183 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s4 5declare <32 x half> @llvm.experimental.constrained.fadd.v32f16(<32 x half>, <32 x half>, metadata, metadata)6declare <32 x half> @llvm.experimental.constrained.fsub.v32f16(<32 x half>, <32 x half>, metadata, metadata)7declare <32 x half> @llvm.experimental.constrained.fmul.v32f16(<32 x half>, <32 x half>, metadata, metadata)8declare <32 x half> @llvm.experimental.constrained.fdiv.v32f16(<32 x half>, <32 x half>, metadata, metadata)9declare <32 x half> @llvm.experimental.constrained.sqrt.v32f16(<32 x half>, metadata, metadata)10declare <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f16(<8 x half>, metadata)11declare <16 x float> @llvm.experimental.constrained.fpext.v16f32.v16f16(<16 x half>, metadata)12declare <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f64(<8 x double>, metadata, metadata)13declare <16 x half> @llvm.experimental.constrained.fptrunc.v16f16.v16f32(<16 x float>, metadata, metadata)14declare <32 x half> @llvm.experimental.constrained.fma.v32f16(<32 x half>, <32 x half>, <32 x half>, metadata, metadata)15declare <32 x half> @llvm.experimental.constrained.ceil.v32f16(<32 x half>, metadata)16declare <32 x half> @llvm.experimental.constrained.floor.v32f16(<32 x half>, metadata)17declare <32 x half> @llvm.experimental.constrained.trunc.v32f16(<32 x half>, metadata)18declare <32 x half> @llvm.experimental.constrained.rint.v32f16(<32 x half>, metadata, metadata)19declare <32 x half> @llvm.experimental.constrained.nearbyint.v32f16(<32 x half>, metadata, metadata)20 21define <32 x half> @f2(<32 x half> %a, <32 x half> %b) #0 {22; CHECK-LABEL: f2:23; CHECK: # %bb.0:24; CHECK-NEXT: vaddph %zmm1, %zmm0, %zmm025; CHECK-NEXT: ret{{[l|q]}}26 %ret = call <32 x half> @llvm.experimental.constrained.fadd.v32f16(<32 x half> %a, <32 x half> %b,27 metadata !"round.dynamic",28 metadata !"fpexcept.strict") #029 ret <32 x half> %ret30}31 32define <32 x half> @f4(<32 x half> %a, <32 x half> %b) #0 {33; CHECK-LABEL: f4:34; CHECK: # %bb.0:35; CHECK-NEXT: vsubph %zmm1, %zmm0, %zmm036; CHECK-NEXT: ret{{[l|q]}}37 %ret = call <32 x half> @llvm.experimental.constrained.fsub.v32f16(<32 x half> %a, <32 x half> %b,38 metadata !"round.dynamic",39 metadata !"fpexcept.strict") #040 ret <32 x half> %ret41}42 43define <32 x half> @f6(<32 x half> %a, <32 x half> %b) #0 {44; CHECK-LABEL: f6:45; CHECK: # %bb.0:46; CHECK-NEXT: vmulph %zmm1, %zmm0, %zmm047; CHECK-NEXT: ret{{[l|q]}}48 %ret = call <32 x half> @llvm.experimental.constrained.fmul.v32f16(<32 x half> %a, <32 x half> %b,49 metadata !"round.dynamic",50 metadata !"fpexcept.strict") #051 ret <32 x half> %ret52}53 54define <32 x half> @f8(<32 x half> %a, <32 x half> %b) #0 {55; CHECK-LABEL: f8:56; CHECK: # %bb.0:57; CHECK-NEXT: vdivph %zmm1, %zmm0, %zmm058; CHECK-NEXT: ret{{[l|q]}}59 %ret = call <32 x half> @llvm.experimental.constrained.fdiv.v32f16(<32 x half> %a, <32 x half> %b,60 metadata !"round.dynamic",61 metadata !"fpexcept.strict") #062 ret <32 x half> %ret63}64 65define <32 x half> @f10(<32 x half> %a) #0 {66; CHECK-LABEL: f10:67; CHECK: # %bb.0:68; CHECK-NEXT: vsqrtph %zmm0, %zmm069; CHECK-NEXT: ret{{[l|q]}}70 %ret = call <32 x half> @llvm.experimental.constrained.sqrt.v32f16(71 <32 x half> %a,72 metadata !"round.dynamic",73 metadata !"fpexcept.strict") #074 ret <32 x half > %ret75}76 77define <8 x double> @f11(<8 x half> %a) #0 {78; CHECK-LABEL: f11:79; CHECK: # %bb.0:80; CHECK-NEXT: vcvtph2pd %xmm0, %zmm081; CHECK-NEXT: ret{{[l|q]}}82 %ret = call <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f16(83 <8 x half> %a,84 metadata !"fpexcept.strict") #085 ret <8 x double> %ret86}87 88define <8 x half> @f12(<8 x double> %a) #0 {89; CHECK-LABEL: f12:90; CHECK: # %bb.0:91; CHECK-NEXT: vcvtpd2ph %zmm0, %xmm092; CHECK-NEXT: vzeroupper93; CHECK-NEXT: ret{{[l|q]}}94 %ret = call <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f64(95 <8 x double> %a,96 metadata !"round.dynamic",97 metadata !"fpexcept.strict") #098 ret <8 x half> %ret99}100 101define <32 x half> @f13(<32 x half> %a, <32 x half> %b, <32 x half> %c) #0 {102; CHECK-LABEL: f13:103; CHECK: # %bb.0:104; CHECK-NEXT: vfmadd213ph %zmm2, %zmm1, %zmm0105; CHECK-NEXT: ret{{[l|q]}}106 %res = call <32 x half> @llvm.experimental.constrained.fma.v32f16(<32 x half> %a, <32 x half> %b, <32 x half> %c,107 metadata !"round.dynamic",108 metadata !"fpexcept.strict") #0109 ret <32 x half> %res110}111 112define <16 x float> @f14(<16 x half> %a) #0 {113; CHECK-LABEL: f14:114; CHECK: # %bb.0:115; CHECK-NEXT: vcvtph2psx %ymm0, %zmm0116; CHECK-NEXT: ret{{[l|q]}}117 %ret = call <16 x float> @llvm.experimental.constrained.fpext.v16f32.v16f16(118 <16 x half> %a,119 metadata !"fpexcept.strict") #0120 ret <16 x float> %ret121}122 123define <16 x half> @f15(<16 x float> %a) #0 {124; CHECK-LABEL: f15:125; CHECK: # %bb.0:126; CHECK-NEXT: vcvtps2phx %zmm0, %ymm0127; CHECK-NEXT: ret{{[l|q]}}128 %ret = call <16 x half> @llvm.experimental.constrained.fptrunc.v16f16.v16f32(129 <16 x float> %a,130 metadata !"round.dynamic",131 metadata !"fpexcept.strict") #0132 ret <16 x half> %ret133}134 135define <32 x half> @strict_vector_fceil_v32f16(<32 x half> %f) #0 {136; CHECK-LABEL: strict_vector_fceil_v32f16:137; CHECK: # %bb.0:138; CHECK-NEXT: vrndscaleph $10, %zmm0, %zmm0139; CHECK-NEXT: ret{{[l|q]}}140 %res = call <32 x half> @llvm.experimental.constrained.ceil.v32f16(<32 x half> %f, metadata !"fpexcept.strict") #0141 ret <32 x half> %res142}143 144define <32 x half> @strict_vector_ffloor_v32f16(<32 x half> %f) #0 {145; CHECK-LABEL: strict_vector_ffloor_v32f16:146; CHECK: # %bb.0:147; CHECK-NEXT: vrndscaleph $9, %zmm0, %zmm0148; CHECK-NEXT: ret{{[l|q]}}149 %res = call <32 x half> @llvm.experimental.constrained.floor.v32f16(<32 x half> %f, metadata !"fpexcept.strict") #0150 ret <32 x half> %res151}152 153define <32 x half> @strict_vector_ftrunc_v32f16(<32 x half> %f) #0 {154; CHECK-LABEL: strict_vector_ftrunc_v32f16:155; CHECK: # %bb.0:156; CHECK-NEXT: vrndscaleph $11, %zmm0, %zmm0157; CHECK-NEXT: ret{{[l|q]}}158 %res = call <32 x half> @llvm.experimental.constrained.trunc.v32f16(<32 x half> %f, metadata !"fpexcept.strict") #0159 ret <32 x half> %res160}161 162define <32 x half> @strict_vector_frint_v32f16(<32 x half> %f) #0 {163; CHECK-LABEL: strict_vector_frint_v32f16:164; CHECK: # %bb.0:165; CHECK-NEXT: vrndscaleph $4, %zmm0, %zmm0166; CHECK-NEXT: ret{{[l|q]}}167 %res = call <32 x half> @llvm.experimental.constrained.rint.v32f16(<32 x half> %f,168 metadata !"round.dynamic", metadata !"fpexcept.strict") #0169 ret <32 x half> %res170}171 172define <32 x half> @strict_vector_fnearbyint_v32f16(<32 x half> %f) #0 {173; CHECK-LABEL: strict_vector_fnearbyint_v32f16:174; CHECK: # %bb.0:175; CHECK-NEXT: vrndscaleph $12, %zmm0, %zmm0176; CHECK-NEXT: ret{{[l|q]}}177 %res = call <32 x half> @llvm.experimental.constrained.nearbyint.v32f16(<32 x half> %f,178 metadata !"round.dynamic", metadata !"fpexcept.strict") #0179 ret <32 x half> %res180}181 182attributes #0 = { strictfp }183