195 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mcpu=skx | FileCheck %s --check-prefixes=CHECK,CHECK-SKX3; RUN: llc < %s -mcpu=knl | FileCheck %s --check-prefixes=CHECK,CHECK-KNL4 5target triple = "x86_64-unknown-unknown"6 7define <8 x i64> @test1(<8 x i64> %m, <8 x i64> %a, <8 x i64> %b) {8; CHECK-SKX-LABEL: test1:9; CHECK-SKX: # %bb.0: # %entry10; CHECK-SKX-NEXT: vpsllq $63, %zmm0, %zmm011; CHECK-SKX-NEXT: vpmovq2m %zmm0, %k112; CHECK-SKX-NEXT: vpblendmq %zmm1, %zmm2, %zmm0 {%k1}13; CHECK-SKX-NEXT: retq14;15; CHECK-KNL-LABEL: test1:16; CHECK-KNL: # %bb.0: # %entry17; CHECK-KNL-NEXT: vpsllq $63, %zmm0, %zmm018; CHECK-KNL-NEXT: vptestmq %zmm0, %zmm0, %k119; CHECK-KNL-NEXT: vpblendmq %zmm1, %zmm2, %zmm0 {%k1}20; CHECK-KNL-NEXT: retq21entry:22 %m.trunc = trunc <8 x i64> %m to <8 x i1>23 %ret = select <8 x i1> %m.trunc, <8 x i64> %a, <8 x i64> %b24 ret <8 x i64> %ret25}26 27; This is a very contrived test case to trick the legalizer into splitting the28; v16i1 masks in the select during type legalization, and in so doing extend them29; into two v8i64 types. This lets us ensure that the lowering code can handle30; both formulations of vselect. All of this trickery is because we can't31; directly form an SDAG input to the lowering.32define <16 x double> @test2(<16 x float> %x, <16 x float> %y, <16 x double> %a, <16 x double> %b) {33; CHECK-LABEL: test2:34; CHECK: # %bb.0: # %entry35; CHECK-NEXT: vxorps %xmm6, %xmm6, %xmm636; CHECK-NEXT: vcmpltps %zmm0, %zmm6, %k037; CHECK-NEXT: vcmpltps %zmm6, %zmm1, %k138; CHECK-NEXT: korw %k1, %k0, %k139; CHECK-NEXT: vblendmpd %zmm2, %zmm4, %zmm0 {%k1}40; CHECK-NEXT: kshiftrw $8, %k1, %k141; CHECK-NEXT: vblendmpd %zmm3, %zmm5, %zmm1 {%k1}42; CHECK-NEXT: retq43entry:44 %gt.m = fcmp ogt <16 x float> %x, zeroinitializer45 %lt.m = fcmp olt <16 x float> %y, zeroinitializer46 %m.or = or <16 x i1> %gt.m, %lt.m47 %ret = select <16 x i1> %m.or, <16 x double> %a, <16 x double> %b48 ret <16 x double> %ret49}50 51define <16 x i64> @test3(<16 x i8> %x, <16 x i64> %a, <16 x i64> %b) {52; CHECK-SKX-LABEL: test3:53; CHECK-SKX: # %bb.0:54; CHECK-SKX-NEXT: vptestnmb %xmm0, %xmm0, %k155; CHECK-SKX-NEXT: vpblendmq %zmm1, %zmm3, %zmm0 {%k1}56; CHECK-SKX-NEXT: kshiftrw $8, %k1, %k157; CHECK-SKX-NEXT: vpblendmq %zmm2, %zmm4, %zmm1 {%k1}58; CHECK-SKX-NEXT: retq59;60; CHECK-KNL-LABEL: test3:61; CHECK-KNL: # %bb.0:62; CHECK-KNL-NEXT: vpxor %xmm5, %xmm5, %xmm563; CHECK-KNL-NEXT: vpcmpeqb %xmm5, %xmm0, %xmm064; CHECK-KNL-NEXT: vpmovsxbd %xmm0, %zmm065; CHECK-KNL-NEXT: vptestmd %zmm0, %zmm0, %k166; CHECK-KNL-NEXT: vpblendmq %zmm1, %zmm3, %zmm0 {%k1}67; CHECK-KNL-NEXT: kshiftrw $8, %k1, %k168; CHECK-KNL-NEXT: vpblendmq %zmm2, %zmm4, %zmm1 {%k1}69; CHECK-KNL-NEXT: retq70 %c = icmp eq <16 x i8> %x, zeroinitializer71 %ret = select <16 x i1> %c, <16 x i64> %a, <16 x i64> %b72 ret <16 x i64> %ret73}74 75define <16 x i64> @test4(<16 x i16> %x, <16 x i64> %a, <16 x i64> %b) {76; CHECK-SKX-LABEL: test4:77; CHECK-SKX: # %bb.0:78; CHECK-SKX-NEXT: vptestnmw %ymm0, %ymm0, %k179; CHECK-SKX-NEXT: vpblendmq %zmm1, %zmm3, %zmm0 {%k1}80; CHECK-SKX-NEXT: kshiftrw $8, %k1, %k181; CHECK-SKX-NEXT: vpblendmq %zmm2, %zmm4, %zmm1 {%k1}82; CHECK-SKX-NEXT: retq83;84; CHECK-KNL-LABEL: test4:85; CHECK-KNL: # %bb.0:86; CHECK-KNL-NEXT: vpxor %xmm5, %xmm5, %xmm587; CHECK-KNL-NEXT: vpcmpeqw %ymm5, %ymm0, %ymm088; CHECK-KNL-NEXT: vpmovsxwd %ymm0, %zmm089; CHECK-KNL-NEXT: vptestmd %zmm0, %zmm0, %k190; CHECK-KNL-NEXT: vpblendmq %zmm1, %zmm3, %zmm0 {%k1}91; CHECK-KNL-NEXT: kshiftrw $8, %k1, %k192; CHECK-KNL-NEXT: vpblendmq %zmm2, %zmm4, %zmm1 {%k1}93; CHECK-KNL-NEXT: retq94 %c = icmp eq <16 x i16> %x, zeroinitializer95 %ret = select <16 x i1> %c, <16 x i64> %a, <16 x i64> %b96 ret <16 x i64> %ret97}98 99define <16 x i64> @test5(<16 x i32> %x, <16 x i64> %a, <16 x i64> %b) {100; CHECK-LABEL: test5:101; CHECK: # %bb.0:102; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1103; CHECK-NEXT: vpblendmq %zmm1, %zmm3, %zmm0 {%k1}104; CHECK-NEXT: kshiftrw $8, %k1, %k1105; CHECK-NEXT: vpblendmq %zmm2, %zmm4, %zmm1 {%k1}106; CHECK-NEXT: retq107 %c = icmp eq <16 x i32> %x, zeroinitializer108 %ret = select <16 x i1> %c, <16 x i64> %a, <16 x i64> %b109 ret <16 x i64> %ret110}111 112define <32 x i32> @test6(<32 x i8> %x, <32 x i32> %a, <32 x i32> %b) {113; CHECK-SKX-LABEL: test6:114; CHECK-SKX: # %bb.0:115; CHECK-SKX-NEXT: vptestnmb %ymm0, %ymm0, %k1116; CHECK-SKX-NEXT: vpblendmd %zmm1, %zmm3, %zmm0 {%k1}117; CHECK-SKX-NEXT: kshiftrd $16, %k1, %k1118; CHECK-SKX-NEXT: vpblendmd %zmm2, %zmm4, %zmm1 {%k1}119; CHECK-SKX-NEXT: retq120;121; CHECK-KNL-LABEL: test6:122; CHECK-KNL: # %bb.0:123; CHECK-KNL-NEXT: vpxor %xmm5, %xmm5, %xmm5124; CHECK-KNL-NEXT: vpcmpeqb %ymm5, %ymm0, %ymm0125; CHECK-KNL-NEXT: vextracti128 $1, %ymm0, %xmm5126; CHECK-KNL-NEXT: vpmovsxbd %xmm5, %zmm5127; CHECK-KNL-NEXT: vptestmd %zmm5, %zmm5, %k1128; CHECK-KNL-NEXT: vpmovsxbd %xmm0, %zmm0129; CHECK-KNL-NEXT: vptestmd %zmm0, %zmm0, %k2130; CHECK-KNL-NEXT: vpblendmd %zmm1, %zmm3, %zmm0 {%k2}131; CHECK-KNL-NEXT: vpblendmd %zmm2, %zmm4, %zmm1 {%k1}132; CHECK-KNL-NEXT: retq133 %c = icmp eq <32 x i8> %x, zeroinitializer134 %ret = select <32 x i1> %c, <32 x i32> %a, <32 x i32> %b135 ret <32 x i32> %ret136}137 138define <32 x i32> @test7(<32 x i16> %x, <32 x i32> %a, <32 x i32> %b) {139; CHECK-SKX-LABEL: test7:140; CHECK-SKX: # %bb.0:141; CHECK-SKX-NEXT: vptestnmw %zmm0, %zmm0, %k1142; CHECK-SKX-NEXT: vpblendmd %zmm1, %zmm3, %zmm0 {%k1}143; CHECK-SKX-NEXT: kshiftrd $16, %k1, %k1144; CHECK-SKX-NEXT: vpblendmd %zmm2, %zmm4, %zmm1 {%k1}145; CHECK-SKX-NEXT: retq146;147; CHECK-KNL-LABEL: test7:148; CHECK-KNL: # %bb.0:149; CHECK-KNL-NEXT: vextracti64x4 $1, %zmm0, %ymm5150; CHECK-KNL-NEXT: vpxor %xmm6, %xmm6, %xmm6151; CHECK-KNL-NEXT: vpcmpeqw %ymm6, %ymm5, %ymm5152; CHECK-KNL-NEXT: vpmovsxwd %ymm5, %zmm5153; CHECK-KNL-NEXT: vptestmd %zmm5, %zmm5, %k1154; CHECK-KNL-NEXT: vpcmpeqw %ymm6, %ymm0, %ymm0155; CHECK-KNL-NEXT: vpmovsxwd %ymm0, %zmm0156; CHECK-KNL-NEXT: vptestmd %zmm0, %zmm0, %k2157; CHECK-KNL-NEXT: vpblendmd %zmm1, %zmm3, %zmm0 {%k2}158; CHECK-KNL-NEXT: vpblendmd %zmm2, %zmm4, %zmm1 {%k1}159; CHECK-KNL-NEXT: retq160 %c = icmp eq <32 x i16> %x, zeroinitializer161 %ret = select <32 x i1> %c, <32 x i32> %a, <32 x i32> %b162 ret <32 x i32> %ret163}164 165define <64 x i16> @test8(<64 x i8> %x, <64 x i16> %a, <64 x i16> %b) {166; CHECK-SKX-LABEL: test8:167; CHECK-SKX: # %bb.0:168; CHECK-SKX-NEXT: vptestnmb %zmm0, %zmm0, %k1169; CHECK-SKX-NEXT: vpblendmw %zmm1, %zmm3, %zmm0 {%k1}170; CHECK-SKX-NEXT: kshiftrq $32, %k1, %k1171; CHECK-SKX-NEXT: vpblendmw %zmm2, %zmm4, %zmm1 {%k1}172; CHECK-SKX-NEXT: retq173;174; CHECK-KNL-LABEL: test8:175; CHECK-KNL: # %bb.0:176; CHECK-KNL-NEXT: vextracti64x4 $1, %zmm0, %ymm5177; CHECK-KNL-NEXT: vpxor %xmm6, %xmm6, %xmm6178; CHECK-KNL-NEXT: vpcmpeqb %ymm6, %ymm5, %ymm5179; CHECK-KNL-NEXT: vpcmpeqb %ymm6, %ymm0, %ymm0180; CHECK-KNL-NEXT: vpmovsxbw %xmm0, %ymm6181; CHECK-KNL-NEXT: vextracti128 $1, %ymm0, %xmm0182; CHECK-KNL-NEXT: vpmovsxbw %xmm0, %ymm0183; CHECK-KNL-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm0184; CHECK-KNL-NEXT: vpternlogq {{.*#+}} zmm0 = zmm3 ^ (zmm0 & (zmm1 ^ zmm3))185; CHECK-KNL-NEXT: vpmovsxbw %xmm5, %ymm1186; CHECK-KNL-NEXT: vextracti128 $1, %ymm5, %xmm3187; CHECK-KNL-NEXT: vpmovsxbw %xmm3, %ymm3188; CHECK-KNL-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1189; CHECK-KNL-NEXT: vpternlogq {{.*#+}} zmm1 = zmm4 ^ (zmm1 & (zmm2 ^ zmm4))190; CHECK-KNL-NEXT: retq191 %c = icmp eq <64 x i8> %x, zeroinitializer192 %ret = select <64 x i1> %c, <64 x i16> %a, <64 x i16> %b193 ret <64 x i16> %ret194}195