176 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx512fp16,avx512vl -O3 | FileCheck %s --check-prefixes=CHECK,X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512fp16,avx512vl -O3 | FileCheck %s --check-prefixes=CHECK,X644 5declare <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i1(<16 x i1>, metadata, metadata)6declare <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i1(<16 x i1>, metadata, metadata)7declare <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i8(<16 x i8>, metadata, metadata)8declare <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i8(<16 x i8>, metadata, metadata)9declare <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i16(<16 x i16>, metadata, metadata)10declare <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i16(<16 x i16>, metadata, metadata)11declare <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i32(<8 x i32>, metadata, metadata)12declare <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i32(<8 x i32>, metadata, metadata)13declare <4 x half> @llvm.experimental.constrained.sitofp.v4f16.v4i64(<4 x i64>, metadata, metadata)14declare <4 x half> @llvm.experimental.constrained.uitofp.v4f16.v4i64(<4 x i64>, metadata, metadata)15declare <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i64(<8 x i64>, metadata, metadata)16declare <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i64(<8 x i64>, metadata, metadata)17 18define <16 x half> @sitofp_v16i1_v16f16(<16 x i1> %x) #0 {19; CHECK-LABEL: sitofp_v16i1_v16f16:20; CHECK: # %bb.0:21; CHECK-NEXT: vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero22; CHECK-NEXT: vpsllw $15, %ymm0, %ymm023; CHECK-NEXT: vpsraw $15, %ymm0, %ymm024; CHECK-NEXT: vcvtw2ph %ymm0, %ymm025; CHECK-NEXT: ret{{[l|q]}}26 %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i1(<16 x i1> %x,27 metadata !"round.dynamic",28 metadata !"fpexcept.strict") #029 ret <16 x half> %result30}31 32define <16 x half> @uitofp_v16i1_v16f16(<16 x i1> %x) #0 {33; X86-LABEL: uitofp_v16i1_v16f16:34; X86: # %bb.0:35; X86-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}{1to4}, %xmm0, %xmm036; X86-NEXT: vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero37; X86-NEXT: vcvtw2ph %ymm0, %ymm038; X86-NEXT: retl39;40; X64-LABEL: uitofp_v16i1_v16f16:41; X64: # %bb.0:42; X64-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm043; X64-NEXT: vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero44; X64-NEXT: vcvtw2ph %ymm0, %ymm045; X64-NEXT: retq46 %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i1(<16 x i1> %x,47 metadata !"round.dynamic",48 metadata !"fpexcept.strict") #049 ret <16 x half> %result50}51 52define <16 x half> @sitofp_v16i8_v16f16(<16 x i8> %x) #0 {53; CHECK-LABEL: sitofp_v16i8_v16f16:54; CHECK: # %bb.0:55; CHECK-NEXT: vpmovsxbw %xmm0, %ymm056; CHECK-NEXT: vcvtw2ph %ymm0, %ymm057; CHECK-NEXT: ret{{[l|q]}}58 %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i8(<16 x i8> %x,59 metadata !"round.dynamic",60 metadata !"fpexcept.strict") #061 ret <16 x half> %result62}63 64define <16 x half> @uitofp_v16i8_v16f16(<16 x i8> %x) #0 {65; CHECK-LABEL: uitofp_v16i8_v16f16:66; CHECK: # %bb.0:67; CHECK-NEXT: vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero68; CHECK-NEXT: vcvtw2ph %ymm0, %ymm069; CHECK-NEXT: ret{{[l|q]}}70 %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i8(<16 x i8> %x,71 metadata !"round.dynamic",72 metadata !"fpexcept.strict") #073 ret <16 x half> %result74}75 76define <16 x half> @sitofp_v16i16_v16f16(<16 x i16> %x) #0 {77; CHECK-LABEL: sitofp_v16i16_v16f16:78; CHECK: # %bb.0:79; CHECK-NEXT: vcvtw2ph %ymm0, %ymm080; CHECK-NEXT: ret{{[l|q]}}81 %result = call <16 x half> @llvm.experimental.constrained.sitofp.v16f16.v16i16(<16 x i16> %x,82 metadata !"round.dynamic",83 metadata !"fpexcept.strict") #084 ret <16 x half> %result85}86 87define <16 x half> @uitofp_v16i16_v16f16(<16 x i16> %x) #0 {88; CHECK-LABEL: uitofp_v16i16_v16f16:89; CHECK: # %bb.0:90; CHECK-NEXT: vcvtuw2ph %ymm0, %ymm091; CHECK-NEXT: ret{{[l|q]}}92 %result = call <16 x half> @llvm.experimental.constrained.uitofp.v16f16.v16i16(<16 x i16> %x,93 metadata !"round.dynamic",94 metadata !"fpexcept.strict") #095 ret <16 x half> %result96}97 98define <8 x half> @sitofp_v8i32_v8f16(<8 x i32> %x) #0 {99; CHECK-LABEL: sitofp_v8i32_v8f16:100; CHECK: # %bb.0:101; CHECK-NEXT: vcvtdq2ph %ymm0, %xmm0102; CHECK-NEXT: vzeroupper103; CHECK-NEXT: ret{{[l|q]}}104 %result = call <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i32(<8 x i32> %x,105 metadata !"round.dynamic",106 metadata !"fpexcept.strict") #0107 ret <8 x half> %result108}109 110define <8 x half> @uitofp_v8i32_v8f16(<8 x i32> %x) #0 {111; CHECK-LABEL: uitofp_v8i32_v8f16:112; CHECK: # %bb.0:113; CHECK-NEXT: vcvtudq2ph %ymm0, %xmm0114; CHECK-NEXT: vzeroupper115; CHECK-NEXT: ret{{[l|q]}}116 %result = call <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i32(<8 x i32> %x,117 metadata !"round.dynamic",118 metadata !"fpexcept.strict") #0119 ret <8 x half> %result120}121 122define <4 x half> @sitofp_v4i64_v4f16(<4 x i64> %x) #0 {123; CHECK-LABEL: sitofp_v4i64_v4f16:124; CHECK: # %bb.0:125; CHECK-NEXT: vcvtqq2ph %ymm0, %xmm0126; CHECK-NEXT: vzeroupper127; CHECK-NEXT: ret{{[l|q]}}128 %result = call <4 x half> @llvm.experimental.constrained.sitofp.v4f16.v4i64(<4 x i64> %x,129 metadata !"round.dynamic",130 metadata !"fpexcept.strict") #0131 ret <4 x half> %result132}133 134define <4 x half> @uitofp_v4i64_v4f16(<4 x i64> %x) #0 {135; CHECK-LABEL: uitofp_v4i64_v4f16:136; CHECK: # %bb.0:137; CHECK-NEXT: vcvtuqq2ph %ymm0, %xmm0138; CHECK-NEXT: vzeroupper139; CHECK-NEXT: ret{{[l|q]}}140 %result = call <4 x half> @llvm.experimental.constrained.uitofp.v4f16.v4i64(<4 x i64> %x,141 metadata !"round.dynamic",142 metadata !"fpexcept.strict") #0143 ret <4 x half> %result144}145 146define <8 x half> @sitofp_v8i64_v8f16(<8 x i64> %x) #1 {147; CHECK-LABEL: sitofp_v8i64_v8f16:148; CHECK: # %bb.0:149; CHECK-NEXT: vcvtqq2ph %ymm1, %xmm1150; CHECK-NEXT: vcvtqq2ph %ymm0, %xmm0151; CHECK-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]152; CHECK-NEXT: vzeroupper153; CHECK-NEXT: ret{{[l|q]}}154 %result = call <8 x half> @llvm.experimental.constrained.sitofp.v8f16.v8i64(<8 x i64> %x,155 metadata !"round.dynamic",156 metadata !"fpexcept.strict") #0157 ret <8 x half> %result158}159 160define <8 x half> @uitofp_v8i64_v8f16(<8 x i64> %x) #1 {161; CHECK-LABEL: uitofp_v8i64_v8f16:162; CHECK: # %bb.0:163; CHECK-NEXT: vcvtuqq2ph %ymm1, %xmm1164; CHECK-NEXT: vcvtuqq2ph %ymm0, %xmm0165; CHECK-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]166; CHECK-NEXT: vzeroupper167; CHECK-NEXT: ret{{[l|q]}}168 %result = call <8 x half> @llvm.experimental.constrained.uitofp.v8f16.v8i64(<8 x i64> %x,169 metadata !"round.dynamic",170 metadata !"fpexcept.strict") #0171 ret <8 x half> %result172}173 174attributes #0 = { strictfp }175attributes #1 = { strictfp "min-legal-vector-width"="256" "prefer-vector-width"="256" }176