191 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 <16 x half> @llvm.experimental.constrained.fadd.v16f16(<16 x half>, <16 x half>, metadata, metadata)6declare <16 x half> @llvm.experimental.constrained.fsub.v16f16(<16 x half>, <16 x half>, metadata, metadata)7declare <16 x half> @llvm.experimental.constrained.fmul.v16f16(<16 x half>, <16 x half>, metadata, metadata)8declare <16 x half> @llvm.experimental.constrained.fdiv.v16f16(<16 x half>, <16 x half>, metadata, metadata)9declare <16 x half> @llvm.experimental.constrained.sqrt.v16f16(<16 x half>, metadata, metadata)10declare <16 x half> @llvm.experimental.constrained.fma.v16f16(<16 x half>, <16 x half>, <16 x half>, metadata, metadata)11declare <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f16(<4 x half>, metadata)12declare <8 x float> @llvm.experimental.constrained.fpext.v8f32.v8f16(<8 x half>, metadata)13declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f64(<4 x double>, metadata, metadata)14declare <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f32(<8 x float>, metadata, metadata)15declare <16 x half> @llvm.experimental.constrained.ceil.v16f16(<16 x half>, metadata)16declare <16 x half> @llvm.experimental.constrained.floor.v16f16(<16 x half>, metadata)17declare <16 x half> @llvm.experimental.constrained.trunc.v16f16(<16 x half>, metadata)18declare <16 x half> @llvm.experimental.constrained.rint.v16f16(<16 x half>, metadata, metadata)19declare <16 x half> @llvm.experimental.constrained.nearbyint.v16f16(<16 x half>, metadata, metadata)20 21define <16 x half> @f2(<16 x half> %a, <16 x half> %b) #0 {22; CHECK-LABEL: f2:23; CHECK: # %bb.0:24; CHECK-NEXT: vaddph %ymm1, %ymm0, %ymm025; CHECK-NEXT: ret{{[l|q]}}26 %ret = call <16 x half> @llvm.experimental.constrained.fadd.v16f16(<16 x half> %a, <16 x half> %b,27 metadata !"round.dynamic",28 metadata !"fpexcept.strict") #029 ret <16 x half> %ret30}31 32define <16 x half> @f4(<16 x half> %a, <16 x half> %b) #0 {33; CHECK-LABEL: f4:34; CHECK: # %bb.0:35; CHECK-NEXT: vsubph %ymm1, %ymm0, %ymm036; CHECK-NEXT: ret{{[l|q]}}37 %ret = call <16 x half> @llvm.experimental.constrained.fsub.v16f16(<16 x half> %a, <16 x half> %b,38 metadata !"round.dynamic",39 metadata !"fpexcept.strict") #040 ret <16 x half> %ret41}42 43define <16 x half> @f6(<16 x half> %a, <16 x half> %b) #0 {44; CHECK-LABEL: f6:45; CHECK: # %bb.0:46; CHECK-NEXT: vmulph %ymm1, %ymm0, %ymm047; CHECK-NEXT: ret{{[l|q]}}48 %ret = call <16 x half> @llvm.experimental.constrained.fmul.v16f16(<16 x half> %a, <16 x half> %b,49 metadata !"round.dynamic",50 metadata !"fpexcept.strict") #051 ret <16 x half> %ret52}53 54define <16 x half> @f8(<16 x half> %a, <16 x half> %b) #0 {55; CHECK-LABEL: f8:56; CHECK: # %bb.0:57; CHECK-NEXT: vdivph %ymm1, %ymm0, %ymm058; CHECK-NEXT: ret{{[l|q]}}59 %ret = call <16 x half> @llvm.experimental.constrained.fdiv.v16f16(<16 x half> %a, <16 x half> %b,60 metadata !"round.dynamic",61 metadata !"fpexcept.strict") #062 ret <16 x half> %ret63}64 65 66define <16 x half> @f10(<16 x half> %a) #0 {67; CHECK-LABEL: f10:68; CHECK: # %bb.0:69; CHECK-NEXT: vsqrtph %ymm0, %ymm070; CHECK-NEXT: ret{{[l|q]}}71 %ret = call <16 x half> @llvm.experimental.constrained.sqrt.v16f16(72 <16 x half> %a,73 metadata !"round.dynamic",74 metadata !"fpexcept.strict") #075 ret <16 x half > %ret76}77 78define <4 x double> @f11(<4 x half> %a) #0 {79; CHECK-LABEL: f11:80; CHECK: # %bb.0:81; CHECK-NEXT: vcvtph2pd %xmm0, %ymm082; CHECK-NEXT: ret{{[l|q]}}83 %ret = call <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f16(84 <4 x half> %a,85 metadata !"fpexcept.strict") #086 ret <4 x double> %ret87}88 89define <4 x half> @f12(<4 x double> %a) #0 {90; CHECK-LABEL: f12:91; CHECK: # %bb.0:92; CHECK-NEXT: vcvtpd2ph %ymm0, %xmm093; CHECK-NEXT: vzeroupper94; CHECK-NEXT: ret{{[l|q]}}95 %ret = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f64(96 <4 x double> %a,97 metadata !"round.dynamic",98 metadata !"fpexcept.strict") #099 ret <4 x half> %ret100}101 102define <16 x half> @f13(<16 x half> %a, <16 x half> %b, <16 x half> %c) #0 {103; CHECK-LABEL: f13:104; CHECK: # %bb.0:105; CHECK-NEXT: vfmadd213ph %ymm2, %ymm1, %ymm0106; CHECK-NEXT: ret{{[l|q]}}107 %res = call <16 x half> @llvm.experimental.constrained.fma.v16f16(<16 x half> %a, <16 x half> %b, <16 x half> %c,108 metadata !"round.dynamic",109 metadata !"fpexcept.strict") #0110 ret <16 x half> %res111}112 113define <8 x float> @f14(<8 x half> %a) #0 {114; CHECK-LABEL: f14:115; CHECK: # %bb.0:116; CHECK-NEXT: vcvtph2psx %xmm0, %ymm0117; CHECK-NEXT: ret{{[l|q]}}118 %ret = call <8 x float> @llvm.experimental.constrained.fpext.v8f32.v8f16(119 <8 x half> %a,120 metadata !"fpexcept.strict") #0121 ret <8 x float> %ret122}123 124define <8 x half> @f15(<8 x float> %a) #0 {125; CHECK-LABEL: f15:126; CHECK: # %bb.0:127; CHECK-NEXT: vcvtps2phx %ymm0, %xmm0128; CHECK-NEXT: vzeroupper129; CHECK-NEXT: ret{{[l|q]}}130 %ret = call <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f32(131 <8 x float> %a,132 metadata !"round.dynamic",133 metadata !"fpexcept.strict") #0134 ret <8 x half> %ret135}136 137define <16 x half> @fceilv16f16(<16 x half> %f) #0 {138; CHECK-LABEL: fceilv16f16:139; CHECK: # %bb.0:140; CHECK-NEXT: vrndscaleph $10, %ymm0, %ymm0141; CHECK-NEXT: ret{{[l|q]}}142 %res = call <16 x half> @llvm.experimental.constrained.ceil.v16f16(143 <16 x half> %f, metadata !"fpexcept.strict") #0144 ret <16 x half> %res145}146 147define <16 x half> @ffloorv16f16(<16 x half> %f) #0 {148; CHECK-LABEL: ffloorv16f16:149; CHECK: # %bb.0:150; CHECK-NEXT: vrndscaleph $9, %ymm0, %ymm0151; CHECK-NEXT: ret{{[l|q]}}152 %res = call <16 x half> @llvm.experimental.constrained.floor.v16f16(153 <16 x half> %f, metadata !"fpexcept.strict") #0154 ret <16 x half> %res155}156 157 158define <16 x half> @ftruncv16f16(<16 x half> %f) #0 {159; CHECK-LABEL: ftruncv16f16:160; CHECK: # %bb.0:161; CHECK-NEXT: vrndscaleph $11, %ymm0, %ymm0162; CHECK-NEXT: ret{{[l|q]}}163 %res = call <16 x half> @llvm.experimental.constrained.trunc.v16f16(164 <16 x half> %f, metadata !"fpexcept.strict") #0165 ret <16 x half> %res166}167 168define <16 x half> @frintv16f16(<16 x half> %f) #0 {169; CHECK-LABEL: frintv16f16:170; CHECK: # %bb.0:171; CHECK-NEXT: vrndscaleph $4, %ymm0, %ymm0172; CHECK-NEXT: ret{{[l|q]}}173 %res = call <16 x half> @llvm.experimental.constrained.rint.v16f16(174 <16 x half> %f,175 metadata !"round.dynamic", metadata !"fpexcept.strict") #0176 ret <16 x half> %res177}178 179define <16 x half> @fnearbyintv16f16(<16 x half> %f) #0 {180; CHECK-LABEL: fnearbyintv16f16:181; CHECK: # %bb.0:182; CHECK-NEXT: vrndscaleph $12, %ymm0, %ymm0183; CHECK-NEXT: ret{{[l|q]}}184 %res = call <16 x half> @llvm.experimental.constrained.nearbyint.v16f16(185 <16 x half> %f,186 metadata !"round.dynamic", metadata !"fpexcept.strict") #0187 ret <16 x half> %res188}189 190attributes #0 = { strictfp }191