216 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 %s --check-prefixes=CHECK3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=CHECK4 5declare <8 x half> @llvm.experimental.constrained.fadd.v8f16(<8 x half>, <8 x half>, metadata, metadata)6declare <8 x half> @llvm.experimental.constrained.fsub.v8f16(<8 x half>, <8 x half>, metadata, metadata)7declare <8 x half> @llvm.experimental.constrained.fmul.v8f16(<8 x half>, <8 x half>, metadata, metadata)8declare <8 x half> @llvm.experimental.constrained.fdiv.v8f16(<8 x half>, <8 x half>, metadata, metadata)9declare <8 x half> @llvm.experimental.constrained.sqrt.v8f16(<8 x half>, metadata, metadata)10declare half @llvm.experimental.constrained.fptrunc.f16.f32(float, metadata, metadata)11declare half @llvm.experimental.constrained.fptrunc.f16.f64(double, metadata, metadata)12declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float>, metadata, metadata)13declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float>, metadata, metadata)14declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f64(<2 x double>, metadata, metadata)15declare float @llvm.experimental.constrained.fpext.f32.f16(half, metadata)16declare double @llvm.experimental.constrained.fpext.f64.f16(half, metadata)17declare <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half>, metadata)18declare <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half>, metadata)19declare <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f16(<2 x half>, metadata)20declare <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half>, <8 x half>, <8 x half>, metadata, metadata)21 22define <8 x half> @f2(<8 x half> %a, <8 x half> %b) #0 {23; CHECK-LABEL: f2:24; CHECK: # %bb.0:25; CHECK-NEXT: vaddph %xmm1, %xmm0, %xmm026; CHECK-NEXT: ret{{[l|q]}}27 %ret = call <8 x half> @llvm.experimental.constrained.fadd.v8f16(<8 x half> %a, <8 x half> %b,28 metadata !"round.dynamic",29 metadata !"fpexcept.strict") #030 ret <8 x half> %ret31}32 33define <8 x half> @f4(<8 x half> %a, <8 x half> %b) #0 {34; CHECK-LABEL: f4:35; CHECK: # %bb.0:36; CHECK-NEXT: vsubph %xmm1, %xmm0, %xmm037; CHECK-NEXT: ret{{[l|q]}}38 %ret = call <8 x half> @llvm.experimental.constrained.fsub.v8f16(<8 x half> %a, <8 x half> %b,39 metadata !"round.dynamic",40 metadata !"fpexcept.strict") #041 ret <8 x half> %ret42}43 44define <8 x half> @f6(<8 x half> %a, <8 x half> %b) #0 {45; CHECK-LABEL: f6:46; CHECK: # %bb.0:47; CHECK-NEXT: vmulph %xmm1, %xmm0, %xmm048; CHECK-NEXT: ret{{[l|q]}}49 %ret = call <8 x half> @llvm.experimental.constrained.fmul.v8f16(<8 x half> %a, <8 x half> %b,50 metadata !"round.dynamic",51 metadata !"fpexcept.strict") #052 ret <8 x half> %ret53}54 55define <8 x half> @f8(<8 x half> %a, <8 x half> %b) #0 {56; CHECK-LABEL: f8:57; CHECK: # %bb.0:58; CHECK-NEXT: vdivph %xmm1, %xmm0, %xmm059; CHECK-NEXT: ret{{[l|q]}}60 %ret = call <8 x half> @llvm.experimental.constrained.fdiv.v8f16(<8 x half> %a, <8 x half> %b,61 metadata !"round.dynamic",62 metadata !"fpexcept.strict") #063 ret <8 x half> %ret64}65 66define <8 x half> @f10(<8 x half> %a) #0 {67; CHECK-LABEL: f10:68; CHECK: # %bb.0:69; CHECK-NEXT: vsqrtph %xmm0, %xmm070; CHECK-NEXT: ret{{[l|q]}}71 %sqrt = call <8 x half> @llvm.experimental.constrained.sqrt.v8f16(72 <8 x half> %a,73 metadata !"round.dynamic",74 metadata !"fpexcept.strict") #075 ret <8 x half > %sqrt76}77 78define <8 x half> @f11(<2 x double> %a0, <8 x half> %a1) #0 {79; CHECK-LABEL: f11:80; CHECK: # %bb.0:81; CHECK-NEXT: vcvtsd2sh %xmm0, %xmm0, %xmm082; CHECK-NEXT: vmovsh {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]83; CHECK-NEXT: ret{{[l|q]}}84 %ext = extractelement <2 x double> %a0, i32 085 %cvt = call half @llvm.experimental.constrained.fptrunc.f16.f64(double %ext,86 metadata !"round.dynamic",87 metadata !"fpexcept.strict") #088 %res = insertelement <8 x half> %a1, half %cvt, i32 089 ret <8 x half> %res90}91 92define <2 x double> @f12(<2 x double> %a0, <8 x half> %a1) #0 {93; CHECK-LABEL: f12:94; CHECK: # %bb.0:95; CHECK-NEXT: vcvtsh2sd %xmm1, %xmm1, %xmm196; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]97; CHECK-NEXT: ret{{[l|q]}}98 %ext = extractelement <8 x half> %a1, i32 099 %cvt = call double @llvm.experimental.constrained.fpext.f64.f16(half %ext,100 metadata !"fpexcept.strict") #0101 %res = insertelement <2 x double> %a0, double %cvt, i32 0102 ret <2 x double> %res103}104 105define <8 x half> @f13(<8 x half> %a, <8 x half> %b, <8 x half> %c) #0 {106; CHECK-LABEL: f13:107; CHECK: # %bb.0:108; CHECK-NEXT: vfmadd213ph %xmm2, %xmm1, %xmm0109; CHECK-NEXT: ret{{[l|q]}}110 %res = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> %a, <8 x half> %b, <8 x half> %c,111 metadata !"round.dynamic",112 metadata !"fpexcept.strict") #0113 ret <8 x half> %res114}115 116define <2 x double> @f15(<2 x half> %a) #0 {117; CHECK-LABEL: f15:118; CHECK: # %bb.0:119; CHECK-NEXT: vcvtph2pd %xmm0, %xmm0120; CHECK-NEXT: ret{{[l|q]}}121 %ret = call <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f16(122 <2 x half> %a,123 metadata !"fpexcept.strict") #0124 ret <2 x double> %ret125}126 127define <2 x half> @f16(<2 x double> %a) #0 {128; CHECK-LABEL: f16:129; CHECK: # %bb.0:130; CHECK-NEXT: vcvtpd2ph %xmm0, %xmm0131; CHECK-NEXT: ret{{[l|q]}}132 %ret = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f64(133 <2 x double> %a,134 metadata !"round.dynamic",135 metadata !"fpexcept.strict") #0136 ret <2 x half> %ret137}138 139define <8 x half> @f17(<4 x float> %a0, <8 x half> %a1) #0 {140; CHECK-LABEL: f17:141; CHECK: # %bb.0:142; CHECK-NEXT: vcvtss2sh %xmm0, %xmm0, %xmm0143; CHECK-NEXT: vmovsh {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]144; CHECK-NEXT: ret{{[l|q]}}145 %ext = extractelement <4 x float> %a0, i32 0146 %cvt = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %ext,147 metadata !"round.dynamic",148 metadata !"fpexcept.strict") #0149 %res = insertelement <8 x half> %a1, half %cvt, i32 0150 ret <8 x half> %res151}152 153define <4 x float> @f18(<4 x float> %a0, <8 x half> %a1) #0 {154; CHECK-LABEL: f18:155; CHECK: # %bb.0:156; CHECK-NEXT: vcvtsh2ss %xmm1, %xmm1, %xmm1157; CHECK-NEXT: vmovss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]158; CHECK-NEXT: ret{{[l|q]}}159 %ext = extractelement <8 x half> %a1, i32 0160 %cvt = call float @llvm.experimental.constrained.fpext.f32.f16(half %ext,161 metadata !"fpexcept.strict") #0162 %res = insertelement <4 x float> %a0, float %cvt, i32 0163 ret <4 x float> %res164}165 166define <2 x float> @f19(<2 x half> %a) #0 {167; CHECK-LABEL: f19:168; CHECK: # %bb.0:169; CHECK-NEXT: vxorps %xmm1, %xmm1, %xmm1170; CHECK-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]171; CHECK-NEXT: vcvtph2psx %xmm0, %xmm0172; CHECK-NEXT: ret{{[l|q]}}173 %ret = call <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(174 <2 x half> %a,175 metadata !"fpexcept.strict") #0176 ret <2 x float> %ret177}178 179define <4 x float> @f20(<4 x half> %a) #0 {180; CHECK-LABEL: f20:181; CHECK: # %bb.0:182; CHECK-NEXT: vcvtph2psx %xmm0, %xmm0183; CHECK-NEXT: ret{{[l|q]}}184 %ret = call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(185 <4 x half> %a,186 metadata !"fpexcept.strict") #0187 ret <4 x float> %ret188}189 190define <2 x half> @f21(<2 x float> %a) #0 {191; CHECK-LABEL: f21:192; CHECK: # %bb.0:193; CHECK-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero194; CHECK-NEXT: vcvtps2phx %xmm0, %xmm0195; CHECK-NEXT: ret{{[l|q]}}196 %ret = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(197 <2 x float> %a,198 metadata !"round.dynamic",199 metadata !"fpexcept.strict") #0200 ret <2 x half> %ret201}202 203define <4 x half> @f22(<4 x float> %a) #0 {204; CHECK-LABEL: f22:205; CHECK: # %bb.0:206; CHECK-NEXT: vcvtps2phx %xmm0, %xmm0207; CHECK-NEXT: ret{{[l|q]}}208 %ret = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(209 <4 x float> %a,210 metadata !"round.dynamic",211 metadata !"fpexcept.strict") #0212 ret <4 x half> %ret213}214 215attributes #0 = { strictfp }216