268 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx | FileCheck %s --check-prefix=AVX13; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefix=AVX24; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512bw,avx512vl,avx512f | FileCheck %s --check-prefix=AVX5125 6; PR37427 - https://bugs.llvm.org/show_bug.cgi?id=374277 8define <8 x i32> @eq_zero(ptr %p, <8 x i32> %x, <8 x i32> %y) {9; AVX1-LABEL: eq_zero:10; AVX1: # %bb.0:11; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero12; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm313; AVX1-NEXT: vpcmpeqb %xmm3, %xmm2, %xmm214; AVX1-NEXT: vpmovsxbd %xmm2, %xmm315; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]16; AVX1-NEXT: vpmovsxbd %xmm2, %xmm217; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm218; AVX1-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm019; AVX1-NEXT: retq20;21; AVX2-LABEL: eq_zero:22; AVX2: # %bb.0:23; AVX2-NEXT: vpmovzxbd {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero24; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm325; AVX2-NEXT: vpcmpeqd %ymm3, %ymm2, %ymm226; AVX2-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm027; AVX2-NEXT: retq28;29; AVX512-LABEL: eq_zero:30; AVX512: # %bb.0:31; AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero32; AVX512-NEXT: vptestnmb %xmm2, %xmm2, %k133; AVX512-NEXT: vpblendmd %ymm0, %ymm1, %ymm0 {%k1}34; AVX512-NEXT: retq35 %load = load <8 x i8>, ptr %p36 %cmp = icmp eq <8 x i8> %load, zeroinitializer37 %sel = select <8 x i1> %cmp, <8 x i32> %x, <8 x i32> %y38 ret <8 x i32> %sel39}40 41define <4 x i64> @ne_zero(ptr %p, <4 x i64> %x, <4 x i64> %y) {42; AVX1-LABEL: ne_zero:43; AVX1: # %bb.0:44; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero45; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm346; AVX1-NEXT: vpcmpeqw %xmm3, %xmm2, %xmm247; AVX1-NEXT: vpcmpeqd %xmm3, %xmm3, %xmm348; AVX1-NEXT: vpxor %xmm3, %xmm2, %xmm249; AVX1-NEXT: vpmovsxwq %xmm2, %xmm350; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]51; AVX1-NEXT: vpmovsxwq %xmm2, %xmm252; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm253; AVX1-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm054; AVX1-NEXT: retq55;56; AVX2-LABEL: ne_zero:57; AVX2: # %bb.0:58; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero59; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm360; AVX2-NEXT: vpcmpgtq %ymm3, %ymm2, %ymm261; AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm062; AVX2-NEXT: retq63;64; AVX512-LABEL: ne_zero:65; AVX512: # %bb.0:66; AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero67; AVX512-NEXT: vptestmw %xmm2, %xmm2, %k168; AVX512-NEXT: vpblendmq %ymm0, %ymm1, %ymm0 {%k1}69; AVX512-NEXT: retq70 %load = load <4 x i16>, ptr %p71 %cmp = icmp ne <4 x i16> %load, zeroinitializer72 %sel = select <4 x i1> %cmp, <4 x i64> %x, <4 x i64> %y73 ret <4 x i64> %sel74}75 76define <16 x i16> @sgt_zero(ptr %p, <16 x i16> %x, <16 x i16> %y) {77; AVX1-LABEL: sgt_zero:78; AVX1: # %bb.0:79; AVX1-NEXT: vmovdqa (%rdi), %xmm280; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm381; AVX1-NEXT: vpcmpgtb %xmm3, %xmm2, %xmm282; AVX1-NEXT: vpmovsxbw %xmm2, %xmm383; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]84; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm285; AVX1-NEXT: vandnps %ymm1, %ymm2, %ymm186; AVX1-NEXT: vandps %ymm2, %ymm0, %ymm087; AVX1-NEXT: vorps %ymm1, %ymm0, %ymm088; AVX1-NEXT: retq89;90; AVX2-LABEL: sgt_zero:91; AVX2: # %bb.0:92; AVX2-NEXT: vpmovsxbw (%rdi), %ymm293; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm394; AVX2-NEXT: vpcmpgtw %ymm3, %ymm2, %ymm295; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm096; AVX2-NEXT: retq97;98; AVX512-LABEL: sgt_zero:99; AVX512: # %bb.0:100; AVX512-NEXT: vpxor %xmm2, %xmm2, %xmm2101; AVX512-NEXT: vpcmpltb (%rdi), %xmm2, %k1102; AVX512-NEXT: vpblendmw %ymm0, %ymm1, %ymm0 {%k1}103; AVX512-NEXT: retq104 %load = load <16 x i8>, ptr %p105 %cmp = icmp sgt <16 x i8> %load, zeroinitializer106 %sel = select <16 x i1> %cmp, <16 x i16> %x, <16 x i16> %y107 ret <16 x i16> %sel108}109 110define <8 x i32> @slt_zero(ptr %p, <8 x i32> %x, <8 x i32> %y) {111; AVX1-LABEL: slt_zero:112; AVX1: # %bb.0:113; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero114; AVX1-NEXT: vpmovsxbd %xmm2, %xmm3115; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]116; AVX1-NEXT: vpmovsxbd %xmm2, %xmm2117; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2118; AVX1-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm0119; AVX1-NEXT: retq120;121; AVX2-LABEL: slt_zero:122; AVX2: # %bb.0:123; AVX2-NEXT: vpmovsxbd (%rdi), %ymm2124; AVX2-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm0125; AVX2-NEXT: retq126;127; AVX512-LABEL: slt_zero:128; AVX512: # %bb.0:129; AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero130; AVX512-NEXT: vpmovb2m %xmm2, %k1131; AVX512-NEXT: vpblendmd %ymm0, %ymm1, %ymm0 {%k1}132; AVX512-NEXT: retq133 %load = load <8 x i8>, ptr %p134 %cmp = icmp slt <8 x i8> %load, zeroinitializer135 %sel = select <8 x i1> %cmp, <8 x i32> %x, <8 x i32> %y136 ret <8 x i32> %sel137}138 139define <4 x double> @eq_zero_fp_select(ptr %p, <4 x double> %x, <4 x double> %y) {140; AVX1-LABEL: eq_zero_fp_select:141; AVX1: # %bb.0:142; AVX1-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero143; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3144; AVX1-NEXT: vpcmpeqb %xmm3, %xmm2, %xmm2145; AVX1-NEXT: vpmovsxbq %xmm2, %xmm3146; AVX1-NEXT: vpsrld $16, %xmm2, %xmm2147; AVX1-NEXT: vpmovsxbq %xmm2, %xmm2148; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2149; AVX1-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0150; AVX1-NEXT: retq151;152; AVX2-LABEL: eq_zero_fp_select:153; AVX2: # %bb.0:154; AVX2-NEXT: vpmovzxbq {{.*#+}} ymm2 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero,mem[2],zero,zero,zero,zero,zero,zero,zero,mem[3],zero,zero,zero,zero,zero,zero,zero155; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3156; AVX2-NEXT: vpcmpeqq %ymm3, %ymm2, %ymm2157; AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0158; AVX2-NEXT: retq159;160; AVX512-LABEL: eq_zero_fp_select:161; AVX512: # %bb.0:162; AVX512-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero163; AVX512-NEXT: vptestnmb %xmm2, %xmm2, %k1164; AVX512-NEXT: vblendmpd %ymm0, %ymm1, %ymm0 {%k1}165; AVX512-NEXT: retq166 %load = load <4 x i8>, ptr %p167 %cmp = icmp eq <4 x i8> %load, zeroinitializer168 %sel = select <4 x i1> %cmp, <4 x double> %x, <4 x double> %y169 ret <4 x double> %sel170}171 172define <8 x float> @ne_zero_fp_select(ptr %p, <8 x float> %x, <8 x float> %y) {173; AVX1-LABEL: ne_zero_fp_select:174; AVX1: # %bb.0:175; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero176; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3177; AVX1-NEXT: vpcmpeqb %xmm3, %xmm2, %xmm2178; AVX1-NEXT: vpcmpeqd %xmm3, %xmm3, %xmm3179; AVX1-NEXT: vpxor %xmm3, %xmm2, %xmm2180; AVX1-NEXT: vpmovsxbd %xmm2, %xmm3181; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]182; AVX1-NEXT: vpmovsxbd %xmm2, %xmm2183; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2184; AVX1-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm0185; AVX1-NEXT: retq186;187; AVX2-LABEL: ne_zero_fp_select:188; AVX2: # %bb.0:189; AVX2-NEXT: vpmovzxbd {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero190; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3191; AVX2-NEXT: vpcmpgtd %ymm3, %ymm2, %ymm2192; AVX2-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm0193; AVX2-NEXT: retq194;195; AVX512-LABEL: ne_zero_fp_select:196; AVX512: # %bb.0:197; AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero198; AVX512-NEXT: vptestmb %xmm2, %xmm2, %k1199; AVX512-NEXT: vblendmps %ymm0, %ymm1, %ymm0 {%k1}200; AVX512-NEXT: retq201 %load = load <8 x i8>, ptr %p202 %cmp = icmp ne <8 x i8> %load, zeroinitializer203 %sel = select <8 x i1> %cmp, <8 x float> %x, <8 x float> %y204 ret <8 x float> %sel205}206 207define <4 x double> @sgt_zero_fp_select(ptr %p, <4 x double> %x, <4 x double> %y) {208; AVX1-LABEL: sgt_zero_fp_select:209; AVX1: # %bb.0:210; AVX1-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero211; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3212; AVX1-NEXT: vpcmpgtb %xmm3, %xmm2, %xmm2213; AVX1-NEXT: vpmovsxbq %xmm2, %xmm3214; AVX1-NEXT: vpsrld $16, %xmm2, %xmm2215; AVX1-NEXT: vpmovsxbq %xmm2, %xmm2216; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2217; AVX1-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0218; AVX1-NEXT: retq219;220; AVX2-LABEL: sgt_zero_fp_select:221; AVX2: # %bb.0:222; AVX2-NEXT: vpmovsxbq (%rdi), %ymm2223; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3224; AVX2-NEXT: vpcmpgtq %ymm3, %ymm2, %ymm2225; AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0226; AVX2-NEXT: retq227;228; AVX512-LABEL: sgt_zero_fp_select:229; AVX512: # %bb.0:230; AVX512-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero231; AVX512-NEXT: vpxor %xmm3, %xmm3, %xmm3232; AVX512-NEXT: vpcmpgtb %xmm3, %xmm2, %k1233; AVX512-NEXT: vblendmpd %ymm0, %ymm1, %ymm0 {%k1}234; AVX512-NEXT: retq235 %load = load <4 x i8>, ptr %p236 %cmp = icmp sgt <4 x i8> %load, zeroinitializer237 %sel = select <4 x i1> %cmp, <4 x double> %x, <4 x double> %y238 ret <4 x double> %sel239}240 241define <8 x float> @slt_zero_fp_select(ptr %p, <8 x float> %x, <8 x float> %y) {242; AVX1-LABEL: slt_zero_fp_select:243; AVX1: # %bb.0:244; AVX1-NEXT: vpmovsxwd 8(%rdi), %xmm2245; AVX1-NEXT: vpmovsxwd (%rdi), %xmm3246; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2247; AVX1-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm0248; AVX1-NEXT: retq249;250; AVX2-LABEL: slt_zero_fp_select:251; AVX2: # %bb.0:252; AVX2-NEXT: vpmovsxwd (%rdi), %ymm2253; AVX2-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm0254; AVX2-NEXT: retq255;256; AVX512-LABEL: slt_zero_fp_select:257; AVX512: # %bb.0:258; AVX512-NEXT: vpxor %xmm2, %xmm2, %xmm2259; AVX512-NEXT: vpcmpgtw (%rdi), %xmm2, %k1260; AVX512-NEXT: vblendmps %ymm0, %ymm1, %ymm0 {%k1}261; AVX512-NEXT: retq262 %load = load <8 x i16>, ptr %p263 %cmp = icmp slt <8 x i16> %load, zeroinitializer264 %sel = select <8 x i1> %cmp, <8 x float> %x, <8 x float> %y265 ret <8 x float> %sel266}267 268