775 lines · c
1// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 52// RUN: %clang_cc1 -triple arm64-none-linux-gnu -target-feature +neon -target-feature +fullfp16 -target-feature +v8.2a\3// RUN: -flax-vector-conversions=none -disable-O0-optnone -emit-llvm -o - %s \4// RUN: | opt -S -passes=mem2reg,sroa \5// RUN: | FileCheck --check-prefix=UNCONSTRAINED %s6// RUN: %clang_cc1 -triple arm64-none-linux-gnu -target-feature +neon -target-feature +fullfp16 -target-feature +v8.2a\7// RUN: -ffp-exception-behavior=maytrap -DEXCEPT=1 \8// RUN: -flax-vector-conversions=none -disable-O0-optnone -emit-llvm -o - %s \9// RUN: | opt -S -passes=mem2reg,sroa \10// RUN: | FileCheck --check-prefix=CONSTRAINED --implicit-check-not=fpexcept.maytrap %s11 12// REQUIRES: aarch64-registered-target13 14// Test that the constrained intrinsics are picking up the exception15// metadata from the AST instead of the global default from the command line.16// Any cases of "fpexcept.maytrap" in this test are clang bugs.17 18#if EXCEPT19#pragma float_control(except, on)20#endif21 22#include <arm_neon.h>23 24// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vsqrt_f16(25// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {26// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]27// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>28// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>29// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x half>30// UNCONSTRAINED-NEXT: [[VSQRT_I:%.*]] = call <4 x half> @llvm.sqrt.v4f16(<4 x half> [[TMP2]])31// UNCONSTRAINED-NEXT: ret <4 x half> [[VSQRT_I]]32//33// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vsqrt_f16(34// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]]) #[[ATTR0:[0-9]+]] {35// CONSTRAINED-NEXT: [[ENTRY:.*:]]36// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>37// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>38// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x half>39// CONSTRAINED-NEXT: [[VSQRT_I:%.*]] = call <4 x half> @llvm.experimental.constrained.sqrt.v4f16(<4 x half> [[TMP2]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2:[0-9]+]]40// CONSTRAINED-NEXT: ret <4 x half> [[VSQRT_I]]41//42float16x4_t test_vsqrt_f16(float16x4_t a) {43 return vsqrt_f16(a);44}45 46// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vsqrtq_f16(47// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]]) #[[ATTR0]] {48// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]49// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>50// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>51// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x half>52// UNCONSTRAINED-NEXT: [[VSQRT_I:%.*]] = call <8 x half> @llvm.sqrt.v8f16(<8 x half> [[TMP2]])53// UNCONSTRAINED-NEXT: ret <8 x half> [[VSQRT_I]]54//55// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vsqrtq_f16(56// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]]) #[[ATTR0]] {57// CONSTRAINED-NEXT: [[ENTRY:.*:]]58// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>59// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>60// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x half>61// CONSTRAINED-NEXT: [[VSQRT_I:%.*]] = call <8 x half> @llvm.experimental.constrained.sqrt.v8f16(<8 x half> [[TMP2]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]62// CONSTRAINED-NEXT: ret <8 x half> [[VSQRT_I]]63//64float16x8_t test_vsqrtq_f16(float16x8_t a) {65 return vsqrtq_f16(a);66}67 68// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_f16(69// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {70// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]71// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>72// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>73// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>74// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>75// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>76// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>77// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>78// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>79// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>80// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]])81// UNCONSTRAINED-NEXT: ret <4 x half> [[TMP9]]82//83// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_f16(84// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {85// CONSTRAINED-NEXT: [[ENTRY:.*:]]86// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>87// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>88// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>89// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>90// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>91// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>92// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>93// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>94// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>95// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]96// CONSTRAINED-NEXT: ret <4 x half> [[TMP9]]97//98float16x4_t test_vfma_f16(float16x4_t a, float16x4_t b, float16x4_t c) {99 return vfma_f16(a, b, c);100}101 102// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_f16(103// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {104// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]105// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>106// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>107// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>108// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>109// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>110// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>111// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>112// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>113// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>114// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]])115// UNCONSTRAINED-NEXT: ret <8 x half> [[TMP9]]116//117// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_f16(118// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {119// CONSTRAINED-NEXT: [[ENTRY:.*:]]120// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>121// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>122// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>123// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>124// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>125// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>126// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>127// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>128// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>129// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]130// CONSTRAINED-NEXT: ret <8 x half> [[TMP9]]131//132float16x8_t test_vfmaq_f16(float16x8_t a, float16x8_t b, float16x8_t c) {133 return vfmaq_f16(a, b, c);134}135 136// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_f16(137// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {138// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]139// UNCONSTRAINED-NEXT: [[FNEG_I:%.*]] = fneg <4 x half> [[B]]140// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>141// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG_I]] to <4 x i16>142// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>143// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>144// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>145// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>146// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>147// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>148// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>149// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]])150// UNCONSTRAINED-NEXT: ret <4 x half> [[TMP9]]151//152// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_f16(153// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {154// CONSTRAINED-NEXT: [[ENTRY:.*:]]155// CONSTRAINED-NEXT: [[FNEG_I:%.*]] = fneg <4 x half> [[B]]156// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>157// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG_I]] to <4 x i16>158// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>159// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>160// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>161// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>162// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>163// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>164// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>165// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]166// CONSTRAINED-NEXT: ret <4 x half> [[TMP9]]167//168float16x4_t test_vfms_f16(float16x4_t a, float16x4_t b, float16x4_t c) {169 return vfms_f16(a, b, c);170}171 172// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_f16(173// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {174// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]175// UNCONSTRAINED-NEXT: [[FNEG_I:%.*]] = fneg <8 x half> [[B]]176// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>177// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG_I]] to <8 x i16>178// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>179// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>180// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>181// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>182// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>183// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>184// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>185// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]])186// UNCONSTRAINED-NEXT: ret <8 x half> [[TMP9]]187//188// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_f16(189// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {190// CONSTRAINED-NEXT: [[ENTRY:.*:]]191// CONSTRAINED-NEXT: [[FNEG_I:%.*]] = fneg <8 x half> [[B]]192// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>193// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG_I]] to <8 x i16>194// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>195// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>196// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>197// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>198// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>199// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>200// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>201// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]202// CONSTRAINED-NEXT: ret <8 x half> [[TMP9]]203//204float16x8_t test_vfmsq_f16(float16x8_t a, float16x8_t b, float16x8_t c) {205 return vfmsq_f16(a, b, c);206}207 208// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_lane_f16(209// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {210// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]211// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>212// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>213// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>214// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>215// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>216// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>217// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>218// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <4 x i32> <i32 3, i32 3, i32 3, i32 3>219// UNCONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>220// UNCONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>221// UNCONSTRAINED-NEXT: [[FMLA2:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[FMLA]], <4 x half> [[LANE]], <4 x half> [[FMLA1]])222// UNCONSTRAINED-NEXT: ret <4 x half> [[FMLA2]]223//224// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_lane_f16(225// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {226// CONSTRAINED-NEXT: [[ENTRY:.*:]]227// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>228// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>229// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>230// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>231// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>232// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>233// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>234// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <4 x i32> <i32 3, i32 3, i32 3, i32 3>235// CONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>236// CONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>237// CONSTRAINED-NEXT: [[FMLA2:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[FMLA]], <4 x half> [[LANE]], <4 x half> [[FMLA1]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]238// CONSTRAINED-NEXT: ret <4 x half> [[FMLA2]]239//240float16x4_t test_vfma_lane_f16(float16x4_t a, float16x4_t b, float16x4_t c) {241 return vfma_lane_f16(a, b, c, 3);242}243 244// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_lane_f16(245// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {246// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]247// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>248// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>249// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>250// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>251// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>252// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>253// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>254// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>255// UNCONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>256// UNCONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>257// UNCONSTRAINED-NEXT: [[FMLA2:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[FMLA]], <8 x half> [[LANE]], <8 x half> [[FMLA1]])258// UNCONSTRAINED-NEXT: ret <8 x half> [[FMLA2]]259//260// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_lane_f16(261// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {262// CONSTRAINED-NEXT: [[ENTRY:.*:]]263// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>264// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>265// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>266// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>267// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>268// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>269// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>270// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>271// CONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>272// CONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>273// CONSTRAINED-NEXT: [[FMLA2:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[FMLA]], <8 x half> [[LANE]], <8 x half> [[FMLA1]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]274// CONSTRAINED-NEXT: ret <8 x half> [[FMLA2]]275//276float16x8_t test_vfmaq_lane_f16(float16x8_t a, float16x8_t b, float16x4_t c) {277 return vfmaq_lane_f16(a, b, c, 3);278}279 280// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_laneq_f16(281// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {282// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]283// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>284// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>285// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>286// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>287// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>288// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>289// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>290// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>291// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>292// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <4 x i32> <i32 7, i32 7, i32 7, i32 7>293// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[LANE]], <4 x half> [[TMP7]], <4 x half> [[TMP6]])294// UNCONSTRAINED-NEXT: ret <4 x half> [[TMP9]]295//296// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_laneq_f16(297// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {298// CONSTRAINED-NEXT: [[ENTRY:.*:]]299// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>300// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>301// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>302// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>303// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>304// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>305// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>306// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>307// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>308// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <4 x i32> <i32 7, i32 7, i32 7, i32 7>309// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[LANE]], <4 x half> [[TMP7]], <4 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]310// CONSTRAINED-NEXT: ret <4 x half> [[TMP9]]311//312float16x4_t test_vfma_laneq_f16(float16x4_t a, float16x4_t b, float16x8_t c) {313 return vfma_laneq_f16(a, b, c, 7);314}315 316// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_laneq_f16(317// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {318// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]319// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>320// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>321// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>322// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>323// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>324// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>325// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>326// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>327// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>328// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <8 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>329// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[LANE]], <8 x half> [[TMP7]], <8 x half> [[TMP6]])330// UNCONSTRAINED-NEXT: ret <8 x half> [[TMP9]]331//332// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_laneq_f16(333// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {334// CONSTRAINED-NEXT: [[ENTRY:.*:]]335// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>336// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>337// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>338// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>339// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>340// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>341// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>342// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>343// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>344// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <8 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>345// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[LANE]], <8 x half> [[TMP7]], <8 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]346// CONSTRAINED-NEXT: ret <8 x half> [[TMP9]]347//348float16x8_t test_vfmaq_laneq_f16(float16x8_t a, float16x8_t b, float16x8_t c) {349 return vfmaq_laneq_f16(a, b, c, 7);350}351 352// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_n_f16(353// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {354// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]355// UNCONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <4 x half> poison, half [[C]], i32 0356// UNCONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <4 x half> [[VECINIT]], half [[C]], i32 1357// UNCONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <4 x half> [[VECINIT1]], half [[C]], i32 2358// UNCONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <4 x half> [[VECINIT2]], half [[C]], i32 3359// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>360// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>361// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[VECINIT3]] to <4 x i16>362// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>363// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>364// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>365// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>366// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>367// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>368// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]])369// UNCONSTRAINED-NEXT: ret <4 x half> [[TMP9]]370//371// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfma_n_f16(372// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {373// CONSTRAINED-NEXT: [[ENTRY:.*:]]374// CONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <4 x half> poison, half [[C]], i32 0375// CONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <4 x half> [[VECINIT]], half [[C]], i32 1376// CONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <4 x half> [[VECINIT1]], half [[C]], i32 2377// CONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <4 x half> [[VECINIT2]], half [[C]], i32 3378// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>379// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[B]] to <4 x i16>380// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[VECINIT3]] to <4 x i16>381// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>382// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>383// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>384// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>385// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>386// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>387// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]388// CONSTRAINED-NEXT: ret <4 x half> [[TMP9]]389//390float16x4_t test_vfma_n_f16(float16x4_t a, float16x4_t b, float16_t c) {391 return vfma_n_f16(a, b, c);392}393 394// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_n_f16(395// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {396// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]397// UNCONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <8 x half> poison, half [[C]], i32 0398// UNCONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <8 x half> [[VECINIT]], half [[C]], i32 1399// UNCONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <8 x half> [[VECINIT1]], half [[C]], i32 2400// UNCONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <8 x half> [[VECINIT2]], half [[C]], i32 3401// UNCONSTRAINED-NEXT: [[VECINIT4:%.*]] = insertelement <8 x half> [[VECINIT3]], half [[C]], i32 4402// UNCONSTRAINED-NEXT: [[VECINIT5:%.*]] = insertelement <8 x half> [[VECINIT4]], half [[C]], i32 5403// UNCONSTRAINED-NEXT: [[VECINIT6:%.*]] = insertelement <8 x half> [[VECINIT5]], half [[C]], i32 6404// UNCONSTRAINED-NEXT: [[VECINIT7:%.*]] = insertelement <8 x half> [[VECINIT6]], half [[C]], i32 7405// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>406// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>407// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[VECINIT7]] to <8 x i16>408// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>409// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>410// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>411// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>412// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>413// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>414// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]])415// UNCONSTRAINED-NEXT: ret <8 x half> [[TMP9]]416//417// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmaq_n_f16(418// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {419// CONSTRAINED-NEXT: [[ENTRY:.*:]]420// CONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <8 x half> poison, half [[C]], i32 0421// CONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <8 x half> [[VECINIT]], half [[C]], i32 1422// CONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <8 x half> [[VECINIT1]], half [[C]], i32 2423// CONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <8 x half> [[VECINIT2]], half [[C]], i32 3424// CONSTRAINED-NEXT: [[VECINIT4:%.*]] = insertelement <8 x half> [[VECINIT3]], half [[C]], i32 4425// CONSTRAINED-NEXT: [[VECINIT5:%.*]] = insertelement <8 x half> [[VECINIT4]], half [[C]], i32 5426// CONSTRAINED-NEXT: [[VECINIT6:%.*]] = insertelement <8 x half> [[VECINIT5]], half [[C]], i32 6427// CONSTRAINED-NEXT: [[VECINIT7:%.*]] = insertelement <8 x half> [[VECINIT6]], half [[C]], i32 7428// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>429// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[B]] to <8 x i16>430// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[VECINIT7]] to <8 x i16>431// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>432// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>433// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>434// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>435// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>436// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>437// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]438// CONSTRAINED-NEXT: ret <8 x half> [[TMP9]]439//440float16x8_t test_vfmaq_n_f16(float16x8_t a, float16x8_t b, float16_t c) {441 return vfmaq_n_f16(a, b, c);442}443 444// UNCONSTRAINED-LABEL: define dso_local half @test_vfmah_lane_f16(445// UNCONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {446// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]447// UNCONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <4 x half> [[C]], i32 3448// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = call half @llvm.fma.f16(half [[B]], half [[EXTRACT]], half [[A]])449// UNCONSTRAINED-NEXT: ret half [[TMP0]]450//451// CONSTRAINED-LABEL: define dso_local half @test_vfmah_lane_f16(452// CONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {453// CONSTRAINED-NEXT: [[ENTRY:.*:]]454// CONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <4 x half> [[C]], i32 3455// CONSTRAINED-NEXT: [[TMP0:%.*]] = call half @llvm.experimental.constrained.fma.f16(half [[B]], half [[EXTRACT]], half [[A]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]456// CONSTRAINED-NEXT: ret half [[TMP0]]457//458float16_t test_vfmah_lane_f16(float16_t a, float16_t b, float16x4_t c) {459 return vfmah_lane_f16(a, b, c, 3);460}461 462// UNCONSTRAINED-LABEL: define dso_local half @test_vfmah_laneq_f16(463// UNCONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {464// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]465// UNCONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <8 x half> [[C]], i32 7466// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = call half @llvm.fma.f16(half [[B]], half [[EXTRACT]], half [[A]])467// UNCONSTRAINED-NEXT: ret half [[TMP0]]468//469// CONSTRAINED-LABEL: define dso_local half @test_vfmah_laneq_f16(470// CONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {471// CONSTRAINED-NEXT: [[ENTRY:.*:]]472// CONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <8 x half> [[C]], i32 7473// CONSTRAINED-NEXT: [[TMP0:%.*]] = call half @llvm.experimental.constrained.fma.f16(half [[B]], half [[EXTRACT]], half [[A]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]474// CONSTRAINED-NEXT: ret half [[TMP0]]475//476float16_t test_vfmah_laneq_f16(float16_t a, float16_t b, float16x8_t c) {477 return vfmah_laneq_f16(a, b, c, 7);478}479 480// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_lane_f16(481// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {482// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]483// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>484// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <4 x half> [[B]]485// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG]] to <4 x i16>486// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>487// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>488// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>489// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>490// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>491// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <4 x i32> <i32 3, i32 3, i32 3, i32 3>492// UNCONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>493// UNCONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>494// UNCONSTRAINED-NEXT: [[FMLA2:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[FMLA]], <4 x half> [[LANE]], <4 x half> [[FMLA1]])495// UNCONSTRAINED-NEXT: ret <4 x half> [[FMLA2]]496//497// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_lane_f16(498// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {499// CONSTRAINED-NEXT: [[ENTRY:.*:]]500// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>501// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <4 x half> [[B]]502// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG]] to <4 x i16>503// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>504// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>505// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>506// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>507// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>508// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <4 x i32> <i32 3, i32 3, i32 3, i32 3>509// CONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>510// CONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>511// CONSTRAINED-NEXT: [[FMLA2:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[FMLA]], <4 x half> [[LANE]], <4 x half> [[FMLA1]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]512// CONSTRAINED-NEXT: ret <4 x half> [[FMLA2]]513//514float16x4_t test_vfms_lane_f16(float16x4_t a, float16x4_t b, float16x4_t c) {515 return vfms_lane_f16(a, b, c, 3);516}517 518// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_lane_f16(519// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {520// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]521// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>522// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <8 x half> [[B]]523// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG]] to <8 x i16>524// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>525// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>526// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>527// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>528// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>529// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>530// UNCONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>531// UNCONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>532// UNCONSTRAINED-NEXT: [[FMLA2:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[FMLA]], <8 x half> [[LANE]], <8 x half> [[FMLA1]])533// UNCONSTRAINED-NEXT: ret <8 x half> [[FMLA2]]534//535// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_lane_f16(536// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {537// CONSTRAINED-NEXT: [[ENTRY:.*:]]538// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>539// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <8 x half> [[B]]540// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG]] to <8 x i16>541// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[C]] to <4 x i16>542// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>543// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>544// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>545// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>546// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <4 x half> [[TMP6]], <4 x half> [[TMP6]], <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>547// CONSTRAINED-NEXT: [[FMLA:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>548// CONSTRAINED-NEXT: [[FMLA1:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>549// CONSTRAINED-NEXT: [[FMLA2:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[FMLA]], <8 x half> [[LANE]], <8 x half> [[FMLA1]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]550// CONSTRAINED-NEXT: ret <8 x half> [[FMLA2]]551//552float16x8_t test_vfmsq_lane_f16(float16x8_t a, float16x8_t b, float16x4_t c) {553 return vfmsq_lane_f16(a, b, c, 3);554}555 556// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_laneq_f16(557// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {558// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]559// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>560// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <4 x half> [[B]]561// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG]] to <4 x i16>562// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>563// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>564// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>565// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>566// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>567// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>568// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>569// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <4 x i32> <i32 7, i32 7, i32 7, i32 7>570// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[LANE]], <4 x half> [[TMP7]], <4 x half> [[TMP6]])571// UNCONSTRAINED-NEXT: ret <4 x half> [[TMP9]]572//573// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_laneq_f16(574// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {575// CONSTRAINED-NEXT: [[ENTRY:.*:]]576// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>577// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <4 x half> [[B]]578// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG]] to <4 x i16>579// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>580// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>581// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>582// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>583// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>584// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>585// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>586// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <4 x i32> <i32 7, i32 7, i32 7, i32 7>587// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[LANE]], <4 x half> [[TMP7]], <4 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]588// CONSTRAINED-NEXT: ret <4 x half> [[TMP9]]589//590float16x4_t test_vfms_laneq_f16(float16x4_t a, float16x4_t b, float16x8_t c) {591 return vfms_laneq_f16(a, b, c, 7);592}593 594// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_laneq_f16(595// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {596// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]597// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>598// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <8 x half> [[B]]599// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG]] to <8 x i16>600// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>601// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>602// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>603// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>604// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>605// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>606// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>607// UNCONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <8 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>608// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[LANE]], <8 x half> [[TMP7]], <8 x half> [[TMP6]])609// UNCONSTRAINED-NEXT: ret <8 x half> [[TMP9]]610//611// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_laneq_f16(612// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {613// CONSTRAINED-NEXT: [[ENTRY:.*:]]614// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>615// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <8 x half> [[B]]616// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG]] to <8 x i16>617// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[C]] to <8 x i16>618// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>619// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>620// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>621// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>622// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>623// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>624// CONSTRAINED-NEXT: [[LANE:%.*]] = shufflevector <8 x half> [[TMP8]], <8 x half> [[TMP8]], <8 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>625// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[LANE]], <8 x half> [[TMP7]], <8 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]626// CONSTRAINED-NEXT: ret <8 x half> [[TMP9]]627//628float16x8_t test_vfmsq_laneq_f16(float16x8_t a, float16x8_t b, float16x8_t c) {629 return vfmsq_laneq_f16(a, b, c, 7);630}631 632// UNCONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_n_f16(633// UNCONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {634// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]635// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <4 x half> [[B]]636// UNCONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <4 x half> poison, half [[C]], i32 0637// UNCONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <4 x half> [[VECINIT]], half [[C]], i32 1638// UNCONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <4 x half> [[VECINIT1]], half [[C]], i32 2639// UNCONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <4 x half> [[VECINIT2]], half [[C]], i32 3640// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>641// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG]] to <4 x i16>642// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[VECINIT3]] to <4 x i16>643// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>644// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>645// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>646// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>647// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>648// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>649// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]])650// UNCONSTRAINED-NEXT: ret <4 x half> [[TMP9]]651//652// CONSTRAINED-LABEL: define dso_local <4 x half> @test_vfms_n_f16(653// CONSTRAINED-SAME: <4 x half> noundef [[A:%.*]], <4 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {654// CONSTRAINED-NEXT: [[ENTRY:.*:]]655// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <4 x half> [[B]]656// CONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <4 x half> poison, half [[C]], i32 0657// CONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <4 x half> [[VECINIT]], half [[C]], i32 1658// CONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <4 x half> [[VECINIT1]], half [[C]], i32 2659// CONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <4 x half> [[VECINIT2]], half [[C]], i32 3660// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <4 x half> [[A]] to <4 x i16>661// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <4 x half> [[FNEG]] to <4 x i16>662// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <4 x half> [[VECINIT3]] to <4 x i16>663// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>664// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>665// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>666// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x half>667// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x half>668// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>669// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> [[TMP7]], <4 x half> [[TMP8]], <4 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]670// CONSTRAINED-NEXT: ret <4 x half> [[TMP9]]671//672float16x4_t test_vfms_n_f16(float16x4_t a, float16x4_t b, float16_t c) {673 return vfms_n_f16(a, b, c);674}675 676// UNCONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_n_f16(677// UNCONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {678// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]679// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <8 x half> [[B]]680// UNCONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <8 x half> poison, half [[C]], i32 0681// UNCONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <8 x half> [[VECINIT]], half [[C]], i32 1682// UNCONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <8 x half> [[VECINIT1]], half [[C]], i32 2683// UNCONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <8 x half> [[VECINIT2]], half [[C]], i32 3684// UNCONSTRAINED-NEXT: [[VECINIT4:%.*]] = insertelement <8 x half> [[VECINIT3]], half [[C]], i32 4685// UNCONSTRAINED-NEXT: [[VECINIT5:%.*]] = insertelement <8 x half> [[VECINIT4]], half [[C]], i32 5686// UNCONSTRAINED-NEXT: [[VECINIT6:%.*]] = insertelement <8 x half> [[VECINIT5]], half [[C]], i32 6687// UNCONSTRAINED-NEXT: [[VECINIT7:%.*]] = insertelement <8 x half> [[VECINIT6]], half [[C]], i32 7688// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>689// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG]] to <8 x i16>690// UNCONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[VECINIT7]] to <8 x i16>691// UNCONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>692// UNCONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>693// UNCONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>694// UNCONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>695// UNCONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>696// UNCONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>697// UNCONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]])698// UNCONSTRAINED-NEXT: ret <8 x half> [[TMP9]]699//700// CONSTRAINED-LABEL: define dso_local <8 x half> @test_vfmsq_n_f16(701// CONSTRAINED-SAME: <8 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], half noundef [[C:%.*]]) #[[ATTR0]] {702// CONSTRAINED-NEXT: [[ENTRY:.*:]]703// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg <8 x half> [[B]]704// CONSTRAINED-NEXT: [[VECINIT:%.*]] = insertelement <8 x half> poison, half [[C]], i32 0705// CONSTRAINED-NEXT: [[VECINIT1:%.*]] = insertelement <8 x half> [[VECINIT]], half [[C]], i32 1706// CONSTRAINED-NEXT: [[VECINIT2:%.*]] = insertelement <8 x half> [[VECINIT1]], half [[C]], i32 2707// CONSTRAINED-NEXT: [[VECINIT3:%.*]] = insertelement <8 x half> [[VECINIT2]], half [[C]], i32 3708// CONSTRAINED-NEXT: [[VECINIT4:%.*]] = insertelement <8 x half> [[VECINIT3]], half [[C]], i32 4709// CONSTRAINED-NEXT: [[VECINIT5:%.*]] = insertelement <8 x half> [[VECINIT4]], half [[C]], i32 5710// CONSTRAINED-NEXT: [[VECINIT6:%.*]] = insertelement <8 x half> [[VECINIT5]], half [[C]], i32 6711// CONSTRAINED-NEXT: [[VECINIT7:%.*]] = insertelement <8 x half> [[VECINIT6]], half [[C]], i32 7712// CONSTRAINED-NEXT: [[TMP0:%.*]] = bitcast <8 x half> [[A]] to <8 x i16>713// CONSTRAINED-NEXT: [[TMP1:%.*]] = bitcast <8 x half> [[FNEG]] to <8 x i16>714// CONSTRAINED-NEXT: [[TMP2:%.*]] = bitcast <8 x half> [[VECINIT7]] to <8 x i16>715// CONSTRAINED-NEXT: [[TMP3:%.*]] = bitcast <8 x i16> [[TMP0]] to <16 x i8>716// CONSTRAINED-NEXT: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>717// CONSTRAINED-NEXT: [[TMP5:%.*]] = bitcast <8 x i16> [[TMP2]] to <16 x i8>718// CONSTRAINED-NEXT: [[TMP6:%.*]] = bitcast <16 x i8> [[TMP3]] to <8 x half>719// CONSTRAINED-NEXT: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x half>720// CONSTRAINED-NEXT: [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>721// CONSTRAINED-NEXT: [[TMP9:%.*]] = call <8 x half> @llvm.experimental.constrained.fma.v8f16(<8 x half> [[TMP7]], <8 x half> [[TMP8]], <8 x half> [[TMP6]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]722// CONSTRAINED-NEXT: ret <8 x half> [[TMP9]]723//724float16x8_t test_vfmsq_n_f16(float16x8_t a, float16x8_t b, float16_t c) {725 return vfmsq_n_f16(a, b, c);726}727 728// UNCONSTRAINED-LABEL: define dso_local half @test_vfmsh_lane_f16(729// UNCONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {730// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]731// UNCONSTRAINED-NEXT: [[CONV:%.*]] = fpext half [[B]] to float732// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg float [[CONV]]733// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = fptrunc float [[FNEG]] to half734// UNCONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <4 x half> [[C]], i32 3735// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = call half @llvm.fma.f16(half [[TMP0]], half [[EXTRACT]], half [[A]])736// UNCONSTRAINED-NEXT: ret half [[TMP1]]737//738// CONSTRAINED-LABEL: define dso_local half @test_vfmsh_lane_f16(739// CONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <4 x half> noundef [[C:%.*]]) #[[ATTR0]] {740// CONSTRAINED-NEXT: [[ENTRY:.*:]]741// CONSTRAINED-NEXT: [[CONV:%.*]] = call float @llvm.experimental.constrained.fpext.f32.f16(half [[B]], metadata !"fpexcept.strict") #[[ATTR2]]742// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg float [[CONV]]743// CONSTRAINED-NEXT: [[TMP0:%.*]] = call half @llvm.experimental.constrained.fptrunc.f16.f32(float [[FNEG]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR2]]744// CONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <4 x half> [[C]], i32 3745// CONSTRAINED-NEXT: [[TMP1:%.*]] = call half @llvm.experimental.constrained.fma.f16(half [[TMP0]], half [[EXTRACT]], half [[A]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]746// CONSTRAINED-NEXT: ret half [[TMP1]]747//748float16_t test_vfmsh_lane_f16(float16_t a, float16_t b, float16x4_t c) {749 return vfmsh_lane_f16(a, b, c, 3);750}751 752// UNCONSTRAINED-LABEL: define dso_local half @test_vfmsh_laneq_f16(753// UNCONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {754// UNCONSTRAINED-NEXT: [[ENTRY:.*:]]755// UNCONSTRAINED-NEXT: [[CONV:%.*]] = fpext half [[B]] to float756// UNCONSTRAINED-NEXT: [[FNEG:%.*]] = fneg float [[CONV]]757// UNCONSTRAINED-NEXT: [[TMP0:%.*]] = fptrunc float [[FNEG]] to half758// UNCONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <8 x half> [[C]], i32 7759// UNCONSTRAINED-NEXT: [[TMP1:%.*]] = call half @llvm.fma.f16(half [[TMP0]], half [[EXTRACT]], half [[A]])760// UNCONSTRAINED-NEXT: ret half [[TMP1]]761//762// CONSTRAINED-LABEL: define dso_local half @test_vfmsh_laneq_f16(763// CONSTRAINED-SAME: half noundef [[A:%.*]], half noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) #[[ATTR0]] {764// CONSTRAINED-NEXT: [[ENTRY:.*:]]765// CONSTRAINED-NEXT: [[CONV:%.*]] = call float @llvm.experimental.constrained.fpext.f32.f16(half [[B]], metadata !"fpexcept.strict") #[[ATTR2]]766// CONSTRAINED-NEXT: [[FNEG:%.*]] = fneg float [[CONV]]767// CONSTRAINED-NEXT: [[TMP0:%.*]] = call half @llvm.experimental.constrained.fptrunc.f16.f32(float [[FNEG]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR2]]768// CONSTRAINED-NEXT: [[EXTRACT:%.*]] = extractelement <8 x half> [[C]], i32 7769// CONSTRAINED-NEXT: [[TMP1:%.*]] = call half @llvm.experimental.constrained.fma.f16(half [[TMP0]], half [[EXTRACT]], half [[A]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR2]]770// CONSTRAINED-NEXT: ret half [[TMP1]]771//772float16_t test_vfmsh_laneq_f16(float16_t a, float16_t b, float16x8_t c) {773 return vfmsh_laneq_f16(a, b, c, 7);774}775