, %eax216; SSE-NEXT: orl %ecx, %eax217; SSE-NEXT: retq218;219; AVX1-LABEL: v32i16:220; AVX1: # %bb.0:221; AVX1-NEXT: vpcmpgtw %xmm3, %xmm1, %xmm8222; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3223; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1224; AVX1-NEXT: vpcmpgtw %xmm3, %xmm1, %xmm1225; AVX1-NEXT: vpcmpgtw %xmm2, %xmm0, %xmm3226; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2227; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0228; AVX1-NEXT: vpcmpgtw %xmm2, %xmm0, %xmm0229; AVX1-NEXT: vpcmpgtw %xmm7, %xmm5, %xmm2230; AVX1-NEXT: vpand %xmm2, %xmm8, %xmm2231; AVX1-NEXT: vextractf128 $1, %ymm7, %xmm7232; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm5233; AVX1-NEXT: vpcmpgtw %xmm7, %xmm5, %xmm5234; AVX1-NEXT: vpand %xmm5, %xmm1, %xmm1235; AVX1-NEXT: vpacksswb %xmm1, %xmm2, %xmm1236; AVX1-NEXT: vpcmpgtw %xmm6, %xmm4, %xmm2237; AVX1-NEXT: vpand %xmm2, %xmm3, %xmm2238; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm3239; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm4240; AVX1-NEXT: vpcmpgtw %xmm3, %xmm4, %xmm3241; AVX1-NEXT: vpand %xmm3, %xmm0, %xmm0242; AVX1-NEXT: vpacksswb %xmm0, %xmm2, %xmm0243; AVX1-NEXT: vpmovmskb %xmm0, %ecx244; AVX1-NEXT: vpmovmskb %xmm1, %eax245; AVX1-NEXT: shll $16, %eax246; AVX1-NEXT: orl %ecx, %eax247; AVX1-NEXT: vzeroupper248; AVX1-NEXT: retq249;250; AVX2-LABEL: v32i16:251; AVX2: # %bb.0:252; AVX2-NEXT: vpcmpgtw %ymm2, %ymm0, %ymm0253; AVX2-NEXT: vpcmpgtw %ymm3, %ymm1, %ymm1254; AVX2-NEXT: vpcmpgtw %ymm6, %ymm4, %ymm2255; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0256; AVX2-NEXT: vpcmpgtw %ymm7, %ymm5, %ymm2257; AVX2-NEXT: vpand %ymm2, %ymm1, %ymm1258; AVX2-NEXT: vpacksswb %ymm1, %ymm0, %ymm0259; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]260; AVX2-NEXT: vpmovmskb %ymm0, %eax261; AVX2-NEXT: vzeroupper262; AVX2-NEXT: retq263;264; AVX512F-LABEL: v32i16:265; AVX512F: # %bb.0:266; AVX512F-NEXT: vextracti64x4 $1, %zmm1, %ymm4267; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm5268; AVX512F-NEXT: vpcmpgtw %ymm4, %ymm5, %ymm4269; AVX512F-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0270; AVX512F-NEXT: vextracti64x4 $1, %zmm3, %ymm1271; AVX512F-NEXT: vextracti64x4 $1, %zmm2, %ymm5272; AVX512F-NEXT: vpcmpgtw %ymm1, %ymm5, %ymm1273; AVX512F-NEXT: vpand %ymm1, %ymm4, %ymm1274; AVX512F-NEXT: vpcmpgtw %ymm3, %ymm2, %ymm2275; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm0276; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0277; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0278; AVX512F-NEXT: kmovw %k0, %ecx279; AVX512F-NEXT: vpmovsxwd %ymm1, %zmm0280; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0281; AVX512F-NEXT: kmovw %k0, %eax282; AVX512F-NEXT: shll $16, %eax283; AVX512F-NEXT: orl %ecx, %eax284; AVX512F-NEXT: vzeroupper285; AVX512F-NEXT: retq286;287; AVX512BW-LABEL: v32i16:288; AVX512BW: # %bb.0:289; AVX512BW-NEXT: vpcmpgtw %zmm1, %zmm0, %k1290; AVX512BW-NEXT: vpcmpgtw %zmm3, %zmm2, %k0 {%k1}291; AVX512BW-NEXT: kmovd %k0, %eax292; AVX512BW-NEXT: vzeroupper293; AVX512BW-NEXT: retq294 %x0 = icmp sgt <32 x i16> %a, %b295 %x1 = icmp sgt <32 x i16> %c, %d296 %y = and <32 x i1> %x0, %x1297 %res = bitcast <32 x i1> %y to i32298 ret i32 %res299}300 301define i16 @v16i32(<16 x i32> %a, <16 x i32> %b, <16 x i32> %c, <16 x i32> %d) {302; SSE-LABEL: v16i32:303; SSE: # %bb.0:304; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm8305; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm9306; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm10307; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm11308; SSE-NEXT: pcmpgtd %xmm4, %xmm0309; SSE-NEXT: pcmpgtd %xmm5, %xmm1310; SSE-NEXT: pcmpgtd %xmm6, %xmm2311; SSE-NEXT: pcmpgtd %xmm7, %xmm3312; SSE-NEXT: pcmpgtd {{[0-9]+}}(%rsp), %xmm11313; SSE-NEXT: pand %xmm0, %xmm11314; SSE-NEXT: pcmpgtd {{[0-9]+}}(%rsp), %xmm10315; SSE-NEXT: pand %xmm1, %xmm10316; SSE-NEXT: packssdw %xmm10, %xmm11317; SSE-NEXT: pcmpgtd {{[0-9]+}}(%rsp), %xmm9318; SSE-NEXT: pand %xmm2, %xmm9319; SSE-NEXT: pcmpgtd {{[0-9]+}}(%rsp), %xmm8320; SSE-NEXT: pand %xmm3, %xmm8321; SSE-NEXT: packssdw %xmm8, %xmm9322; SSE-NEXT: packsswb %xmm9, %xmm11323; SSE-NEXT: pmovmskb %xmm11, %eax324; SSE-NEXT: # kill: def $ax killed $ax killed $eax325; SSE-NEXT: retq326;327; AVX1-LABEL: v16i32:328; AVX1: # %bb.0:329; AVX1-NEXT: vpcmpgtd %xmm2, %xmm0, %xmm8330; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2331; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0332; AVX1-NEXT: vpcmpgtd %xmm2, %xmm0, %xmm0333; AVX1-NEXT: vpcmpgtd %xmm3, %xmm1, %xmm2334; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3335; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1336; AVX1-NEXT: vpcmpgtd %xmm3, %xmm1, %xmm1337; AVX1-NEXT: vpcmpgtd %xmm6, %xmm4, %xmm3338; AVX1-NEXT: vpand %xmm3, %xmm8, %xmm3339; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm6340; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm4341; AVX1-NEXT: vpcmpgtd %xmm6, %xmm4, %xmm4342; AVX1-NEXT: vpand %xmm4, %xmm0, %xmm0343; AVX1-NEXT: vpackssdw %xmm0, %xmm3, %xmm0344; AVX1-NEXT: vpcmpgtd %xmm7, %xmm5, %xmm3345; AVX1-NEXT: vpand %xmm3, %xmm2, %xmm2346; AVX1-NEXT: vextractf128 $1, %ymm7, %xmm3347; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm4348; AVX1-NEXT: vpcmpgtd %xmm3, %xmm4, %xmm3349; AVX1-NEXT: vpand %xmm3, %xmm1, %xmm1350; AVX1-NEXT: vpackssdw %xmm1, %xmm2, %xmm1351; AVX1-NEXT: vpacksswb %xmm1, %xmm0, %xmm0352; AVX1-NEXT: vpmovmskb %xmm0, %eax353; AVX1-NEXT: # kill: def $ax killed $ax killed $eax354; AVX1-NEXT: vzeroupper355; AVX1-NEXT: retq356;357; AVX2-LABEL: v16i32:358; AVX2: # %bb.0:359; AVX2-NEXT: vpcmpgtd %ymm3, %ymm1, %ymm1360; AVX2-NEXT: vpcmpgtd %ymm2, %ymm0, %ymm0361; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0362; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1363; AVX2-NEXT: vpcmpgtd %ymm7, %ymm5, %ymm2364; AVX2-NEXT: vpcmpgtd %ymm6, %ymm4, %ymm3365; AVX2-NEXT: vpackssdw %ymm2, %ymm3, %ymm2366; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3367; AVX2-NEXT: vpand %xmm3, %xmm1, %xmm1368; AVX2-NEXT: vpand %xmm2, %xmm0, %xmm0369; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0370; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]371; AVX2-NEXT: vpmovmskb %xmm0, %eax372; AVX2-NEXT: # kill: def $ax killed $ax killed $eax373; AVX2-NEXT: vzeroupper374; AVX2-NEXT: retq375;376; AVX512F-LABEL: v16i32:377; AVX512F: # %bb.0:378; AVX512F-NEXT: vpcmpgtd %zmm1, %zmm0, %k1379; AVX512F-NEXT: vpcmpgtd %zmm3, %zmm2, %k0 {%k1}380; AVX512F-NEXT: kmovw %k0, %eax381; AVX512F-NEXT: # kill: def $ax killed $ax killed $eax382; AVX512F-NEXT: vzeroupper383; AVX512F-NEXT: retq384;385; AVX512BW-LABEL: v16i32:386; AVX512BW: # %bb.0:387; AVX512BW-NEXT: vpcmpgtd %zmm1, %zmm0, %k1388; AVX512BW-NEXT: vpcmpgtd %zmm3, %zmm2, %k0 {%k1}389; AVX512BW-NEXT: kmovd %k0, %eax390; AVX512BW-NEXT: # kill: def $ax killed $ax killed $eax391; AVX512BW-NEXT: vzeroupper392; AVX512BW-NEXT: retq393 %x0 = icmp sgt <16 x i32> %a, %b394 %x1 = icmp sgt <16 x i32> %c, %d395 %y = and <16 x i1> %x0, %x1396 %res = bitcast <16 x i1> %y to i16397 ret i16 %res398}399 400define i16 @v16f32(<16 x float> %a, <16 x float> %b, <16 x float> %c, <16 x float> %d) {401; SSE-LABEL: v16f32:402; SSE: # %bb.0:403; SSE-NEXT: movaps {{[0-9]+}}(%rsp), %xmm8404; SSE-NEXT: movaps {{[0-9]+}}(%rsp), %xmm9405; SSE-NEXT: movaps {{[0-9]+}}(%rsp), %xmm10406; SSE-NEXT: movaps {{[0-9]+}}(%rsp), %xmm11407; SSE-NEXT: cmpltps %xmm0, %xmm4408; SSE-NEXT: cmpltps %xmm1, %xmm5409; SSE-NEXT: cmpltps %xmm2, %xmm6410; SSE-NEXT: cmpltps %xmm3, %xmm7411; SSE-NEXT: cmpltps {{[0-9]+}}(%rsp), %xmm11412; SSE-NEXT: andps %xmm4, %xmm11413; SSE-NEXT: cmpltps {{[0-9]+}}(%rsp), %xmm10414; SSE-NEXT: andps %xmm5, %xmm10415; SSE-NEXT: packssdw %xmm10, %xmm11416; SSE-NEXT: cmpltps {{[0-9]+}}(%rsp), %xmm9417; SSE-NEXT: andps %xmm6, %xmm9418; SSE-NEXT: cmpltps {{[0-9]+}}(%rsp), %xmm8419; SSE-NEXT: andps %xmm7, %xmm8420; SSE-NEXT: packssdw %xmm8, %xmm9421; SSE-NEXT: packsswb %xmm9, %xmm11422; SSE-NEXT: pmovmskb %xmm11, %eax423; SSE-NEXT: # kill: def $ax killed $ax killed $eax424; SSE-NEXT: retq425;426; AVX1-LABEL: v16f32:427; AVX1: # %bb.0:428; AVX1-NEXT: vcmpltps %ymm0, %ymm2, %ymm0429; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2430; AVX1-NEXT: vcmpltps %ymm1, %ymm3, %ymm1431; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3432; AVX1-NEXT: vcmpltps %ymm4, %ymm6, %ymm4433; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm6434; AVX1-NEXT: vpand %xmm6, %xmm2, %xmm2435; AVX1-NEXT: vcmpltps %ymm5, %ymm7, %ymm5436; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm6437; AVX1-NEXT: vpand %xmm6, %xmm3, %xmm3438; AVX1-NEXT: vandps %xmm5, %xmm1, %xmm1439; AVX1-NEXT: vpackssdw %xmm3, %xmm1, %xmm1440; AVX1-NEXT: vandps %xmm4, %xmm0, %xmm0441; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0442; AVX1-NEXT: vpacksswb %xmm1, %xmm0, %xmm0443; AVX1-NEXT: vpmovmskb %xmm0, %eax444; AVX1-NEXT: # kill: def $ax killed $ax killed $eax445; AVX1-NEXT: vzeroupper446; AVX1-NEXT: retq447;448; AVX2-LABEL: v16f32:449; AVX2: # %bb.0:450; AVX2-NEXT: vcmpltps %ymm1, %ymm3, %ymm1451; AVX2-NEXT: vcmpltps %ymm0, %ymm2, %ymm0452; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0453; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1454; AVX2-NEXT: vcmpltps %ymm5, %ymm7, %ymm2455; AVX2-NEXT: vcmpltps %ymm4, %ymm6, %ymm3456; AVX2-NEXT: vpackssdw %ymm2, %ymm3, %ymm2457; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3458; AVX2-NEXT: vpand %xmm3, %xmm1, %xmm1459; AVX2-NEXT: vpand %xmm2, %xmm0, %xmm0460; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0461; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]462; AVX2-NEXT: vpmovmskb %xmm0, %eax463; AVX2-NEXT: # kill: def $ax killed $ax killed $eax464; AVX2-NEXT: vzeroupper465; AVX2-NEXT: retq466;467; AVX512F-LABEL: v16f32:468; AVX512F: # %bb.0:469; AVX512F-NEXT: vcmpltps %zmm0, %zmm1, %k1470; AVX512F-NEXT: vcmpltps %zmm2, %zmm3, %k0 {%k1}471; AVX512F-NEXT: kmovw %k0, %eax472; AVX512F-NEXT: # kill: def $ax killed $ax killed $eax473; AVX512F-NEXT: vzeroupper474; AVX512F-NEXT: retq475;476; AVX512BW-LABEL: v16f32:477; AVX512BW: # %bb.0:478; AVX512BW-NEXT: vcmpltps %zmm0, %zmm1, %k1479; AVX512BW-NEXT: vcmpltps %zmm2, %zmm3, %k0 {%k1}480; AVX512BW-NEXT: kmovd %k0, %eax481; AVX512BW-NEXT: # kill: def $ax killed $ax killed $eax482; AVX512BW-NEXT: vzeroupper483; AVX512BW-NEXT: retq484 %x0 = fcmp ogt <16 x float> %a, %b485 %x1 = fcmp ogt <16 x float> %c, %d486 %y = and <16 x i1> %x0, %x1487 %res = bitcast <16 x i1> %y to i16488 ret i16 %res489}490 491define i64 @v64i8(<64 x i8> %a, <64 x i8> %b, <64 x i8> %c, <64 x i8> %d) {492; SSE-LABEL: v64i8:493; SSE: # %bb.0:494; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm8495; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm9496; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm10497; SSE-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm11498; SSE-NEXT: pcmpgtb %xmm4, %xmm0499; SSE-NEXT: pcmpgtb %xmm5, %xmm1500; SSE-NEXT: pcmpgtb %xmm6, %xmm2501; SSE-NEXT: pcmpgtb %xmm7, %xmm3502; SSE-NEXT: pcmpgtb {{[0-9]+}}(%rsp), %xmm11503; SSE-NEXT: pand %xmm0, %xmm11504; SSE-NEXT: pcmpgtb {{[0-9]+}}(%rsp), %xmm10505; SSE-NEXT: pand %xmm1, %xmm10506; SSE-NEXT: pcmpgtb {{[0-9]+}}(%rsp), %xmm9507; SSE-NEXT: pand %xmm2, %xmm9508; SSE-NEXT: pcmpgtb {{[0-9]+}}(%rsp), %xmm8509; SSE-NEXT: pand %xmm3, %xmm8510; SSE-NEXT: pmovmskb %xmm11, %eax511; SSE-NEXT: pmovmskb %xmm10, %ecx512; SSE-NEXT: shll $16, %ecx513; SSE-NEXT: orl %eax, %ecx514; SSE-NEXT: pmovmskb %xmm9, %edx515; SSE-NEXT: pmovmskb %xmm8, %eax516; SSE-NEXT: shll $16, %eax517; SSE-NEXT: orl %edx, %eax518; SSE-NEXT: shlq $32, %rax519; SSE-NEXT: orq %rcx, %rax520; SSE-NEXT: retq521;522; AVX1-LABEL: v64i8:523; AVX1: # %bb.0:524; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm8525; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm9526; AVX1-NEXT: vpcmpgtb %xmm8, %xmm9, %xmm8527; AVX1-NEXT: vpcmpgtb %xmm3, %xmm1, %xmm1528; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3529; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm9530; AVX1-NEXT: vpcmpgtb %xmm3, %xmm9, %xmm3531; AVX1-NEXT: vpcmpgtb %xmm2, %xmm0, %xmm0532; AVX1-NEXT: vextractf128 $1, %ymm7, %xmm2533; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm9534; AVX1-NEXT: vpcmpgtb %xmm2, %xmm9, %xmm2535; AVX1-NEXT: vpand %xmm2, %xmm8, %xmm2536; AVX1-NEXT: vpcmpgtb %xmm7, %xmm5, %xmm5537; AVX1-NEXT: vpand %xmm5, %xmm1, %xmm1538; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm5539; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm7540; AVX1-NEXT: vpcmpgtb %xmm5, %xmm7, %xmm5541; AVX1-NEXT: vpand %xmm5, %xmm3, %xmm3542; AVX1-NEXT: vpcmpgtb %xmm6, %xmm4, %xmm4543; AVX1-NEXT: vpand %xmm4, %xmm0, %xmm0544; AVX1-NEXT: vpmovmskb %xmm0, %eax545; AVX1-NEXT: vpmovmskb %xmm3, %ecx546; AVX1-NEXT: shll $16, %ecx547; AVX1-NEXT: orl %eax, %ecx548; AVX1-NEXT: vpmovmskb %xmm1, %edx549; AVX1-NEXT: vpmovmskb %xmm2, %eax550; AVX1-NEXT: shll $16, %eax551; AVX1-NEXT: orl %edx, %eax552; AVX1-NEXT: shlq $32, %rax553; AVX1-NEXT: orq %rcx, %rax554; AVX1-NEXT: vzeroupper555; AVX1-NEXT: retq556;557; AVX2-LABEL: v64i8:558; AVX2: # %bb.0:559; AVX2-NEXT: vpcmpgtb %ymm2, %ymm0, %ymm0560; AVX2-NEXT: vpcmpgtb %ymm3, %ymm1, %ymm1561; AVX2-NEXT: vpcmpgtb %ymm6, %ymm4, %ymm2562; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0563; AVX2-NEXT: vpcmpgtb %ymm7, %ymm5, %ymm2564; AVX2-NEXT: vpand %ymm2, %ymm1, %ymm1565; AVX2-NEXT: vpmovmskb %ymm0, %ecx566; AVX2-NEXT: vpmovmskb %ymm1, %eax567; AVX2-NEXT: shlq $32, %rax568; AVX2-NEXT: orq %rcx, %rax569; AVX2-NEXT: vzeroupper570; AVX2-NEXT: retq571;572; AVX512F-LABEL: v64i8:573; AVX512F: # %bb.0:574; AVX512F-NEXT: vextracti64x4 $1, %zmm1, %ymm4575; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm5576; AVX512F-NEXT: vpcmpgtb %ymm4, %ymm5, %ymm4577; AVX512F-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0578; AVX512F-NEXT: vextracti64x4 $1, %zmm3, %ymm1579; AVX512F-NEXT: vextracti64x4 $1, %zmm2, %ymm5580; AVX512F-NEXT: vpcmpgtb %ymm1, %ymm5, %ymm1581; AVX512F-NEXT: vpand %ymm1, %ymm4, %ymm1582; AVX512F-NEXT: vpcmpgtb %ymm3, %ymm2, %ymm2583; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm0584; AVX512F-NEXT: vpmovmskb %ymm0, %ecx585; AVX512F-NEXT: vpmovmskb %ymm1, %eax586; AVX512F-NEXT: shlq $32, %rax587; AVX512F-NEXT: orq %rcx, %rax588; AVX512F-NEXT: vzeroupper589; AVX512F-NEXT: retq590;591; AVX512BW-LABEL: v64i8:592; AVX512BW: # %bb.0:593; AVX512BW-NEXT: vpcmpgtb %zmm1, %zmm0, %k1594; AVX512BW-NEXT: vpcmpgtb %zmm3, %zmm2, %k0 {%k1}595; AVX512BW-NEXT: kmovq %k0, %rax596; AVX512BW-NEXT: vzeroupper597; AVX512BW-NEXT: retq598 %x0 = icmp sgt <64 x i8> %a, %b599 %x1 = icmp sgt <64 x i8> %c, %d600 %y = and <64 x i1> %x0, %x1601 %res = bitcast <64 x i1> %y to i64602 ret i64 %res603}604;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:605; AVX12: {{.*}}606